PROCESS DESIGN

UNDERSTANDING THE PRODUCT

AND PROCESS
LIFECYCLE APPROACH TO PROCESS
VALIDATION FDA STAGE 1
FDA Lifecycle Approach to Process ValidationWhat, Why, and How? ............................................................... 3
The Forgotten Origins of Quality by Design ....................................................................................................... 14
Understanding Physicochemical Properties for Pharmaceutical Product
Development and ManufacturingDissociation, Distribution/Partition, and Solubility ....................................... 19
Understanding Physicochemical Properties for Pharmaceutical Product Development
and Manufacturing II: Physical and Chemical Stability and Excipient Compatibility ........................................... 30
Patent Potential .............................................................................................................................................. 42
First Steps in Experimental DesignThe Screening Experiment ........................................................................ 46
First Steps in Experimental Design II: More on Screening Experiments .............................................................. 54
A Further Step in Experimental Design (III): The Response Surface ................................................................... 63
Estimation: Knowledge Building with Probability Distributions .......................................................................... 71
Understanding Hypothesis Testing Using Probability Distributions .................................................................... 86
PQ=Confirmation ........................................................................................................................................... 102
 Paul L. Pluta

FDA Lifecycle Approach

to Process Validation
What, Why, and How?
Paul L. Pluta

t
FDA has provided recommendations for the

PQ Forum provides a mechanism for validation
general lifecycle and stages 1, 2, and 3. Specific
practitioners to share information about Stage
expectations are discussed.
2 process qualification in the validation lifecycle.
t
Stage 1Process Design may be generally
Information about supporting activities such as
described as process understanding. Stage
equipment and analytical validation is shared.
1 work is ultimately reflected in the master
The information provided should be helpful and
production record and control records.
practical so as to enable application in actual work
t
Stage 2Process Qualification may be
situations.
described as validation performance. This
Reader comments, questions, and suggestions
stage comprises demonstration of final process
are needed to help us fulfill our objective for this
performance by means of conformance lots.
column. Please contact column coordinator Paul
Stage 2 confirms the development work of Stage
Pluta at paul.pluta@comcast.net or managing editor
1 Process Design.
Susan Haigney at shaigney@advanstar.com with
t
Stage 2 specific recommendations are provided
comments, suggestions, or topics for discussion.
for design of a facility and qualification of
utilities and equipment, process performance
KEY POINTS qualification (PPQ), PPQ protocol, and PPQ
The following key points are discussed:
protocol execution and report.
t
The US Food and Drug Administration issued
t
Stage 3Continued Process Verification
Process Validation: General Principles and Practices
may be simply described as maintaining
in January 2011, which has given widespread
validation. This stage comprises the ongoing
visibility to the lifecycle approach concept.
commercial manufacturing of the product
t
The process validation guidance integrates
under the same or equivalent conditions as
strategy and approaches to provide a
demonstrated in Stage 2 Process Qualification.
comprehensive approach to validation. Three
t
The integration of development work, process
stages in the lifecycle approach are identified.
conformance, and continuing verification
The lifecycle concept links development,
provides assurance the product or process will
validation performance, and product or process
consistently remain in control throughout the
maintenance in a state of control during routine
entire product lifetime.
commercial production.
t
The lifecycle approach integrates various
t
Understanding the sources of variation and
strategies, approaches, and expectations that
control of variation commensurate with risk is a
had been mentioned in multiple previously
key component of the lifecycle approach.
published documents, guidelines, and
presentations for many years.

[
For more Author
ABOUT THE AUTHOR
Paul L. Pluta, Ph.D., is a pharmaceutical scientist with extensive industrial development,
information,
manufacturing, and management experience. Dr. Pluta is also an adjunct associate professor at the
go to
University of Illinois-Chicago College of Pharmacy. Dr. Pluta may be contacted by e-mail at paul.
gxpandjvt.com/bios
pluta@comcast.net.

PROCESS VALIDATION Process Design 3

Paul L. Pluta
t
The concepts identified in the respective
stages of the FDA process validation
guidanceunderstanding, performance, and
maintenanceserve as a model for all areas of
validation and qualification.
t
The new guidance affects many areas of site
validation programs including organizational
aspects, validation performance specifics, risk
analysis, training, and documentation.
t
Senior and functional management support
is needed to transition organizations to the
lifecycle approach to validation. Risk analysis
is key to development and prioritization of a
suitable program that will be embraced and
supported.
INTRODUCTION
The US Food and Drug Administration issued
Process Validation: General Principles and Practices
(1) in January 2011. This guidance has given
widespread visibility to the lifecycle approach
concept. Validation managers are now responding
to questions and comments about the guidance
from their colleagues. The following discusses
these and other areas of concern raised by
attendees at validation meetings in Montreal
(2010), Philadelphia (2010), and Amsterdam
(2011). These are relevant hands-on questions
from people that face validation problems every
day. Topics addressed in this discussion include
the following:
t
What is different about the lifecycle approach?
What is its emphasis compared to the 1987 FDA
process validation guidance (2)?
t
Why the lifecycle approach? Is it really a new
approach?
t
Should the lifecycle approach be applied to
other areas of validation and qualification?
What about using the lifecycle approach to
other processes and to equipment, HVAC,
computer systems, and other qualifications?
t
How does the guidance affect our current
validation programs? What areas need to
be modified to be compliant with the new
guidance?
THE LIFECYCLE APPROACH
The January 2011 process validation guidance
(1) has integrated information, strategy,
and approaches discussed in various US
and international documents to provide a
comprehensive approach to validation (i.e., the
lifecycle approach). The guidance provides specific
and detailed recommendations for each stage of the
lifecycle approach.
4 PROCESS VALIDATION Process Design
The definition of process validation stated in the
2011 guidance is as follows:
Process validation is defined as the collection
and evaluation of data, from the process design
stage throughout production, which establishes
scientific evidence that a process is capable of
consistently delivering quality product. Process
validation involves a series of activities taking place
over the lifecycle of the product and process.
The guidance describes process validation
activities in the following three stages:
t
Stage 1Process Design: The commercial
process is defined during this stage based on
knowledge gained through development and
scale-up activities.
t
Stage 2Process Qualification: During this
state, the process design is confirmed as
being capable of reproducible commercial
manufacturing.
t
Stage 3Continued Process Verification:
Ongoing assurance is gained during routine
production that the process remains in a state of
control.
These sections of the 2011 guidance clearly
identify the key difference between the lifecycle
approach compared to validation in the 1987 FDA
guidance. The 2011 lifecycle approach to process
validation encompasses product and process
activities beginning in development and continuing
throughout the commercial life of the product. The
1987 definition and subsequent discussion in the
guidance placed major emphasis on the validation
protocol, testing, results, and documentationwhat
is now considered to be Stage 2 in the lifecycle
approach. Development work and post-validation
monitoring were not emphasized in the 1987
guidance.
Approach to Process
ValidationStages 1, 2, and 3
The approach to process validation stated in the
2011 guidance clearly emphasizes contemporary
concepts and expectations for pharmaceutical
manufacturing. The manufacturer should
have great confidence that the performance of
manufacturing will consistently produce active
pharmaceutical ingredients (APIs) and drug
products meeting expected attributes. This
confidence is obtained from objective information
and data from laboratory, pilot, and commercial-
scale studies (i.e., the work of Stage 1). After
completion of Stage 1 development, Stage 2
Process Qualification confirms the work of Stage
1. After successful Stage 2 performance, Stage
3 Continued Process Verification maintains the
validated state. The guidance states:

The lifecycle concept links product and process
development, qualification of the commercial
manufacturing process, and maintenance of
the process in a state of control during routine
commercial production. This guidance supports
process improvement and innovation through
sound science.
Successful validation depends on knowledge
and understanding from product and process
development. Specific key areas mentioned in the
guidance include the following:
t
Understanding the sources of variation
t
Detect the presence and degree of variation
t
Understanding the impact of variation on the
process and ultimately on product attributes
t
Control the variation in a manner
commensurate with the risk it represents to the
process and product.
FDA Recommendations
The 2011 guidance discusses several areas
and provides specific details. These include
recommendations for the general lifecycle and
stages 1, 2, and 3. The entire recommendations
section of the guidance is provided online at FDA.
gov.
General considerations. These considerations
are applicable to all stages in the lifecycle. For
example, an integrated team approach that includes
expertise from multiple disciplines and project
plans is recommended. The support of senior
management is termed essential. Other general
topics discussed include the initiation of studies
to further understand product and process during
the lifecycle, attribute evaluation, and the need for
higher levels of control for parameters associated
with higher risk.
Stage 1process design. This stage may be
generally described as process understanding.
Studies are conducted during this stage to develop
and characterize product and process. The work of
Stage 1 should be commensurate with the identified
or expected risk for the product and process.
Stage 1 recommendations address development
activities that will ultimately be reflected in the
master production record and control records. The
guidance clearly states the goal of stage 1: To
design a process suitable for routine commercial
manufacturing that can consistently deliver a
product that meets its quality attributes. The
following two topics are discussed:
t
Building and capturing process knowledge
and understanding. This section discusses
the role of product development and uses
terminology common to the quality-by-design
(QbD) initiativequality attributes, design of
experiments (DOE) studies, and so on.
Paul L. Pluta
t
Establishing a strategy for process control. This
section addresses reducing input variation,
adjustment for input variation during
processing, and related topics.
Stage 2process qualification. This
stage may be simply described as validation
performance. This stage is most similar to
the traditional definition and performance
of validation. The testing of Stage 2 should be
commensurate with the risk identified for the
product and process.
Stage 2 comprises demonstration of commercial
process performance by means of conformance lots.
This stage confirms the development work of Stage
1. Successful stage 2 performance demonstrates
that the proposed manufacturing process is capable
of reproducible commercial manufacture. Process
performance qualification (PPQ) conformance
lot manufacturing includes increased testing
to demonstrate acceptability of the developed
formulation and process.
The 2011 validation guidance provides several
specific recommendations for the respective stages
of process validation. Validation managers must
become familiar with these requirements and
incorporate them into their site training programs.
The guidance discusses the following in Stage 2.
Facility, utilities, and equipment. The FDA 2011
guidance specifies the following regarding facility,
equipment, and utilities:
t
Utilities and equipment construction materials,
operating principles, and performance
characteristics must be appropriate for their
specific use.
t
Utilities systems and equipment must be
built and correctly installed, according to
manufacturers directions, and then properly
maintained and calibrated.
t
Utility system and equipment must be
qualified to operate in the ranges required in
processing. The equipment should have been
qualified under production-level loads and for
production-level durations. Testing should also
include interventions, stoppage, and start-up as
is expected during routine production.
The 2011 guidance provides specific expectations
for a plan to qualify facility, equipment, and
utilities, as follows:
t
The plan should include risk management to
prioritize activities and documentation
t
The plan should identify
(1) Studies or tests to use
(2) Criteria appropriate to assess outcomes
(3) Timing of qualification activities
PROCESS VALIDATION Process Design 5
Paul L. Pluta
(4) Responsibilities
(5) The procedures for documenting and
approving the qualification
t
Change evaluation policy
t
Documentation of qualification activities
t
Quality assurance (QA) approval of the
qualification plan.
The above is a clear directive to the site
validation approval committee (VAC) as to
FDAs expectations for facilities, equipment, and
utilities qualification.
Process performance qualification. The PPQ is
intended to confirm the process design and
development work and demonstrate that the
commercial manufacturing process performs as
expected. This stage is an important milestone in
the product lifecycle. The PPQ should be based on
sound science and experience. The PPQ should
have a higher level of testing and sampling. The
goal of the PPQ is to demonstrate that the process
is reproducible and will consistently deliver quality
products.
PPQ protocol. A written protocol is essential and
should discuss the following:
t
Manufacturing conditions, process parameters,
process limits, and raw material inputs
t
How data are to be collected and evaluated
t
Testing and acceptance criteria
t
Sampling plan including sampling points and
number of samples
t
Number of samples should demonstrate
statistical confidence
t
Confidence level based on risk analysis
t
Criteria for a rational conclusion of whether the
process is acceptable
t
Statistical methods used to analyze data
t
Provision to address deviations and
non-conformances
t
Design of facilities, qualification of equipment
and facilities
t
Personnel training and qualification
t
Verification of sources of materials and
containers and closures
t
Analytical method validation discussion
t
Approval by appropriate departments and the
quality unit.
PPQ protocol execution and report. Protocol
execution should not start until the protocol has
been approved. Changes to the approved protocol
must be made according to established procedures.
The routine manufacturing process and procedures
must be followed (i.e., usual conditions, personnel,
materials, environments, etc.). The PQ report
should do the following:
6 PROCESS VALIDATION Process Design
t
Discuss and cross-reference all aspects of the
protocol
t
Summarize and analyze data
t
Evaluate unexpected observations and
additional data not specified in the protocol
t
Discuss deviations and non-conformances
t
Describe corrective actions
t
State a clear conclusion whether the process
is validated or if not, what should be done to
validate the process
t
Be approved by appropriate departments and
the quality unit.
Stage 3Continued process verification.
This stage may be simply described as maintaining
validation, or maintaining the validated state.
Maintenance activities of Stage 3 should be
commensurate with the risk identified for the
product and process.
Assuming good development of the process,
identification of potential variation, and control of
same, the manufacturer must maintain the process
under control over the product lifetime (i.e., the
work of Stage 3). This control must accommodate
expected changes in materials, equipment,
personnel, and other changes throughout the
commercial life of the product based on risk
analysis.
Stage 3 comprises the ongoing commercial
manufacturing of the product under the same
or equivalent conditions as demonstrated in
Stage 2. This phase continues throughout the
entire commercial life of the product or process.
Specific topics discussed in this section include the
following:
t
Ongoing program to collect and analyze process
data, including process trends, incoming
materials, in-process material, and finished
products
t
Statistical analysis of data by trained personnel
t
Procedures defining trending and calculations
t
Evaluation of inter-batch and intra-batch
variation
t
Evaluation of parameters and attributes at
PPQ levels until variability estimates can be
established
t
Adjustment of monitoring levels based on the
above
t
Timely assessment of defect complaints, out-of-
specification (OOS) findings, deviations, yield
variations, and other information
t
Periodic discussion with production and
quality staff on process performance
t
Process improvement changes
t
Facilities, utilities, and equipment must be
maintained to ensure process control.

WHY THE LIFECYCLE APPROACH?
For manufacturing processes to be truly validated,
each of the stages must be addressed and
integrated. This integration of development work,
process conformance, and continuing verification
provides assurance that the product or process
will consistently remain in control throughout the
entire product lifecycle. Process validation must
not be considered a one-time event or a focused
one-time task performed just prior to commercial
launch that emphasizes only the manufacture of
three conformance lots. Acceptable manufacture of
three conformance batches must not be interpreted
as completion of validation. These lots cannot
truly represent the future manufacturing process
with unexpected and unpredictable changes.
Conformance lots are often inadvertently biased
(i.e., they may utilize well-characterized and
controlled API and excipients, be manufactured
under well-controlled conditions, be monitored
by expert individuals, and performed by most
experienced or well-trained personnelall best-
case conditions). It is highly unrealistic to contend
that the manufacture of three conformance lots
under best-case conditions conclusively predicts
successful manufacturing over the product lifetime.
True process validation must be a process that is
never completed and is always ongoing.
Is This Really a New Approach?
The lifecycle approach to process validation is not
really a new approach or a new concept (3). In an
interview with FDA investigator Kristen Evans
published in the Journal of Validation Technology in
February 2000, the investigator commented on the
failure of manufacturers to recognize a lifecycle
approach to validation (see Sidebar).
The three-stage lifecycle description of process
validation as discussed in the FDA process
validation guidance integrates various strategies,
approaches, and expectations that had been
mentioned in several published documents,
guidelines, and presentations. FDA representatives
have openly discussed the lifecycle approach to
process validation for several years (4,5,6). The
draft process validation guidance that formally
introduced the lifecycle approach for industry
comment was published in 2008 (7). The lifecycle
approach overcomes the checklist approach to
process validation, whereby, process validation is
considered to be a one-time event. Encouraging
comprehensive process understanding improves
root cause analysis when manufacturing problems
occur. Successfully manufacturing three validation
lots without sufficient process understanding does
not provide good assurance that the manufacturing
Paul L. Pluta
Excerpt from an interview with FDA investi-
gator Kristen Evans published in the Journal
of Validation Technology, Volume 6, No. 2,
February 2000.
Q. What are some of the major process vali-
dation problems you have seen during your
inspections of manufacturing facilities in the
United States?
A. I think, as a whole, the failure to recognize
the lifecycle approach to validation. We see
many firms, for whatever reason, thinking
that once they complete their prospective
three-batch validation, thats the end and
theyre on their way. I like to say that pro-
spective validation is not the end. Its not the
beginning of the end; it is hopefully the end
of the beginning. But, clearly, its an ongo-
ing process. It requires a concerted effort to
really maintain confidence in the process and
to be able to demonstrate that at any given
time. So, when we conduct our inspections,
we want to know how the firm gives itself,
and therefore us, the confidence that a given
process on that day is under control. And
youre not simply saying, Well, we validated
it a few years ago, or Were going to do
our annual review in a couple of months,
and that will show us, but rather, systems
are in place at any given time to show from a
big picture that its validated. As opposed to
general problems discussed in the previous
paragraph, a more specific problem that we
see is a lack of scientific rationale in the pro-
tocols and acceptance criteria. At least there
is a lack of documentation of such rationale,
which is what were expecting to see. We
want the process to be there, that youve
come up with a scientific study, a protocol
this is what youre attempting to show, and
this is why, and this is how its going to be
evaluated, and then just simply executing
that. Thats documentation of the scientific
rationale.
Note: The above comments are the per-
sonal opinions of Mr. Evans and are not FDA
policy.
PROCESS VALIDATION Process Design 7
Paul L. Pluta
process will consistently yield an acceptable product
throughout the product commercial life.
The September 2006 FDA Quality Systems
Approach to Pharmaceutical CGMP Regulations (8)
clearly discusses expectations for maintenance of
the validated state. In discussing performance and
monitoring of operations, the regulations state,
An important purpose of implementing a quality
systems approach is to enable a manufacturer to
more efficiently and effectively validate, perform,
and monitor operations and ensure that the
controls are scientifically sound and appropriate.
Further, Although initial commercial batches
can provide evidence to support the validity and
consistency of the process, the entire product
lifecycle should be addressed by the establishment
of continual improvement mechanisms in the
quality system. Thus, in accordance with the quality
systems approach, process validation is not a one-
time event, but an activity that continues through a
products life. This document also discusses trend
analysis, corrective action and preventive action
(CAPA), change control, and other quality systems
programs.
The FDA Pharmaceutical cGMPs for the 21st
Century-a Risk-Based Approach (9), states the
following:
We have begun updating our current thinking
on validation under a Cross-Agency Process
Validation workgroup led by CDERs Office of
Compliance Coordinating Committee with
participation from CDER, CBER, ORA, and CVM.
In March of this year, FDA began this process
issuing a compliance policy guide (CPG) entitled
Process Validation Requirements for Drug Products
and Active Pharmaceutical Ingredients Subject to
Pre-Market Approval (CPG 7132c.08, Sec 490.100)
(10). The CPG stresses the importance of rational
experimental design and ongoing evaluation of
data. The document also notes that achieving and
maintaining a state of control for a process begins
at the process development phase and continues
throughout the commercial phase of a products
lifecycle. The CPG incorporates risk-based
approaches with respect to inspectional scrutiny;
use of advanced technologies, and by articulating
more clearly the role of conformance batches in
the product lifecycle. The document clearly signals
that a focus on three full-scale production batches
would fail to recognize the complete story on
validation.
In the 2004 revision of FDA Compliance
Policy Guide Sec. 490.100, Process Validation
Requirements for Drug Products and Active
Pharmaceutical Ingredients Subject to Pre-
Market Approval (CPG 7132c.08), expectations
8 PROCESS VALIDATION Process Design
for validated processes are clearly stated. Before
commercial distribution begins, a manufacturer
is expected to have accumulated enough data
and knowledge about the commercial production
process to support post-approval product
distribution. Normally, this is achieved after
satisfactory product and process development,
scale-up studies, equipment and system
qualification, and the successful completion of the
initial conformance batches. Conformance batches
(sometimes referred to as validation batches and
demonstration batches) are prepared to demonstrate
that, under normal conditions and defined ranges
of operating parameters, the commercial scale
process appears to make acceptable product. Prior
to the manufacture of the conformance batches
the manufacturer should have identified and
controlled all critical sources of variability. FDA
has removed reference to manufacture of three lots
as a requirement for validation in this document.
The process validation guidance is consistent
with FDA QbD principles. The various QbD
presentations and publications strongly encourage
demonstrations of process understanding for both
API and drug product (11,12,13,14). In the 2006
FDA Perspective on the Implementation of Quality by
Design (QbD), a QbD system is defined as follows:
t
The API or drug product is designed to meet
patient needs and performance requirements
t
The process is designed to consistently meet
critical quality attributes
t
The impact of starting raw materials and
process parameters on quality is well
understood
t
The process is evaluated and updated to allow
for consistent quality over time
t
Critical sources of process variability are
identified and controlled
t
Appropriate control strategies are developed.
The various FDA guides to inspections (15,16,17),
all issued during the 1990s, emphasized the
development phase of the validated process
and associated documentation. This included
documented experiments, data, results, control
of the physical characteristics of the excipients,
particle size testing of multi-source excipients
and determination of critical process parameters.
Development data serves as the foundation for the
manufacturing procedures, and variables should be
identified in the development phase. Raw materials
were identified as a source of lot-to-lot variation,
as were equipment or processes that could impact
product effectiveness or product characteristics (i.e.,
the validated state must be maintained).

Some of the key concepts in the 2011 process
validation guidance were originally mentioned in
the FDA 1987 guidance. For example, the 1987
guidance states, ...adequate product and process
design...quality, safety, and effectiveness must
be designed and built into the product... and
During the research and development (R&D)
phase, the desired product should be carefully
defined in terms of its characteristics, such as
physical, chemical, electrical, and performance
characteristics. In addition to discussing actual
validation protocols, the document mentions
several post-validation considerations, as follows:
...quality assurance system in place which requires
revalidation whenever there are changes in
packaging, formulation, equipment, or processes
which could impact product effectiveness or
product characteristics, and whenever there
are changes in product characteristics. The
quality assurance procedures should establish
the circumstances under which revalidation is
required.
The 2011 process validation guidance clearly states
its consistency with International Conference on
Harmonisation (ICH) Q8, Q9, and Q10 documents
(18). These documents provide current global
thinking on various aspects of the product lifecycle
from development through commercialization. They
provide a comprehensive and integrated approach
to product development and manufacturing to be
conducted over the lifecycle of the product. ICH Q8
discusses information for regulatory submission
in the ICH M4 Common Technical Document
format (19). ICH Q8 describes a comprehensive
understanding of the product and manufacturing
process that is the basis for future commercial
manufacturing including QbD concepts. ICH Q9
provides a systematic approach to quality risk
management through various risk assessment tools.
ICH Q9 also suggests application of risk management
methods to specific functions and business processes
in the organization. ICH Q10 complements Q8 and
Q9, and discusses the application of the various
quality system elements during the product lifecycle.
Elements of the quality system include process
performance monitoring, CAPA, change control, and
management review. These quality system elements
are applied throughout the various phases of the
product lifecycle.
The 2000 ICH Q7 Good Manufacturing Practice
Guide for Active Pharmaceutical Ingredients (20)
discusses activities conducted prior to and post
validation. For example, ICH Q7 states that critical
parameters or attributes should be identified during
development, and these critical process parameters
should be controlled and monitored. Non-critical
Paul L. Pluta
parameters should not be included in validation.
Regarding post validation, there should be periodic
review of validated systems.
Medical Device Validation Guidance
Although the 2011 process validation guidance
does not apply to medical devices, medical device
documents espouse an equivalent comprehensive
approach to process validation. In the January
2004 Global Harmonization Task Force (GHTF)
Study Group 3, Quality Management SystemsProcess
Validation Guidance (21), activities conducted during
product or process development to understand
the process are described. For example, The use
of statistically valid techniques such as screening
experiments to establish key process parameters
and statistically designed experiments to optimize
the process can be used during this phase. This
document also describes activities conducted post-
validation to maintain the product or process. For
example, Maintaining a state of validation by
monitoring and control including trend analysis,
changes in processes or product, and continued
state of control of potential input variation such as
raw materials. Tools including statistical methods,
process capability, control charts, design of
experiments, risk analysis, and other concepts are
described.
The 1997 FDA Medical Device Quality Systems
Manual (22) further emphasizes activities to be
conducted post validation. It states, Process and
product data should be analyzed to determine what
the normal range of variation is for the process
output. Knowing what is the normal variation
of the output is crucial in determining whether
a process is operating in a state of control and is
capable of consistently producing the specified
output. Process and product data should also
be analyzed to identify any variation due to
controllable causes. Appropriate measures should
be taken to eliminate controllable causes of
variation...Whether the process is operating in a
state of control is determined by analyzing day-to-
day process control data and finished device test
data for conformance with specifications and for
variability.
The 1997 Guide to Inspections of Medical Device
Manufacturers (23) states, It is important to
remember that the manufacturer needs to maintain
a validated state. Any change to the process,
including changes in procedures, equipment,
personal, etc. needs to be evaluated to determine
the extent of revalidation necessary to assure
the manufacturer that they still have a validated
process.
PROCESS VALIDATION Process Design 9
Paul L. Pluta
APPLYING THE LIFECYCLE APPROACH
The concepts identified in the respective stages
of the FDA process validation guidanceprocess
design (understanding), process qualification
(performance), and continued process verification
(maintaining validation)serve as a model for all
areas of validation and qualification. Although
not specifically mentioned in the FDA guidance,
the sequence of understanding, performance,
and maintaining the validated state is certainly
applicable and desirable for other processes in
pharmaceutical manufacturing including packaging,
cleaning, analytical, and so on. Further applying
this sequence to equipment qualification, HVAC,
computer systems, and other areas is also appropriate
and desirable. Presentations on these associated
topics at validation meetings have already been
structured according to this model. The installation
qualification-operational qualification-performance
qualification (IQ/OQ/PQ) model (24) and the
ASTM E2500 (25) model are consistent with
understanding, qualifying, and maintaining
qualification through calibration, preventive
maintenance, change control, and associated
activities. Applying the stages 1, 2, and 3 sequence of
activities to all validation and qualification unifies
the site approach to project management activities,
standardizes expectations, facilitates training, and
generally simplifies organizational thinking.
THE AFFECT ON CURRENT
VALIDATION PROGRAMS
A major concern of validation practitioners gets to
the bottom lineHow does the 2011 guidance
affect current validation programs, and how can the
new guidance be implemented?
Organizational Aspects
The lifecycle approach to process validation
requires commitment from many areas in the
organization. The lifecycle approach must become
part of organizational strategy. This will require a
comprehensive and continuing view of validation
rather than focus on the performance of the usual
three conformance lotsand job done. Many
firms organize their operations in distinct silos
(e.g., R&D, manufacturing, and quality). The silos
create barriers to communication and cooperation.
The R&D organization develops the product. After
development is completed, the product is transferred
to manufacturing. Commercial operations
personnel adjust the process and make it ready for
validation and routine production. The validation
function coordinates process validation. After
the conformance lots are successfully completed,
the validation effort is finished. Manufacturing
10 PROCESS VALIDATION Process Design
then continues routine commercial production
with oversight by the quality unit or the qualified
person (QP). Often there is minimal ongoing
constructive interaction between R&D, validation,
manufacturing, and quality during the product
lifetime.
The lifecycle approach to validation is
clearly different than the above described
situation. Product R&D and technical support
should approach their work as supporting the
entire product lifecycle including commercial
manufacturing. They must be involved in
monitoring and maintenance of the validated
state. Their work should provide the technical
basis or justification for all aspects of
manufacturing including any changes and
necessary improvements. The validation group
should coordinate the process qualification stage
of manufacturing based on technical development
work, and should participate in determining
the ongoing control strategy. The validated state
must be maintained through process monitoring,
technical data evaluation, and change control.
Manufacturing fixes or tweaks should be
evaluated by technical people, and should ideally
be supported by data or sound technical judgment
whenever possible. R&D should be involved in
process improvements and provide the technical
justification for these improvements. Organizations
should foster development of a continuous
business process beginning in R&D and continuing
throughout the entire product lifecycle with
ongoing collaboration and communication among
all relevant organizational areas. The lifecycle
approach to process validation must become a
comprehensive organizational effort.
Validation Performance Specifics
The 2011 guidance describes many specific details
and expectations for Stage 2 and Stage 3. Validation
and quality managers should evaluate their practices
and procedures regarding these specifics. FDA
recommendations for Stage 2 PPQ protocol-related
activities are substantial. FDA recommendations for
Stage 3 post-validation monitoring are significantly
different from a traditional Annual Product Review
approach. Deficiencies in site programs should be
identified and corrective actions or improvements
prioritized. Risk to the patient and to the
organization should be considered in prioritization.
Risk Analysis
Risk assessment has a critical role in all of the
activities described herein. All activities conducted
in the organization should be conducted with risk
in mind. ICH Q9 describes various risk assessment

methods and potential applications of risk
assessment. There are numerous applications of risk
management used during the entire process validation
lifecycle. Examples cited in ICH Q9 relevant to process
validation include product and process development,
facilities and equipment design, hygiene aspects
in facilities, qualification of equipment, facility, or
utilities, cleaning of equipment and environmental
control, calibration and preventive maintenance,
computer systems and computer controlled
equipment, and so on. In brief, risk assessment helps
to identify the most important potential problems
in all three stages of process validation, and then
addresses these problems appropriately. There should
be consistency between the risk-based activities in all
three stages of process validation. Risk management
must become pervasive in the organization.
Training
The issuance of the 2011 FDA guidance requires
appropriate training for all involved in validation-
related activities. All involved in validation and
validation- or qualification-related activities must
be aware of the 2011 process validation guidance
and concepts therein. Personnel who previously
considered themselves to be apart or distant from
commercial product validation (e.g., development
scientists) must now be included in validation
training.
Especially important are personnel who
write validation plans, protocols, results, and
associated documents. These writers must think
comprehensively, incorporating pre-validation
development information as well as considerations for
post-validation maintenance of the validated state into
their validation documentation.
Also critically important for training are the site
VAC members. There must be a clear understanding
and agreement among VAC members and the
validation group as to their functions and
responsibilities. Clearly stating the responsibilities
of the VAC provides focus and expectations for the
VAC. Clearly stating the responsibilities of the VAC
provides clear expectations for those submitting
protocols, validation plans, and other documents
for VAC review and approval. VAC members must
maintain awareness and compliance with the 2011
process validation guidance. The VAC members
should consider themselves to be a surrogate FDA
(or other regulatory agency) auditor. The VAC should
assume responsibility for site preparedness for
future regulatory audits of the validation function.
Future audits will certainly include concepts and
recommendations stated in the 2011 process
validation guidance.
Paul L. Pluta
Terminology
The terminology associated with the various
phases of validation has had minor variations over
the years. The 2011 process validation guidance
describes process design, process qualification, and
continued process verification stages in the validation
lifecycle. Stage 2 Process Qualification includes
PPQ manufacturing of commercial lots. The 1987
FDA validation guidance describes installation and
operational qualification, process performance
qualification, and product performance qualification.
Products lots manufactured in the process
qualification phase were termed conformance lots.
PPQ batches have also been named demonstration
lots, qualification lots, PQ lots, and validation
lots, in past years. Stage 2 process qualification
phase also includes equipment, facilities, and utilities
qualification.
While the variety of terminology used may
cause difficulties in communicating, the intent
of all validation programs is the same: Sequential
process understanding, validation performance,
and maintaining the validated state as described
herein comprise the validation lifecycle continuum.
Validation programs addressing these phases of
the product or process lifecycle, no matter what
specific terminology is used or how categorized
in documentation, will meet the expectations
robustness, repeatability, and reliability for validated
process. Regulatory investigators are knowledgeable
and able to interpret different organizational
terminology as long as the sequence of process
understanding, validation performance, and
maintaining the validated state are demonstrated.
Documentation
All work associated with process validation in
all stages of the validation lifecycle must be
documented. This includes product and process
design, experimental and development studies for
process understanding, risk analysis in development,
designed experiments, process parameter
optimization, validation and qualification protocols,
and process monitoring to maintain the validated
state. Development scientists must understand that
their work is integral to the validation lifecycle.
Development reports may be requested in regulatory
audits. Summary documents are recommended,
especially when multiple documents must be
integrated by the reader. All work associated with
equipment, facilities, and utilities qualification and
analytical validation must also be documented.
Document quality is important; in many cases,
documents are reviewed literally years after they
are written and long after authors have moved on
to new careers inside or outside of the company. All
PROCESS VALIDATION Process Design 11
Paul L. Pluta
associated documents must be readily available.
Documents are often required to be quickly retrieved
in regulatory audits. Document storage in an easily
accessible centralized location is recommended.
Analytical
The guidance briefly discusses expectations for
analytical methodology in process validation. It
states that process knowledge depends on accurate
and precise measuring techniques. Analytical areas
supporting early Stage 1 R&D work must be aware that
their methods and data may be subject to inspection in
validation audits. Test methods must be scientifically
sound (e.g., specific, sensitive, accurate), suitable,
and reliable. Analytical instruments must function
reliably. Analytical method development reports
must be available for auditor review. Procedures
for analytical methods, equipment maintenance,
documentation practices, and calibration practices
should be documented or described. Current good
manufacturing practice CFR 210 and 211 must be
followed as appropriate for batch release of commercial
lots.
Management Support
The support of senior management and the respective
functional management of affected areas in the
organization is critical to implementing the lifecycle
approach. Management in the organization must
become familiar with the 2011 validation guidance
and its ramifications. Transitioning organizations
to the lifecycle approach to validation cannot be
completed without management support. Employees
provide what management expects. Validation and
quality professionals should help their management
to assess the status of their organizations. Deficiencies
must be corrected and enhancements implemented.
Validation and quality professionals should prioritize
activities based on risk to patient and organization.
Economic impact must also be considered. A balance
of risk, cost, and compliance considerations is key to
development of a suitable validation program that will
be embraced and supported.
CONCLUSIONS
FDA issued Process Validation: General Principles and
Practices in January 2011, which has given widespread
visibility to the lifecycle approach concept. This
comprehensive approach to validation has significant
and far-reaching ramifications. Three stages in the
lifecycle approach have been identified. The lifecycle
concept links development, validation performance,
and maintenance of the process in a state of control
during routine commercial production.
12 PROCESS VALIDATION Process Design
FDA has provided recommendations for the general
lifecycle and stages 1, 2, and 3 including specific
details for each of the stages. Stage 1Process
Design may be generally described as process
understanding. Stage 1 work will ultimately be
reflected in the master production record and control
records. Stage 2Process Qualification may be
described as validation performance. This stage
comprises demonstration of commercial process
performance by means of conformance lots and
confirms the development work of Stage 1. Stage
2 details are also provided for design of a facility
and qualification of utilities and equipment, PPQ,
PPQ protocol, and PPQ protocol execution and
report. Stage 3Continued Process Verification may
be simply described as maintaining validation.
This stage comprises the ongoing commercial
manufacturing of the product under the same or
equivalent conditions as demonstrated in Stage 2.
Stage 3 continues throughout the entire commercial
life of the product or process. Understanding
the sources of variation and control of variation
commensurate with risk should be applied during all
stages of validation. The integration of development
work, process conformance, and continuing
verification provides assurance the product or process
will consistently remain in control throughout the
entire product lifetime.
The lifecycle approach is not a new concept. This
approach as described in the guidance integrates
various strategies, approaches, and expectations that
had been mentioned in several previously published
documents, guidelines, and presentations. The
concepts identified in the respective stages of the
FDA process validation guidanceunderstanding,
performance, and maintenanceserve as a model for
all areas of validation.
Implementation of the lifecycle approach in site
validation programs has significant ramifications
for the organization. Organizational functions
previously distant from commercial processes
are now integral to ongoing performance. Post-
validation monitoring of process performance
including timely responsiveness to data trends is an
expectation. The lifecycle approach affects many
areas of validation programs including organizational
aspects, validation performance guidance specifics,
risk analysis, training, and documentation. Senior
and functional management support is needed to
transition organizations to the lifecycle approach to
validation. Risk analysis is key to development and
prioritization of a suitable validation program that
will be embraced and supported.
 Paul L. Pluta

REFERENCES 17. FDA, Guide to Inspections, Oral Solutions and Suspensions,

1. FDA, Guidance for Industry, Process Validation: General August 1994.
Principles and Practices, January 2011. 18. www.ICH.org
2. FDA, Guideline on General Principles of Process Validation, 19. ICH, M4. The Common Technical Document for the
May 1987. Registration of Pharmaceuticals for Human Use, www.ich.
3. Evans, Kristen and Vincent, D., Ask the FDA: Questions org.
a nd A nswers w it h Cur rent a nd For mer A genc y 20. ICH, Q7. Good Manufacturing Practice Guide for Active
Personnel. Critical Factors Which Affect Conducting Pharmaceutical Ingredients, November 2000.
Effective Process Validation: One FDA Investigators 21. GHTF, Study Group 3. Quality Management Systems
View, Journal of Validation Technology, Volume 6, #2, Process Validation Guidance. Edition 2, January 2004.
February 2000. 22. FDA, Medical Device Quality Systems Manual, January
4. Mc Nally, Grace E ., Lifec yc le Approac h Process 1997.
Validation, GMP by the Sea, Cambridge, MD. August 26, 23. FDA, Guide to Inspections of Medical Device Manufacturers,
2008. Process Validation, 21 CFR 820.75, December 9, 1997.
5. Mc Nally, Grace E ., Lifec yc le Approac h Process 24. ISPE G A M P Te st i ng SIG ., Ke y P r i nc iple s a nd
Validation, GMP by the Sea, Cambridge, MD. August 29, C on side r at ion s for Te s t i ng of G x P S y s t e m s ,
2007. Pharmaceutical Engineering, Vol. 25, No. 6, Nov.-Dec.
6. Famulare, Joseph, Benefits of a Pharmaceutical Quality 2005.
System, PDA/FDA Joint Conference, Bethesda, MD, 25. ASTM, ASTM E 2500-07, Standard Guide for Specification,
November 2, 2007. De sig n , an d Ve r if i c at io n of Ph ar m a c e ut ic al an d
7. FDA, Draft Guidance for Industry, Process Validation: Biopharmaceutical Manufacturing Systems and equipment,
General Principles and Practices, November 2008. August 2007. JVT
8. FDA, Quality Systems Approach to Pharmaceutical CGMP
Regulations, September 2006. ARTICLE ACRONYM LISTING
9. FDA, Pharmaceutical cGMPs for the 21st CenturyA Risk- API Active Pharmaceutical Ingredient
Based Approach, Final Report-Fall 2004, September 2004. CAPA Corrective Action and Preventive Action
10. FDA, Compliance Policy Guide 7132c.08. Section DOE Design of Experiments
490.100. Process Validation Requirements for Drug FDA US Food and Drug Administration
Products and Active Pharmaceutical Ingredients Subject ICH International Conference on Harmonisation
to Pre-Market Approval. IQ Installation Qualification
11. Nasr, Moheb, FDA Perspective on the Implementation OOS Out of Specification
of Quality by Design (QbD), 9th APIC/CEFIC, Prague, OQ Operational Qualification
Czech Republic, October 10, 2006. PPQ Process Performance Qualification
12. Yu, Lawrence X., Robert Lionberger, Michael C. Olson, PQ Performance Qualification
Gordon Johnston, Gary Buehler, and Helen Winkle, QbD Quality by Design
Quality by Design for Generic Drugs, PharmTech.com, QP Quality Person
October 2, 2009. R&D Research and Development
13. Winkle, Helen H., Implementing Quality by Design, VAC Validation Approval Committee
PDA/FDA Joint Regulatory Conference, September 24,
2007.
14. Yu, Lawrence X., Pharmaceutical Quality by Design:
Product and Process Development, Understanding,
and Control, Pharmaceutical Research online, January
10, 2008.
15. FDA, Guide to Inspections of Oral Solid Dosage Forms.
Pre/Post Approval Issues for Development and Validation,
January 1994.
16. FDA, Guide to Inspections of Topical Drug Products, July
1994.

Originally published in the Spring 2011 issue of The Journal of Validation Technology

PROCESS VALIDATION Process Design 13

John McConnell, Brian K. Nunnally, and Bernard McGarvey

The Forgotten Origins of

Quality by Design
John McConnell, Brian K. Nunnally, and Bernard McGarvey

t
QbD then proceeds with identification of critical

Analysis and Control of Variation is dedicated
to revealing weaknesses in existing approaches to
understanding, reducing, and controlling variation
and to recommend alternatives that are not only
based on sound science, but also which demonstrably
work. Case studies will be used to illustrate both
problems and successful methodologies. The
objective of the column is to combine sound science
with proven practical advice.
Reader comments, questions, and suggestions
will help us fulfill our objective for this column.
Case studies illustrating the successful reduction
or control of variation submitted by readers are
most welcome. Please send your comments and
suggestions to column coordinator John McConnell
at john@wysowl.com.au or journal coordinating
editor Susan Haigney at shaigney@advanstar.com.
KEY POINTS DISCUSSED
The following key points are discussed:
t
Quality by design can be traced to the original
work of Dr. Joseph M. Juran.
t
Juran believed that nearly all problems (e.g.,
deviations, quality defects, etc.) are built into
the system, and that they can be traced to
inadequate design of the process. Therefore, if
addressed during the design phase, they can be
largely designed out.
t
QbD starts with the quality target product
profile. Juran calls this defining the customer
needs.
quality attributes (CQAs). Juran calls this
development of product features. These should
be reviewed by both customer and producer.
t
Internal service groups must feel competition
from outsourced services to continually improve
their services, reduce variation, and positively
impact product throughout the product lifecycle.
t
Once CQAs have been defined, the
manufacturing process to meet these CQAs can
be designed including technology and human
components
t
Sampling and sample handling variability
should be quantified and minimized.
t
Capability (CpK) is the measure of a process
ability to meet its targets or specifications.
t
The final aspect of QbD is designing a control
strategy, which Juran called developing process
controls.
t
Understanding the impact of variation, and
designing it out wherever possible is an
important part of QbD.
t
The dominant variables of a process can be
described as set-up, components, time, and
worker elements.
t
Knowing which is likely to be dominant
significantly impacts on design.
t
Designing a control strategy that ensures a robust
process depends on understanding the sources of
variation for each step (or assay) and controlling
them. Understanding which category a variable
helps to determine how best to attack variability.

[
ABOUT THE AUTHORS
For more Author John McConnell is owner and director at Wysowl Pty Ltd. in Queensland, Australia. He may be
information, reached at john@wysowl.com.au. Brian K. Nunnally, Ph.D., is in charge of process validation at
go to
Pfizer in Sanford, NC. He may be reached at nunnalb@wyeth.com. Bernard McGarvey, Ph.D., is
Process Modeling Group Leader, Process Engineering Center at Eli Lilly and Company in Indianapolis,
gxpandjvt.com/bios
IN. He may be reached at mcgarvey_bernard@lilly.com.

14 PROCESS VALIDATION Process Design


t
Jurans work, the basis for QbD, is now described
in ICH Q8 (R2), Pharmaceutical Development.
INTRODUCTION
The US Food and Drug Administration and the
International Conference on Harmonisation (ICH)
have embraced the concept of building quality into
pharmaceutical manufacturing processes. This
has been called quality by design (QbD). Many
pharmaceutical scientists are not aware of the origins
of QbD, nor are they familiar with its originator, Dr.
Joseph Juran.
The QbD concept was originally published in
1985. The seminal work is Juran on Quality by Design
(1). This book outlines the reasons and methodology
for planning quality into a manufacturing process.
Juran did not use pharmaceuticals or medical devices
in his book, but all of the principles are present and
have been adapted into ICH Q8 (R2), Pharmaceutical
Development (2).
BASICS OF QBD
For Juran, quality problems are planned that way
(i.e., they are built into the system) (1). He believed
that all quality defects were a result of a poorly
planned quality system. The Juran trilogy is shown
in Figure 1. We have shown this as an interlocking
sequence, as we are sure Juran believed it to be.
The corollary to this, and the most important
aspect for the pharmaceutical and medical device
industry, is that the quality problems can be
prevented by understanding the process better for
the determination of design space, identification
of critical quality attributes, and defining an
appropriate control strategy.
Quality Target Product Profile
QbD starts with an understanding of the quality
target product profile (QTPP). This is defined as a
prospective summary of quality characteristics of
a drug product (or medical device, drug substance,
etc.) that ideally will be achieved to ensure the
desired quality, taking into account safety and
efficacy of the drug product (or medical device,
drug substance, etc.) (2). Juran would call this
defining the customer needs. Juran states, the goal
should be customer satisfaction rather than mere
conformance to stated needs (1). Translated for the
pharmaceutical industry (note: medical devices are
included but will not be further explicitly discussed),
this would mean a focus on producing medicines
that improve the lives of the patients we serve. Our
industry is not involved in the production of widgets;
we produce life-saving and life-enhancing medicines
designed to improve, save, and elongate the lives of
our customers. In this light, the best way to serve
John McConnell, Brian K. Nunnally, and Bernard McGarvey
Figure 1: The Juran trilogy (adapted from 1).
Quality Quality
Planning Control
Quality
Improve-
ment
our customers is to be a customer, as Juran would
probably say. We need to think like a patient and
not forget the patient in all of our work. The quality
cannot be built into the drug until the parameters
needed are defined (e.g., the route of administration,
dosage, strength, etc.)
Critical Quality Attributes
Once the QTTP is defined, work on defining the
critical quality attributes (CQAs) can commence.
Juran would call this process the development of
product features. His advice rings true with todays
development. Product development requires
not only functional expertise; it also requires
the use of a body of quality-related know-how
(1). Development scientists need a thorough
understanding of manufacturing and quality as their
data and experiments form the basis of the processes
and methods used during commercial manufacture.
Juran would be supportive of the efforts made by
manufacturing to have influence on the development
process and by development to better understand the
needs of commercial manufacturing and the product
released to the marketplace. It is the job of both to
ensure the planned approach includes work to guard
against external failures. These include product and
process design-related issues, carryover from previous
processes, and degradation (2). The carryover of
previous processes (think toolbox technologies) can
contribute to future quality issues by embedding these
into new processes.
The CQAs are those product (or drug substance)
characteristics having an impact on product quality
and should be studied and controlled (2). Juran
advises that this criticality needs to be reviewed by
PROCESS VALIDATION Process Design 15
John McConnell, Brian K. Nunnally, and Bernard McGarvey
Figure 2: Pharmaceutical adapted Juran pro-
cess model (adapted from 1).
QTPP
Inputs
CQAs
Product
Process
Development
Process
Outputs
Features
both customer and supplier (1). This means both
the patient and the business are stakeholders. The
patient benefits from a robust process that ensures
those attributes central to the safety and efficacy of
the drug are predictable. Likewise, the business
benefits through reduction of cost and increase in
volume (see Littles Law [3,4]), lowered frequency of
deviations, improved utilization of human capital
and fixed assets, and less rework and waste. The
CQAs will be dependent on the various inputs
to the processJuran predicted this as well (1).
Minimizing the variability in these inputs (including
process parameters and raw materials, sampling, and
analytical) will improve the ability to determine and
maintain the CQAs.
Internal Monopolies
In his discussions of the development of product
features, Juran has an interesting sidebar on internal
monopolies. Internal monopolies were once the
rule in the pharmaceutical industry. Services such
as clinical trial management, active pharmaceutical
ingredient (API) manufacture, and testing were
all done in-house by the companys employees.
With the increase in outsourcing, there is really
no such thing as an internal monopoly. All of the
services mentioned previously are being performed
by contract organizations. Juran predicted the
beginnings of this when he advised that competitive
information from outside suppliers of similar
services could be acquired to ensure the highest
quality, speed, and value are being obtained (1).
This put pressure (used in its least pejorative sense)
on the internal monopoly to demonstrate its
competitiveness. This benchmarking is critical to the
survival of the internal monopoly. A word of caution
on outsourcing: While it seems to be all benefit
with no cost, this is not always the case. Part of
the agreement between the company and the service
providers has to include tough variability reduction
(per Demings approach to quality) of operating
16 PROCESS VALIDATION Process Design
parameters, output characteristics, and assays.
Variability should be identified and controlled to
ensure patient and business success. This is not
an end, but should be considered to be a process
encompassing the entire lifecycle history of the
product and process.
Critical Process Parameters
Once CQAs have been defined, the manufacturing
process to meet these CQAs can be selected or
designed (2). This process includes both technology
as well as the human components (1). The
human factor is built into the process and must be
considered in the design to error proof the human
element. Too many deviations are seen that are
blamed on human root causes when, in fact,
these are process design elements that could have
designed the problem out. We have been told about
companies where up to 50% of the root causes are
human error. For these situations, we predict these
same 50% of the deviations will happen again due to
inadequate corrective and preventative actions.
Juran offered that process design should be goal
oriented, systemic, capable, and legitimate (1). A
pharmaceutical adapted Juran model of process
development is shown in Figure 2. A process cannot
meet a target not specified. The following are
examples:
t
What are the quality goals to be met?
t
How much variation can be tolerated from
the process?
t
Sampling?
t
Analytical methods?
t
What output is needed?
t
How can the process be scaled to meet demand
(higher or lower)?
All of these questions (and more) must be
discussed to prepare a comprehensive set of
goals for the process. The process is a series of
interconnected systems; each one having an
impact on the other. For instance, variability
in analytical results used by the process as part
of the manufacture (e.g., used by the process to
determine how the next step will be run) have a
larger impact on the overall variability than those
that are not used to run the process. It is important
to remember that the laboratory tests the sample,
not the process. Sampling (and sample handling)
variability cannot be ignored and should be
quantified and minimized as it ultimately affects
the overall variability of the process. Capability is
crucial to meeting the process goals. Capability is
the measure of a process ability to meet its targets
or specifications. The key aspect of capability (CpK)
is variability, as shown in the following equation:

where
USL is the upper specification limit
LSL is the lower specification limit
x is the mean
is a standard deviation.
There are only three ways to improve capability,
as follows:
t
Move the specification away from the
mean. This is a difficult task to accomplish
in the pharmaceutical industry without
significant data and lengthy regulatory agency
discussions.
t
Move the mean away from the specification.
For double-sided specifications, this only
works if the mean is not centered.
t
Decrease the variability. Given the fact
that this is multiplied by three, this has the
greatest power to improve capability. In an
extreme understatement, Juran suggested this
evaluation had merit (2). More than this, it
is a critical part of process understanding.
It is important to remember that it is impossible
to know or understand the capability of an
unpredictable process. Finally, the process
must be legitimate. Juran states, ...if there is the
approval of those to whom responsibility has been
delegated (1). This is not a common concern in the
pharmaceutical industry where clear accountability
is often present.
These are critical steps that are important to the
determination of the CQAs. The planning associated
with these steps should also include planning of the
quality system. This plan includes alarm strategy,
redundancy, and design to improve robustness and
decrease variability. The final process including
manufacturing, sampling, and analytical should be
designed and demonstrated to be in control (e.g.,
predictable), appropriate for its intended use, and
capable of meeting the defined CQAs (and thus the
QTPP).
Control Strategy
The final aspect is designing a control strategy
(2) that Juran called developing process controls
(1). This should be done in concert with process
development. The concept of design space comes
into play at this point. So much has been written
and presented on design space that we will not
discuss this in great detail in this article. Design
space is the relationship of the process inputs to the
John McConnell, Brian K. Nunnally, and Bernard McGarvey
CQAs (2). Juran did not utilize the terminology, but
did describe the different types of variables present
in a process. He described the concept of dominant
variables, defined as those variables that are more
important than others. The variables described
include the following:
t
Set-up dominant variables. These are
variables affecting processes that are dependent
on the setup and confirmation of the setup to
provide stability and reproducibility. Equipment
and supplier differences can contribute to
variability for these operations. Automated
assays and packaging machines fall into the set-
up dominant category.
t
Time dominant variables. These are variables
affecting processes that change over time and
need to be adjusted based on evaluation of
control parameters. An example of this variable
is ultra filtration operations. It is important to
be vigilant against over-control (e.g., needlessly
adjusting the process based on the last result
obtained).
t
Component dominant variables. These are
variables affecting processes in which the main
variable is the quality of the inputs or materials.
The variability in drug substance can be a large
source of variability for a drug product process.
Many chemical reactions fall in this category
as well. In these instances reducing variability
upstream is critical (Littles law).
t
Worker dominant variables. These are
variables affecting processes where the quality
attributes are determined by the skill of the
worker. Many types of assays fit into this
category. Non-automated processes can be
included here as well. Ensuring operators
or analysts performing the same functions
or on different shifts have identical practices
is important to reducing variability in this
category.
Designing a control strategy that ensures a robust
process depends on understanding the sources of
variation for each step (or assay) and controlling
them. Understanding which category a variable
helps to determine how best to attack variability.
The closest comparison to the design space concept
is what Juran referred to as the Data Base (1). He
described this as a compendium of lessons learned
from human experiences. He referred to the exercise
of reviewing the information as the Santayana
Review named after George Santayana who stated,
those who cannot remember the past are doomed
to repeat it (1). The purpose of this process is to
improve decision-making by using the lessons of
the past. For the pharmaceutical industry, the
PROCESS VALIDATION Process Design 17
John McConnell, Brian K. Nunnally, and Bernard McGarvey

design space is reviewed and documented into REFERENCES

process experience reports that are summarized in
the development history section of the common
technical document format of the dossier. Having
this information codified in a single location is
critical to resolve process upsets and to continue
process improvements post marketing.
FINAL THOUGHTS
The concept initially developed by Joseph Juran
is the origin of the modern QbD movement in
the pharmaceutical industry. All of the relevant
parts are there; long before they were published in
the ICH. A review of these lessons forms the first
principles understanding of QbD.
1. Juran, Joseph M., Juran on Quality by Design, The Free
Press, 1992.
2. ICH, Q8 (R2), Pharmaceutical Development, 2009.
3. Nunnally, Brian K, and J.S. McConnell, Six Sigma in the
Pharmaceutical Industry, Taylor and Francis, 2007.
4. McConnell, John, Brian K. Nunnally, and Bernard
McGarvey, Having It All, Journal of Validation Technology,
Vol. 16, No. 2, Spring 2010. JVT
ARTICLE ACRONYM LISTING
CQA Critical Quality Attributes
FDA US Food and Drug Administration
ICH International Conference on Harmonisation
QbD Quality by Design
QTPP Quality Target Product Profile

Originally published in the Summer 2010 issue of The Journal of Validation Technology

18 PROCESS VALIDATION Process Design

 Deliang Zhou

Understanding Physicochemical
Properties for Pharmaceutical
Product Development and
ManufacturingDissociation,
Distribution/Partition, and Solubility
Deliang Zhou

Welcome to Product and Process Design.
Product and Process Design discusses scientific and
technical principles associated with pharmaceutical
product development useful to practitioners in
validation and compliance. We intend this column to
be a useful resource for daily work applications. The
primary objective for this feature: Useful information.
Information developed during product and
process design is fundamental to drug and
product development and all subsequent activities
throughout the product lifecycle. The quality-by-
design (QbD) initiative encourages understanding
of product and process characteristics and process
control based on risk analysis. Process design
is the first stage of pharmaceutical process
validation as described in the recent US Food
and Drug Administration process validation draft
guidance. This stage comprises activities that
support product and process knowledge leading
to a consistent manufacturing process throughout
the product lifecycle. Product development
work provides much of the information in the
product and process design stage, including active
pharmaceutical ingredient information, dosage
form characteristics, quality attributes, and so
on. Work conducted in development supporting
product and process design stage must be based
on scientific and technical principles. Disciplines
supporting this work include basic chemistry and
pharmaceutics; pharamcokinetics including drug
absorption, metabolism, distribution, and excretion;
biopharmaceutics information, and so on. The
principles associated with these areas are broad
and complex. Also, the technical language and
mathematics associated therein may be esoteric and
intimidating. This column addresses topics relevant
to product and process design with these difficulties
in mind. It is our challenge to present these topics
clearly and in a meaningful way so that our readers
will be able to understand and apply the principles
in their daily work applications.
Reader comments, questions, and suggestions are
needed to help us fulfill our objective for this column.
Please send your comments and suggestions to column
coordinator Yihong Qiu at qiu.yihong@abbott.com
or to journal coordinating editor Susan Haigney at
shaigney@advanstar.com.

[
For more Author ABOUT THE AUTHOR
information, Deliang Zhou, Ph.D., is an associated research investigator in Global Formulation Sciences, Global
go to Pharmaceutical R&D at Abbott Laboratories. He may be reached at deliang.zhou@abbott.com.
gxpandjvt.com/bios Yihong Qiu, Ph.D., is the column coordinator of Product and Process Design. Dr. Qiu is a re-
search fellow and associate director in Global Pharmaceutical Regulatory Affairs CMC, Global Pharma-
ceutical R&D at Abbott Laboratories. He may be reached at qui.yihong@abbott.com.

PROCESS VALIDATION Process Design 19

Deliang Zhou
KEY POINTS
The following are key points addressed in this
discussion:
t

6OEFSTUBOEJOH
UIF
QIZTJDPDIFNJDBM
QSPQFSUJFT

of the active pharmaceutical ingredient (API)
is fundamental to pharmaceutical product
development, manufacturing, and stability
t

.PTU
QIBSNBDFVUJDBMT
BSF
PSHBOJD
NPMFDVMFT

compound is the basis for applications of the
many of which are weak acids or weak bases
that partially dissociate in aqueous solution
t

%JTTPDJBUJPO
NBZ
CF
DIBSBDUFSJ[FE
CZ

acidity dissociation constant (K a) or basicity
dissociation constant (K b), and their
corresponding pK a and pKb
t

#VGGFS
TZTUFNT
TVDI
BT
DJUSBUF
BOE
QIPTQIBUF

buffers that resist pH changes upon the
addition of small quantities of acid or base are
based on dissociation behavior or weak acids
and weak bases
t

"DJECBTF
EJTTPDJBUJPO
IBT
TJHOJGJDBOU
JOGMVFODF

on almost all other physicochemical properties
including salt formation, API purification,
formulation, processibility, solubility, stability,
mechanical properties, and biopharmaceutical
performance
t

%JTUSJCVUJPO
PS
QBSUJUJPO
SFGFST
UP
UIF
SFMBUJWF

solubility of a drug between aqueous and non-
aqueous immiscible liquids
t

%JTUSJCVUJPO
PS
QBSUJUJPO
JT
DIBSBDUFSJ[FE
CZ

logD or logP
t

5IF
EJTUSJCVUJPOQBSUJUJPO
QIFOPNFOPO
JT

important in manufacturing and laboratory
extraction processes to separate organic
and inorganic compounds. Partition
is the foundation underlying various
chromatographic techniques
t

%JTUSJCVUJPOQBSUJUJPO
JT
JNQPSUBOU
JO

determining the biological, pharmacological,
pharmaceutical, and toxicological activities of
a drug molecule. It is widely used in designing
drug candidates to possess appropriate
absorption, distribution, metabolism, and
excretion (ADME) properties
t

n-Octanol/water is the most commonly used
system to characterize drug partitioning
t

"
TPMVUJPO
JT
B
TJOHMFQIBTF
NJYUVSF
DPOTJTUJOH

of two or more components, usually solute and
solvent
t

"
TPMVUJPO
JO
XIJDI
UIF
TPMVUF
JT
BCPWF

the solubility limit (supersaturated) is not
thermodynamically stable and will ultimately
lead to solute crystallization
t

5IF
6OJUFE
4UBUFT
1IBSNBDPQFJB
IBT
EFGJOFE

solubility definitions such as very soluble,
freely soluble, slightly soluble, and other terms
20 PROCESS VALIDATION Process Design
t

.PMFDVMBS
JOUFSBDUJPOT
JNQPSUBOU
JO
UIF

solution process include Van der Waals forces,
hydrogen bonds, and ionic interactions
t

5FNQFSBUVSF
HSFBUMZ
JOGMVFODFT
UIF
TPMVCJMJUZ

of a drug molecule and is always noted when
solubility values are reported
t

5IF
TPMVCJMJUZ
QSPEVDU
PG
BO
JOTPMVCMF

common ion effect
t

5IF
BRVFPVT
TPMVCJMJUZ
PG
B
XFBL
BDJE
PS
CBTF
JT

strongly pH-dependent due to relative species
abundances at different pH
t

5IF
TPMVCJMJUZ
PG
QPMZNPSQIT
DBO
CF

significantly different because they have
different free energies. The metastable
polymorph can convert to the more stable
polymorphs through solution-mediated
transformation. The solubility difference
among various solid forms is one of the driving
forces responsible for phase changes during
pharmaceutical processing.
t

4PMVCJMJUZ
QMBZT
B
DFOUSBM
SPMF
JO
B
OVNCFS
PG

issues related to drug dissolution, absorption,
formulation and process development and
manufacture, as well as stability
t

4PMVCJMJUZ
JT
PG
QBSBNPVOU
TJHOJGJDBODF
UP
UIF

biological performance of a drug molecule.
The physiology of the gastrointestinal
(GI) tract plays an important role in drug
solubilization, including the pH environment
and solubilization by bile salts. pH change is
particularly relevant for weakly acidic or basic
drugs. Solubility changes along the GI tract
could also cause changes such as formation of
salt, free acid or base, and hydrate, amorphous,
and polymorph conversion.
t

4PMVCJMJUZ
JT
FYUFOTJWFMZ
VUJMJ[FE
JO

crystallization and purification of active
pharmaceutical ingredients and intermediates.
Experimental conditions, such as pH, solvent,
antisolvent, temperature, and other factors may
be modified to optimize processes, control
solid forms, and maximize yields.
t

4PMVCJMJUZ
JT
POF
PG
UIF
VOEFSMZJOH
SFBTPOT

during many process-induced phase
transformations such as hydrate formation
during wet granulation and amorphous
transition during drying.
INTRODUCTION
Physicochemical properties are the foundation
for pharmaceutical product development,
manufacturing, and stability. They also provide
insights to the biopharmaceutical performance of
a drug molecule. A general appreciation of these

subjects is important for scientists and engineers
involved in pharmaceutical product development
and manufacturing.
This column discusses basic physical properties
including: Dissociation of weak acids and bases,
distribution and partitioning between aqueous and
non-aqueous liquids, and solubility. This column
starts with basic definitions, follows with important
considerations, and completes each discussion with
practical relevance. Particular attention is given to
the contributions of dissociation to distribution and
solubility. The significance of these basic properties
is highlighted in the contexts of biopharmaceutics,
pharmaceutical development, manufacturing, and
stability.
DISSOCIATION AND THE DISSOCIATION
CONSTANT
The vast majority of pharmaceuticals are essentially
organic molecules, a large fraction of which are
weak acids or weak bases. Typical acidic groups
include carboxylic acid (COOH), sulfonic acid
(SO3H), sulfonamide (SO2NH), and phenols
(C6H5OH). Primary, secondary, and tertiary
amines, as well as pyridines and other aromatic
amines, are the most frequently encountered basic
groups. These molecules tend to dissociate partially
in aqueous solutions, where equilibrium exists
between the various ionized and unionized species.
The Ionization Equilibrium
The dissociation of a monoprotic weak acid, HA,
can be represented as follows:
HA H+ + A
Similar to chemical equilibrium, the dissociation
equilibrium is defined as follows:
[H+][A]
Ka =
[HA]
Where Ka is known as the acidity dissociation
constant or simply as the ionization constant, and the
bracket [ ] represents concentration of a particular
species at equilibrium. The concentrations, instead
of activities, are used throughout because of their
simplicity, and appropriateness for the purpose of
this general discussion. Similarly to the notion
of pH, pKa is more commonly used to denote
the negative logarithm of the acidity dissociation
constant. The pKa value, often a small positive
number, is more convenient because most Ka values
are so small. It is obvious that the acidity decreases
with pKa. A typical organic carboxylic acid has pKa ~
5, while phenols and sulfonamides are weaker, with
typical pKa values of 8-10.
Deliang Zhou
Similar treatments can be applied to a weak base:
B + H 2O BH+ + OH
In this case, the proton is transferred from the
water to the base, generating a hydroxyl ion. The
basicity dissociation constant is then as follows:
[BH+][OH]
Kb =
[B]
By convention, the concentration term of water
is dropped because it is in large excess and its
concentration does not change appreciably. It is
also well known that water itself is very weakly
dissociable to produce a proton and a hydroxyl ion
as follows:
H 2O H+ + OH
The self-dissociation constant of water is often
designated as Kw = [H+] [OH], which equals to 1
10 14 at 25C. Substituting the Kw in the previous
equation, one obtains the following:
Kw[BH+] Kw
Kb = =
[H+][B] Ka
where Ka is the acidity constant of BH+, the
conjugate acid of the base B, as shown below:
BH+ H+ + B
[H+][B]
Ka =
[BH+]
Therefore, the basicity constant of a weak
base can be represented by the acidity constant
of its conjugate acid as: pKa = pKw pKb, as is
conventionally done. Basicity increases with
increasing pK a, contrary to the acidity.
Acids and bases can be categorized as monoprotic,
diprotic, or polyprotic based on the number of
protons they can accept or donate. Multiple ion
equilibriums exist simultaneously for a polyprotic
electrolyte, and each has its own dissociation
equilibrium. For example, phosphoric acid
dissociates in the following three steps:
Ka1
H 3PO4 H+ + H 2 PO4
Ka2
H 2 PO4 H+ + HPO42
Ka3
HPO42 H+ + PO43
These stepwise pKas for phosphoric acid are 2.21,
7.21, and 12.7, respectively. The acid becomes
progressively weaker.
PROCESS VALIDATION Process Design 21
Deliang Zhou
Figure 1: Species distribution as a function of
pH for succinic acid.
1.0
0.8
H2 Suc
H Suc
2
Suc
0.6
Fraction
0.4
0.2
0
0 2 4 6 8 10
pH
Species distribution and buffer capacity are direct
results from dissociation and are discussed next.
Species Distribution In Solution
In an aqueous solution of a weak electrolyte various
species, such as the unionized parent, the proton,
and the ion, coexist. The relative fraction of each
species depends on its pKa value and the solution pH,
which can be calculated based on the dissociation
equilibrium. For example, consider a monoprotic
acid HA:
log
( )
[A]
[HA]
= pH - pKa
Therefore, the ionized species will dominate at
pH > pKa, while the unionized neutral molecule
will dominate at pH < pKa. To preferably target a
certain species, either ionized or unionized, we
often follow the rule of two pH units from pKa,
because the ratio is either 100 to 1 or 1 to 100. This
is particularly useful in separation science such as
high performance liquid chromatography (HPLC),
because mixed mode partition can cause peak
broadening or even shouldering. Similar equations
can be worked out for bases. An example species
distribution is shown in Figure 1 for succinic acid.
The hydrogen succinate displays a maximum
around pH of 5.
Buffer Capacity Of Weak Electrolyte
Solutions
Weak acids and bases are often used as buffering
agents (e.g., citrate buffer, phosphate buffer). A
buffered solution, namely, refers to its ability
to resist pH changes upon the addition of small
quantities of acid or base. Quantitatively, buffer
capacity, , is the ratio the added strong base (or
acid) to the resulted pH change.
22 PROCESS VALIDATION Process Design
For example, consider the following equation for a
weak acid:
pH = pKa + log
( )
[A]
[HA]
= pKa + log
[base form]
[acid form]
This equation is known as the buffer equation or
the Henderson-Hasselbalch equation, which describes
how pH changes with the ratio of its species
concentrations. The right most part of the equation
is more general as it also applies to a weak base.
It can be demonstrated that the maximum buffer
capacity occurs at pH = pK a, where concentrations
of the acid and base forms are equal. The exact
buffer capacity can be obtained by differentiating
12
the Henderson-Hasselbalch equation:
[H+]Ka
= d[base]/dpH = 2.303 Ctot =
([H+]Ka)2
This is the exact reason that most buffers are
used around their pKa values in order to achieve
their intended function. Unlike the strong acid and
base buffers, a weak acid and base buffer allows
the buffering pH to be maintained while adjusting
the buffer capacity by changing the total buffer
concentration, Ctot.
Significance Of Dissociation
Acid/base dissociation is a simple concept, yet
has significant influences on almost all other
physicochemical properties. Salt formation is
well built upon the dissociation concept, which
has been extensively exploited during active
pharmaceutical ingredient (API) manufacture for
purpose of purification, processibility, and other
aspects of pharmaceutical development such as
solubility, stability, mechanical properties, and
biopharmaceutical performance. A general rule
of thumb for salt formation is that the difference
in pK a between the drug and the counter ion
should be at least three to increase the chance of
success. Dissociation can also significantly impact
stability, because the ionized and unionized species
usually have different intrinsic reactivitytheir
respective rates of degradation may be significantly
different. The more stable form of the API can thus
be selected for product development. One or both
of these forms can be catalyzed by the proton or
hydroxyl ions, as evidence from the pH-stability
profiles. Buffering agents have been frequently
used to maintain a formulation at the desired
pH for optimal stability, which is particularly
important for parenteral and other liquid
formulations.

DISTRIBUTION AND PARTITION
When an immiscible liquid, such as hexane or
n-octanol, is added to an aqueous drug solution, the
drug molecules will distribute themselves in each
of the two immiscible phases. This phenomenon is
known as distribution or partition. At equilibrium,
the ratio of the concentrations in each phase is
constant, which is a characteristic of the drug
molecule itself and the nature of the two phases.
This ratio is called the distribution coefficient of the
drug as follows:
D = CO/CW
where C o and Cw are the equilibrium concentration
of molecule in the oil and in the water,
respectively. The distribution coefficient can also
be approximated as the solubility ratio. Because
distribution coefficients are usually large for
organic molecules, they are often converted to
logarithms, known as LogD. Following the
convention, the natures of the partitioning phases
are often noted, such as LogD n-octanol / water or LogD
n-octanol / pH 7.4 buffer
. be possible if their LogD values are sufficiently
The partition coefficient, LogP, is usually
reserved to the partitioning of the neutral or
unionized species, which can be deemed as the
intrinsic distribution coefficient of a molecule
(as to be shown, LogD is pH-dependent with
ionizable molecules). Another term, called cLogP,
refers to calculated LogP value based on molecular
fragmentation algorithms.
LogD For Weak Electrolytes
An ionized species has much less affinity to the
oil phase than the corresponding neutral species.
It could be well assumed that only the unionized
species distributes in both phases while the ionized
species is concentrated in the aqueous phase. The
pH-distribution relationship can then be derived.
For example, the following equation holds for a
monoprotic weak base:
LogD = LogP + log
( Ka
[H+] + Ka ) pharmaceutical, and toxicological activities of a
Figure 2 shows such a plot for a weak organic
base with a LogP of 4 and a pK a of 9. LogD
decreases linearly with pH with a slope of negative
unity at pH < pK a. It even becomes negative below
pH 5, which renders itself practically not extractable
by oil phase and can be utilized in separation.
Applications Of Distribution/Partition
The distribution/partition phenomenon is common
and has been utilized to a great extent by various
industrial and laboratory extraction processes.
Organic compounds can be readily separated
Deliang Zhou
Figure 2: Example pH - LogD profile for a
weak base with pK a of 9.
4
2
LogD
0
-2
-4
0 2 4 6 8 10 12
pH
from inorganics by extracting with an immiscible
solvent. Separation of organic mixtures can
different, or by manipulating their LogD values via
pH modification.
Partition is the foundation underlying various
chromatographic techniques. In HPLC, a solute
molecule is repeatedly partitioned between the
stationary (column) and the mobile (eluent) phases.
The number of equivalent partitioning process is so
large that minute differences in the solute molecules
can be magnified and thus successfully separated.
The pH of the mobile phase is an important factor
to optimize the interactions between solute and the
stationary phase and between the solute and the
mobile phase, because ionized and unionized species
have enormously different LogD values. Both single
mode and mixed mode elution have been utilized in
modern HPLC development.
Distribution/partition also plays an important
role in determining the biological, pharmacological,
drug molecule.
The distribution/partition characteristics are
directly related to the lipophilicity of a molecule.
Lipophicility characterizes the affinity of a
molecule to lipid, fat, or oil. It is well known
that the cell membrane has the phospholipid
bilayer structure, which consists of various
phosphoglycerides. There is a similarity between
water/oil partitioning of a molecule and its ability
to penetrate cell membrane. As a perquisite for a
molecule to function biologically, the molecule
itself has to be able to pass the cell membrane
and penetrate into the cell. Indeed, Hansch et al.
(1) established a correlation between n-octanol
PROCESS VALIDATION Process Design 23
Deliang Zhou
and water partition coeffienct and biological
activities in the era of quantitative structure activity
relationship (QSAR). Since then, the n-octanol/
water distribution/partition coefficient has been
well established in the pharmaceutical industry and
other related fields.
In pharmaceutical sciences, distribution/
partition coefficient is now widely used in
designing drug candidates to possess appropriate
absorption, distribution, metabolism, and excretion
(ADME) properties. The majority and the most
convenient route of oral administration of a drug
requires the drug molecule to be able to cross the
gut wall membrane and get absorbed into the blood
stream, and similar requirements hold true for
routes of administration other than intravenous
applications. In the absence of carrier mediated
absorption (nutrients or nutrient-like molecules)
or paracelluar diffusion (small hydrophilic
molecules), passive diffusion in which the drug
crosses the bilayer of the intestinal epithelium
remains the primary mechanism for most drug
absorption. A molecule needs to be sufficiently
lipophilic to be able to cross the gastrointestinal
(GI) membrane. Therefore, an appropriate LogD
value is required for passive permeability in the GI
epithelium. After a drug gets absorbed into the
blood stream, the drug molecules are then carried
to the various tissues and distributed throughout
the body. This process can be deemed similarly
as partition. It is generally true that a hydrophilic
drug is distributed primarily in the plasma, while a
lipophilic molecule is more extensively distributed
in peripheral tissues and organs. It has been well
known that sufficient lipophilicity is necessary for a
drug molecule to pass the thick lipid bilayer called
blood-brain-barrier (BBB), a tight layer of glial cells
surrounding the capillary endothelial cells in the
brain and spinal cord. Therefore, the distribution
coefficient is of vast importance in designing
molecules to have pharmacological activities in the
brain. A recent survey (2) on drug metabolism has
also suggested that lipophilic molecules are usually
more extensively metabolized in the body, while
hydrophilic ones are often excreted unchanged in
the urine or bile, relating to its ability to access the
enzymes responsible for metabolism. Therefore,
the LogD/LogP values can provide tremendous
insight on the drugs relevant fates in the body.
Distribution/partition has also been useful in
environmental sciences. Hydrophobic chemicals
tend to stay in the environment longer and are more
difficult to clean up. They also tend to preferably
get into various living species. Hence, the partition
coefficient is an important consideration in
environmental regulations.
24 PROCESS VALIDATION Process Design
Partitioning Systems
For historical reasons, the n-octanol/water partition
coefficient has prevailed compared to other solvent
systems in the pharmaceutical fields. The octanol/
water system was initially thought to model
essential properties of typical biological membrane.
However, a thorough literature review has not
resulted in compelling arguments. Partition
between water and other immiscible solvents,
such as hexane, cyclohexane, heptane, isooctane,
dodecane, hexadecane, and chloroform have also
appeared in the research literature, but far less
frequently.
From a physicochemical property point of
view, the n-octanol molecule is characterized by
a hydrophobic backbone (i.e., the C8 group) and a
hydrophilic hydroxyl head. The hydroxyl group can
be both a hydrogen-bond donor and a hydrogen-
bond acceptor. In addition, water-saturated octanol
has ~25 mol % water, that has led to the suggestion
of the 1:4 tetrameric structure. However, it was
later determined by X-ray diffraction analysis
that a cluster structure is more evident where ~16
octanol molecules point their hydroxyl groups
to the water cluster by forming an extensive
hydrogen-bonding network (3). Therefore, a cluster
structures interaction with a drug molecule can
be complicated and may not capture all the lipid
permeation characteristics of a molecule.
The proposition of using multiple systems to
better characterize biological membrane modeling
has also appeared, such as the critical quartet
system (4). However, there has not been much
development in this area. The quartet system is
composed of n-octanol (a hydrogen-bond acceptor
and donor), chloroform (a hydrogen-bond donor),
alkane (an inert medium), and propylene glycol
dipelargonate (a hydrogen-bond acceptor).
SOLUBILITY
A solution is a single-phase mixture consisting of
two or more components. A binary solution is
composed of only two components, while a ternary
solution is composed of three components. The
liquid component is usually called the solvent,
and the dissolved component is the solute. But the
distinction is not absolute and becomes vague when
two liquid components are concerned.
A drug may not be dissolved in a solvent at
any concentration. Indeed, above the solubility
limit, a solution will not be thermodynamically
stable. Still widely employed today as the most
reliable method, solubility can be determined
by equilibrating excess solids with a solvent.
Equilibrium is deemed reached when the solution
concentration reaches an asymptotethe dissolving

Table: USP designations on solubility.
Part of solvent required Solubility in mg solute
for 1 part of solute per mL of solvent
Less than 1 part >1 g/mL
1-10 parts 100-1000 mg/mL
10-30 parts 33.3-100 mg/mL
30-100 parts 10-33.3 mg/mL
100-1000 parts 1-10 mg/mL
1000-10,000 parts 0.1-1 mg/mL
More than 10,000 parts <0.1 mg/mL
solvent contains the maximum amount of solute.
At this point, the solution is said to be saturated
with the solute and the solution is known as
a saturated solution. The solution concentration
equals the solubility in a saturated solution. In the
language of thermodynamics, the free energy of the
solute in solution equals that of the solid solute. A
solution is unsaturated when concentration is less
than the solubility, while a supersaturated solution
refers to one whose concentration is higher than the
solubility. One might be curious as how a solution
could become supersaturated because macroscopic
dissolution ceases once its concentration reaches
the solubility. Well, there are many scenarios where
this could happen. For example, temperature can
significantly affect solubility. Therefore, one can
purposely create an unsaturated or supersaturated
solution by merely adjusting the temperature of the
solution.
A supersaturated solution is thermodynamically
unstable and will lead to solute crystallization
when adequate time is given. The reason is that the
free energy of solute molecules in a supersaturated
solution is always higher than those in the
solid phase, and the crystallization process is
spontaneous by going from a higher free energy
state to a lower free energy state. However, rate of
crystallization (kinetics) could vary. Crystallization
needs to start with nucleation, a process where
nuclei are formed, that serves as the sites for
crystal growth. A supersaturated solution may
be kinetically stable for a long duration without
crystallization. Parent crystals, dusts, and other
exogenous materials can in many cases serve as the
nucleation sites and are often utilized in laboratory
and industrial crystallization processes.
Solubility is often reported in units such as
weight by volume (w/v, e.g., mg/mL), weight-by-
weight (w/w), molar fraction, molarity, or molarlity.
These units can be inter-converted when adequate
information is given. The most frequently used
Deliang Zhou
USP terminology
Very soluble
Freely soluble
Soluble
Sparingly soluble
Slightly soluble
Very slightly soluble
Practically insoluble
units for drugs are mg/mL or g/mL, because
solubility usually falls in these concentration
ranges. It should also be noted that the United
States Pharmacopeia (USP) has a special solubility
designation system, as shown in the Table.
Solubility And Molecular Interactions
Lets take a look at schematic solution formation
process at the microscopic level. According to
Hesss law, energy change of a process is path
independent, only the initial and final states being
of importance. This path independence is true for
all state functions, such as enthalpy, free energy,
and entropy. Therefore, the solution formation can
be deemed to be composed of the following three
essential steps:
t
Step 1. A solute molecule is liberated from
its crystal lattice. Work is needed to overcome
the lattice energy. This process could be
approximated as the melting of the solute
crystal. A high melting temperature and
melting enthalpy usually imply high lattice
energy.
+
t
Step 2. A solvent cavity is created, large
enough to hold the solute molecule. This
process needs to overcome the cohesive
attractions among solvent molecules.
t

Step 3. Finally, a solution is formed by placing
the solute molecule into the solvent cavity created
in step 2. Energy is released resulting from
PROCESS VALIDATION Process Design 25
Deliang Zhou
Figure 3: A pH-solubility profile for a weak acid
with pKa of 4 and intrinsic solubility of 1 g/mL.
1.E+07
1.E+06
1.E+05
Solubility
1.E+04
1.E+03
1.E+02
1.E+01
1.E+00
0 2 4 6 8 10
pH
solvent-solute interaction. Hopefully this energy
is large enough to compensate the work needed in
the previous two steps. It should be pointed out,
though, that the energy in the first two steps need
not be entirely compensated, because the entropy
gained after mixing is also favorable, and can
negate part of the energy consumed during steps
1 and 2.
+ high melting points in their corresponding
Tremendous insight can be gained on solubility
by just considering these processes. The solubility
of a compound can be low due to: high melting
point; high melting enthalpy; lack of solute-
solvent interaction; and too strong solvent-solvent
cohesion.
For a specific compound and a specific solid
form, the melting temperature and enthalpy is
given and can not be changed. However, the
solvent-solute interaction and solvent-solvent
cohesion can be modulated and an appropriate
solvent system may be designed to achieve desired
solubility. For this reason, the following molecular
interactions are important:
t

Van der Waals forces. Van der Waals forces
include both dipole-dipole, dipole-induced
dipole, as well as induced dipole-induced dipole
interactions. A molecule having permanent
dipole tends to interact with another dipole by
pointing to opposite pole and this interaction
is known as dipole-dipole or Keesom forces.
Permanent dipoles are also able to induce an
electric dipole in nonpolar but polarizable
neighboring molecules, and this interaction
26 PROCESS VALIDATION Process Design
is called the dipole-induced dipole or Debye
force. Finally, transient polarization exists in
nonpolar molecules because of the electronic
cloud movements in the molecule. This induced
dipole-induced dipole interaction is called the
dispersion or London force. All these van der
Waals forces decrease quickly with distance (1/
r6) and are usually in the magnitude of 1 10
kcal/mol. On an individual basis, they are weak
and follow the rank order Keesom force > Debye
force > dispersion force. However, they cannot
be ignored for a collection of molecules. Even
the weakest dispersion force is sufficient to cause
the condensation of nonpolar gas molecules and
crystallization of nonpolar liquids.
t

Hydrogen bonds. Hydrogen bonds refer
to the interaction between a hydrogen atom
and a strongly electronegative atom such as
fluorine, oxygen, and nitrogen. The nature
of hydrogen-bond is largely electrostatic
interaction between the small size and large
electrostatic field of the hydrogen atom, and the
strong electronegativity of the acceptor atoms.
Hydrogen bond requires certain direction, is
a fairly strong intermolecular interaction (2-7
kcal/mol), and is always in addition to the van
der Waals forces. It exists in water, hydrogen
fluoride, and ammonium, and is responsible
for their abnormally high boiling points and
homologous series.
t

Ionic interactions. Ionic interaction is
the electrostatic interaction (attraction and
repulsion) between ions of the opposite or same
charges. They are the primary interactions
responsible for ionic crystals. Solubility of
ionizable molecules also depends partially
on the ion-dipole and ion-induced dipole
interactions between the solute and solvent. A
polar solvent has high dielectric constant (e.g.,
water) that can reduce the ion-ion interaction.
Therefore, ionizable compounds such as salts
usually are more soluble in polar solvents than
in non-polar solvents.
The empirical rule on solubility states like
dissolves like. The rule itself is a bit vague;
however, one may make judgment calls by
examining the interactions stated previously. A
polar molecule is likely more soluble in polar
solvents. When a drug molecule is capable of
forming hydrogen bonding, a solvent capable of
satisfying such requirements is likely to provide
improved solubility. In the absence of any polar
group, a drug molecule is probably more soluble in
a nonpolar solvent such as alkanes and benzene.

A drug molecule that contains both polar and
nonpolar groups may require a solvent or a
mixed solvent capable of providing simultaneous
interactions to both parts of the molecule.
Effect Of Temperature
Temperature greatly influences the solubility of a
drug molecule; therefore, the temperature is always
noted when solubility values are reported.
The influence of temperature on solubility
depends on the heat of solution. Solubility
decreases with temperature when the heat of
solution is negative (i.e., when the dissolution is
an exothermic process). Solubility increases with
temperature if the solution process absorbs heat
(endothermic). This observation is consistent with
the Le Chateliers principle on equilibrium. The
magnitude of the heat of solution determines the
steepness of solubility change when temperature
is altered. The quantitative relationship is well
captured by the Vant Hoff equation.
Some molecules have a more complicated
temperature-solubility relationship. For example,
phenol shows an upper consolute temperature
(UCP) of 66.8C, above which it can be mixed with
water at any ratio. Triethylamine demonstrates a
lower consolute temperature (LCT) in water of ~18,
below which the two are miscible in all portions.
Nicotine-water system shows both an UCP and
LCT. The existence of UCP and LCT has something
to do with the solute-solvent interactions and
their temperature dependences. For example, the
hydrogen-bonding interaction could be destroyed at
higher temperature; therefore, an LCT could form.
Sometimes solute can form complexes with
solvent called solvates. It is then the solvated form
that controls the equilibrium solubility. A solvate
is usually less soluble in the same solvent than the
unsolvated form due to thermodynamic reasons. A
solvate may become unstable and dissociate above
a certain temperature. Therefore, the temperature-
solubility line may break its trend. Sodium sulfate
is an example of this type of interaction.
Solubility Product And The Common Ion
Effect
When a salt is dissolved, it dissociates into the
respective anion and cation. At equilibrium, the
product of the concentrations of these ions is
constant. This is known as the solubility product,
often represented as K sp. For example, the
dissolution of silver chloride is as follows:
AgCl(s) Ag+ (aq.) + Cl (aq.)
Deliang Zhou
where s and aq. represent solids and aqueous,
respectively. Its solubility product is as follows:
Ksp = [Ag+][Cl]
Therefore, the solubility of silver chloride can be
reduced if additional chloride (e.g., NaCl) or silver
(e.g., silver nitrate) ions are introduced into the
solution. This phenomenon is known as the common
ion effect.
pH-Solubility Profile
The aqueous solubility of a weak acid or base is
strongly pH-dependent due to relative species
abundances at different pH. The concentration
ratio between the ionized and unionized species is
related via the Henderson-Hasselbalch equation,
as discussed previously. The overall solubility can
then be derived for a weak acid or base, noting that
the unionized species is in equilibrium with the
excess solid (i.e., at the intrinsic solubility S0). For
example, the total solubility of a weak acid is as
follows:
S = S 0 (1 + 10pHpKa)
A typical pH-solubility plot is shown in Figure 3
for a weak acid with pK a of 4 and intrinsic solubility
of 10 g/mL. Below pH 4, solubility is essentially
constant and is equal to the intrinsic solubility.
However, above pH 4, the solubility increases
exponentially with pH. The solubility will not,
however, increase monotonically without a limit.
Salt formation could also take place; therefore,
the solubility reaches a plateau (shown in Figure
3 as the blue dash line), the value of which is
determined by the solubility product of the salt.
The broken red line is just an extension of the
pH-solubility profile if a salt is not formed.
Solubility Of Polymorphs And Solvates
Polymorphism is the existence of multiple crystal
forms by the same chemical entity. They exist as
different molecular packings in the crystal lattice.
The reasons leading to polymorphism are complex
but are related to the ability of a molecule to
efficiently pack itself in a unit cell with different
conformation, orientation, or hydrogen-bonding
motifs. Crystal hydrate is the incorporation
of water molecules in a drug crystal unit cell.
Similarly, a solvate could be formed. In addition,
amorphous form refers to the non-crystalline solid
when the long-range order in a crystal is destroyed.
All these can be collectively called the solid forms
of a drug molecule.
The solubility of polymorphs can be significantly
different because they have different free energies.
PROCESS VALIDATION Process Design 27
Deliang Zhou
For example, the solubility of ritonavir Form II is
about 1/4 of that of Form I at 5C (5). The most
stable form has the least free energy and lowest
solubility. The free energy difference between two
polymorphs is related to the following solubility
ratio:
G(I II) = RT ln(SI/SII)
The metastable polymorph can convert to the
more stable polymorphs through solution-mediated
transformation, where the solution concentration
is higher than the solubility of more stable form
thus providing the thermodynamic driving force
for conversion. The ritonavir Form I to Form II
transition in the liquid-filled capsules is such an
example. Still, the amorphous form, being the
most energetic, has the highest solubility.
Solvates usually have less solubility in the
same solvent than the unsolvated form because
equilibrium dictates so due to the large excess
of solvent. Therefore, a solvated form is often
discovered during the crystallization in a particular
solvent. There also exists a critical solvent activity,
above which the solvate is more stable while
below which the unsolvated form is more stable.
Specifically for hydrates, the critical relative humidity
(RH) or water activity comes to play a significant
role, because moisture is ubiquitous during storage
and because water is the most frequently used
solvent in wet granulation. It should also be noted
that a solvate has higher solubility in other miscible
solvents, because the dissociated solvent molecules
drag the drug molecules into solution.
The solubility difference among various solid
forms is one of the driving forces responsible for
phase changes during pharmaceutical processing.
Significance Of Solubility
Solubility plays a central role in a number of issues
related to drug dissolution, absorption, formulation
and process development and manufacture, as well
as stability.
Solubility is of paramount significance to the
biological performance of a drug molecule. Drug
molecules, when presented as oral solid dosage
forms, need first to dissolve in the GI fluid before
absorption can take place. The dissolution rate of a
solid is proportional to its solubility, diffusivity, and
surface area. Therefore, the higher the solubility,
the faster a molecule can get into solution, and the
higher its concentration in the GI fluid. In turn,
this leads to a higher rate of absorption, because
the passive diffusion is driven by the concentration
gradient across the GI membrane.
28 PROCESS VALIDATION Process Design
The physiology of the GI tract plays an important
role in drug solubilization, including the pH
environment and the solubilization by various
bile salts. For example, the human stomach is
characterized by a low pH of 1.5-2 under fed
conditions and pH 2-6 under fasted conditions.
The pH increases going down to duodenum (pH
4.5-5.5), small intestine (pH 6-6.5), ileum (pH 7-8),
and colon (pH~5.5-7). The small intestine is also
characterized by a high concentration of bile salts
and the large surface area attributing to its villi and
micro-villi structures. The small intestine serves
as the primary site of drug absorption. On the
contrary, the large intestine does not possess the
villi structure. It also lacks the fluid to solubilize
a drug and is challenging for drug absorption,
except for certain highly soluble and highly
permeable molecules. The change in pH, bile salt
concentration, fluid contents, motility, as well as
the residence time, all affect the solubility and in
vivo dissolution of a drug molecule, and influences
its oral absorption. The pH change is particularly
relevant for weakly acidic or basic drugs. Solubility
changes along the GI tract could also cause
solid form changes, such as salt formation,
parent formation, hydrate formation, as well as
polymorphic conversion, all complicating the drug
absorption process.
Appropriate solubility and solvent system are
also required to develop other formulations such as
parenteral, subcutaneous, transdermal, and aerosol
formulations.
Solubility is extensively utilized in crystallization
and purification of active pharmaceutical
ingredients and/or their intermediates.
Experimental conditions, such as pH, solvent,
antisolvent, temperature, and other factors may be
modified individually or in combination in order to
optimize the process. Solubility plays an important
role in controlling the solid forms as well as the
yields.
Amorphous forms have been extensively utilized
to improve drug bioavailability because they have
higher apparent solubility (6).
Solubility is one of the underlying reasons for
many process-induced phase transformations (7).
For example, hydrate formation can occur during
wet granulation. The hydrate may then partially
or fully dehydrate during drying, resulting in
formations of partially amorphous or less ordered
molecular form. A highly water-soluble, low
dose drug may completely dissolve during wet
granulation and turn into amorphous upon drying.
All these changes can have profound effects on
product attributes such as stability, dissolution, and
in vivo performance. These phase changes can be

scale, time, and equipment dependent due to their
kinetic nature, which is worrisome from the quality
control point of view. The solubility and solubility
differences among the various solid forms are
responsible for, and are the keys to, understanding
and resolving these transformations as they provide
the thermodynamic driving forces for change.
SUMMARY
Dissociation, distribution/partition, and
solubility constitute the fundamental physical
properties of a drug molecule. They influence,
directly or indirectly, almost every aspect in
drug development: Absorption, distribution,
metabolism and excretion, formulation
development, processing development, stability,
product manufacture, and regulatory concerns.
Many issues during pharmaceutical development
can be related to characteristics in these basic
properties in one way or another. Awareness
and knowledge of these subjects is important
to understand and solve various issues during
pharmaceutical development, manufacturing, and
stability.
REFERENCES
1. Hansc h, C. and L eo, A. Substituent Constants for
Correlation Analysis in Chemistry and Biolog y, Wiley-
Interscience, New York, 1979.
2. Custodio, J. M.; Wu, C.Y. and Benet, L. Z., Predicting
Drug Disposition, Absorption/Elimination/Transporter
Interplay and the Role of Food on Drug Absorption,
Advanced Drug Delivery Reviews 60, 717-733, 2008.
3. Franks, N. P., Abraham, M. H., and Lieb, W. R .,
Molecular Organization of Liquid n-Octanol: An X-ray
Diffraction Analysis, Journal of Pharmaceutical Sciences,
82, 707-712, 1993.
Originally published in the Spring 2009 issue of The Journal of Validation Technology
Deliang Zhou
4. Leahy, D. E.; Morris, J. J.; Taylor, P. J. and Wait, A. R.,
Membranes and Their Models: Towards a Rational
Choice of Partitioning System, Pharmacochem. Libr.
16, 75-82, 1991.
5. Bauer, J.; Spanton, S.; Henry, R.; Quick, J.; Dziki, W.;
Porter, W. and Morris, J., Ritonavir: An Extraordinary
E x a mple of C on for m at ion a l Pol y mor ph i sm ,
Pharmaceutical Research, 18, 859-866, 2001.
6. Law, D.; Schmitt, E. A.; Marsh, K. C.; et al., Ritonavir-
PEG 8000 Amorphous Solid Dispersions: In vitro and
in vivo Evaluations, Journal of Pharmaceutical Sciences,
93, 563-570, 2004.
7. Zhang, G. G. Z.; Law, D.; Schmitt, E. A. and Qiu, Y.,
Phase Transformation Considerations During Process
Development and Manufacture of Solid Oral Dosage
Forms, Advanced Drug Delivery Reviews, 56, 371-390,
2004. JVT
ARTICLE ACRONYM LISTING
ADME Absorption, Distribution, Metabolism,
Excretion
API Active Pharmaceutical Ingredient
BBB Blood-Brain-Barrier
FDA US Food and Drug Administration
GI Gastrointestinal
HPLC High Performance Liquid Chromatography
Ka Acidity Dissociation Constant
Kb Basicity Dissociation Constant
LCT Lower Consolute Temperature
QSAR Quantitative Structure Activity Relationship
RH Relative Humidity
UCP Upper Consolute Temperature
USP United States Pharmacopeia
PROCESS VALIDATION Process Design 29
Deliang Zhou

Understanding Physicochemical Properties

for Pharmaceutical Product Development
and Manufacturing II:

Physical and Chemical Stability

and Excipient Compatibility
Deliang Zhou

Product and Process Design discusses scientific and t

1IBSNBDFVUJDBM
QSPDFTTJOH
TVDI
BT
DPNNJOVUJPO

technical principles associated with pharmaceutical
product development useful to practitioners in
validation and compliance. We intend this column to
be a useful resource for daily work applications. The
primary objective for this feature: Useful information.
Reader comments, questions, and suggestions are
needed to help us fulfill our objective for this column.
Please send your comments and suggestions to column
coordinator Yihong Qiu at qiu.yihong@abbott.com or
to coordinating editor Susan Haigney at shaigney@
advanstar.com.
KEY POINTS
The following key points are discussed in this article:
t

4UBCJMJUZ
PG
QIBSNBDFVUJDBM
QSPEVDUT

JODMVEJOH
degradation. Three primary mechanisms exist
physical and chemical stability, is a core quality
attribute potentially impacting efficacy and safety
t
1IBTF
USBOTGPSNBUJPO
JT
B
DPNNPO
GPSN
PG
QIZTJDBM

instability. Polymorphic transition; solvation and
desolvation; salt and parent conversion or salt and
salt exchange; and amorphization and devitrification
are the common types of phase transformations.
Phase transformation can occur via solid-state, melt,
solution, or solution-mediated mechanisms.
(miling), compaction, granulation, drying, and
coating may induce partial or complete phase
conversion, which could lead to inconsistent drug
product quality if not understood and properly
controlled
t

%SVH
EFHSBEBUJPO
NBZ
CF
DBUFHPSJ[FE
BT

thermolytic, oxidative, and photolytic
t

)ZESPMZTJT
BDDPVOUT
GPS
UIF
NBKPSJUZ
PG
SFQPSUFE

drug degradations and it is common for a broad
category of organic molecules derived from
weak functional groups such as carboxylic acids.
Moisture, temperature, and pH may greatly impact
the rate of hydrolysis.
t

0YJEBUJPO
PG
ESVHT
JT
UIF
OFYU
HSFBUFTU
DBVTF
PG

for oxidative degradations: nucleophilic and
electrophilic, electron transfer, and autoxidation.
Nucleophilic and electrophilic oxidations are
typically mediated by peroxides. Transition metal
catalyzes oxidation via electron transfer process.
Autoxidation involves free-radical initiated chain
reactions. A single free-radical can cause oxidation
of many drug molecules. Autoxidation is often
autocatalytic and non-Arrhenius.

[
For more Author ABOUT THE AUTHOR
information, Deliang Zhou, Ph.D., is an associated research investigator in Global Formulation Sciences, Global
go to Pharmaceutical R&D at Abbott Laboratories. He may be reached at deliang.zhou@abbott.com.
gxpandjvt.com/bios Yihong Qiu, Ph.D., is the column coordinator of Product and Process Design. Dr. Qiu is a
research fellow and associate director in Global Pharmaceutical Regulatory Affairs CMC, Global
Pharmaceutical R&D at Abbott Laboratories. He may be reached at qui.yihong@abbott.com.

30 PROCESS VALIDATION Process Design


t

1IPUPMZUJD
EFHSBEBUJPO
PDDVST
POMZ
XIFO
MJHIU
JT

absorbed. Excited states of drug molecules may
have enhanced reactivity. Photolytic degradation
is complex but can usually be mitigated by
packaging.
t

3FBDUJPO
PSEFS

DBUBMZTJT

BOE
Q)SBUF
QSPGJMF
BSF
related to the change of the physical, chemical, and
some of the basic concepts in chemical kinetics,
which are helpful to the basic understanding of drug
degradation
t

%SVH
EFHSBEBUJPO
JO
TPMJE
EPTBHF
GPSNT
JT

often determined by the surface characteristics
of the active pharmaceutical ingredient (API)
and excipient particles, or collectively the
microenvironment
t

%FGFDUT

EJTPSEFSFE
PS
BNPSQIPVT
DPOUFOUT
JO
"1*

uniformity (suspensions in particular), alteration in
and excipients pose significant concerns on drug
degradations
t

.PJTUVSF
QSPGPVOEMZ
BGGFDUT
DIFNJDBM
TUBCJMJUZ
CZ

direct involvement in a reaction or by catalyzing
a reaction via increasing molecular mobility and
facilitating the creation of microenvironments.
Moisture preferentially penetrates into amorphous
or disordered regions and may greatly enhance
chemical degradation.
t

1SPDFTTJOEVDFE
QIBTF
USBOTGPSNBUJPO
PG
"1*

is often a leading cause of chemical instability,
particularly when significant amorphous content is
generated
t

$PNQBUJCJMJUZ
JT
BO
JNQPSUBOU
GBDUPS
GPS
FYDJQJFOU

selection. Timely identification of excipient
incompatibility can result in significant savings in
time and resources. Excipients can directly react
with drug molecules, which may be judiciously
avoided based on the chemical structures
and physicochemical properties of the drug
molecules and excipients. Excipients can also
enhance drug degradation by creating a favorable
microenvironment such as pH, moisture, metal
ions, or simply by providing large accessible
surfaces.
t

'VOEBNFOUBM
VOEFSTUBOEJOH
PG
HFOFSBM
BTQFDUT

of chemical degradation as well as specific
physicochemical properties of a drug molecule is key
to its stabilization. Stability issues may be mitigated
by proper solid form selection. Formulation and
engineering approaches may also be used to mitigate
a stability issue.
t

7BMJEBUJPO
BOE
DPNQMJBODF
QFSTPOOFM
TIPVME
CF

aware of the physical and chemical properties
of the active drugs for which they have
responsibility. They should become especially
knowledgeable about active drugs and products
that may be highly prone to phase transformation
and chemical degradation, and apply this
knowledge in process control and change
management.
Deliang Zhou
INTRODUCTION
Stability is a core quality attribute for any viable drug
product. While drug molecules tend to degrade over
time like other chemicals, stability of drug products is
broader than chemical degradation alone. Any aspect
biopharmaceutical properties of a drug product could
be classified as instability.
Generally speaking, instability of pharmaceuticals
includes but is not limited to: loss of active ingredient
potency, increase in levels of impurities, change in
appearance (e.g., color, shape), change in mechanical
strength (e.g., tablet hardening/softening), change
in dissolution/release rate, change in content
bioavailability, or change in other pharmaceutical
elegances. These instabilities may impact the
handling, purity, bioavailability, safety, and efficacy
of a drug product, or may be merely cosmetic changes
which can lead to poor patient acceptance.
This column discusses the key stability concepts
in drug products, with an emphasis on the basic
physical and chemical principles applicable to all
stages of product development and manufacture.
More in-depth treatments on these topics may be
found in many textbooks and review articles (1-6).
The role of physical changes in chemical
degradation is often inadequately acknowledged.
Physical stability often contributes greatly to the
chemical stability of a drug product and cannot be
ignored. This discussion begins with an introduction
of the basic concepts and general mechanisms on
physical and chemical stability of drugs, followed by
excipient compatibility, and finally a brief overview
of remedies to instability.
PHYSICAL STABILITY OF DRUGS
In principle, any stability issue not involving
chemical changes can be considered physical in
nature. On the surface, the physical instability may
manifest itself in various forms, such as appearance,
mechanical strength of dosage forms, content
uniformity, dissolution rate, and bioavailability.
These phenomena may arise from various root
causes; phase transformation is frequently the leading
cause of these problems.
A drug molecule may exist in different solid forms,
including polymorphs, salts, solvates (hydrates), and
amorphous phases. A primary task of preformulation
investigation is to select an appropriate solid form
which bears viable properties in various aspects.
Depending on the specific circumstances, a solid
form may be selected with particular emphasis on
its ability to improve one or more characteristics of
the drug molecule, such as solubility and dissolution
rate, melting, polymorphism, purity, mechanical
PROCESS VALIDATION Process Design 31
Deliang Zhou
properties, manufacturability, and chemical stability.
For example, the bioavailability of a poorly water-
soluble compound may be greatly improved by using
an amorphous phase. Salts have been conventionally
used to improve properties such as purity, melting
temperature, crystallinity, hygroscopicy, and/or
chemical stability, as well as the bioavailability of the
parent form. A number of properties (e.g., solubility
and dissolution, spectroscopic, mechanical, chemical)
have been shown to be affected by polymorphism.
When modification of a particular property is aimed
via solid form selection, properties in other aspects
should be balanced to facilitate pharmaceutical
development. It should be understood that a solid
form change during product manufacturing and
storage could defeat the primary purpose for its
selection.
Types of Phase Transformations
The following paragraphs describe the types of phase
transformations.
Polymorphic Transition. When a molecule
can be packed differently in the crystal lattice,
those crystalline forms are called polymorphs.
Polymorphic transition refers to the conversion
between polymorphs. Polymorphs differ in free
energy and could impact melting, hygroscopicity,
solubility and dissolution, bioavailability, chemical
stability, mechanical, morphological, and flow
properties. Any two polymorphs are related
thermodynamically as either enantiotropes
or monotropes. Enantiotropy exists when
one polymorph is more stable below a critical
temperature (transition temperature, Tt) and the other
more stable above Tt. In monotropy, one polymorph
is always more stable. Under any circumstances,
there is always a most (more) stable polymorph.
Solvation and Desolvation. This is related
to the conversion between the solvated and
unsolvated crystal forms, or solvated forms of
different stoichiometry, with hydration/dehydration
being the most typical. Similarly to polymorphs,
differently solvated forms usually differ in
various physicochemical properties. This type
of form conversion can affect chemical stability,
pharmaceutical elegance, and other quality attributes.
Particularly, the differences in apparent solubility and
dissolution rate could impact on the oral absorption
for many poorly or borderline soluble compounds.
An important concept for hydration and dehydration
is the critical relative humidity (RH), below which
the anhydrous form is more stable and above which
the hydrate is more stable. Similarly, critical RH also
exists between hydrates of different stoichiometry.
This concept is vital to the stability of hydrates
because moisture has to be dealt with everywhere:
32 PROCESS VALIDATION Process Design
active pharmaceutical ingredient (API) manufacture,
formulation development, drug product manufacture,
and storage. Changes in the environmental humidity
could inadvertently cause hydration or dehydration
to occur. In aqueous solutions, the hydrated form is
usually less soluble than the anhydrous form because
of the nature of equilibrium. The solubility difference
between hydrated and anhydrous forms is the driving
force behind solution-mediated hydrate formation,
which could occur during various wet processes or
during in vivo dissolution.
Salt and Parent Conversions or Salt and
Salt Exchange. Salts may be converted to their
parent forms or to salts of different counterion
during processing and storage. The propensity of
salt form conversion is determined by pKa, crystal
lattice energy, solubility and solubility relationship,
and other experimental conditions. For example,
salts of very weak acids or bases may readily convert
back to the parent forms upon exposure to moisture
or water. The potential of salt conversion during
wet granulation and other wet processes is more
significant than other dry processes and should
be watched closely. Similar to other solid form
conversions, change of a salt to its parent, or to
different salt forms, can profoundly alter the quality
attributes of a drug product.
Amorphization and Devitrification.
When the long-range order in a crystal lattice
is destroyed, an amorphous phase is then
generated. Partially or fully amorphous phases
may be generated (amorphization) by various
methods including many unit operations such
as milling, melting, wet granulation, drying, and
compaction. Amorphous phases have higher free
energy than their crystalline counterparts and are
often exploited to improve the bioavailability of
poorly water-soluble compounds. However, the
reversion to crystalline forms (devitrification)
is thermodynamically favorable based on the
thermodynamic principle. When the use of an
amorphous phase is desired, the physical stability
becomes critical because crystallization could
lead to a decrease in bioavailability and thus lose
the intended advantage of the amorphous state.
When crystalline forms are desired, however, the
formation of even a small amount of amorphous
content during various processes is also a concern
because this may bring undesired effects to drug
products such as enhanced degradation.
Mechanisms of Phase Transformations
The conversion of metastable form to stable form
is thermodynamically favored, which poses a
concern for any system where a metastable form
is employed. The transformation of a stable

form to a metastable form, on the other hand, is
thermodynamically unfavorable, and is possible
only when there is an energy input from the
environment. The transformation kinetics in both
cases, however, vary greatly, which are compound-
specific and depend on the experimental
conditions.
The primary mechanisms for phase conversions
include solid-state, melt, solution, and solution-
mediated transformations.
Solid-State. Certain phase transitions occur in
the solid-state without passing through intervening
transient liquid or vapor phases. These solid-state
transitions are generally nucleated by defects or strain.
The physical characteristics of the solids, such as
particle size, morphology, crystallinity, mechanical
treatments, and presence of impurities may greatly
affect the rate of such transformations.
Melt. A solid form is destroyed upon melting
and may not regenerate when cooled. The melt
is often supercooled to generate an amorphous
phase, which may or may not crystallize during
further cooling. Even when it crystallizes, it
usually goes through metastable forms, and may
or may not convert to the most stable form on the
experimental time scale. A potential change in
crystalline form is therefore likely. The final solid
phase may depend on: thermodynamic stability
among different forms; the relative nucleation and
crystal growth rate of each form; and the kinetics of
form conversion. The solid phases formed in such a
process are usually accompanied by relatively high
amount of disorders. Impurities or excipients are
also likely to affect the kinetics of crystallization
and phase transformation. Presence of excipients
may cause eutectic melting at temperatures much
lower than the melting of pure API and give rise to
potential phase transformations.
Solution. When solvents are involved in
processing, solid API may dissolve partially or
completely. Subsequent solvent removal may cause
the dissolved portion to change its solid form similar
to a melt. It is important to note that formation of a
metastable form (crystalline or amorphous) is often
likely, especially when the original solid form is
completely dissolved. Particular attention should be
paid to drug molecules that have high solubility in the
processing solvents.
Solution-Mediated. When a metastable solid
form comes to contact a solvent during processing,
it may convert to more stable forms due to
solubility difference between the metastable form
and the stable form. As opposed to the solution
mechanism, the solution-mediated mechanism
only allows the transition from a meta-stable form
to a more stable form.
Deliang Zhou
PHASE TRANSFORMATIONS DURING
PHARMACEUTICAL PROCESSING
Phase transformations can occur in a number of
pharmaceutical unit operations, including milling,
wet granulation/spray drying/lyophilization, drying,
and compaction, and therefore may impact drug
product manufacturing and performance (7). Because
of their very kinetic nature, the extents of these
phase transformations may be time and equipment
dependent. These phase transformations, when not
controlled properly, often lead to inconsistent product
quality.
The application of mechanical and possibly the
accompanying thermal stresses during operations such
as milling, dry granulation, and compaction, may
lead to phase transformation such as polymorphic
transition, dehydration and desolvation, or
vitrification via the solid-state or melt mechanisms.
The rate and extent of these phase transitions will
depend on characteristics of the original solid phase
and the energy input from these processes. Caffeine,
sulfabenzamide, and maprotiline hydrochloride
have been reported to undergo polymorphic
transformations during compression.
A number of wet processes, such as wet granulation,
lyophilization, spray drying, and coating, may induce
phase transitions via solution or solution-mediated
transformation. Important factors governing the
likeliness and the extent of conversion include:
solubility of the drug substance in the liquid, duration
in contacting with liquid, other operation parameters
such as temperature, drying conditions, and the
specific properties of the drug molecule. The presence
of other excipients may also significantly influence
the propensity and rate of a phase transformation.
All the aforementioned phase transitions, such as
polymorphic conversion, hydration and dehydration,
salt/parent conversion or salt/salt exchange, and
vitrification and crystallization may occur. For
example, a highly water-soluble drug may completely
dissolve in granulating liquid and subsequently
produce amorphous phase during drying that may
or may not convert back to the original solid form.
The formation of a hydrate is also likely to occur in
wet granulation. Subsequent drying may dehydrate
the formed hydrate. In any solid-solid conversion;
however, it is very likely that a less ordered form will be
produced even if conversion is close to completeness.
Therefore, various amounts of amorphous content may
exist in the end product, which could cause further
concerns on its chemical stability. In film coating, due
to the relatively short contacting time between solid
surface and liquid, the likelihood of solid form change
is much lower. However, the potential for phase
transformation can be significant when an active is
present in the coating solution.
PROCESS VALIDATION Process Design 33
Deliang Zhou
CHEMICAL DEGRADATION OF DRUGS
Chemical degradation represents one of the most
important stability aspects of pharmaceuticals,
because inadequate stability not only limits the
shelflife of a drug product but also potentially
impacts efficacy and patient safety.
An important aspect of preformulation
characterization is to examine the chemical stability
of new drug candidates, assess the impact of stability
on pharmaceutical development and processing, and
design strategies to stabilize an unstable molecule. The
understanding of commonly encountered degradation
pathways, basic concepts in chemical kinetics, and
characteristics of chemical degradation in drug
products are key to the overall success of formulation
development and scale up.
Common Pathways of Drug Degradation
Drug degradation can be generally divided
into thermolytic, oxidative, and photolytic. By
examining structural features of a drug molecule,
possible degradation routes and products may be
predicted to a certain extent. These may then aid in
the design and execution of stability studies.
Thermolytic Degradation
Thermolytic degradation refers to those that are driven
by heat or greatly influenced by temperature, to which
the following Arrhenius relationship often applies:
k = Aexp(Ea/RT)
where k is the rate constant, T is the absolute temperature,
Ea is the activation energy, A is the frequency factor,
and R is the universal gas constant. Any degradation
mechanism can be considered thermolytic if high
temperature enhances the rate.
Typical activation energies for drug degradation
fall in the range of 12-24 kcal/mol (8). The Q10 rule
states that the rate of a chemical reaction increases
approximately two- to threefold for each 10 C rise
in temperature. The exact value of this increase may
be determined based on the value of the activation
energies and is the theoretical basis behind the
International Conference on Harmonisation (ICH)
conditions.
Hydrolysis forms a subset of thermolytic
degradation reactions and is the most common
drug degradation pathway. Hydrolysis accounts for
more than half of the reported drug degradation.
Derivatives of relatively weakly-bonding groups such
as esters, amides, anhydride, imides, ethers, imines,
oximes, hydrazones, semicarbazones, lactams,
lactones, thiol esters, sulfonates, sulfonamides, and
acetals can undergo hydrolysis both in solution
and in the solid state in the presence of water. In
34 PROCESS VALIDATION Process Design
particular, the presence of hydrogen or hydroxyl ions
likely catalyzes the hydrolytic reactions. Each of
these may have different reactivity and may require
different conditions.
Transacylation is also likely when appropriate
other functional groups (i.e., alcohols, amines,
esters, etc.) are present in the environment, either as
solvent or solvent residuals, or more commonly, as
excipients or impurities.
Other thermolytic degradation pathways include
rearrangement, isomerization and epimerization,
cyclization, decarboxylation, hydration/dehydration,
and dimerization and polymerization.
Oxidative Degradation
Oxidation accounts for 20-30% of reported drug
degradations and is secondary only to hydrolysis.
Oxidation can proceed with three primary
mechanisms: electrophilic/nucleophilic, electron
transfer, and autoxidation.
A drug molecule may be oxidized by electrophilic
attack from peroxides, which is the typical scenario
of nucleophilic/electrophilic mechanism. Similarly,
the electron transfer process shares certain features
of the nucleophilic/electrophilic process, except that
an electron is transferred from a low electron affinity
donor (e.g., drug molecule) to an oxidizing agent via
the mediation of transition metal. Fe3+ and Cu2+ are
commonly used to probe such mechanisms.
The autoxidation process involves the initiation of
free radicals, which propagates through reactions with
oxygen and drug molecule to form oxidation products.
Because of the complexity of the reaction mechanism,
non-Arrhenius behavior is frequently observed. The
three stages of autoxidation may be represented as the
following:
t
*OJUJBUJPO
In \

RH In H 
R \
t

1SPQBHBUJPO

R \

O2 ROO \






(fast)
ROO \

RH ROOH 
R \





(rate-limiting)
t
5FSNJOBUJPO
R \

R \
R R
R \

ROO \
ROOR
Free radicals may be generated by homolytic
cleavage of a chemical bond via thermal, photolytic, or
chemical processes. A drug free radical is formed when

one of its hydrogen atoms is abstracted by the initial
free radical. The formed drug free radical then quickly
reacts with oxygen to form a peroxy free radical. The
latter can abstract a hydrogen atom from another drug
molecule so that the drug molecule gets oxidized
(forming hydroperoxide) and a new drug free radical
is re-generated. This chain reaction can continue until
the free radical is terminated. A single free radical
can, in principle, cause the oxidation of many drug
molecules in its lifetime. Therefore, the autoxidation
is typically auto-catalytic in nature: its initial rate may
be low; however, the rate increases quickly due to the
gradual buildup of the free radicals.
Multiple oxidation mechanisms can co-exist. For
example, peroxide can undergo homolytic dissociation
at high temperature to form free radicals; the presence
of trace amount of transition metal could react with
peroxides (e.g., oxidation degradants) and form
free radicals. It may be true that multiple oxidative
processes occur simultaneously in many cases.
Photolytic Degradation
Chemical reactions may occur upon light absorption
because light carries certain energy:
E = hv = hc/
where h is Planks constant, c is the speed of light,
is the frequency, and is the wavelength. The
Grotthuss-Draper law states that photochemical
reaction can occur only after light is absorbed. A
simplistic calculation indicates that the energy
associated with UV-visible light corresponds roughly
to the bond energies in typical organic molecules. For
example, the weakest single bond is roughly ~35 kcal/
mol (e.g., OO bond) and the strongest single bond
corresponds to ~100 kcal/mol (e.g., CH bond).
Photophysical processes refer to the changes in
molecular orbitals after light absorption (excitation).
Properties of the excited molecules are expected to
differ from those of ground states and are generally
more reactive. For example, the radical-like structures
of some of the excited states make (photo) oxidation
favorable. For more detailed discussions on the fates of
a molecule upon light absorption and the potentially
increased chemical reactivity, readers may refer to the
textbook by Turro (9).
Photolytic degradation is directly initiated by
light absorption; therefore, temperature has a
negligible effect. In fact, some photo reactions can
occur even at absolute zero. Photolytic degradation
is not uncommon but may be minimized during
manufacturing, shipping, and storing of drug products
by appropriate packaging. However, mechanisms of
photolytic degradations are usually more complex.
Deliang Zhou
Figure: pH-rate profile of hydrolysis of aspirin
(adapted from reference 10).
-3
-4
log Kobs (s-1)
-5
-6
-7
0 2 4 6 8 10
pH
Basic Concepts in Chemical Kinetics
The following sections describe the basic concepts of
chemical kinetics.
Reaction Order. The rate of a reaction is often
used to describe the relationship between reaction rate
and the concentration of various species. For example,
the hydrolysis of aspirin under acidic conditions is
first-order with respect to both aspirin and hydrogen
ion, but is an overall second-order reaction:
d[Aspirin]
= k[Aspirin][H+]
dt
Degradation of most pharmaceuticals are second-
order in nature, because degradation usually involves
two molecules to react, one of which being the drug
molecule itself. However, the concentration of the
other component (i.e., the hydrogen ion, hydroxyl
ion, or the buffer species) is usually in large excess and
can be considered as constant throughout. Therefore,
apparent first-order reactions are often reported for
most pharmaceuticals. The apparent rate constant,
in this case, includes the contribution from the
concentration of the other reactants. Apparent zero-
order degradations may arise in cases where drug
concentrations are maintained as a constant (e.g.,
suspensions).
Catalysis. A catalyst is a substance that influences
the rate of a reaction without itself being consumed. A
positive catalyst promotes a reaction while a negative
catalyst demotes a reaction.
Catalysis occurs by changing the activation energy
of a reaction but not changing the thermodynamic
nature of the reaction. A catalyst can only influence
the rate, but not the equilibrium of the reaction. In
some cases, a reaction product can catalyze the rate,
which is often termed as autocatalysis. The free-radical
initiated oxidation certainly is such an example.
PROCESS VALIDATION Process Design 35
Deliang Zhou
Hydrogen ions and/or hydroxyl ions are
often involved directly in the degradation of
pharmaceuticals. When the concentration of hydrogen
ion or hydroxyl ion appears in the rate equation, the
reaction is said to be subject to specific acid-base
catalysis. Drug degradations are often determined in
buffered solutions and studied by monitoring the drug
itself. As a result, the degradation kinetics is often
apparent first-order with the apparent rate constant in
the following form:
kobs = k0 
kH[H]

kOH [OH]
For a reaction subject to specific acid catalysis, a plot
of the logarithm of the apparent first-order rate constant
with respect to pH gives a straight line of slope of 1,
while a specific base catalysis generates a straight line of
slope of +1. When both are present, a V or U-shaped more complex than those in solution and have some
pH-rate profile is often observed.
General acid-base catalysis occurs when the
buffering components catalyze a reaction. Either the
acidic or basic components of the buffer, or both,
can catalyze a reaction. General acid-base catalysis
often causes deviation of the rate-pH profile from the
expected unit slopes.
pH-Rate Profiles. The pH dependence of the rate
constant of degradation of a compound can be concisely
represented by a plot of kobs vs. pH.
In general, rate of drug degradation can be represented
by summing up all possible pathways, including
intrinsic, specific acid-base catalysis, general acid-base
catalysis, etc., as follows:
kobs = k0 
kH[H]

kOH [OH] 
k1 [bufferspecies 1] 

k2 [bufferspecies 2] 
\\\

k0 
4
ki Ci
i
The pH-rate profile provides a summary of the
primary features of a specific degradation. Specific acid-
base catalysis is designated by the straight line with slope
of negative or positive unity, while general acid-base
catalysis may be indicated by the apparent deviation of
the slopes. Commonly, many drug molecules are weak
acid or weak base. Ionized species may have different
reactivity, which is often revealed as a sigmoid in the
rate-pH profile.
For example, hydrolysis of aspirin (10) (see Figure)
is subject to specific acid catalysis at pH < 2, and
specific base catalysis at pH >10. The ionization
causes a difference in reactivity in pH range of 2.5-4.
The broad shoulder in the pH range of 4.5-9 is caused
by the intramolecular catalysis:
36 PROCESS VALIDATION Process Design
O
O
H
O O
H
CH3
O
pH-rate profiles can provide tremendous insights on
the nature of a reaction and can serve as a very useful
tool in developing solution formulations, lyophilized
products, as well as conventional solid oral dosage
forms.
DEGRADATION IN SOLID DOSAGE FORMS
Drug degradation in solid dosage forms are certainly
unique features related to the state of the matter.
Topochemical reactions are a class of reactions that
are directly related to the molecular packing in the
crystal lattice. Certain molecular rearrangements
are required in order for a reaction to take place
truly in the solid state (i.e., in the crystal lattice).
For example, the photo-dimerization of cinnamic
acid derivatives requires certain minimum distance
between the double bonds. As a result, only
form of p-methylcinnamic acid is feasible for this
dimerization in solid state (11). Topochemical
reactions can be identified if a reaction does not occur
in solution or is much slower in solution or if the
reaction products are different from those obtained
in solution. For example, the rearrangement of
methyl p-dimethylaminobenzenesulfonate to the
p-trimethylammoniumbenzenesulfonate zwitterion
shows a reaction rate that is 25 times slower in its
melt than in the solid state (12). Because the crystal
structure of a solid phase determines its chemical
reactivity in the case of topochemical reaction,
solid-state forms (e.g., polymorphs, solvates, salts,
co-crystals) frequently exhibit different reactivity.
Degradations of most drug products, however,
are not topochemical in nature. Drug degradations
occur mostly around the surface of a drug particle.
Therefore, the surface characteristics of API
particles and the excipientsor collectively, the
microenvironmentare of vital importance to the
understanding and remedy of drug degradation.
Solids often contain defects or strains, where
molecules have higher free energy. These high energy
sites usually serve as the initiation (nucleation) sites
of a reaction. Therefore, disorders in crystals may
be a key point in the understanding of solid-state
reactions of drugs. To extend this concept further,
the amorphous phase is highly energetic, is expected,
and has been found to enhance chemical reactivity.
Often crystalline API is used in drug development.

However, a small amount of disordered or
amorphous content is apparent, particularly after
various pharmaceutical operations such as milling,
wet granulation, drying, and compaction. Even when
the degree of crystallinity is not affected, a solid-state
reaction can be enhanced by rendering larger surface
area (smaller particle size) because of the increased
concentrations of possible defects. Regardless of
how perfect the surface may be, surface itself can
be deemed as a type of defect based on the notion
of crystal orders. It is not uncommon that small
API particles (particularly via milling) exacerbate a
stability problem.
Moisture has a profound effect on drug
degradation in solid dosage forms, either as a
reactant (e.g., hydrolysis) or as a promoter or both.
The catalytic role of water in drug degradation is
related to its ability as an excellent plasticizer. Water
greatly increases molecular mobility of reactants
and enhances drug degradation and makes it more
solution-like. In the worst scenario, water can
form a saturated solution and maximize the rate of
degradation.
Various mechanisms exist for water molecules
to interact with a solid. Water molecules can be
incorporated into crystal lattice through hydrogen
bonding and/or Van de Waals interactions. Generally,
lattice water is not a cause of chemical instability.
Some compounds rely on the interaction of water to
form a stable crystal thus improving their chemical
stability. Water can be also adsorbed to the surface as
a monolayer and as multilayers. Water molecules in
the monolayer may behave significantly differently
than those in the second or third layers. Water
molecules beyond these 2-3 layers are expected to
behave like bulk water. A more thorough discussion
on water-solid interactions in pharmaceutical systems
can be found elsewhere (13).
Tightly bound water (e.g., hydrate, monolayer)
have decreased activity and are therefore not
detrimental to chemical stability. The loosely bound
water (other than the monolayer) or free water is
believed to enhance a chemical reaction. In addition,
pharmaceutical solids usually contain various defects
and various degrees of disordered, amorphous
regions. Water molecules are preferentially absorbed
into the interior of these regions (sorption). Because
water is an efficient plasticizer, it significantly
increases molecular mobility in the amorphous phase
and, therefore, enhances degradation. A simplified
calculation indicates that the typical moisture content
in pharmaceuticals (e.g., a few percent) far exceeds
the estimate by mode of adsorption (14, 15). Hence,
water sorption into the disordered or amorphous
regions is a realistic concern for the stability of
pharmaceutical dosage forms. Greatly enhanced
Deliang Zhou
degradation of ABT-232 was reported when wet
granulated (16).
The presence of moisture also facilitates the
creation of microenvironment for drug degradation.
The pH of the microenvironment is often a key
parameter that greatly influences drug degradation.
The pH-rate profile can provide significant insight
in this case. Microenvironmental pH could be
altered by solid forms, such as salts, or by buffering
agents, such as citric acid and sodium phosphate, or
other excipients such as magnesium stearate. For
example, stability of moexipril (17) was improved
by wet granulation with basic buffering agents. This
modification of microenvironmental pH is one of the
modes of drug-excipient interactions.
The change of API solid phases in a dosage form
can have profound impacts on its chemical stability.
Even when topochemical reaction is not observed,
the surface characteristics of solid forms can differ
significantly, which affects its ability to adsorb,
interact, and react. More importantly, solid-form
conversion in dosage forms is often incomplete,
and leaves significant amount of disordered
or amorphous regions behind. The extent of
amorphous content can be significant in the case of
wet granulation where a soluble API may dissolve
completely and remain as amorphous upon drying,
or hydrate formation may occur but dehydrate upon
drying while leaving behind a significant portion
of amorphous phase. Amorphous content may also
result from other processing steps such as milling and
compaction. All these process-induced amorphous
content, in combination with typical moisture
contents, could be detrimental to the chemical
stability of pharmaceutical dosage forms, especially
when the dose strength is low. The process-induced
phase transformations often depend on equipment,
time, operational conditions, as well as raw material
attributes, which makes scale up as well as quality
control challenging.
EXCIPIENT COMPATIBILITY
Excipient selection is of great importance to
accomplish the target product profile and critical
quality attributes. Excipient functionality and
compatibility with the active drug are two important
considerations.
Incompatibility may be referred to as the
undesirable interactions between drug and one or
more components in formulation. Incompatibility
may adversely alter the product quality attributes,
including physical, chemical, biopharmaceutical, or
other properties.
Excipient compatibility studies are usually
conducted in the early phase of drug product
development. Their objective is to further
PROCESS VALIDATION Process Design 37
Deliang Zhou
characterize the stability profile of the drug molecule
in typical formulation design, identify potential
material incompatibility, characterize the degradants,
understand the mechanisms of degradation, and
provide guidance to formulation development.
A carefully planned and executed compatibility
study may result in significant savings in time and
resources. In addition, as expected by quality-by-
design (QbD) principles, excipient compatibility
information is required by the regulatory agencies to
justify the selection of formulation components and
to assess the overall risk.
It does not mean, however, that one absolutely
cannot use an incompatible excipient. In some
cases, a small amount of incompatible excipient may
be acceptable based on other benefits it provides.
There are also situations where there is simply no
other alternate ingredient and one has to live with
an incompatible excipient. However, it is still
essential to understand the associated risks and their
mitigations.
Direct Reactions Between Drug and
Excipient
Probably the most known excipient incompatibility
is the Maillard reaction between a primary or
secondary amine and a reducing sugar such as
lactose and glucose, resulting in the production of
browning spots. From a mechanistic point of view,
the reactivity of the amine is related to its electronic
density. Therefore, salt formation of the amines
usually improves stability. Like most solid-state
reactions, amorphous characteristics of both the
amines and the reducing sugar enhance the reactivity,
a concern when using spray-dried lactose.
Interactions between acidic drugs and basic
excipients and vice versa have been reported.
Examples include the interactions between
indomethacin and sodium bicarbonate, ibuprofen
with magnesium oxide, and citric acid/tartaric acid
with sodium bicarbonate, the latter of which is
utilized in effervescent tablets.
Drug-conterion interactions may also occur.
Seproxetine maleate was reported to form a Michael
addition product. In a broad sense, drug-buffer
interactions can be grouped into this category.
Examples of the latter include the interactions
between dexamethasone 21-phosphate and sodium
bisulfite, and between epinephrine and sodium
bisulfite.
Transacylation is another major group of drug-
excipient interaction. Carboxylic moiety can
react with alcohols, esters, or other carboxylic
derivatives to form esters, amides, etc., whether
presented in excipients or in drug molecules. For
example, aspirin can undergo transesterification
38 PROCESS VALIDATION Process Design
with a number of drugs including acetaminophen,
codeine, sulfadiazine, and polyethylene glycol (PEG).
Norfloxacin is reported to react with magnesium
stearate to form stearoyl amide (18). Similarly,
duloxetine reacts with hydroxypropyl methylcellulose
acetate succinate (HPMCAS) (celluosic coating
polymer) to form succinamide.
Impurities in excipients can be a major concern
to cause incompatibilities. It is well known that
traces of peroxides in povidone, crospovidone,
hydroxypropyl cellulose (HPC), dicalcium
phosphase, PEG, polysorbates, and a number of
modified excipients containing polyoxyethylene
units are responsible for the oxidation of a number
of drugs including ceronapril and raloxifene.
Traces of low molecular weight aldehydes such as
formaldehyde and/or acetaldehyde may exist in a
number of excipients such as starch, polysorbate,
glycols, povidone, crospovidone, ethylcellulose,
which could condensate with amine drugs. The
5-hydroxylmethyl-2-furfuraldehyde impurity in
lactose poses a similar concern. Transition metals
may be present in a number of excipients. As
mentioned previously, these may catalyze the
oxidations of many drugs. In principle, it is very
beneficial to have some basic understanding on the
production, isolation, and purification process of
each excipient in order to anticipate the possible
impurities and the potential risks to drug product
stability.
Complexes between drug and excipients have also
been reported, which could impact the dissolution,
solubility, and bioavailability of a drug product.
Tetracycline is reported to interact with CaCO3
to form an insoluble complex. Diclofenac and
salicylic acid have also been reported to complex
with Eudragit polymers. The complexation between
weakly basic drugs (e.g., amines) and anionic
super-disintegrants such as croscarmellose sodium
and sodium starch glycolate has been reported;
metaformin-croscarmellose sodium is an example.
Drug Degradations Enhanced by
Excipients
Drug degradation usually occurs on the surface or
the interfaces; therefore, the surface characteristics
of API particles and excipients may be of vital
importance. Direct reactions between a drug
molecule and excipient do not frequently occur and
may be anticipated based on knowledge of drug
degradation pathways and understanding of the
drug molecule. However, the presence of an inert
excipient could lead to adsorption of drug molecules
to the surface of excipient particles. Drug molecules
adsorbed onto a surface are expected to have higher
mobility and increased reactivity. This increase

in accessible surface area is expected to contribute

IZESPMZTJT

PYJEBUJPO

PS

to increased degradation even for an otherwise
inert excipient. The excipient may be considered
a heterogeneous catalyst in this case. Often low
strength formations are reported to have increased

Q)


degradation attributed to the increased accessible
surface area to drug ratio. In addition, the presence
of excipient might also change the moisture,
microenvironment pH, and other characteristics
which are important considerations regarding drug
product stability.
Microenvironmental pH may be an important
factor on drug degradation in dosage forms. Many
excipients are acid or basic, such as citric acid,
tartaric acid, stearic acid, sodium bicarbonate,
sodium carbonate, magnesium oxide, and
magneiusm stearate. In addition, many otherwise
neutral excipients may be weakly acidic or weakly
basic due to the processes used in the production
and treatments thereof. All these can cause an
acidic or basic microenvironment around the drug
particles, which is facilitated by the presence of
moisture. Depending on the particular case, a drug
molecule may be more or less stable in a particular
pH environment. For example, reduced stability
of aspirin was noted in the presence of acidic
excipients such as stearic acid or basic excipients
such as talc, calcium succinate or calcium
carbonate. Bezazepines hydrolysis was enhanced
in the presence of microcrystalline cellulose,
dicalcium phosphate, magnesium stearate, and
sodium stearyl furate. Moexipril and lansoprazole
(19) have been reported to be stabilized by basic
excipients.
The moisture content and other attributes of
excipients may also influence drug degradation.
Generally, it is believed high moisture content from
excipients may be detrimental to chemical stability.
However, in a number of cases, improved stability
has been attributed to more hygroscopic excipients
that may preferentially absorb moisture in the
formulation. The characteristics of moisture may
differ from one excipient to another as suggested
by spectroscopic evidence, which could potentially
explain how excipients affect stability differently.
General Approaches to Stabilization
Once instability is observed, efforts should be
made to identify the degradation products which
may suggest possible degradation mechanisms.
Anticipation of the nature of degradation based
on molecular structure and prior knowledge
can greatly assist the efforts and avoid excipient
incompatibilities. Preformulation degradation
studies are essential to obtain useful information.
Relevant questions include the following:
Deliang Zhou
t

8IBU
JT
UIF
UZQF
FH
others) of degradation?
t
*T
B
HFOFSBM
NFDIBOJTN
BWBJMBCMF
t
8IBU
GBDUPST
BSF
HFOFSBMMZ
JOGMVFOUJBM
t

*T
UIFSF
B
HFOFSBM
TUBCJMJ[BUJPO
BQQSPBDI
FH
anti-oxidant)?
t

5P
XIBU
FYUFOU
JT
UIF
SBUF
PG
EFHSBEBUJPO
JT

affected by moisture and/or temperature?
t
$PVME
BOZ
PG
UIF
QSPDFTTJOH
TUFQT
QMBZ
B
SPMF
Because fixing a stability issue in late development
phase could be costly both in terms of project delays
and resources, early identification and remediation
of stability issues is beneficial and preferred.
Degradation studies are vital in the preformulation
assessments of a drug candidate. These studies often
employ forced degradation studies (i.e., exposure to
strong acidic pH, strong basic pH, various oxidants,
light, etc.) in order to delineate the intrinsic stability
profile, potential degradants, and degradation
pathways. By utilizing information obtained from
the preformulation studies, instability concerns can
often be addressed and mitigated before formulation
activities start. In the authors experience, selection
of viable solid forms can be made based on stability
in the solid state. Solid forms, particular salt forms,
allow one to modify the surface characteristics of
the API, such as crystallinity, moisture sorption,
surface pH, and solubility, all of which may play
a significant role in drug degradation. Based
on this principle, a number of solid forms have
been successfully selected to provide viable
stability profile for otherwise somewhat labile
compounds.
A good understanding on the degradation
mechanism or even the general features of the
class of the reactions can be very helpful to
the stabilization of a formulation. Hydrolytic
degradations have been frequently minimized
by modification of the microenvironment pH via
acidic or basic excipients. Antioxidants have been
routinely used to alleviate oxidative degradations.
In the latter case, however, the optimal antioxidant
can only be selected based on the understanding of
oxidation mechanisms. For example, free-radical
scavengers such as butylated hydroxytoluene (BHT)
are better for autoxidation, while ascorbic acid may
be more suitable for nucleophilic/electrophilic type
of oxidation. Combination of antioxidants may be
used and have been reported to be more effective,
particularly when there is a mix of oxidative
processes. When metal ions play a significant role
in oxidation, a chelate (such as EDTA or citric acid)
may be used.
When necessary, engineering approaches can be
used alone or in addition to those mentioned above.
PROCESS VALIDATION Process Design 39
Deliang Zhou
For example, coating, double granulation, and
bi-layer tablet approaches may be used to separate
incompatible components in a formulation. When
moisture plays a significant role, packaging or
packaging with desiccant is routinely used to
achieve viable shelf life of a drug product. For
drug molecules that react slowly with oxygen,
packaging combined with an oxygen scavenger can
be effective. Packaging has been routinely used to
alleviate the photolytic (light) degradation of many
drug products.
PHASE TRANSFORMATIONS, CHEMICAL
DEGRADATION, AND EXCIPIENT
COMPATIBILITYIMPLICATIONS FOR
VALIDATION AND COMPLIANCE
Validation and compliance personnel should
be knowledgeable of the physical and chemical
properties of the active drugs for which they have
responsibility. They should be especially aware
of active drugs and products that may be highly
prone to phase transformation and chemical
degradation (i.e., they must be aware of high
risk drugs and products). This information is
routinely determined during the pharmaceutical
development process and is often filed in regulatory
submissions. The manufacturing and processing
of high-risk drugs should receive heightened
vigilance by compliance personnel to minimize
the risk of phase transformations and/or chemical
degradation. Any changes to the manufacturing
process of susceptible APIs and products should be
carefully evaluated. Referring to previous excipient
compatibility studies is highly recommended
prior to any changes in formulation. Changes
to high risk processes such as wet granulation,
milling, and processes with solvents should receive
heightened scrutiny as part of the change control
program. Laboratory evaluations of proposed
changes may be necessary in advance of large-scale
changes. Validation of high-risk process changes
should include appropriate testing to assure the
acceptability of the process change. Fundamental
knowledge obtained during development should be
available to validation and compliance personnel
throughout the product lifecycle and should be
utilized as necessary to evaluate formulation and
process changes.
SUMMARY
Stability of drug product is complex due to the
multi-particulate and heterogeneous nature of
pharmaceutical products. Drug degradation
in solid dosage forms is mostly determined
by the surface characteristics of both the API
and the excipient particles. Solid-state forms
40 PROCESS VALIDATION Process Design
provide a viable means to the modulation of API
characteristics and thus can significantly affect its
chemical stability. Phase transformation during
pharmaceutical processing is often a likely cause of
chemical degradation of the drug product even if
the starting API is stable. In particular, the small
amount of disordered or amorphous content, in
combination with typical moisture absorption in
pharmaceutical products, could greatly impact drug
degradation. Excipients influence drug degradation
by direct interactions or modification of the
microenvironment. A thorough understanding
of the physicochemical principles is critical to the
anticipation, identification, and remediation of
stability issues during formulation development,
manufacturing, and storage of drug products.
REFERENCES
1. Baertschi, S. W., Pharmaceutical Stress Testing. Predicting
Drug Degradation, Taylor & Francis Group: Boca Raton,
Florida, 2005.
2. Yoshioka, S. and Stella, V. J., Stability of Drugs and Dosage
Forms, Kluwer Academic/Plenum Publishers: New York,
NY 10013, 2000.
3. Zhou, D.; Porter, W. and Zhang, G. G. Z., Drug Stability
and Stability Studies, Developing Solid Oral Dosage Forms -
Pharmaceutical Theory & Practice, Qiu, Y., Chen, Y., Zhang,
G., Liu, L., Porter, W., Eds.; Academic Press: San Diego,
CA, 2009, pp 87-124.
4. Narang, A. S.; Rao, V. M. and Raghavan, K. S., Excipient
Compatibility, Developing Solid Oral Dosage Forms
Pharmaceutical Theory & Practice; Qiu, Y., Chen, Y., Zhang,
G., Liu, L., Porter, W., Eds.; Academic Press: San Diego,
CA, 2009, pp 125-145.
5. Hovorka, S. W. and Schneich, C., Oxidative Degradation
of Pharmaceuticals: Theory, Mechanisms and Inhibition,
J Pharm Sci 2001, 90, 253-269.
6. Waterman, K. C.; Adami, R. C.; A lsante, K. M.;
Hong, J.; Landis, M. S.; Lombardo, F. and Roberts,
C. J., Stabilization of Pharmaceuticals to Oxidative
Degradation, Pharm Dev Technol, 2002, 7, 1 - 32.
7. Zhang, G. G. Z.; Law, D.; Schmitt, E. A. and Qiu, Y.,
Phase Transformation Considerations During Process
Development and Manufacture of Solid Oral Dosage
Forms, Adv Drug Del Rev 2004, 56, 371-390.
8. Connors, K. A., Chemical Kinetics: The Study of Reaction
Rates in Solution, Willey-VCH Publishers: New York, 1990,
p 191.
9. Turro, N. J., Modern Molecular Photochemisty, University
Science Books: Sausalito, CA, 1991.
10. Garrett, E. R., Prediction of Stability in Pharmaceutical
Preparations. IV. The Interdependence of Solubility and
Rate in Saturated Solutions of Acylsalicylates, J Am Pharm
Ass 1957, 46, 584-586.
 Deliang Zhou

11. Schmidt, G. M. J., Topochemistry. III. The Crystal Angiotensin-converting Enzyme Inhibitor, Moexipril
Chemistry of Some Trans-cinnamic Acids, J Chem Soc Hydrochloride: Dry Powder vs Wet Granulation, Pharm
1964, 2014-2021. Res 1990, 7, 379-83.
12. Sukenik, C. N.; Bonopace, J. A.; Mandel, N. S.; Bergman, 18. Florey, K., Analytical Profiles of Drug Substances, Academic
R. C.; Lau, P. Y. and Wood, G., Enhancement of a Press: New York, 1991, p 588.
Chemical Reaction Rate by Proper Orientation of Reacting 19. Tabata, T.; Makino, T.; Kashihara, T.; Hirai, S.; Kitamori, N.
Molecules in the Solid State, J Am Chem Soc 1975, 97, and Toguchi, H., Stabilization of a New Antiulcer Drug
5290-5291. (lansoprazole) in the Solid Dosage Forms, Drug Dev Ind
13. Zografi, G., States of Water Associated with Solids, Drug Pharm 1992, 18, 1437-47. JVT
Dev Ind Pharm 1988, 14, 1905-1926.
14. Ahlneck, C. and Zografi, G., The Molecular Basis of ARTICLE ACRONYM LISTING
Moisture Effects on the Physical and Chemical Stability API Active Pharmaceutical Ingredient
of Drugs in the Solid State, Int J Pharm 1990, 62, 87-95. BHT Butylated Hydroytoluene
15. Kontny, M. J.; Grandolfi, G. P. and Zografi, G., Water HPC Hydroxypropyl Cellulose
Vapor Sorption of Water-soluble Substances: Studies HPMCAS Hydroxypropyl Methylcellulose Acetate
of Cr ystalline Solids Below their Critical Relative Succinate
Humidities, Pharm Res 1987, 4, 104-12. ICH International Conference on Harmonisation
16. Wardrop, J.; Law, D.; Qiu, Y.; Engh, K.; Faitsch, L. and Ling, PEG Polyethylene Glycol
C., Influence of Solid Phase and Formulation Processing QbD Quality by Design
on Stability of Abbott-232 Tablet Formulations, J Pharm RH Relative Humidity
Sci 2006, 95, 2380-92.
17. Gu, L.; Strickley, R. G.; Chi, L. H. and Chowhan, Z. T.,
Drug-excipient Incompatibility Studies of the Dipeptide

Originally published in the Summer 2009 issue of The Journal of Validation Technology

PROCESS VALIDATION Process Design 41

John F. Bauer
Patent Potential
John F. Bauer
Pharmaceutical Solids discusses scientific principles
associated with pharmaceutical solids useful to
practitioners in validation and compliance. We
intend this column to be a useful resource for daily
work applications. The key objective for this column:
Usefulness.
Reader comments, questions, and suggestions
are needed to help us fulfill our objective for this
column. Case studies illustrating principles associated
with pharmaceutical solids submitted by readers
are most welcome. Please send your comments and
suggestions to column coordinator John Bauer at
Consultjbnow@gmail.com or to journal coordinating
editor Susan Haigney at shaigney@ advanstar.com.
KEY POINTS
The following key points are discussed in this article:
t
Pharmaceutical compounds with specific physical
properties may be patentable
t
It is important to show a correlation between the
specific physical property and a useful benefit in
manufacturing, stability, or some other area
t
Although polymorphism is a common
phenomenon, the existence of stable polymorphs
is not obvious
t
It is important to identify an analytical method
in the patent that can be used to demonstrate
infringement.
INTRODUCTION
The various physical properties of pharmaceuticals and
their impact on the behavior and usefulness of these
materials have been discussed in earlier installments of
Pharmaceutical Solids. It is in some cases painfully
obvious that having an active pharmaceutical ingredient
(API) or individual excipients with the wrong set of
physical properties can negatively impact the ability
to successfully manufacture a dosage form or delivery
For more Author
information,
go to
gxpandjvt.com/bios [ ABOUT THE AUTHOR
John F. Bauer, Ph.D., is president of Consult JB LLC Pharmaceutical Consultants. Dr. Bauer has
more than 30 years of pharmaceutical industry experience, including work in solid-state chemistry,
analytical chemistry, stability, pharmaceutics, regulatory CMC, patents, and litigation. He may be
reached at Consultjbnow@gmail.com.
42 PROCESS VALIDATION Process Design
system. Furthermore, there are situations when having
a particular property or set of properties can greatly
increase the quality and throughput of the manufacturing
process. When such examples of favorable ranges for a
physical property exist, they may qualify for patenting. In
this column we will investigate the concept of patenting
materials with specific physical properties.
Constitutional Basis
Article I, Section B, Clause 8 of the Constitution of the
United States of America states, The Congress shall have
power to promote the progress of science and useful
arts by securing for limited times to the authors and
inventors the exclusive right to their respective writings
and discoveries. This is the basis of todays patent
system in the United States. However, the constitution
directs neither the criteria to be used in judging
patentability nor does it define the limited time to be
granted the inventor.
United States Constitution Sec. 101 35 clarifies
the criteria somewhat when stating whoever invents
or discovers any new and useful process, machine,
manufacture, or composition of matter, or any new
and useful improvements thereof, may obtain a patent
therefore.
In the committee notes accompanying the patent
act of 1952, it is noted that congress intended that the
statutory subject matter include anything under the
sun that is made by man. As broad as this statement
sounds, there are some things that are specifically not
included, such as the following:
t
Laws of nature. Sir Isaac Newton would not have
been allowed to patent gravity.
t
Physical phenomena. This exclusion would have
prevented Sir Edmund Hillary from patenting any
rock formations he found on Mt. Everest.
t
Abstract ideas. Thus denying Shirley MacLaine
patents on reincarnation and aliens.

PATENTABILITY
Pharmaceuticals with unique physical properties fit the
requirements for patentability. The patent criteria are
that the discovery or invention must be new, useful,
and non-obvious. It is easily understood that a new
chemical entity never prepared before would fit these
criteria. Consequently, the first patent usually filed for
a new drug is for the compound itself and is called a
composition of matter patent. This is generally filed
as soon as possible after the initial discovery in order to
prohibit others from developing the same compound.
When granted, the patent gives the patentee the right
to exclude others from making, using, selling, offering
to sell, or importing the invention for a limited period
of time. In the mid 1990s, the time frame of exclusivity
was redefined in the United States such that patents
filed prior to June 1995 carry 17 years of exclusivity
starting at the date of issuance, and patents filed after
June of 1995 received 20 years exclusivity from the date
of filing.
Because development of a pharmaceutical from
discovery to market can average 10 years, the actual
exclusivity of marketed drug is significantly shorter
than the 20 years defined by Congress.
For this reason it is important that scientists
throughout the development chain take note of any
truly new, useful, and non-obvious aspects of the
compound and/or product that might be patentable.
PATENTS BASED ON PHYSICAL
PROPERTIES
The physical properties of a pharmaceutical have been
considered to fall into the category of patent-potential
material. Probably the most obvious and most widely
patented physical aspect of pharmaceuticals is the
crystal form.
Although polymorphism is coming to be seen as
a reasonably common and widespread phenomenon
among organic compounds, it is not obvious. There
is not presently a way of predicting for any particular
compound whether or not other stable crystal forms
(such as polymorphs or solvates) exist. Several
computer programs are in commercial use that
can estimate the energy of various crystal packings
and speculate as to whether other polymorphs of a
compound are possible. However, these programs
cannot demonstrate that another form can be made
and remain stable under normal storage conditions.
Because of this, the discovery of a new polymorph
continues to meet the nonobvious requirement.
Once it is produced for the first time, it is certainly
newso the remaining qualification is usefulness.
As described in an earlier column (1) various
polymorphs and solvents, especially hydrates, can
John F. Bauer
possess significantly different properties from each
other. These physical properties are the usual basis of
their usefulness and patentability. Additionally, these
physical properties can, in and of themselves, be the
subject of additional patents. As the Figure illustrates,
patents can be thought of as umbrellas that protect the
innovators discovery. They can be of various sizes in
so far as they protect the product to varying extents.
The composition of matter patent on a pharmaceutical
provides the greatest protection, because it excludes
others from using the compound in any way and
in any form. As a subset of that protection, one can
potentially receive patents protecting individual
aspects of the compound such as its solid form and
individual physical properties. These umbrellas do not
provide such broad protection but can often extend
slightly beyond the composition of matter patent and
thus extend exclusivity, depending on when their
benefits are discovered.
EXAMPLE PATENTS ON PHYSICAL
PROPERTIES
The following are examples of patents on physical
properties of pharmaceuticals.
Ranitidine
In the mid 1990s, GlaxoSmithKlines ulcer treatment
Zantac was the beneficiary of a second polymorph.
Ranitidine hydrochloride, the active ingredient in
Zantac was found to exist in a second polymorphic
form. This second form was easier to manufacture and,
therefore, had a demonstratable usefulness. The patent
(2) on the second form extended Glaxos exclusivity on
Zantac.
Terazosin
A similar, more involved situation existed with
Abbotts antihypertensive drug Hytrin. The original
composition of matter patent for terazosin, the active
ingredient, was on an anhydrate (3)a form in which
there were no water molecules involved in the crystal
structure. This form had high solubility but was
also hygroscopic and tended to absorb water from
the atmosphere making it difficult to handle during
manufacturing. Later a hydrate form was discovered
that was new, non-obvious, and had the usefulness of
not being hygroscopic. Abbott obtained a patent on
the hydrate form (4). The drawback of this form was
lowered solubility. Still later a second anhydrate form
was found that was not hygroscopic and had increased
solubility; Abbott was also able to obtain a patent on
this form (5). This is a good example of how changes in
physical form can lead to beneficial changes in physical
properties that qualify for patenting.
PROCESS VALIDATION Process Design 43
John F. Bauer
Figure: Depiction of how supplemental patents can
extend product exclusivity.
Composition of Matter
Solid Form
Individual PP
Exclusivity Time
Pp = physical property
Ritonavir
The previous examples involve different crystal forms.
There are times when the non-crystalline or amorphous
form can also be useful. In most situations, amorphous
drug has non-favorable properties (6). It is hygroscopic,
often less stable, and can be hard to handle in
manufacturing. However, in the case of ritonavir,
an important protease inhibitor for the treatment of
AIDS, the crystalline forms are not bioavailable from
the solid state and must be dosed in solution. The
amorphous form, on the other hand, is very soluble
and is bioavailable in a solid oral dosage form; this
property presents a strong case for patent eligibility.
These examples emphasize the necessity of continually
observing the manufacturing and formulation
processes with a view for unique properties of the solid
forms that can be useful and can present potential
patent possibilities.
Sirolimus
Although solid form is the most obvious and
predominant basis for physical property patents,
other characteristics can also qualify if the non-
obvious and usefulness criteria can be met. Particle
size is another well known physical characteristic
that can affect manufacturing and performance
of pharmaceutical preparations. However, some
patent examiners would consider some properties
related to particle size such as increase dissolution
rate with smaller particle size as being obvious to
one skilled in the art. More individual and unique
relationships between ones individual process and
these properties must be established. This is usually
done by comparison with drug having other particle
size ranges. A good time to obtain these data is when
establishing the design space for specifications,
during development, process justification studies,
and studies supporting process validation. In spite of
44 PROCESS VALIDATION Process Design
this possible objection, patents have been obtained
with claims relating specifically to particle size and/or
surface area. For example, Wyeth has a patent on the
macrolide antibiotic sirolimus, the active ingredient
in Rapamycin. The patent claims sirolimus particles
having a D90 value of from about 2 microns to about
1 micron (7). D90 means 90% of the material is
within this particle size range.
Wyeth based the usefulness and non-obviousness
for its claim on the observation that compositions
of Sirolimus with conventional excipients
show unpredictable dissolution rates, irregular
bioavailability, and stability problems. These present
very real obstacles to processing. Using drug within the
specified particle size range minimizes these problems.
Noticing this correlation during process development
allowed Wyeth to obtain exclusivity on this valuable
manufacturing advantage.
Nifedipine
Surface area can also be the basis for patenting and
product exclusivity as in the case of Bayers patent on
nifedipine capsules. The claim reads an effective
amount of nifedipine crystals with a specific surface
area of 1.0 to 4 m2/g (8).
As mentioned previously, an ideal time to monitor
for patent potential of physical properties is while
establishing the design space and specifications during
process justification. The reason for this is that patent
claims should not be limited only to the parameters
used in manufacturing, but should cover the entire
range of parameters that achieve the stated usefulness
and non-obviousness cited in the patent. A patent
that can be easily circumvented provides very little
exclusivity.
ANALYTICAL CONSIDERATIONS
Another important aspect to be considered is the
ability to demonstrate infringement. One may
be able to obtain a patent on a valuable property.
However, if there are no clear cut analytical methods
that can demonstrate infringement, then the patent
protection will be minimal. For example a patent
on a drug patch that claims 70% of drug is in the
solid state on the patch and 30% is dissolved may
not be easily enforceable because of the analytical
challenge of demonstrating such a distribution. Once
a patent is issued it is the responsibility of the patent
holder to enforce the exclusivity. A patent does not
give the inventor the right to market the product. It
only excludes others from using the invention. In
the area of pharmaceuticals, the potential marketer
of a drug must certify to the US Food and Drug
Administration that its product and process do not
infringe anyone elses patents in what is called a notice
of certification. Neither the FDA nor the US patent

department has any obligation to help the inventor
demonstrate infringement or enforce exclusivity rights.
It is, therefore, very important to prepare a patent
application with proof of infringement in mind. As
mentioned before, a patent is only an impressive
government document if it cannot be enforced. An
analytical method should be described in the body
of the patent that is demonstrative of and specific for
the property cited in the patent claim. Some examples
are x-ray powder diffraction or solid-state nuclear
magnetic resonance for detecting small amounts
of a claimed crystal form in a drug sample. In the
case of crystal form patents, analytical sensitivity is
important. X-ray powder diffraction can generally
detect about 1% of a second form. Particle size and
surface area especially require very specific method
descriptions concerning instruments and operating
parameters because results for these properties can vary
significantly with technique and sample preparation. If
Originally published in the Autumn 2009 issue of The Journal of Validation Technology
John F. Bauer
this kind of pre-thought and care are put into the patent
application preparation, long and often unsuccessful
patent defenses can be avoided, and patents on physical
properties can be extremely beneficial.
REFERENCES
1. Bauer, John F., Pharmaceutical Solids: PolymorphismA
Critical Consideration in Pharmaceutical Development,
Manufacturing, and Stability, Journal of Validation
Technology, Autumn 2008.
2. United States Patent 4672133, June 1987.
3. United States Patent 4,026,894, May 1977.
4. United States Patent 4,251,532, February 1981.
5. United States Patent 5,294,615, March 1994.
6. United States Patent 5,294,615, March 1994.
7. Bauer, John F., Pharmaceutical Solids-The Amorphous
Phase, Journal of Validation Technology, Summer 2009.
8. United States Patent 5264446, November 1993. JVT
PROCESS VALIDATION Process Design 45
John A. Wass
First Steps in
Experimental Design
The Screening Experiment
John A. Wass
Statistical Viewpoint addresses principles of
statistics useful to practitioners in compliance and
validation. We intend to present these concepts in a
meaningful way so as to enable their application in
daily work situations.
Reader comments, questions, and suggestions are
needed to help us fulfill our objective for this column.
Please send your comments to coordinating editor
Susan Haigney at shaigney@advanstar.com.
KEY POINTS
The following key points are discussed in this article:
t
Design of experiments (DOE) consists of three
basic stages: screening (to identify important
factors), response surface methodology (to define
the optimal space), and model validation (to
confirm predictions
t
A critical preliminary step in the screening stage is
for subject matter experts to identify the key list of
factors that may influence the process
t
A DOE design consists of a table with rows
that represent experimental trials and columns
(vectors) that give the corresponding factor levels.
In a DOE analysis, the factor level columns are
used to estimate the corresponding factor main
effects
t
Interaction columns in a design are formed as the
dot product of two other columns. In a DOE
analysis, the interaction columns are used to
estimate the corresponding interaction effects
[
For more Author ABOUT THE AUTHOR
information, John A. Wass, Ph.D, is a consulting statistician with Quantum Cats Consulting in the Chicago
go to area, as well as a contributing editor at Scientific Computing and administrator of a regional
gxpandjvt.com/bios statistical software group. He may be reached by e-mail at john.wass@tds.net.
46 PROCESS VALIDATION Process Design
t
When two design columns are identical, the
corresponding factors or interactions are aliased
and their corresponding effects cannot be
distinguished
t
A desirable feature of a screening design is
orthogonality in which the vector products of any
two main effect or interaction columns sum to
zero. Orthogonality means that all estimates can
be obtained independently of one another
t
DOE software provides efficient screening designs
with columns that are not aliased and from which
orthogonal estimates can be obtained
t
Full-factorial designs include all combinations of
factor levels and provide a predictive model that
includes main effects and all possible interactions
t
Fractional factorial (screening) designs include
fewer trials and may be more efficient than the
corresponding full factorial design
t
The concept of aliasing is one of the tools that
can be used to construct efficient, orthogonal,
screening designs
t
Center points are often included in screening
designs to raise the efficiency and to assess lack of
model fit due to curvature
t
The order of running and testing experimental
trials is often randomized to protect against the
presence of unknown lurking variables
t
Blocking variables (such as day or run or session)
may be included in a design to raise the design
efficiency

t
Factor effects in screening designs may be missed
because they were not included in the screening
experiment, because they were not given
sufficiently wide factor ranges, because the design
was underpowered for those factors, because trial
order was not properly randomized or blocked, or
because of an inadequate model.
INTRODUCTION
In days of old (i.e., the authors undergraduate
years), we were introduced to the joys of manual
calculations and analysis of variance (ANOVA).
Experimenters would change one factor at a time
and identify what they felt were optimal processing
conditions. With the advent of personal computers
and the dissemination of more efficient techniques by
statisticians, problems of increasing complexity were
solved. This not only enlightened the basic researcher,
but permitted scientists and engineers to design more
robust products and processes. In fact, the statistical
design of experiments (DOE) has been called the most
cost-effective quality and productivity optimization
method known. In this brief introduction, we
will concentrate on practical aspects and keep
mathematical theory to a minimum.
The advent of DOE brought a modicum of order
to the wild west of one-factor at a time changes. The
technique has many variations but consists of the
following three basic stages:
t
Screeningto exclude extraneous effects
considered as noise
t
Response surface methodologyto finely define
the optimal result space
t
Model validationto confirm predictions.
Each is quite important. In this paper we will
concentrate on the first stage, screening design and
analysis.
There are many commercial software packages,
either standalone or modules, within general statistics
programs that will support DOE. Some of these are
listed later in this discussion. Each has its unique
strengths and weaknesses. For this paper, JMP8 has
been used. The principles will be the same for most
programs; although, the user interfaces, algorithms,
and output will vary.
THEORY
The literature of DOE is replete with names such as
full factorial, fractional factorial, runs, power, levels,
and interactions. In addition, we have categorical
and continuous factors and a variety of design names.
Fortunately, in screening we usually confine ourselves
to the fractional factorial designs. Unfortunately, as
John A. Wass
with everything in real-life, there is a price to pay for
every extra bit of information required.
We can show an experimental design as a table.
An example is presented in Figure 1. Each row in the
table corresponds to an experimental trial. The table
columns indicate the levels of the experimental factors.
For screening designs we usually consider only two
levels, usually coded +/-1. In this notation, + represents
high and represents low. We may also include
other columns that indicate interactions among the
factors. The columns giving the experimental factor
levels permit us to estimate main effects and the
interaction columns permit us to estimate interaction
effects. We will say more about main effects and
interaction effects below. In addition to factor level
and interaction columns, we may record one or more
columns of measured variable values that result from
each trial. We refer to these dependent variables as the
experimental responses.
Screening designs are useful as they are a practical
compromise between cost and information. Their
main contribution is in suggesting which of many
factors that may impact a result are actually the most
important. Because screening designs require fewer
runs, they are far less costly than the more informative
full-factorial designs where the practitioner uses all
combinations of factor levels. It has been suggested
that no more than 25% of the total budget for DOE
be spent on the screening runs. Screening runs are
usually a prelude to further experimentation, namely
the response surface and confirmatory runs, where
specific information is gained around target (desired)
outcomes.
Key Assumption For Screening Studies
In screening designs, we make the assumption that
our real-world processes are driven by only a few
factors, the others being relatively unimportant. This
usually works quite well but it is a crucial assumption
that requires careful consideration by subject matter
experts. Also keep in mind that the fractional
factorial designs may be upgraded to full factorial
designs (main effects plus all interactions) if there are
only a few main effects. This allows us to observe
interactions at a reasonable cost.
Number Of Runs
With screening designs, responses are taken only for
a small fraction of the total possible combinations to
reduce the number of runs and thus cost. The total
number of runs is calculated by raising the number of
levels to the power of the number of factors (e.g., for
three factors at two levels each we have runs = 2^3 =
2x2x2 = 8). This is actually a full factorial design as
PROCESS VALIDATION Process Design 47
John A. Wass
we are testing all combinations of factor levels. Full
factorial designs allow us to build predictive models
that include the main effects of each factor as well
as interactions. This brings us to three important
concepts of these models: interaction (the effects
of one factor on another), orthogonality (all factors
are independent of one another), and aliasing
(when the effects due to multiple factors cannot be
distinguished).
Interactions
One of the more important things that practitioners
need to know about is that main factors may affect each
other in ways known and unknown (i.e., interaction
among effects). For example, the interaction of two
reagents in a chemical process may be a significant
driver of the overall process (think enzyme and
substrate). In deciding which are important, statistically
and physically, it is necessary to consult with the bench
scientists and technicians to get a handle on what
is already known and suspected to be important to
the process. Too few factors risk missing something
important. Including too many factors will render the
screening more costly and lead to a lack of orthogonality
due to aliasing. Aliasing occurs when two columns
in the design (referred to by statisticians as a vector of
the input space) are identical or when one column is
identical to another formed from the interaction of
two columns (i.e., the vector or dot product of two our estimate of the other effect. Orthogonal designs
columns).
Figure 1 presents a design table for a full factorial
design in three factors. This design requires eight trials
(rows) and has three factors (A, B, C) for which we can
estimate main effects. It is also possible to estimate
all possible two-way interactions with this design (AB,
AC, BC) as well as the single three-way interaction ABC
that is shown in the design table. If later, the design is
augmented with a fourth factor D (all runs not shown
below), we have a problem. Now the contribution
to any measured outcome (effect) from ABC is
indistinguishable from D and, therefore, we do not
know if the driver was D or ABC.
The A, B, and C columns give the levels (coded +/-) of
four experimental design factors. The ABC interaction
column is formed as the dot product of the A, B, and
C columns. Notice how each row in the ABC column
is formed as a product of the corresponding levels of
A, B, and C. In the analysis of such a design, the ABC
interaction column would be used to estimate the
corresponding ABC interaction effect.
The sums of the levels of the A, B, C, and ABC
columns all equal zero. This means that the design is
balanced. Balanced designs give more precise estimates
of main and interaction effects than unbalanced
designs.
48 PROCESS VALIDATION Process Design
Figure 1: Aliasing in a simple design.
Run Factors
A B C ABC D
1 - - - - -
2 - - + + +
3 - + - + +
4 - + + - -
5 + - - + +
6 + - + - -
7 + + - - -
8 + + + + +
Sum 0 0 0 0 0
Orthogonality
Further, the dot product of any two of the columns
A, B, C, or ABC will also sum to zero (try it and
see). This more subtle design characteristic is called
orthogonality and is critical to good experimental
design. To understand why orthogonality is so
important, we return to our concept of aliasing.
Aliasing is the extreme absence of orthogonality.
It is impossible to separately estimate the effects
corresponding to two design columns that are aliased.
In a sense, such estimates are 100% correlated (the
statistical term is confounded). In contrast, when
two design columns are orthogonal, the corresponding
effect estimates have zero correlation. This means that
errors in estimating one effect do not, on average, bias
prevent us from accidentally confusing the effects of
two different factors.
DOE software provides us with orthogonal (or
nearly orthogonal) screening designs in which the
main effect and interaction columns are not aliased.
These allow us to estimate the corresponding effects
without worrying about possible correlations. This
lack of correlation usually implies that the estimates
are independent.
Figure 2 gives a good mental image of the value of
orthogonality and the independence it provides in
our estimates. Let the three mutually perpendicular
(orthogonal) axes X, Y, and Z represent dimensions
along which three estimates from some experimental
design may lie. The results of an experiment may
then be indicated as a point in that three dimensional
space. If we repeat the experiment many times,
the resulting estimates will form a cluster of points
in space, centered about the true effects being
estimated. With orthogonal designs, the cluster of
points will be spherical in shape indicating a lack of
correlation (or independence) among the estimates.
In designs with aliasing, the points will fall along a
one-dimensional line or two-dimensional plane that
indicate a complete correlation of three or two effects,
respectively. In between these extremes will be designs
that will produce ellipsoid-shaped clusters, whose axes
 John A. Wass

are not parallel to X, Y, and Z. Such designs are less Figure 2: Orthogonality.
efficient than fully orthogonal designs and may result
in misinterpretation of screening results. Y
Similarly, when the columns in our experimental
design are unaliased and when the vector products of
any two columns sum to zero, the corresponding effect
estimates are mutually independent. This means that
random errors in estimating one effect will not (on X
average) bias the estimate of another effect. Designing
for orthogonality is excellent protection against
aliasing. Z

Aliasing
The concept of aliasing is also useful in constructing fast generation of an experimental sheet showing the
efficient fractional factorial designs. The design in factors and their levels, plus a column for the results.
Figure 1 includes an ABC interaction column. In This design has three input factors (X1-X3) and
the real world, two-way interaction effects may well a single output (Y). The levels indicate low (-1),
be present, but three-way interactions such as ABC median (0), and high (1) levels of the factor. The
are often considered unlikely to be important. This trials for which all factors are at their median level
is referred to as the sparcity of effects principle. are called center points and may represent target or
Based on our knowledge of the process, we may control factor settings. The 000 points are the center
be willing to assume that the ABC interaction points and are useful for testing linearity in the
effect is not present. In that case we might consider process. It is a simple and inexpensive way to check
including factor D in our experiment and setting for curvature. We now have nine design points, each
its levels to those indicated by ABC. In this way we replicated twice to yield 18 runs. This design is a
can estimate the main effects of all four factors, full factorial design in which each of the nine trials
although we sacrifice our ability to learn about in the full factorial is replicated twice. Replication
the ABC interaction. This is an example of using
aliasing as a design tool. We indicate this aliasing Figure 3: A typical screening design.
mathematically by writing ABC=D. This is
not meant to imply that the effects of D are the
same as the interaction of ABC, but only that our
experiment cannot distinguish the main effect of Pattern X1 X2 X3 Y
D from the ABC interaction. We have now created 1 --- -1 -1 -1 *
a design with four factors in only eight trials (a 2 --+ -1 -1 1 *
full factorial design with four factors would have
3 000 0 0 0 *
required 2^4=16 trials).
As this is an introduction we will not delve 4 -+- -1 1 -1 *
into the more advanced concepts of saturation, 5 +-+ 1 -1 1 *
efficiency, foldover, and the more complex 6 ++- 1 1 -1 *
screening designs. Any of the textbooks cited in the 7 +++ 1 1 1 *
reference section may be consulted for these topics.
8 +-+ 1 -1 1 *
GENERAL TECHNIQUES 9 +-- 1 -1 -1 *
In most modern software it is a straightforward 10 --+ -1 -1 1 *
matter to design an experiment and analyze the 11 -++ -1 1 1 *
results (with a little practice and much consulting
12 000 0 0 0 *
of the manual). Its the decision as to what factors
to include that is critical. It is strongly advised 13 -+- -1 1 -1 *
that you not throw in everything and the kitchen 14 +-- 1 -1 -1 *
sink for fear of missing something. This is where 15 -++ -1 1 1 *
consultation with subject matter experts like the 16 ++- 1 1 -1 *
process engineers and bench scientists familiar with
17 +++ 1 1 1 *
the process or product is crucial.
Figure 3 is an example of an experimental design 18 --- -1 -1 -1 *
produced by the JMP software. The software allows

PROCESS VALIDATION Process Design 49

John A. Wass
is sometimes needed to provide sufficient statistical
power. These factor levels are obviously coded but
the actual numbers could be entered and more
factors screened against multiple outputs, if desired.
An Example
The medical diagnostics industry makes great use
of DOE for myriad products. In this example, we
will look at the design and analysis of a clinical
laboratory kit pack for an unspecified analyte.
We wish to determine and minimize the product
variability given certain known conditions of reagent
concentration and physical/environmental factors for
the chemical reactions occurring in the kit pack. The
input factors are as follows:
t
Reagent 1
t
Reagent 2
t
Enzyme
t
Temperature
t
Mixing speed.
In this example, the two reagents are combined
with the enzyme in a reaction vessel at a specified
temperature. The mixture is agitated at several
mixing speeds and the researcher wants to know how
important each of the factors (i.e., reagents, enzyme,
temp, and mix speed) is to the concentration of the
final product.
All are held at either low, median, or high
levels. The single output (response) variable is the
concentration of the product. We will minimize the
variability of the concentration measurement (i.e.,
for the given inputs we wish to hold the measured
concentration to be within a stated reference range
with minimal variability). This reference range could
be from the clinical range for that particular analyte
in human fluids or dictated by the company quality
document.
With five factors, the full factorial design would
2^5=32 trials; however, that is about twice as many
trials as our budget will allow. Also, our statistician
informs us that the addition of center points in the
design will increase efficiency and replicates are
necessary to evaluate random error. Our experiment
will require four sessions to complete, and we are
concerned that experimental conditions may vary
from session to session. We would like to design
the experiment in some way so that any session
differences are not misconstrued as factor effects. So
our statistician suggests the use of blocking in our
design. Blocking is the arranging of experimental
units (the runs) in groups, or blocks, that are similar to
one another.
We block our designs to reduce random noise
between sessions (in this case) as the experiment was
carried out over several time periods and we expect the
50 PROCESS VALIDATION Process Design
data within any one block to be more homogeneous
than that between blocks. Thus, greater precision
is obtained by doing within-block comparisons, as
inter-block differences are eliminated. Also runs are
randomized within a block, and block order may
be randomized if necessary. Note that although
this screen is designed for maximal efficiency and
precision, only main effects order effects may be
estimated. After gaining management approval for the
cost, we obtain the design using JMP software. Note
that the block column here indicates the session in
which to perform each trial (see Figure 4).
Notice that we have five factors and that a full
factorial design would require 2^5 or 32 runs yet our
design only has 18. The reduction comes in an aspect
of DOE called design efficiency. This is a measure
of the efficiency of any given design in covering the
design space, compared to a 100% efficient design
that would cover all of the points needed to extract
maximal information about the dependency of the
results upon the input factors. Modern DOE software
employs highly efficient algorithms based upon
some very complex mathematics that will reduce the
number of points needed while minimally impacting
efficiency.
Having designed the experiment, we go into the lab
and collect the data then return to the JMP design table
and fill in the Y column of the table (see Figure 5).
The data can be analyzed in JMP. The first step
is to specify the model. The model can then be
automatically generated as an effect screen using JMPs
standard least squares platform. The model contains
a list of all the factor effects that we wish (and are
able) to estimate. Our estimates will include the five
main factor effects, the block effect and the two-way
interaction between reagents 1 and 2 (see Figure 6).
Running this model shows us an actual vs. predicted
plot for the product (see Figure 7).
As the p-value from the analysis of variance shows
that the model is very significant (p<< 0.01), the
R-squared is 92% and the estimate of the standard
deviation of the process noise (RMSE) is quite low,
we are initially happy with the model fit. However,
Figure 8 shows the parameter estimates that test the
significance of the individual factors to the overall
model.
Based upon the usual p-value cutoff of 0.05, only the
effects of the two reagents were statistically significant.
The intercept is usually not of interest in the physical
sense of the model as it only gives information about
the 0 point of the x-axis.
As the chemists suspected, the reagents are quite
important to the product and the mix speed was not.
However, the lack of an enzyme and temperature effects
is surprising and may indicate a flaw in the design.
Either the wrong values of those factors were chosen
 John A. Wass

Figure 4: Example problem screening design.

Pattern Block Reagent 1 Reagent 2 Enzyme Temp MixSpeed Y

1 00000 1 0 0 0 0 0 *
2 +--++ 1 1 -1 -1 1 1 *
3 ++--- 1 1 1 -1 -1 -1 *
4 -++-+ 1 -1 1 1 -1 1 *
5 --++- 1 -1 -1 1 1 -1 *
6 +-+-- 2 1 -1 1 -1 -1 *
7 +++++ 2 1 1 1 1 1 *
8 ----+ 2 -1 -1 -1 -1 1 *
9 -+-+- 2 -1 1 -1 1 -1 *
10 -++-+ 3 -1 1 1 -1 1 *
11 ++--- 3 1 1 -1 -1 -1 *
12 00000 3 0 0 0 0 0 *
13 +--++ 3 1 -1 -1 1 1 *
14 --++- 3 -1 -1 1 1 -1 *
15 ----+ 4 -1 -1 -1 -1 1 *
16 -+-+- 4 -1 1 -1 1 -1 *
17 +-+-- 4 1 -1 1 -1 -1 *
18 +++++ 4 1 1 1 1 1 *

(the range was not sufficiently wide to see the effect),
the design was underpowered for those factors (Power
in this case is taken as: not enough runs to discern a
difference given that one truly exists), or both. The one
interaction and the block tests cannot truly be discerned
(as to the true effect on product) as we have not the
resolution to do so with this design. We can test the
power, however, for the surprises in the main effects (see
Figure 9).
These are both low and we would need more data to
determine true lack of significance. For guidance on
what constitutes adequate power, see the Tips, Tricks,
and Pitfalls section. The Pareto Plot (see Figure 10)
shows the results graphically.
The prediction profiler (see Figure 11) shows us how
to set factors to maximize our product and the mean
and 95% confidence interval for the mean response.
Our task is now to determine the best settings for
the temperature and enzyme factors and to augment
the design with enough data to ensure adequate power.
Once this is achieved, we can zero in on maximal
setting and minimal variation using response surface
methodology.
t
Plackett-Burman. Resolution is higher than
simple factorial and this is a good technique for
large numbers of input factors. Also more flexible
as the runs need not be a power of 2. All columns
are balanced and pairwise orthogonal.
t
Box-Behnken. Efficient technique for modeling
quantitative three-level factors. Hold some of the
factors at their centerpoints and are easily blocked.
t
Box-Wilson. These are central composite designs
that may be used when significant interaction is
suspected. They have the desirable property of
rotatability where the predicted response may be
estimated with equal variance from any direction
from the center of the design space.
TIPS, TRICKS, AND PITFALLS
Experimental design is an art. There are many
detailed techniques and lessons learned from years of
experience that novices may have to gradually learn
on their ownthe hard way. The following comments
are meant to minimize the pain for experimenters who
wish to take advantage of these modern experimental
tools.

Other Designs Curvature

Fractional and full factorials are not the only screening First and foremost is the fact that screening is a linear
designs available. The following are among the most process and as such does not take curvature and higher
popular: order interactions that generate curvature into account.

PROCESS VALIDATION Process Design 51

John A. Wass

Figure 5: Test data.

Reagent Reagent Mix

Pattern Block 1 2 Enzyme Temp Speed Y
1 00000 1 0 0 0 0 0 2.4
2 +--++ 1 1 -1 -1 1 1 2
3 ++--- 1 1 1 -1 -1 -1 2.3
4 -++-+ 1 -1 1 1 -1 1 1.8
5 --++- 1 -1 -1 1 1 -1 1.5
6 +-+-- 2 1 -1 1 -1 -1 1.9
7 +++++ 2 1 1 1 1 1 3
8 ----+ 2 -1 -1 -1 -1 1 0.9
9 -+-+- 2 -1 1 -1 1 -1 1.6
10 -++-+ 3 -1 1 1 -1 1 1.9
11 ++--- 3 1 1 -1 -1 -1 2.6
12 00000 3 0 0 0 0 0 2.6
13 +--++ 3 1 -1 -1 1 1 2.2
14 --++- 3 -1 -1 1 1 -1 1.4
15 ----+ 4 -1 -1 -1 -1 1 1.1
16 -+-+- 4 -1 1 -1 1 -1 1.8
17 +-+-- 4 1 -1 1 -1 -1 2.1
18 +++++ 4 1 1 1 1 1 3.1

Figure 6: Data analysis. Figure 7: Actual by predicted plot.

2.5
Y Actual

1.5

1 1.5 2 2.5 3
Y Predicted P=0.0019
RSq=0.92 RMSE=0.253

52 PROCESS VALIDATION Process Design


If the model renders a poor fit, and the graph of the
data appears to suggest curvature (points clustered
above and below the line), the experimenter is well
advised to go to a more complex model (e.g., add a
center point or quadratic). Also note that screening
may underestimate the true number of main factors,
depending upon choices and basic knowledge of the
process. Lack-of-fit should also be carefully assessed,
as by its very nature screening will evidence poor lack
of fit for any curved process despite the fact that the
selections of main factors were complete. Bottom line:
there are problems with only looking at first order
effects.
Randomization
One of the more important concepts in all of DOE is
to randomize the order of runs. This averages over
noise that may be due to a specific time of run, type
of chemical prep, temperature, pressure, or other
condition that may be concentrated in space and time.
When runs may be dependent upon long physical
startup times, the experimenter may choose to block
certain factors.
Power
Here we refer to statistical power (i.e., the probability of
detecting a difference if one truly exists). Most software
will perform this test and it should never be ignored. In
certain cases, if the power is found to be unacceptably
low, it might dictate rerunning the entire experiment
with a larger number of runs. Acceptable power is
determined by the experimenter or may be dictated by
company or governmental rules. As a rule of thumb:
for purely research purposes, 70-80% may be adequate.
For process development 90+% is desirable and in
certain critical cases (e.g., HIV diagnostics) 98% may be
the least that is acceptable.
Aliasing
As was mentioned, due to low number of runs and
resolution in most screening designs, we encounter
situations where it is difficult to assign effects
unambiguously to single factors. To get around this
problem the experimenter may consider the following:
select different design generators, increase the number
of runs, or hold one or more factors constant.
Replication
One of the most important concepts in statistics is
replication. The use of replication provides an estimate
of the magnitude of noise. An adequate number of
replicates will ensure a more precise estimate of the
Originally published in the Spring 2010 issue of The Journal of Validation Technology
John A. Wass
variation about a process mean, as well as facilitating
the detection of true differences. For screening, we can
usually afford but a single replicate, and in some cases
it may not be necessary to replicate all of the points in
a design. The heavy replication of points is usually left
to the response surface methodologies where we really
need to get a sharper picture of the reduced design
space. Replication is one of the most underappreciated
necessities of statistical analyses and the one that will
generate the most headaches if ignored.
SOFTWARE
There are numerous software products available to
assist the practitioner in design and analysis of their
experiments. The author has had experience with the
following commercial packages:
t
JMP (www.jmp.com)
t
Design Expert (www.statease.com)
t
MODDE (www.umetrics.com)
t
Unscrambler (www.camo.no)
t
Minitab (www.minitab.com)
t
SYSTAT (www.systat.com)
t
STATISTICA (www.statsoft.com)
t
GenStat (www.vsni.co.uk).
CONCLUSIONS
Modern experimental design is sometimes art as well
as science. It is the objective of this column to acquaint
the reader with the rudiments of the screening design,
introduce them to the nomenclature and supplement
the learning experience with a real-world example.
Furthermore, as the process is not exact, several
rules-of-thumb are given to provide guidance in
those situations where hard and fast rules may not be
available.
GENERAL REFERENCES
S.R. Schmidt and R.G. Launsby, Understanding Industrial
Designed Experiments (4th ed.), Air Academy Press, 1997.
G.E.P. Box, J.S Hunter, and W.G. Hunter, Statistics for
Experimenters (2nd Ed.), Wiley Interscience, 2005.
D. C. Montgomery, Design and Analysis of Experiments (5th ed.),
John Wiley, 2001.
JMP Design of Experiments Guide, Release 7, SAS Institute Inc.,
2007.
ECHIP Reference Manual, Version 6, ECHIP Inc., 1983-1993.
Deming, S.N., Quality by Design (Part 5), Chemtech, pp 118-
126, Feb. 1990.
Deming, S.N., Quality by Design (Part 6), Chemtech, pp 604
607, Oct. 1992.
Deming, S.N., Quality by Design (Part 7), Chemtech, pp 666
673, Nov. 1992. JVT
PROCESS VALIDATION Process Design 53
John A. Wass
First Steps in
Experimental Design II:
More on Screening
Experiments
John A. Wass
Statistical Viewpoint addresses principles of
statistics useful to practitioners in compliance and
validation. We intend to present these concepts in a
meaningful way so as to enable their application in
daily work situations.
The comments, questions, and suggestions of the
readers are needed to help us fulfill our objective
for this column. Please contact our coordinating
editor Susan Haigney at shaigney@advanstar.com
with comments, suggestions, or manuscripts for
publication.
KEY POINTS
The following key points are discussed:
t
Design of experiments (DOE) consists of three
basic stages: screening to identify important
factors, response surface methodology to define
the optimal space, and model validation to
confirm predictions.
t
A critical preliminary step in the screening stage
is for subject matter experts to identify the key
list of factors that may influence the process.
t
A DOE design consists of a table whose rows
represent experimental trials and whose
columns (vectors) give the corresponding
factor levels. In a DOE analysis, the factor level
columns are used to estimate the corresponding
factor main effects.
For more Author
information,
go to
gxpandjvt.com/bios [ ABOUT THE AUTHOR
John Wass is a consulting statistician with Quantum Cats Consulting in the Chicago area as well as
a contributing editor at Scientific Computing and administrator of the Great Lakes JMP Users Group.
He may be reached by e-mail at john.wass@tds.net.
54 PROCESS VALIDATION Process Design
t
Interaction columns in a design are formed as
the dot product of two other columns. In a
DOE analysis, the interaction columns are used
to estimate the corresponding interaction effects.
t
When two design columns are identical, the
corresponding factors or interactions are aliased
and their corresponding effects cannot be
distinguished.
t
A desirable feature of a screening design is
orthogonality in which the vector products of
any two main effect or interaction columns sum
to zero. Orthogonality means that all estimates
can be obtained independently of one another.
t
DOE software provides efficient screening
designs whose columns are not aliased and from
which orthogonal estimates can be obtained.
t
Fractional factorial screening designs include
fewer trials and may be more efficient than the
corresponding full factorial design.
t
The concept of aliasing is one of the tools that
can be used to construct efficient, orthogonal,
screening designs.
t
Center points are often included in screening
designs to raise the efficiency and to provide a
measure of replication error and lack of model
fit.

t
The order of running and testing experimental
trials is often randomized to protect against the
presence of unknown lurking variables.
t
Blocking variables (such as day or run or session)
may be included in a design to raise the design
efficiency.
t
Factor effects in screening designs may be missed
because they were not included in the screening
experiment, because they were not given
sufficiently wide factor ranges, because the design
was underpowered for those factors, because trial
order was not properly randomized or blocked, or
because of an inadequate model.
INTRODUCTION
This article is the second in a series that deals with
the specialized types of screening designs (1). These
designs have been developed to most efficiently accept
many inputs that may or may not be relevant to the
final product and reduce this list to those few that
are most important. Once the results are confirmed,
the analyst proceeds to the response surface designs
to map the fine detail in the area of optimal
response (i.e., decide on the most desirable values
of the inputs to get the optimal output of whatever
is being manufactured or controlled). The three
most important targets usually sought are optimal
concentrations, variance reduction, and robustness.
THEORY
Most screening designs are class III designs, where
main effects are not aliased (confounded with each
other), but the main effects are aliased with the two-
way interactions. Factor levels are briefly discussed
in the following sections. At this point the reader may
wish to review the previous article in the series (1) to
re-examine the importance of randomization and
replication. Randomization ensures the independence
of the observations. Replication assesses variation and
more accurately obtains effect estimates.
Before the models (designs) are run, it may be
advantageous to decide on design level, blocking (if
any), and data transformation (if necessary). Lets
examine transformations first, as this is a common
problem.
Data Transformation
Transformations are usually employed to stabilize
the response variance, make the distribution of the
response variable more normal, or improve the fit
of the model to the data (2). Note that more than
one of these objectives may be simultaneously
achieved, and the transformation is many times done
with one of the power family (y*=y, where is the
John A. Wass
transforming parameter to be determined, e.g., if
=1/2, take the square root of the response variable).
The most useful has been found to be the Box-Cox
procedure that estimates and other model parameters
simultaneously by the method of maximum
likelihood. Modern software does this automatically.
If the analyst prefers to choose the value of , simple
values are preferred as, for example, the real-world
differences between 0.50 and 0.58 may be small but
the square root is much easier to interpret. Also if the
optimal value of is determined to be close to one, no
transformation may be necessary.
Blocking
It is often advantageous to minimize or eliminate
variability contributed by factors of little or no interest
even though they affect the outcome. These nuisance
factors may be reduced by a technique called blocking.
By grouping these nuisance variables and reducing
system variability, the precision of factor (of interest)
comparisons is increased. In the example of the
chemical reaction in our previous article, if several
batches of the substrate are required to run the design,
and there is batch-to-batch variability due to supplier
methodology, we may wish to block the substrate by
supplier, thus reducing the noise from this factor. We
tend to consider a block as a collection of homogeneous
conditions. In this case, we would expect the
difference between different batches to be greater
than those within a single batch (supplier). Please
note, if it is known or highly suspected, that within-
block variability is about the same as between-block
variability, paired analysis of means will be the same
regardless of which design may be used. The use of
blocking here would reduce the degrees of freedom and
lead to a wider confidence interval for the difference of
means.
Figure 1: Screening design and test data.
PROCESS VALIDATION Process Design 55
John A. Wass

Figure 2: Actual by predicted plot.

Parameter Estimates

Term Estimate Std Error t Ratio Prob>|t|

Intercept 2.00375 0.06 33.40 <.0001*
Reagent 1 0.45 0.063246 7.12 0.0001*
Reagent 2 0.3125 0.063246 4.94 0.0011*
Enzyme 0.1375 0.063246 2.17 0.0614
Temp 0.125 0.063246 1.98 0.0835
Mix Speed 0.05 0.063246 0.79 0.4520
Block[1] -0.00375 0.1 -0.04 0.9710
Block[2] -0.15375 0.107703 -1.43 0.1913
Block[3] 0.13625 0.1 1.36 0.2102
Reagent 1*Reagent 2 0.0375 0.063246 0.59 0.5696

Figure 3: Actual by predicted plot.

Parameter Estimates

Term Estimate Std Error t Ratio Prob>|t|

Intercept 2.0111111 0.059471 33.82 <.0001*
Reagent 1 0.45 0.063078 7.13 <.0001*
Reagent 2 0.3125 0.063078 4.95 0.0004*
Enzyme 0.1375 0.063078 2.18 0.0519
Temp 0.125 0.063078 1.98 0.0731
Mix Speed 0.05 0.063078 0.79 0.4447
Reagent 1*Reagent 2 0.0375 0.063078 0.59 0.5642

56 PROCESS VALIDATION Process Design

 John A. Wass

Figure 4: Actual by predicted plot.

Parameter Estimates
Term Estimate Std Error t Ratio Prob>|t|
Intercept 2.1291667 0.02533 84.06 <.0001*
Reagent 1 Biased 0.5 0.029806 16.78 <.0001*
Reagent 2 0.3125 0.021076 14.83 <.0001*
Enzyme 0.1375 0.021076 6.52 0.0003*
Temp[-1] -0.304167 0.030623 -9.93 <.0001*
Temp[0] 0.3583333 0.044432 8.06 <.0001*
Temp[-1]:Mix Speed[-1] Biased -0.1 0.042152 -2.37 0.0494*
Temp[1]:Mix Speed[-1] Zeroed 0 0 . .
Block[1] -0.0575 0.033984 -1.69 0.1345
Block[2] -0.1 0.036505 -2.74 0.0289*
Block[3] 0.0825 0.033984 2.43 0.0456*
Reagent 1*Reagent 2 0.0375 0.021076 1.78 0.1184

Factor Levels eliminated sources of variability and hopefully

The last preliminary item of importance is choosing
factor levels. There are an infinite number of values
for any continuous variable, and a restricted, but
usually large number for categorical variables. In
general, when the objective is to determine the small
number of factors that are important to the outcome
or characterize the process, it is advisable to keep factor
levels lowusually two works well. This is because we
are designing F*k runs (F=factor, k=levels) in a factorial
type experiment and as the levels of each factor rise,
the number of runs increases dramatically. The drama
intensifies further if interactions are included.
TECHNIQUESTHE DESIGNS
The following are three widely-used screening designs:
t
Randomized blocks and fractional factorial designs
t
Nested and split-plot designs
t
Plackett-Burman (P-B) designs.
Randomized Blocks and Fractional
Factorial Designs
As was stated, similar batches of relatively
homogeneous units of data may be grouped. This
grouping restricts complete randomization as the
treatments are only randomized within the block.
By blocking we loose degrees of freedom but have
gained a better understanding of the process. We
cannot always identify these nuisance factors. But by
randomization, we can guard against the effects of
these factors, as their effects are spread or diluted
across the entire experiment.
If we remember our chemical process experiment
from the previous article, we had two reagents with
two levels of an enzyme, temperature, and mix speeds.
We added a center point to check for curvature and
ran a single replicate for each point and blocked across
four days (see Figure 1). We put all main factors plus an
interaction into the model (see Figure 2).
Parameter estimates in Figure 2 told us that the
blocks were not significant, and when we rerun the
model without a block effect, we see the results in
Figure 3.
Although parameter estimates in Figure 3 show
little difference, the enzyme component is closer
to significance. Again this may represent a power
problem or design flaw (we needed a wider enzyme
range). In this example, we may not have needed
to block, but it is always wise to test if an effect is
suspected or anomalous results are encountered.
Fold-over. There is a specialized technique within
this group called fold-over. It is mentioned because
the analyst may find it useful for isolating effects of

PROCESS VALIDATION Process Design 57

John A. Wass
Figure 5: Residual by predicted plot.
interest. It is performed by switching certain signs in
the design systematically to isolate these effects. The
signs are changed (reversed) in certain factors of the
original design to isolate the one of interest in an anti-
aliasing strategy. The name derives from the fact that
it is a fold-over of the original design. The details are
beyond the scope of this introductory article but may
be found in standard references (2, 3).
Latin-square design. Yet another specialized
technique used with fractional factorial designs is the
Latin-square design. This design utilizes the blocking
technique on one or more factors to reduce variation
from nuisance factors. In this case the design is an n
x n where the number of rows equal the number of
columns. It has the desirable property of orthogonality
(independence of the factors, great for simplifying
the math and strengthening the conclusions).
Unfortunately, the row and column arrangement
represent restrictions on randomization, as each cell in
the square contains one of the n letters corresponding
to the treatments, and each letter can occur only once
in each row and column. The statistical model for this
type of design is an effects model and is completely
additive (i.e., there is no interaction between the rows,
columns, and treatments).
Saturated design. One last term that the novice
may bump up against is the concept of a saturated
design, unfortunately all too common in the industrial
world. This refers to a situation where the analyst is
attempting to include many variables in the model and
has few runs (translating ultimately to too few degrees
of freedom) to support the analysis. This allows for
estimation of main effects only. In some cases, all
interactions may be aliased with the main effects,
thus condemning the analyst to missing important
factors and interactions. If it is not possible to increase
the number of runs, it is a good idea to call in subject
matter experts (SMEs) (usually chemists or engineers)
to assist in eliminating variables.
58 PROCESS VALIDATION Process Design
Figure 6: Normal plot.
Nested and Split-Plot Designs
These designs are widely used in many industries.
They introduce the need for random factor designation
and the joys of variance component analysis. The
former refers to those factors that may be taken as
an adequate representative of a larger population.
For example, if two instruments are used in an
experiment to characterize performance because
only two were available, the results may not be
generalized to the population of 100 instruments
that were manufactured and the factor instrument
is not random and, therefore, we classify it as a fixed
effect. If, however, 20 instruments were available and
7 or 8 were chosen at random, the results are much
more likely to represent the population and the factor
may be considered random. As the minimal number
needed may be calculated from sampling theory and
power studies, a statistician may be consulted if there
are no industry recommendations.
Variance component analysis involves the
calculation of expected mean squares of error to
determine how much of the total system variance is
contributed by each term in the model (including the
error term).
Nested design. When levels of an effect B only
occur within a single level of an effect A, then B is said
to be nested within A (4). This may be contrasted
with crossed effects, which are interactions (i.e., the
results of one factor are dependent upon the level of
another factor). Nested designs are sometimes referred
to as hierarchical designs.
In our example of the chemical process, if we only
had several temperatures and mix speeds available,
we might wish to check the effects of using only
certain mix speeds with certain temperatures. This is
easily done by nesting mix speed within temperature,
designated as mix speed [temp]. When the model is
analyzed this way, we get the results in Figure 4.
The fit is better only because we now have
 John A. Wass

categorical variables, less to fit, and many other factors two-way interactions. The great advantage of these
are significant. We have, however, lost degrees of designs is the ability to evaluate many factors with
freedom by nesting terms, and this may negatively few runs. The disadvantages involve the assumptions
affect power. We would then use residual (error) made (i.e., that any interactions are not strong enough
analysis as our diagnostic tool, followed by standard to mask main effects and that any quadratic effects are
checks such as normal probability plots, outlier checks, closely related to factor linear effects). Although these
and plotting the residual versus fitted values (see assumptions usually hold, it is always best to try to
Figures 5 and 6). verify them with any available diagnostics.
Both of the diagnostics in Figures 5 and 6 exhibit Again, for our system the P-B design is seen in Figure
problems (i.e., increasing residuals with predicted 8. The analyses results are seen in Figure 9.
values and many off-axis values on the Normal Plot). It appears that P-B may not be a good design choice
These may be due to singularities during calculations as it requires more runs and is less sensitive to the
(e.g., terms going to infinity or division by zero). We present data structure than simpler designs.
may wish to increase the runs to see if the increasing
degrees of freedom will stabilize the values. SYNOPSIS: BUT WHICH DO I USE?
Split-plot design. In some experiments, due The following provides a pathway for application
to real-world complications, the run order may not and selection of appropriate screening designs. The
be amenable to randomization and we need to use
a generalization of the factorial design called the
split-plot design. The name refers to the historical Table: Fixed effect tests.
origins of the design in agriculture and posits splitting Source Nparm DF DFDen F Ratio Prob > F
some factor into sub-factors due to some problem Reagent1 1 1 2 243.0000 0.0041*
with running a full factorial design or data collection
Reagent2 1 1 2 14.0833 0.0087*
method (e.g., different batches on different days).
Therefore, we are running the experiment as a group Enzyme 1 1 3 0.0388 0.8564
of runs where within each group some factors (or only Temp 1 1 3 0.0580 0.8252
one) remain constant. This is done as it may be very Mix Speed 1 1 2 2.0833 0.2857
difficult or expensive to change these factors between
runs. In our example, we can declare the enzyme prep
and temperature as difficult to change. The software following questions are addressed:
then optimally designs the experiment around 5 t
Which screening design should be used when
plots in just 10 runs, far fewer than even a fractional resources are a consideration?
factorial design (see Figure 7). t
Which screening design should be used when
It declares only the plots as random so they take up flexibility is needed regarding variables to be
all of the variance. The plots are split by the enzyme tested?
and mix speed as these have been declared hard to t
What are the advantages of the respective designs
change and are the subplots. As we have only the one regarding special needs (e.g., reducing noise,
random effect, we test all others as fixed effects (see blocking, and other needs)?
Table).
The results are essentially the same as for the Figure 10 is a flow diagram may be used to answer
fractional factorial, as the design may contain similar these questions.
flaws.
SOFTWARE
Plackett-Burman (P-B) Designs There are numerous software products available to
P-B designs are specialized screening designs where assist the practitioner in design and analysis of their
the number of runs is not required to be powers of experiments. The author has had experience with the
two. If there are funds for extra runs whose number following commercial packages:
does not increase by a power of two, these designs are t
Design Expert (www.statease.com)
ideal as they also generate columns that are balanced t
GenStat (www.vsni.co.uk)
and pairwise orthogonal. They are based upon a very t
JMP (www.jmp.com)
flexible mathematical construct called a Hadamard t
Minitab (www.minitab.com)
matrix where the number of runs increases as a t
MODDE (www.umetrics.com)
multiple of four and thus will increase much more t
STATISTICA (www.statsoft.com)
slowly than the fractional factorial. Note that these t
SYSTAT (www.systat.com)
are Resolution III designs where the main effects are t
Unscrambler (www.camo.no).
not aliased with each other but are aliased with any

PROCESS VALIDATION Process Design 59

John A. Wass

Figure 7: Screening design and test data.

Figure 8: Plackett-Burman design.

60 PROCESS VALIDATION Process Design

 John A. Wass

Figure 9: Plackett-Burman design results.

PROCESS VALIDATION Process Design 61

John A. Wass

Figure 10: Screening design decision matrix.

REFERENCES
1. Wass, John A., Statistical Viewpoint: First
Resources (time/money/personnel) Are/Are Not A Concern Steps in Experimental DesignThe Screening
Experiment, Journal of Validation Technology,
Volume 16, Number 2, Spring 2010.
2. D. C. Montgomery, Design and Analysis of
ARE ARE NOT Experiments (5th ed.), John Wiley, 2001.
Fractional Factorial Full Factorial
Plackett -Burman 3. G.E.P. Box, J.S Hunter and, W.G. Hunter,
Need Flexibility in Run Numbers Statistics for Experimenters (2nd Ed.), Wiley
Interscience, 2005.
4. JMP Design of Experiments Guide, Release 7,
Yes No SAS Institute Inc., 2007.
Plackett -Burman Fractional Factorial

Special Needs
GENERAL REFERENCES
S.R. Schmidt and R.G. Launsby, Understanding
Industrial Designed Experiments (4th ed.), Air
Academy Press, 1997.
Reduce Extraneous Noise Isolate Aliased Effects Allow for embedded Cannot Randomize G.E.P. Box, J.S Hunter and, W.G. Hunter,
Blocking Foldover Levels the run order
Latin Squares Nested Design Split Plot Statistics for Experimenters (2nd Ed.), Wiley
Interscience, 2005.
D. C. Montgomery, Design and Analysis of
CONCLUSIONS Experiments (5th ed.), John Wiley, 2001.
Modern experimental design is sometimes art as well JMP Design of Experiments Guide, Release 7, SAS Institute Inc.,
as science. It is the objective of this column to acquaint 2007.
the reader with the rudiments of the screening design, ECHIP, Reference Manual, Version 6, ECHIP Inc., 1983-1993.
introduce them to the nomenclature, and supplement Deming, S.N., Quality by Design (Part 5), Chemtech, pp 118-
the learning experience with a real-world example. 126, Feb. 1990.
Deming, S.N., Quality by Design (Part 6), Chemtech, pp 604
607, Oct. 1992.
Deming, S.N., Quality by Design (Part 7), Chemtech, pp 666
673, Nov. 1992. JVT

ARTICLE ACRONYM LISTING

DOE Design of Experiments
P-B Plackett-Burman
SME Subject Matter Experts

Originally published in the Winter 2011 issue of The Journal of Validation Technology

62 PROCESS VALIDATION Process Design


A Further Step in
Experimental Design (III):
The Response Surface
John A. Wass
Statistical Viewpoint addresses principles of
statistics useful to practitioners in compliance and
validation. We intend to present these concepts in
a meaningful way so as to enable their application
in daily work situations.
Reader comments, questions, and suggestions
are needed to help us fulfill our objective for this
column. Please send any comments to managing
editor Susan Haigney at shaigney@advanstar.com.
KEY POINTS
The following key points are discussed:
t
Design of experiments (DOE) consists of three
basic stages: screening (to identify important
factors), response surface methodology (to
define the optimal space), and model valida-
tion (to confirm predictions).
t
A critical preliminary step in the screening
stage is for subject matter experts to identify
the key list of factors that might influence the
process.
t
A DOE design consists of a table whose rows
represent experimental trials and whose col-
umns (vectors) give the corresponding factor
levels. In a DOE analysis, the factor level col-
umns are used to estimate the corresponding
factor main effects.
t
Interaction columns in a design are formed as
the dot product of two other columns. In a
DOE analysis, the interaction columns are used
to estimate the corresponding interaction effects.
For more Author
information,
go to
gxpandjvt.com/bios [ ABOUT THE AUTHOR
John A. Wass is a consulting statistician with Quantum Cats Consulting in the Chicago area, as
well as a contributing editor at Scientific Computing, and administrator of the Great Lakes JMP Users
Group. He served for a number of years as a senior statistician at Abbott Laboratories (in both phar-
maceuticals and diagnostics). John may be contacted by e-mail at john.wass@tds.net.
John A. Wass
t
When two design columns are identical, the
corresponding factors or interactions are
aliased and their corresponding effects cannot
be distinguished.
t
The order of running and testing experimental
trials is often randomized to protect against
the presence of unknown lurking variables.
t
Blocking variables (e.g., day or run or session)
may be included in a design to raise the design
efficiency.
t
Factor effects may be missed because they
were not included in the original screening
experiment, because they were not given
sufficiently wide factor ranges, because the
design was underpowered for those factors,
because trial order was not properly random-
ized or blocked, or because of an inadequate
model.
t
Unusual interactions and higher-order effects
occassionaly may be needed to account
for curvature and work around regions of
singularity.
t
Where there are inequality constraints (e.g.,
areas where standard settings will not work),
special designs are needed.
t
The designs may become rather challenging
and a statistician becomes an invaluable part
of the team when considering problems of
non-normal responses, unbalanced data, spe-
cialized covariance structures, and unusual or
unexpected physical or chemical effects.
PROCESS VALIDATION Process Design 63
John A. Wass

Figure 1: Surface point. Figure 3: Central composite design.

2 10 2
Silica 20 Sila
ne
1 30

0 40
220 50
200 220 
180 200
160 180
on

140 3
Abrasi

160

Abrasi
120 140
100
120 -

on
80
60
100 
40 80
10
20
60
40 - 
2
Si
la 30
ne
ca
1
40 1 Sili

50 0
-
Figure 2: Box-Behnken design.

Figure 4: CCD data.

(+1) Reagent Reagent Mix
Pattern 1 2 Enzyme Temp Speed Y

(+1) 1 a0000 -2 0 0 0 0 1.2

2 -+-+- -1 1 -1 1 -1 0.9
3 0A000 0 2 0 0 0 1.35
(-1) 4 0000a 0 0 0 0 -2 1.4
5 A0000 2 0 0 0 0 1.85
(+1) 6 -++++ -1 1 1 1 1 2.6
7 +--+- 1 -1 -1 1 -1 1.5
8 ---++ -1 -1 -1 1 1 1.2
(-1) 9 ++--- 1 1 -1 -1 -1 1.3

(-1) 10 ++++- 1 1 1 1 -1 4
11 000A0 0 0 0 2 0 2.6
12 00a00 0 0 -2 0 0 0.4
13 +-+-- 1 -1 1 -1 -1 2
INTRODUCTION 14 00A00 0 0 2 0 0 1.8
Response surface methodology (RSM) is the devel- 15 000a0 0 0 0 -2 0 0.6
opment of the specific types of special designs to 16 -++-- -1 1 1 -1 -1 1.1
most efficiently accept a small number of inputs
17 -+--+ -1 1 -1 -1 1 0.9
(relative to screening designs) that are known to be
18 0000A 0 0 0 0 2 0.8
relevant to the final product and optimize a pro-
cess result to a desired target (1). Once the results 19 +---+ 1 -1 -1 -1 1 1.5
are confirmed, the analysts load becomes lighter 20 00000 0 0 0 0 0 2.2
(excepting in the case of non-reproducibility or 21 ++-++ 1 1 -1 1 1 2.5
results drifting out of specification). In effect, the 22 --++- -1 -1 1 1 -1 2.3
response surface maps the fine detail in the area of 23 0a000 0 -2 0 0 0 1.5
optimal response (i.e., determines the most desir- 24 +-+++ 1 -1 1 1 1 2.7
able values of the inputs to get the optimal output 25 +++-+ 1 1 1 -1 1 3.3
of whatever is being manufactured, controlled, or 26 00000 0 0 0 0 0 3
studied). The three most important targets usually 27 ----- -1 -1 -1 -1 -1 0.9
sought are optimal concentrations, variance reduc-
28 --+-+ -1 -1 1 -1 1 2
tion, and robustness (2). The adequacy of the model

64 PROCESS VALIDATION Process Design

 John A. Wass

is most often checked by residual analysis, influ- Figure 5: Actual by predicted plot.
ence diagnostics, and lack-of-fit testing (3). JMP 9
4.5
is utilized herein for the design and analysis of an
industrial example (4). 4.0
3.5
THEORY
Many response surface designs are collections of 3.0

Y Actual
specialized statistical and mathematical techniques
2.5
that have been well implemented in software using
efficient algorithms (4, 5). In many real world cases 2.0
the output includes more than one response and
1.5
these need not be continuous functions. Lets exam-
ine the case of a chemical engineer who wishes 1.0
to maximize an important property (y) based on 0.5
given levels of two chemical inputs, (x1) and (x 2).
The desired property is now a function of the two 0.0
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
chemical entities plus error (), as follows:

y = f(x1, x 2) + Y Predicted P=0.0275

RSq=0.63 RMSE=0.6706
The surface is represented by the following:
Summary of Fit
y = f(x1, x 2) RSquare 0.626836
RSquare Adj 0.407328
The response surface is usually displayed graphi-
Root Mean Square Error 0.670639
cally as a smoothly curving surface, a practice that
may obscure the magnitude of local extremes (see Mean of Response 1.764286
Figure 1). Observations (or Sum Wgts) 28
In many problems using RSM, the experimenter
does not know the exact mathematical form of the Analysis of Variance
relationship between the input and output vari- Sum of Mean
ables and, therefore, must find a workable approxi- Source DF Squares Square F Ratio
mation. The first guesstimate is a low order poly- Model 10 12.843426 1.28434 2.8556
nomial (e.g., first order model), as follows: Error 17 7.645859 0.44976 Prob > F
C. Total 27 20.489286 0.0275*
y = 0 + 1x1 + 2 x 2 + + n x n +

Obviously, this is a linear model that will not accom- Lack of Fit
modate curvature. If curvature is suspected, a higher Sum of Mean F Ratio
order polynomial may be tried. The following is a sec- Source DF Squares Square 0.3815
ond order model: Lack of Fit 14 4.8958594 0.349704 Prob > F
Pure Error 3 2.7500000 0.916667 0.9083
y = 0 + 4Ai x i + 4Aii x i2 + 44Aij x i x j +
Total Error 17 7.6458594 Max RSq
where we sum over i, and i<j. The above two models 0.8658
are found to work well in a variety of situations. How-
ever, although these may be reasonable approxima- Another caveat is the importance of using the
tions over the entire design space, they will not be an software tools with an incomplete understanding of
exact fit in all regions. The goal is to find a smaller the problem. It is sometimes tempting to overlook
region of interest where they fit well. A proper experi- expert advice from the chemists or engineers and try
mental design will result in the best estimate of the to force-fit a simple solution. As in the real world, we
model parameters. The response surface is usually fit- often find extremely messy data. It might be neces-
ted by a least squares procedure, and optimal values sary to use a cubic factor, a three-way interaction, or
are pursued through sequential algorithms such as the a highly-customized design. At this point the reader
method of steepest ascent. The details of these algo- may wish to review the previous article in the series
rithms are beyond the scope of this series. (Journal of Validation Technology, Winter 2011) to re-

PROCESS VALIDATION Process Design 65

John A. Wass

Figure 6: Residual by predicted plot. Factorial Designs

These are one of the most common designs used in
situations where there are several factors and the
1.0 experimenter wishes to examine the interaction
effects of these on the response variable(s). The
0.5 class of 2k factorial designs is often used in RSM and
Y Residual

finds wide application in the following three areas

0.0 (6):
t
As a screening design prior to the actual
-0.5
response surface experiments
t
To fit a first-order response surface model
-1.0
t
As a building block to create other response sur-
face designs.
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
Y Predicted For more details on these designs, the reader is
again referred to the article on screening designs in
Figure 7: Sorted parameter estimates. this series.
Term Estimate Std Error t Ratio Prob> |t|
Enzyme 0.5041667 0.136894 3.68 0.0018* D-Optimal Designs
Temp 0.3625 0.136894 2.65 0.0169*
Reagent1 0.3416667 0.136894 2.50 0.0231*
This is a class of designs based on the determinant
Enzyme*Enzyme -0.108594 0.132547 -0.82 0.4240 of a matrix, which has an important relationship to
Reagent2 0.0916667 0.136894 0.67 0.5121 the moment matrix (M):
Enzyme*Temp 0.10625 0.16766 0.63 0.5347
Reagent1*Enzyme 0.06875 0.16766 0.41 0.6869 M = XX/N,
Reagent2*Reagent2 -0.027344 0.132547 -0.21 0.8390
Reagent1*Temp 0.03125 0.16766 0.19 0.8543
Reagent1*Reagent1 -0.002344 0.132547 -0.02 0.9861
where X is the inverse on X, the design matrix of
inputs and N is the number of rows in X.
It is noted that the inverse of the moment matrix,
Figure 8: Normal plot. M-1 = N(XX)-1 contains variances and covariances
4 of the regression coefficients scaled by N divided
+ Enzyme
by the variance. Therefore, control of the moment
3 matrix implies control of the variances and covari-
+ Temp
Normalized Estimates (Orthog t)

+ Reagent1
2 ances and here is the value of these D-optimal
designs (5).
1
+
++ Box-Behnken Designs
0 + ++
These are three-level designs for fitting response
-1 + surfaces. They are made by combining 2k factorials
designs with incomplete block designs. They pos-
-2 sess the advantages of run efficiency (i.e., number
of required runs and rotateability). This last prop-
-3
erty ensures equivalence of variance for all points
-4 in the design space equidistant from the center.
-3 -2 -1 0 1 2 3
The design is constructed by placing each factor,
Normal Quantile
or independent variable, at one of three equally
Blue line is Lenths PSE, from the estimates population. spaced values from the center. By design then, the
Red line is RMSE, Root Mean Squared Error from the residual. estimated variances will depend upon the distance
of the points from the center (Figure 2).
examine the use of data transformations, blocking, It is noticed that the Box-Behnken design has no
and factor levels. points at the vertices or corners of the design space.
This is a plus when it is desired to avoid those areas
TECHNIQUESTHE DESIGNS because of engineering constraints but unfortu-
There are a variety of specialized designs and ana- nately places a higher prediction uncertainty in
lytic techniques available for RSM (6). In this review, those areas.
we will concentrate on four of the more popular (i.e.,
commonly used). One will be chosen to analyze Central Composite Designs
standard but messy industrial data. Central composite design (CCD) is a popular and

66 PROCESS VALIDATION Process Design

 John A. Wass

Figure 9: Prediction profiler.

4
0.858115
3.178847
3
Y

2
1
0
1
Desirability
0.98832

0.5
0
-1

1
-1

1
-1

1
-1

1
0
0.25
0.5
0.75
1
0.978 0.022 1 1
Reagent1 Reagent2 Enzyme Temp Desirability

Figure 10: Surface plot. Figure 11: Surface curvature.

2 1.5 En
1.5 1 0.5 zyme 2 1.5 1 0.5 0 -0.5 -1 -1.5 -2
t1 1 0
-0.5 -1 3 3
en 0.5
ag -1.5
Re 0 -2
-0.5 3
-1
-1.5 2.5 2.5
-2
2.5
3

2.5 2 2 2
Y

2
Y

Y
1.5

1.5 1.5
Y

1.5 1

1 0
2
0 1.5 1 1
2 1
1.5
1 0.5
0.5 0 t1
En n
zy 0 -0.5 age 0.5 0.5
m -0.5
e Re 2 1.5 1 0.5 0 -0.5 -1 -1.5 -2
-1 -1
-1.5 -1.5
-2
have changed due to knowledge gathered from the
efficient design for fitting a second-order model first series of experiments. We now wish to use this
(Figure 3). The CCD most often consists of a 2k new knowledge to optimize the process. The exper-
factorial n runs, 2k axial runs, and nc center runs. iment is designed in JMP 9 software, then run, and
The experimenter must specify the distance from the data appear as seen in Figure 4.
the axial runs to the design center (this is done by The design includes 28 runs with a center point
many software platforms) and the number of cen- as well as face centers on the design hypercube
ter points. As it is desirable to have a reasonably (there are more than three dimensions). When
stable variance at the points of interest, the design these data are analyzed, it is apparent that, while
should have the property of rotateability. the model is significant (analysis of variance
[ANOVA] p = 0.0275) and there is no significant
CCDAn Example of the lack of fit (p= 0.9803), the adjusted R 2 is low
Response Surface Technique (~0.41) and more terms might need to be added to
For an example, the use of the central composite the model (Figure 5).
design is employed in yet another industrial chemi- There is some confidence in the integrity of the
cal process. The chemists are now concerned with a data as the plot of errors (by increasing Y value)
formulation that is only slightly different than that shows randomness rather than any defining pat-
used in the screening design, and concentrations terns that would indicate a problem (Figure 6).

PROCESS VALIDATION Process Design 67

John A. Wass

Figure 12: Data table output. Figure 13: Three reagents and a Reagent1*
Reagent1 Reagent2 Enzyme Y Enzyme interaction.
1 L2 L3 L1 2.1
2 L1 L2 L2 1
3 L1 L2 L1 0.8 3
4 L3 L3 L1 2
5 L1 L1 L3 1.7 2.5
6 L2 L1 L1 1.6

Y Actual
7 L2 L3 L3 2.4
8 L3 L2 L2 2.4 2
9 L3 L1 L2 2.2
10 L1 L2 L3 1.2 1.5
11 L3 L3 L2 2.5
12 L3 L2 L1 1.9
1
13 L2 L3 L2 2.2
14 L3 L1 L3 2.8 1 1.5 2 2.5 3
15 L3 L3 L3 3.1
Y Predicted P=0.0143
RSq=0.97 RMSE=0.217
It appears that from the sorted parameter esti-
mates, what the chemists had long suspected was
true (i.e., the importance of Reagent 1, the enzyme, Summary of Fit
and the reaction temperature to the output). See RSquare 0.967811
Figure 7. This is also seen on the normal plot as RSquare Adj 0.887338
these three parameters are flagged (Figure 8).
Root Mean Square Error 0.21696
The interaction profiles evidence of a (expected)
putative interaction between the enzyme and reac- Mean of Response 1.993333
tion temperature, but this did not appear signifi- Observations (or Sum Wgts) 15
cant from the parameter estimates.
Use of such tools as prediction profilers will Analysis of Variance
allow the optimization of the process, which in Source DF Sum of Mean F Ratio
this case, seen by the position of the dotted red Squares Square 12.0265
lines, is accomplished by maximizing the values Model 10 5.6610476 0.566105
of Reagent 1, enzyme, and temperature. Reagent Error 4 0.1882857 0.047071 Prob > F
2 does not seem important in this respect (flat
C. Total 14 5.8493333 0.0143*
slope) (Figure 9).
The design space is easily visualized on the sur-
face plot, which corroborates the results of the pro- trations for the two reagents and the enzyme. In
filer (Figure 10). most modern software with experimental design
When the figure is rotated, the surface curvature modules, it is possible to roll your own custom
is seen. It is apparent by the absence of lack of fit design to non-standard specifications. Suppose the
that the inclusion of interaction and quadratic new data set consists of the two reagents and the
terms in the model were justified (Figure 11). enzyme, with the reaction temperature and mix
Although there is good evidence that the model is speed held at the pre-determined optima. The soft-
adequate, the low R 2 suggests that input values may ware designs a 15-run experiment, and the analyst
need to change, or more likely, factors may need to fills in the data table output (Figure 12).
be added. There are many ways to model this sys- The model is then analyzed with the three reagents
tem. A custom design may simplify the system. and a Reagent1*Enzyme interaction as seen in Figure
13.
Custom Designing The adjusted R 2 is now ~ 0.89 evidencing a good
The chemists are comfortable with setting the tem- fit, and the ANOVA indicates a significant model
perature at a level that was determined as optimal (p=0.014). In linear regression with standard least
in both theoretic calculations and previous batch squares fitting model, significance is a test of zero
experience. They now wish to zero in on concen- slope (i.e., a steep slope indicating good predictive

68 PROCESS VALIDATION Process Design

 John A. Wass

Figure 14: Effects test.

Source Nparm DF Sum of Squares F Ratio Prob >F
Reagent1 2 2 2.5261587 26.8332 0.0048*
Reagent2 2 2 0.1583810 1.6823 0.2950
Enzyme 2 2 0.9055974 9.6194 0.0296*
Reagent1*Enzyme 4 4 0.1200593 0.6376 0.6632

Figure 15: Residual by predicted plot. Figure 17: Maximal output.

me Rea
Enzy L3 L3 gen
t1
L2
L2
L1
0.2 L1
3

3
0.1 2.5
Y Residual

2.5
2
0
Y
2
1.5

Y
-0.1 1 1.5

0 1
-0.2
0
L3

1 1.5 2 2.5 3
t1

L3
en g

e
ea

L2 m
Y Predicted zy
R

L2 En

Figure 16: Prediction profiler.

3
0.482975

2.5
3.078571

2
Y

1.5
1
0.75 1
Desirability
0.607334

0 0.25

L1

L2

L3

L1

L2

L3

L1

L2

L3

0.25

0.5

0.75

L3 L3 L3
Reagent1 Reagent2 Enzyme Desirability

ability). The effects test confirms that Reagent1
and the enzyme are most important to the reaction
(Figure 14). Chemistry may dictate that Reagent2 is
also important, but any of the concentrations used
in the experiment will work. The exact figure will
be based on convenience and cost. Although there
is no overall significance to the Reagent1*enzyme
interaction, the enzyme is necessary. Opening up
the range of concentrations used would demon-
strate this.

PROCESS VALIDATION Process Design 69

John A. Wass
Again, a quick check of data integrity is evi-
denced by the random pattern of the residual by
predicted plot (Figure 15).
With this model, it is possible to produce an out-
put that overlaps that of the larger, CCD model. And
this was accomplished in just 15 runs (Figure 16).
The response surface here is a stair-step because
of the linear nature of this model with non-contin-
uous inputs. It again illustrates that maximal out-
put is achieved through maximal concentrations
of Reagent1 and the enzyme for the ranges used in
this experiment (Figure 17).
SOFTWARE
There are numerous software products available
to assist the practitioner in design and analysis of
their experiments. The author has had experience
with the following commercial packages:
t
Design Expert (www.statease.com)
t
GenStat (www.vsni.co.uk)
t
JMP (www.jmp.com)
t
Minitab (www.minitab.com)
t
MODDE (www.umetrics.com)
t
STATISTICA (www.statsoft.com)
Originally published in the Autumn 2011 issue of The Journal of Validation Technology
70 PROCESS VALIDATION Process Design
t
SYSTAT (www.systat.com)
t
Unscrambler (www.camo.no).
CONCLUSIONS
Modern experimental design is sometimes art as
well as science. It is the objective of this column
to acquaint the reader with the rudiments of the
screening and response surface designs, introduce
them to the nomenclature, and supplement the
learning experience with real-world examples.
REFERENCES
1. S.R. Schmidt and R.G. Launsby, Understanding Industrial
Designed Experiments (4th ed.), Air Academy Press, 1997.
2. G.E.P. Box, J.S Hunter and, W.G. Hunter, Statistics for
Experimenters (2nd Ed.), Wiley Interscience, 2005.
3. D. C. Montgomery, Design and Analysis of Experiments
(5th ed.), John Wiley,2001.
4. JMP 9 Design of Experiments Guide, SAS Institute Inc.,
2010.
5. ECHIP Reference Manual, Version6, ECHIP Inc., 1983-
1993.
6. R.H. Myers and D.C. Montgomer y, Response Surface
Methodology, John Wiley, 1995. JVT

Estimation: Knowledge
Building with Probability
Distributions
David LeBlond
This column presents statistical principles useful
to practitioners in compliance and validation. We
hope it will become a useful resource. Barriers
to the understanding and use of statistical
principles include the complex mathematics,
abstract ideas, relationship to real applications,
required assumptions, unfamiliar terminology, and
Greek symbols. This column presents statistical
principles with these barriers in mind. It is our
challenge to explain statistical concepts in a
meaningful way so that our readers understand
these important topics and are able to apply
statistical concepts to work situations.
The first issue of Statistical Viewpoint (1)
presented basic probability distribution concepts
and Microsoft Excel tools useful in statistical
calculations. The second (2) reinforced these
concepts and tools through eight scientific decision-
making examples. In this issue, the scientific
knowledge-building process will be illustrated from
a statistical estimation point of view.
Reader comments, questions, and suggestions
are needed to help us fulfill our objective for
this column. Suggestions for future discussion
topics or questions to be addressed are requested.
Discussion topics and case studies demonstrating
applications of statistics to problems in validation
and compliance submitted by readers are also most
welcome. We need your help to make Statistical
Viewpoint a useful resource. Please email your
comments and suggestions to the coordinating
editor at shaigney@advanstar.com.
For more Author
information,
go to ivthome.com
[ ABOUT THE AUTHOR
David LeBlond, Ph.D., has 29 years experience in the pharmaceutical and medical diagnostics
fields. He is currently a principal research statistician supporting analytical and pharmaceutical
development at Abbott. David can be reached at David.LeBlond@abbott.com.
David LeBlond
KEY POINTS DISCUSSED
Good decision-making is a key element in the
assurance of pharmaceutical product safety
and efficacy throughout the product life cycle.
Good decisions require process knowledge.
Process knowledge consists of a statement of a
predictive model for the process plus estimates
of the underlying model parameters. Three types
of estimates are identified: point, interval, and
distributional estimates. Distributional estimates
are most desirable. Process modeling is an
iterative process that requires combining prior
knowledge with new data. Prior knowledge is
subjective, and knowledge experts can disagree.
Modern statistics provides good tools for
managing the modeling activity. This issue of
Statistical Viewpoint presents three knowledge-
building illustrations that highlight the central
role probability distributions can play. They
include estimation of a tablet defect proportion,
unit dose content uniformity, and adverse
event rate monitoring. The illustrations utilize
probability distribution functions available in
Excel, which is readily available and provides
an easy method for calculations. A list of Excel
distribution functions and their corresponding
applications is provided in this paper and in
previous issues of Statistical Viewpoint (1,
2). These tools enable quantifiable estimation
and predictions for common pharmaceutical
problems.
PROCESS VALIDATION Process Design 71
David LeBlond
Figure 1: Expressing process knowledge through a prob-
ability distribution.
Less knowledge
Probability Density
More knowledge
Model parameter or prediction value
KNOWLEDGE BUILDING IS AN
INTEGRAL PART OF PHARMACEUTICAL
DEVELOPMENT
Good decision-making is increasingly recognized
as an essential component in the validation of
pharmaceutical processes during the product life
cycle. ICH Q8 (3) states the following about the
output of pharmaceutical development studies
(italicized emphasis added):
... the applicant should demonstrate
an enhanced knowledge of product
performance ... This understanding can
be gained by application of, for example,
formal experimental designs, process
analytical technology (PAT), and/or prior
knowledge ... It should be recognized that
the level of knowledge gained, and not
the volume of data, provides the basis
for science-based submissions and their
regulatory evaluation.
What is meant by the terms knowledge and
prior knowledge? How can knowledge (prior or
otherwise) be objectively quantified? How do we
track the level of knowledge gained? What role
does data play in the knowledge-building process?
These questions are addressed in the following
paragraphs.
In the previous installments of Statistical
Viewpoint (1,2), it was suggested that a useful
knowledge metric is the certainty by which we
are able to predict the values of random variables
that serve as key indicators of safety and efficacy.
For instance, it is desirable to be able to predict
the major or minor defect proportions in dosage
units, the levels of active or related substances they
contain, and the rate of adverse events patients will
experience using the medication. A related metric
72 PROCESS VALIDATION Process Design
for our state of knowledge about a process is our
degree of certainty about the values of underlying
process parameters. As an illustration, consider the
following simple process model:
Y = intercept + slope*X + error
Here Y is some process outcome or attribute (e.g.,
process yield) that is linearly related to a variable X
(e.g., processing temperature). X is often referred to
as a process set point or control variable. Sometimes
X is referred to as a process parameter; however,
here I would like to avoid using the term parameter
for X.
The error term above might represent
measurement error. As part of our model we might
decide that each measurement of Y will include
a small error and that the error portion in all our
measurements will be random and have a Gaussian
(normal) distribution (1,2) centered on zero with a
standard deviation of sigma.
In statistical modeling, the intercept, slope, and
sigma are referred to as model parameters. I would
like to reserve the word parameters for these
quantities whose values we are interested in knowing.
If we are willing to think of these quantities and
parameters as random variables having probability
distributions then we can take our degree of
certainty about them as inversely proportional to
some measure of spread of their distributions. As
our estimates become increasingly reliable, the
width or shape of probability distributions can be
used to track the knowledge-building progress.
The idea of using probability distributions
through Bayes rule to quantify knowledge is at least
250 years old (4). This idea is illustrated in Figure
1. Less process knowledge is characterized by large
uncertainty in the values of underlying parameters
and predicted values of process performance
measures. When we have more knowledge about
a process, we have more certainty about such
random variables. When the underlying process
mechanism is well characterized and predictable,
process control can be assured. Documenting this
understanding and communicating it effectively
can be a supportive component of validation.
Probability and statistics can help us express our
knowledge-building progress in a quantitative way.
MODELS ARE REQUIRED FOR
KNOWLEDGE BUILDING
Pharmaceutical, product life-cycle decision-making
is guided by models. A model is our (usually naive)
story of how a process behaves. The following
are some examples of pharmaceutical development
models:

t

"
TUBCJMJUZ
NPEFM
UIBU
BTTVNFT
B
QTFVEP[FSP

order chemical reaction
t

"
EPTBHF
VOJGPSNJUZ
NPEFM
UIBU
BTTVNFT
B

normal distribution of levels per dosage unit
t

"
SFTQPOTF
QSFEJDUJPO
FRVBUJPO
UIBU
BTTVNFT

linear and interaction terms
t

"
UFTU
NFUIPE
DBMJCSBUJPO
DVSWF
UIBU
BTTVNFT

a proportional relationship between response
and concentration
t

"
UPYJDJUZ
NPEFM
UIBU
BTTVNFT
B
MPHJTUJD

relationship to dose level
t

"
MPU
BDDFQUBODF
TBNQMJOH
QMBO
UIBU
BTTVNFT

every unit in the lot has an equal probability of
appearing in the sample
t

"
QIBSNBDPLJOFUJD
QFSGVTJPO
NPEFM
UIBU

assumes Michaelis-Menten kinetic elimination
from the liver
t

"
WBMJEBUJPO
QMBO
UIBU
BTTVNFT
GVUVSF
MPUT

and validation lots are derived from the same
population
t

"
DMJOJDBM
USJBM
UIBU
BTTVNFT
UIBU
UIF
MBTU

measurement made on patients who leave the
trial early is a reliable indicator of subsequent
measurements that would have been obtained,
had they remained in the trial
t

"
QIBSNBDPWJHJMBODF
TJHOBM
EFUFDUJPO
TUSBUFHZ

assumes that adverse events follow a Poisson
distribution in exposed populations.
Many of the above models are statistical in
nature and can be expressed in whole or in part
using probability distributions. Many common
probability distributions were described in previous
issues of this series (1, 2). Each distribution
includes a random variable whose values
follow the distribution model, and two or three
additional parameters that specify the location,
spread, or shape of the distribution on the scale
of measurement of the random variable. The
distribution encapsulates our knowledge about the
value of the random variable. Of course models can
also include mechanistic or empirical variables and
can be very complex. Here we limit discussion to
some simple probability distribution models.
If a model is accurate and the parameter values
are known, it can be used to estimate or predict
values of the random variable. Our ability to predict
reliably is one indicator of our knowledge about
the process that generates values of the random
variable. The reliability of our predictions will in
turn depend on our uncertainty in the underlying
model parameters. As model parameter uncertainty
is diminished by scientific investigation and data
gathering, the quality of our predictionsand thus
our process knowledgeincreases. Thus we are
concerned with random variables that are either
process outputs (e.g., whether a tablet defect or
David LeBlond
Figure 2: Process knowledge building.
3. Acquire new data
1. Summarize 2. State process
prior knowledge model
7. Compare 4. Combine prior
confirmatory data knowledge with
with prediction new data to obtain
posterior knowledge
5. Use posterior
6. Acquire knowledge to make
confirmatory predictions of
data future samples
adverse event occurs or the specific drug content
in a tablet) or underlying model parameters (e.g.,
defect proportion, adverse event rate, or the mean
and standard deviation of unit dose levels).
THE SCIENTIFIC KNOWLEDGE-
BUILDING PROCESS IS ITERATIVE
Figure 2 presents a central theme played out often
in scientific endeavors, including those conducted
to manage life cycles of medical therapies and
diagnostics. The seven basic steps are given as
follows:
t

4UFQ

4VNNBSJ[F
QSJPS
LOPXMFEHF
3FDPSE

existing prior process knowledge about
the process under study. This knowledge
comes from previous data, accepted theory,
experience, or expert opinion. Often this
prior knowledge is highly unstructured and
subjective.
t

4UFQ

4UBUF
UIF
QSPDFTT
NPEFM
CBTFE
PO

prior knowledge. This requires a quantitative
description of the mathematical form of the
process model and its underlying parameter
values. At this stage, the model and values of
its underlying parameters are uncertain. We
can characterize our state of knowledge about
underlying model parameters or process output
by the Less Knowledge distribution in Figure 1.
t

4UFQ

"DRVJSF
OFX
EBUB
5P
JNQSPWF
NPEFM

certainty, specific data are gathered to test the
model or obtain improved estimates of the
underlying model parameters. This kind of
study can be thought of as analytic (5) in that
the objective is to improve our estimates of
underlying model parameters.
t

4UFQ

$PNCJOF
QSJPS
LOPXMFEHF
XJUI
OFX

data to obtain posterior knowledge. Both prior
information and new data are used to obtain
improved estimates of underlying model
PROCESS VALIDATION Process Design 73
David LeBlond
Figure 3: Types of estimates.
Point Estimate
Interval Estimate
Distributional Estimate
Parameter Value
parameters. These improved estimates increase
the certainty about the process mechanism and
increase the reliability of our predictions of
process behavior. In quality by design (QbD)
applications (3), we may be concerned with
defining the knowledge or inference space
within which the model applies. At this stage,
our knowledge may be characterized by the
More Knowledge distribution in Figure 1.
t

4UFQ

6TF
QPTUFSJPS
LOPXMFEHF
UP
NBLF

predictions of future samples. Proper scientific
investigation requires that we test our model
against reality. We do this by using the model
to make predictions of process output. It
is important to select conditions that truly
challenge the model.
t

4UFQ


"DRVJSF
DPOSNBUPSZ
EBUB
8F
WFSJGZ

these model predictions by actually running the
process and measuring the output. Examples of
such confirmatory or enumerative (5) studies
are robustness demonstrations, method/
process validations and transfers, comparability
studies, and clinical studies.
t

4UFQ

$PNQBSF
DPOSNBUPSZ
EBUB
XJUI

prediction. The level of agreement is a measure
of our process knowledge. When predicted and
observed output fail to agree, re-evaluate the
original prior assumptions (Step 1), modify the
process model appropriately, and re-investigate it.
The cycle of knowledge building continues
until we have acquired a successful track record
of predicting process output under a range
of conditions. We strive to document and
communicate a comfort level in our ability to
control processes within their respective design
spaces so that the final product will be safe and
effective. While documentation of the confirmatory
studies (steps 6 and 7) is critical, a careful
description of the underlying prior assumptions
(step 1), model logic (step 2), development
experience (steps 3 and 4), and level of model
challenge (step 5) can do much to convey the state
of process knowledge to regulators.
74 PROCESS VALIDATION Process Design
There is no reason for process knowledge
building to stop when the product is approved for
use. Knowledge building can continue throughout
the product life cycle and real world manufacturing
experience may provide the greatest challenge to
our predictive models.
ESTIMATION IS CENTRAL TO THE
KNOWLEDGE-BUILDING PROCESS
In Step 4 of Figure 2, data is used to form improved
estimates of process parameters. A wide variety
of statistical methods such as regression can
be employed. Whatever methodology is used,
the following three general types of estimates
illustrated in Figure 3 can be identified:
t

1PJOU
FTUJNBUF
"
TJOHMF
OVNCFS
UIBU
HJWFT
BO

indication of the location of the parameter
value or process output (X) along its scale
of measurement. Three examples of point
estimates are as follows:
E(X): The expected (or mean) value of X. The
mean is the most common point estimate
but its value can be greatly influenced by the
skew ness (long tailing) of the distribution.
mode(X): The most frequent value of X.
Identified as the maximum of a continuous
probability density function.
median(X): The middle value of X. Identified
as the 50th quantile of a continuous
probability distribution function. The
median is influenced less by skewness.
A point estimate alone is a single number
and conveys little or nothing about the
uncertainty in X. For example, a point
estimate for the potency of a tablet lot might
be 99.8%.
t

*OUFSWBM
FTUJNBUF
5XP
OVNCFST
UIBU
TQFDJGZ

a range (or interval) that gives an indication
of both location and spread of the parameter
value. Two common types of interval estimates
are as follows:


t


DPOEFODF
JOUFSWBM
"O
JOUFSWBM

produced by a given statistical method that
is known to include the true parameter
value 95% of the time. A confidence
interval requires that we consider unknown
parameters to be fixed quantities that
cannot possess a probability distribution.
The interval itself is considered a random
variable whose location and size varies
on repeated application of the specific
statistical method. The use of a confidence

interval focuses attention on the
variability associated with applying the
specific statistical method to hypothetical
data sets, rather than with the uncertainty
in the parameter being estimated from
data at hand.
For example, a 95% interval for the potency of
a tablet lot might be stated as 99.8% +/- 2.2%. If
this interval were a confidence interval, we would
interpret it as the result of one application of a
statistical method known to produce an interval
that contains the true tablet potency 95% of
the time that the method is used. To properly
interpret a confidence interval, we must imagine
a hypothetical repetition of the experimental and
interval estimation process. The confidence level
is based on the properties statistical method when
employed on appropriate types of data.


t


DSFEJCMF
JOUFSWBM
"O
JOUFSWBM
XJUIJO

which a random (uncertain) parameter
value lies with 95% probability. A credible
interval is based on the quantiles of the
probability distribution of the parameter
which is fixed by prior assumptions or
observed data. A credible interval has
a simple interpretation that focuses
attention directly on the uncertainty in the
parameter being estimated.
Returning to our 95% interval example of 99.8%
+/- 2.2%, if this interval was a credible interval
we could make a more direct interpretation that
the true potency of the lot in question has a 95%
probability of being between 97.6 and 102.0%. We
do not need to imagine a hypothetical repeated
application of the statistical method used, but
base our probability level on the observed data
and the model we use to quantify our uncertainty.
When we use non-informative prior assumptions,
the credible interval we estimate is often similar
or even identical to the corresponding confidence
interval. As our prior assumptions become more
informative, a credible interval will generally
become narrower.
Both confidence and credible intervals are
useful. However, because credible intervals relate
more directly to our subject of process knowledge
building, we will use them exclusively here.
t

Distributional estimate: The vertical axis of a
probability density curve is proportional to the
likelihood that the parameter has the value on
the horizontal scale. The following two types
of distributional estimates are distinguished:


t

1 SJPS
EJTUSJCVUJPO
"
EJTUSJCVUJPOBM
FTUJNBUF

based largely on prior assumptions. A prior
distributional estimate of a parameter
might come from step 2 in Figure 2
David LeBlond
Table I: The Beta distribution used in estimating a
binomial proportion, .
Distribution Excel function Distribution Story
Beta P = BETADIST(,A,B,0,1) has a Beta distribution
with parameters A= and
D=EXP(GAMMALN(A+B)- B=-+1
GAMMALN(A)- (see reference 2).
GAMMALN(B))* ^(A A = defect count parameter
-1)*(1- )^(B-1) B = acceptable count pa-
rameter
= BETAINV(P,A,B,0,1)
A=B=1 provides an uninfor-
E () = A /(A+B) mative prior distribution (a
rectangular or uniform dis-
Mode () = (A - 1) / tribution, see reference 2).
(A+B-2)
Median () = BETA-
INV(0.5, A, B,0,1)
and resemble the Less Knowledge
distribution in Figure 1.


t

1PTUFSJPS
EJTUSJCVUJPO
"
EJTUSJCVUJPOBM

estimate based on both prior assumptions
and data. A prior distributional estimate
of a parameter might come from step
4 in Figure 2 and resemble the More
Knowledge curve in Figure 1.
If the distributional estimate is based on a
well-documented, probability distribution, such
as the normal distribution, we can provide the
distributional estimate in a simple statement such
as: The prior (or posterior) estimate of tablet lot
potency follows a normal distribution with a mean
of 98.8% and a standard deviation of 1.1%. If
we want to be more descriptive, or in cases where
the distribution is not well documented, we can
supply a graph of the distributional density (D) as a
function of the parameter value. In fortunate cases
such graphs can be produced using Excel functions.
We will illustrate a number of such distributional
estimates in the examples below.
Note that in going from a point estimate to an
interval estimate to a distributional estimate, the
amount of information contained increases. For
instance, the point estimate contains no indication
of the uncertainty associated with the estimate. It
is common practice to round the values of point
estimates to avoid conveying the impression of false
certainty. In fact the process of rounding only adds
uncertainty to an estimate.
A better approach is to provide an interval
estimate in addition to the point estimate. Interval
estimates convey an aspect of uncertainty by giving
a range within which the parameter value is likely
to lie. Often some confidence or probability level is
associated with the estimate.
PROCESS VALIDATION Process Design 75
David LeBlond
Table II: Knowledge building for a binomial propor-
tion, , using the Beta distribution.
Figure 2 Step Use of the beta distribution functions in Excel
1. Summarize Express knowledge about as a beta distribution
prior knowledge
based on prior experience, data, or expert opinion.
Let n0 be a prior sample size and d0 be the prior
number of defects observed in the prior sample. Then
take parameters of the beta distribution to be A = d0
and B = n0 -d0. In cases where little or nothing is pre-
viously known about , or in cases where decisions
are to be made on data alone, let A = B = 1.
2. State process Assume number defective in a sample of n follows a
model binomial model with unknown proportion .
3. Acquire new Take a random binomial sample of size n and inspect
data to find in this actual sample d defects.
4. Combine prior Express posterior knowledge about as a beta dis-
knowledge with tribution with parameters A=d0+d and B=n0+n-d0 -d.
new data to Obtain point, interval, and distributional estimates of
obtain posterior as described in the text.
knowledge
5. Use posterior The predictive posterior defect
knowledge to proportion = (d0+d)/(n0+n).
make predictions
of future samples
6&7. Acquire Refine model if necessary. Incorporate the posterior
confirmatory knowledge (step 4) as the new prior knowledge.
data and com-
pare it with pre-
diction. Return to
step 1.
However, it is by looking at the distributional
estimate of the parameter in Figure 3 that we realize
that an extremely high value is more likely than
an extremely low value. Such information could
be critical to predicting process performance at
its operating extremes. Often it is our ability to
state the likelihood of performance in the tails
that allows a meaningful process risk assessment.
A distributional estimate summarizes all the
information currently known about a parameters
value. So when available, distributional estimates
are most useful.
USE OF CALIBRATED ESTIMATION
METHODS IS AN IMPORTANT ASPECT OF
KNOWLEDGE BUILDING
In estimating a population location, which point
estimate is best: average, mode, or median?
Similarly for a point estimate population spread,
should we use standard deviation, range, or
something else? There are many possible estimation
methods.
Reliable methodology should be used to obtain
these estimates. Such methods might involve
the way we take samples, obtain our analytical
measurements, or calculate our estimates from
76 PROCESS VALIDATION Process Design
the analytical results. The idea here is that if we
use reliable methods of estimation, then the
parameter values we estimate should, in turn,
be reliable. If we knew, for example, that a given
method of estimation always produced an estimate
within 10% of the true value, that would be
good measure of our state of knowledge about a
parameter whenever its value is estimated by that
method. This point of view places an emphasis
on using reliable estimation methodology as a
critical aspect of scientific knowledge building.
Developing such methodology has been a major
focus of the discipline of statistics. Statisticians refer
to such methods, whose reliability has been well
characterized, as calibrated methods. In analytical
contexts calibration refers to establishing the
traceability and standardization of a measurement
method. As used here, calibration refers to
establishing the performance characteristics of an
estimation method.
PRIOR ASSUMPTIONS ARE THE
FOUNDATION OF THE ITERATIVE
SCIENTIFIC KNOWLEDGE-BUILDING
PROCESS
Some statistical methods are calibrated to
consider only the information contained in the
data at hand and do not consider auxiliary prior
information. Such methods are objective in that
the estimated parameters make minimal prior
assumptions. In particular such methods do
not consider prior knowledge about the values
of model parameters. It may be important
to use such methods in confirmatory trials
(steps 6 and 7 of Figure 2). However, it must be
remembered that all estimation methods depend
on assumptions of one kind or another, and it
is critical to justify them. Notice a pattern in
the list of pharmaceutical models previously
cited. Each model makes assumptions, and
those assumptions serve as the justification for
estimation. So what is the justification for the
modeling assumptions themselves? We base these
assumptions on some prior body of knowledge. A
careful examination will show that the prior body
of knowledge is in turn dependent on its own set
of prior assumptions. In the end, we are drawn
to the conclusion that the assumptions we are
willing to make play a critical role in our scientific
knowledge building.
In calibrating estimation methodology, the
assumption is made that the model used is the correct
model. This is rarely, if ever, the case. In scientific
investigations of a process, we strive to use models
and make assumptions that represent the best
state of our knowledge; yet we are fully aware that
 David LeBlond

these models and assumptions are always only Figure 4: Estimating a binomial proportion () using
approximations of the truth (6). beta distributions.
In addition to objectivity, consider the efficiency
of estimation methodology. For instance, while use 7

A = 1, B =1 A = 3, B =2 A = 2, B =3
of the sample median as an estimator of location 6
A = 20, B =20 A = 20, B =5 A = 5, B =20
makes minimal assumptions about measurement 5
distribution, it may require more measurements

Density
4
to obtain the same size interval estimate than
using a sample average that assumes a Gaussian 3

measurement distribution. Also, if our objective 2

in step 4 of Figure 2 is to build process knowledge, 1

then ignoring prior knowledge about a parameter
0
value makes little sense. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Using objectively calibrated methodology alone 

is not a sufficient metric of our state of process
knowledge. We require some way of incorporating
our prior knowledge into our scientific knowledge Figure 5: Estimation of a proportion defective ().
building. This can be difficult because even
18
knowledge experts can disagree about the reliability Tom's Posterior Density
16
of various prior assumptions. It is challenging
14 Nelly's Posterior Density
to quantify and compare our prior expectations Density 12 Tom's Prior Density
and uncertainties (7), but the result is shared 10
understanding and improved decision-making. Nelly's Prior Density
8
The differences in our prior beliefs are only 6
practically important when they have large effects 4
on our estimates and predictions. Again, the 2
discipline of statistics has provided some good 0
tools that allow us to understand the influence of 0 0.1 0.2 0.3 0.4 0.5

our prior beliefs as we build knowledge through 

modeling and experimentation as in Figure 2.

EXAMPLES
The following examples illustrate three hypothetical
situations in which modern statistical methodology
could be applied to the estimation process of Step
4 in Figure 2. The advantages that distributional
estimates offer are shown. The examples illustrate
how prior knowledge about model parameters
can be incorporated, in a quantitative and helpful
way, into the knowledge-building process. Excel
functions available for applying these approaches
are presented.
These examples are not presented as preferred
methods of analysis, but as illustrations of how
distributional thinking can, in simple cases, help
document and communicate the knowledge-
building process. The particular decision limits
described are purely fictitious and not to be
interpreted as recommended limits or industry-
accepted practice. The statistical methods used
in these examples are presented in what is hoped
will be a practical way; however, the mathematical
justification is not provided. The methodology
is not new, and interested readers can learn
more about these distributional techniques from
introductory textbooks (8, 9, 10).
EXAMPLE 1, BUILDING KNOWLEDGE
ABOUT A TABLET DEFECT PROPORTION
A new coating process may exhibit an unacceptable
defect (blemish) proportion in the resulting tablets.
A proportion above 0.3 (30% of tablets blemished)
is considered unacceptable. Tom and Nelly needed
to estimate the proportion defective (proportion of
blemished tablets, ) of the new process. Tom and
Nelly agreed that, prior to collecting data on the new
process, they would express their prior knowledge
about with beta distributions (see Table I).
The step that Tom and Nelly took in agreeing
to model their prior knowledge quantitatively
using beta distributions may seem unscientific and
arbitrary. However, we must remember that their
objective is to learn about a new process. With
new processes we do not always have the luxury
of directly applicable historical data that might be
available for decision-making with an established
process. Whatever experience, theory, or expert
opinion they can bring to the table will aid in their
learning process. The challenge is to capture this
prior knowledge in a useful way.
Tom and Nelly find that the beta distribution is
useful in modeling proportions for the following
reasons:

PROCESS VALIDATION Process Design 77

David LeBlond
Table III: Distributions useful in estimating a normal mean, , and standard deviation, .
Distribution Excel function
Inverted-Root P =1-GAMMADIST(1/ ,C,1/G,TRUE)
2
Gamma (IRG)
=SQRT(1/GAMMAINV(1-P,C,1/G))
D =2*G^C*EXP(-G/ ^2-GAMMALN(C))/ ^(2*C+1)
E() =EXP(GAMMALN(C-0.5)-
GAMMALN(C))*SQRT(G) , C>0.5
Mode() =SQRT(2*G/(2*C+1))
Median() =SQRT(1/GAMMAINV(0.5,C,1/G))
Normal or D or P =NORMDIST( ,L,Q,*)
Gaussian
= NORMINV(P, L,Q) = L + Q *NORMINV(P,0,1)
E() = Mode() = Median() = L
Location- P=IF(( -T)/U<0,TDIST(-( -T)/U, R,1), 1-TDIST((
Scale Stu- -T)/U,R,1))
dents-t (LSSt)
D=EXP(GAMMALN((R+1)/2)-GAMMALN(R/2))/
(PI()*R)^0.5/(1+(( -T)/U)^2/R)^((R+1)/2)/U
=T+U*IF(P<0.5, -TINV(2*P,R), TINV(2*(1-P),R))
E() = Mode() = Median() = T
*FALSE produces the probability density, D, of observing exactly . TRUE produces the probability, P, of observing x or less (cumulative probability
density function).
t

5IF
CFUB
EJTUSJCVUJPO
NPEFMT
SBOEPN
WBSJBCMFT

that take on values between 0 and 1, the same
range as the proportion,
t

5IF
CFUB
EJTUSJCVUJPO
DBO
UBLF
PO
B
HSFBU
WBSJFUZ

of shapes and thus provides a large selection
of possible distributions for Tom and Nelly to
choose from
t

5IF
UXP
QBSBNFUFST
PG
UIF
CFUB
EJTUSJCVUJPO

"
of distributional reasoning.
and B in Table I, actually have an interpretation
in terms of binomial sampling. To express ones
prior knowledge about , one can imagine a
hypothetical prior sampling experiment in
which A is the prior number of defective tablets
and B the prior number of acceptable tablets in
the hypothetical prior sample of A+B tablets.
Thus the beta distribution is intuitive.
t

5PN
BOE
/FMMZ
XJMM
BTTVNF
UIBU
EFGFDUT

in samples of tablets follow a binomial
distribution (1,2). A beta prior distribution lets
them combine their prior information about
(expressed as a beta distribution) with new data
in a very simple and intuitive way (see below).
t

5IFJS
JNQSPWFE
FTUJNBUF
PG
(the posterior
distribution), that combines information from
both the beta prior and new data, will also have
a beta distribution. In this way, they establish a
consistent learning cycle.
78 PROCESS VALIDATION Process Design
Distribution Story
1/2 2 has a Gamma distribution with parameters C=
and G=
(see reference 2).
C= shape parameter
G= squared scale parameter
has a normal distribution (see reference 2) with
L= location parameter
Q = scale parameter
( -T)/U has a Students-t distribution with R = degrees
of freedom
(see reference 2)
T = location parameter
U = scale parameter
t

*O
DBTFT
XIFSF
UIFSF
JT
WFSZ
MJUUMF
PS
OP
QSJPS

information about , or where they want an
estimate based solely on data, they could select
a beta distribution that is essentially non-
informative. In doing so their posterior estimate
will be consistent with that of other statistical
estimation methods that do not take advantage
(Note: We will not detail the rationale behind
the choice of priors in these examples, but similar
comments apply to all the priors used below.)
Tom and Nelly agreed to use the process given
in Table II to document their knowledge building
for . However, they disagreed as to which beta
distribution best represents their individual prior
beliefs about .
To select their priors, they followed step 1 of
Table II and used the Excel probability density
function (The D= function in Table I) to graph
some probability density function curves (plots of
D vs ) for different values of A and B. These curves
are shown in Figure 4.
When A=1 and B=1, the beta distribution is
a uniform distribution with constant density
between =0 and =1. This uniform distribution
expresses an ignorance (or lack of preference) for
the true value of . Nelly, who has little experience
 David LeBlond

Table IV: Knowledge building for a normal mean, , and standard deviation, .
Figure 2 Step Use of distribution functions in Excel
1. Summarize A. Express knowledge about as an IRG distribution based on prior experience, data, or expert opinion. Let
prior s 0 be a prior point estimate of based on v0 prior degrees of freedom. Take the parameters of the prior IRG
distribution to be C =v0 /2 and G=v0*s0^2/2. In cases where little or nothing is previously known about , or
in cases where decisions are to be based on data alone, let v0 = -1 and choose G near 0 which essentially
results in a flat prior for s (note some Excel functions will fail when C0).

B. Express knowledge about as a normal distribution conditional on . Choose the distribution using prior
experience, data, or expert opinion. Let m0 be a prior estimate of based on a prior sample size of n0. Take
the parameters of the prior normal distribution to be L = m0 and Q = s /sqrt(n0). Since is unknown, use
Q=s0 /sqrt(n0) to plot and verify this prior distribution. In cases where little or nothing is previously known
about m, or in cases where decisions are to be based on data alone, let L = 0 and choose n0 near 0 which
essentially results in a flat prior for .
2. State process Assume data comes from a normal model with mean m and standard deviation .
model

3. Acquire Take a random normal sample of size n and calculate the sample mean, m, and sample standard deviation, .
new data
4. Combine prior Update estimates of s0, v0, and n0 to sn, vn, and nn respectively:
knowledge with vn = v0 + n, nn = n0 + n, sn = v0s02 + (n 1) s2 + n0n (mm0)nnvn
new data to
obtain posterior
knowledge Express posterior knowledge about m as a LSSt distribution with parameters R=vn, T=(n0*m0 + n*m)/nn, and
U=sn /sqrt(nn). Express posterior knowledge about as an inverted root gamma distribution with parameters
C =vn /2 and G=vn*sn^2/2. Obtain point, interval, and distributional estimates of and as described in text.

5. Use posterior The estimated distribution of future individual values will be a LSSt distribution with parameters
knowledge to R=vn, T=(n0*m0 + n*m)/nn, and U=sn*sqrt(1+1/nn).
make predic-
tions of future
samples
6&7. Acquire Take the posterior knowledge (step 4) as the new prior knowledge.
confirmatory
data and com-
pare it with pre-
diction. Return
to step 1.

with this new coating process, felt that this closely
represented her subjective knowledge about the
defect proportion. This is the distribution that
represents a prior sample of n0 =2 containing d 0 =1
prior defect. Nellys prior density distribution is
illustrated in Figure 5.
Tom had used the new coating process on a
previous project and wanted to incorporate his
expert opinion about the defect proportion into
the estimate of p He plotted beta distributions for
a number of combinations of A and B (see Figure
4). Based on his expert opinion, a sample of 30
tablets from the new process should contain about
In step 2 of Table II, they discussed various
assumptions about sampling and independence and
agreed to model the number of defects in a sample
of tablets using a binomial probability distribution
with underlying defect proportion, (2). They
collected an actual random sample of n=200 tablets
from the process and observed d=36 defective
tablets (step 3 of Table II).
Following step 4 of Table II, the A and B
parameters of the posterior distributions for Tom
and Nelly are as follows:
Toms posterior distribution: A beta distribution
with A = d +d = 4 + 36 = 40, and B = n +n-d -d =
0 0 0

4 defective tablets. So he took the beta distribution 30+200-4-36 = 190

with A = 4 and B = 26 as his prior estimate of . Nellys posterior distribution: A beta distribution
Toms prior density distribution is illustrated in with A = d0+d = 1 + 36 = 37, and B = n0+n-d0 -d =
Figure 5. 2+200-1-36 = 165.

PROCESS VALIDATION Process Design 79

David LeBlond

Figure 6: Estimating a normal standard deviation () Both Tom and Nellys posterior density
using inverted-root gamma (IRG) distributions. distributions are plotted in Figure 5 and are
1 seen to match relatively closely. Clearly, the
0.9 v0 = 0.02, s0 = 1 knowledge contributed by the actual sample
0.8 v0 = 1, s0 = 2 of 200 has overwhelmed that contributed by
0.7
Density

v0 = 1, s0 = 10
0.6 prior assumptions. From Table I, Tom and Nelly
v0 = 10, s0 = 2
0.5
v0 = 10, s0 = 10
reported point estimates for (expected value)
0.4 A/(A+B) = 0.174 and 0.183 respectively. These
v0 = 100, s0 = 10
0.3
0.2 expected values give the predicted probability
0.1 that any given tablet sampled from the process
0 will be defective (Step 5 in Table II). In order to
0 2 4 6 8 10 12 14
make a final decision about the acceptability of

the new process, they obtained the following
95% credible intervals for : using the
Figure 7: Estimation of the standard deviation of tablet =BETAINV(P,A,B,0,1) function in Table I and
drug substance content (). setting P to either 0.025 (lower limit) or 0.975
(upper limit).
0.6
Toms 95% credible interval: 0.128 to 0.225
0.5 Dick's Prior Density Nellys 95% credible interval: 0.133 to 0.239
0.4 Jane's Prior Density
Density

Because the upper limits of both intervals were

0.3
Dick's Posterior Density below 0.3 (30% defect proportion), they concluded
0.2 Jane's Posterior Density that the new coating process was acceptable.
0.1
They recorded their knowledge-building process
in a development report that documents not only
0 the data, method of analysis (Tables I and II), and
0 3 6 9 12 15
sigma estimates of , but also their underlying prior
assumptions and beliefs (Figure 5). In this way, a
reviewer of their report acquires an intimate sense
Random Poste- Poste- Dicks Janes Data of the level of understanding and confidence about
Variable rior rior Value Value Only the process and its ability to produce tablets of
Distri- Distri- acceptable quality. The reviewer may have his or her
bution butional
own subjective prior knowledge about , but can see
Model Param-
eter from the report how the acquired data overwhelm
nn 11 11 10 such prior opinions.
sn 3.35 4.33 3.6
LSSt vn=R 12 12 10 Example 2, Building Knowledge About the
T 100.5 100.9 100.5 Uniformity of Individual Capsule Potency
U 1.01 1.31 1.14 Levels
IRG C 6 6 4.5 Drug product capsules were exhibiting unacceptable
G 67.28 112.55 58.65 levels of unit dose potency non-uniformity (relative
standard deviation greater than 6%, individual
95% Credible Interval capsules beyond 90-110%LC). A change to the
Estimate encapsulation process was being considered. Before
Ran- Distri- Table III Equa- Dicks Janes Data validating the change in production, it was decided
dom bution tion Used to Value Value Only to test capsules at pilot scale and use the results to
Vari- Model Obtain Credible estimate the underlying population distribution of
able Interval Esti- capsule potencies made by the process. Dick and
mates
Jane were the members of the team assigned the
LSSt 98.3- 98.0- 98.0-
task of estimating this distribution. They agreed to
=T+U*IF(P<0.5, 102.7 103.7 103.1
use the distributions described in Table III and the
-TINV(2*P,R),
TINV(2*(1-P),R)) associated knowledge building process of Table IV
IRG =SQRT(1/ 2.4- 3.1-7.1 2.5- to address this objective.
GAMMAINV(1- 5.5 6.6 Both Jane and Dick were willing to assume (per
P,C,1/G)) step 2 of Table IV), based on their experience with
such processes, that the capsule potencies would

80 PROCESS VALIDATION Process Design

 David LeBlond

be normally distributed. So the objective required Figure 8: Estimation of mean tablet drug substance content.
them to estimate the distribution mean, , and
0.4
standard deviation, . Given and estimates, and
Dick's Prior Density
assuming no batch-to-batch variation, they could
0.3 Jane's Prior Density
then predict the potencies of future tablets made by
the process.

Density
Dick's Posterior Density
Step 1A of Table IV suggests using an inverted- 0.2
Jane's Posterior Density
root gamma (IRG) distribution to summarize their
prior knowledge and expectations about
(i.e., in 0.1
Table III, IRG functions, they will take X to be
).
The IRG distribution has two parameters, C and G.
0
Jane and Dick examined various IRG distributions
90 95 100 105 110
by plotting the density (D) against
using the
mu
appropriate Excel function in Table III. They

obtained C and G parameter values as described

in step 1A of Table IV by selecting a prior standard
deviation, s0, and a prior degrees of freedom, v0. Figure 9: Predicting individual tablet drug substance content.
The degrees of freedom associated with a standard
deviation is related to the sample size used to 0.12
obtain that standard deviation. As v0 increases D ick 's Prediction

the confidence in our prior estimate s0 rises. The 0.1

J ane's Prediction
resulting distributions are given in Figure 6. 0.08 D ata O nly P rediction
Density

They noticed, as expected, that very low v0 (e.g., 0.02)

produces a nearly flat prior distribution. As v0 rises, 0.06
the distribution spread decreases and becomes more
0.04
bell shaped, centered near the chosen value of s 0.
By examining a number of possible distributions using 0.02
Excel, Dick and Jane arrived at their individual subjective
prior distributions. They also agreed to obtain estimates 0
based on a non-informative prior that contributes little or 85 90 95 100 105 110 115
no information relative to the data they will collect. The
respective prior v and s values were as follows:
0 0 Individual future result

Dicks IRG prior for : S 0 = 3, v0 = 2
Janes IRG prior for : S 0 = 7, v0 = 2
Data Only prior for : S 0 = 0.0001, v0 = -1.
Dick and Janes prior distributions for are
shown in Figure 7. As illustrated in Figure 7, Jane
(who had had some bad experiences with this new
encapsulation process) was less optimistic and
felt that the true would be elevated. Dicks prior
distribution appeared relatively optimistic. The
low number of degrees of freedom for each, v0 =2,
indicate that neither Dick nor Jane had strong prior
knowledge about . The choice of v0 =-1 for the non-
informative, data-only prior seems strange; but it is
mathematically consistent with a prior that shows
no preference for any particular value of .
with the prior estimate of m . Q also requires
knowledge of . Because they did not know s, they
substituted their prior estimate, s 0 , to obtain their
prior distributions. So the prior distribution for
is conditional (or dependent) on the true value
of . This makes sense because is a component
of measurement error. Dick and Jane recognized
that their prior knowledge about was inherently
dependent on the measurement variation, , in the
process. Dick and Jane graphed their priors for
using the appropriate Normal distribution function,
D = NORMDIST(, L, Q, FALSE) to select values
of m0 and n 0 they felt represented their subjective
prior knowledge about . They also included non-
informative (Data Only) values. The respective prior
m and n they chose were as follows:
0 0

Applying Step 1B of Table IV, a normal

distribution will be used to express prior knowledge
about . This normal distribution depends on
the parameters L and Q, as described in Table III.
L is simply the prior estimate of the true . Q is
a prior estimate of the standard error of estimate
of and so is equal to /sqrt(n0). This requires
that we provide a prior sample size, n0, associated
Dicks normal prior for : m0 =100, n 0 =1
Janes normal prior for : m 0 =104, n 0 =1
Data Only prior for : m0 = 0, n 0 = 0.0001
The small prior sample sizes, n0 , for Dick and
Jane indicate that their confidence in their initial
guesses, m0 , was low. The data only prior has

PROCESS VALIDATION Process Design 81

David LeBlond
Table V: The Gamma distribution used in
estimating a Poisson defect per unit rate, .
Distribu- Excel function Distribution Story
tion
Gamma D or P =GAMMADIST( has a Gamma dis-
,V,1/W,*) tribution with
parameters V= and
=GAMMAINV(P,V,1/W) W=
(see reference 2).
E() = V/W V= shape parameter
W= inverse scale pa-
mode () = (V-1)/W rameter
median() =
GAMMAINV(0.5,V,1/W)
* FALSE produces the probability density, D, of observing exactly .
TRUE produces the probability, P, of observing x or less (cumulative
probability density function).
essentially zero sample size and is thus non-
informative. Dick and Janes priors for are shown
in Figure 8.
Clearly, Jane felt the process mean m was greater
than did Dick. Her prior distribution had greater
spread because her prior estimate of , s 0 , was
larger than Dicks. The low sample size for each,
m0 =1, indicated that neither Dick nor Jane had
strong prior knowledge about .
A sample of n=10 capsules was tested and the
observed sample mean, m=100.6 and sample
standard deviation, s=3.6, were calculated (Step 3
of Table IV). These results, along with the above
prior estimates (s 0 , v0 , m0 , and n 0) were used in
step 4 of Table IV to obtain improved (posterior)
distributional estimates of and . A location-
scale Students t (LSSt) and IRG distributions were
used to model the posterior estimates of and
, respectively. The calculated parameters of these
posterior distributions were as follows:
The resulting posterior distributions were
plotted using the appropriate density functions
(D=...) in Table III and are shown in Figures 8 and
7, respectively. Point estimates of and were
obtained from the formulas for expected values of
the respective distributions:
Dicks point estimates: E() = T = 100.5, E() = 3.6
Janes point estimates: E() = T = 100.9, E() = 4.6
Data Only point estimates: E() = T = 100.5, E()
= 4.0
The 95% credible interval estimates of and were
obtained by substituting, into the proper equation in
Table III, either P=0.025 (lower limit) or P=0.975 (upper
limit) plus the appropriate posterior distributional
parameters (R, T, U, C, and G) given above.
82 PROCESS VALIDATION Process Design
Janes 95% credible intervals showed the greatest
spread and Dicks the least with the data only
intervals showing intermediate spread. We note in
passing that the data only 95% credible intervals for
and are identical to traditional 95% confidence
intervals often used. However, the traditional
approach cannot provide distributional estimates of
or .
Dick and Jane used the equations for
density (D=...) in Table III for the LSSt and IRG
distributions in Table III to graph their posterior
distributional estimates for and in Figures 8
and 7, respectively. Jane noted that her posterior
distribution for in Figure 7 was wider than Dicks
and showed substantial probability of being above
6, which would be unacceptable. If Janes prior
assumptions were accurate, there seemed some risk
that batches made by the process would exhibit
unacceptable content uniformity.
To examine this risk further, Dick and Jane
followed step 5 of Table IV to obtain predictive
posterior distributional estimates of the potency
levels of future capsules made by the process.
As indicated in Table IV, these follow an LSSt
distribution with the R, T, and U parameters
calculated as shown. For comparison, the predictive
posterior distribution for the data only, non-
informative, assumption was also obtained. The
three distributions are given in Figure 9.
Dicks distributional estimate showed the least
spread, Janes the greatest, and the data only
estimate showed intermediate spread. Janes prior
assumptions led to a substantial number of capsules
above 110% of target potency. Jane predicted the
percentage of capsules she would expect to be
beyond 90-110%LC using the following Excel
formulas from the LSSt distribution in Table III:
Probability that an individual capsule potency is
< 90%LC =
IF( (90-T)/U<0,TDIST(-(90-T)/U,R,1),1-
TDIST((90-T)/U,R,1) ) = 0.017
Probability that an individual capsule potency
is >110%LC = 1-IF( (110-T)/U<0,TDIST(-(110-T)/
U,R,1),1-TDIST((110-T)/U,R,1) ) = 0.033
Thus Jane predicted that 5% of capsules would be
beyond the desired range of 90 to 110%LC, which
would be unacceptable.
Dick and Jane were disturbed by the fact that
different prior assumptions led to differing
posterior distributional estimates for (see
Figure 7) and also to different predictive posterior
distributions of individual capsule potency levels
(see Figure 9). They felt that the information
provided by the 10 capsules they tested was
 David LeBlond

insufficient relative to their individual subjective Figure 10: Estimating a Poisson event rate () using
prior opinions. Therefore, they decided it would gamma distributions.
be prudent to test another random sample of 1
capsules before making any decision about the new 0.9
0.8 n0=2, r0=2
encapsulation process.
0.7 n0=2, r0=4

Density
EXAMPLE 3, BUILDING KNOWLEDGE 0.6 n0=10, r0=2
ABOUT AN ADVERSE EVENT RATE 0.5 n0=10, r0=4
0.4
Alina and Craig were epidemiologists initiating
0.3
a new pharmaco-vigilance surveillance program
0.2
on spontaneously-reported, quarterly adverse
0.1
event rates. If they discovered strong evidence of a
0
drug-adverse event combination rate greater than
0 2 4 6 8
two events per quarter, they would perform an
investigation. They defined strong evidence as a 
rate with a 95% probability of being greater than
two events per quarter. Figure 11: Estimation of an adverse event rate ()
Prior to examining any quarterly data, they felt
that their subjective prior knowledge about the 0 .8
Poisson event rate, , should be expressed using the 0 .7 Alina's Prior
gamma distribution described in Table V. 0 .6 Craig's Prior
Following step 1 in Table VI, Alina and Craig Density
0 .5 Alina's Poseterior
expressed their prior expectations about by Craig's Posterior
0 .4
choosing n 0 (prior number of quarters examined)
and r0 (prior guess of the true rate ). To familiarize 0 .3

themselves with the shapes of the gamma 0 .2

distribution, they tried a few values of n 0 and r0 as 0 .1
shown in Figure 10. 0
They noted how the distributions moved to the
0 1 2 3 4 5 6 7 8
right as r0 was increased and that the spread of the
distributions decreased as n 0 increased, as would be 
expected.
Alina felt that it would be wise to choose Using Alinas Using Craigs infor-
a non-informative prior distribution for uninformative mative prior
because, a priori, she did not want to prejudice prior
the estimation with her prior expectations. V =n0r0+d 27.0 32.3
Consequently, she chose n 0 =0.001 (no prior W = n0+n 8.0 10.5
quarters examined) with an initial prior guess E( ) = V/W 3.4 3.1
of of r0 =1000 events per quarter. Thus her
expected number of defects in this prior sample
was V = n 0 r0 = 1 with W = n 0 near zero. She used and below the decision limit of two events per
the Excel gamma density function in Table V, quarter.
D=GAMMADIST(r ,1, 1/0.001,FALSE), to graph Rather than choose a strategy with a single prior,
her prior distribution of which appears as a flat, they agreed to estimate from data using both
non-informative prior in Figure 11. priors. If either estimate showed strong evidence
Craig felt, based on his experience, that of being greater than two events per quarter, they
values above 6 are very unlikely and he wanted to would take that as a signal to be investigated. They
incorporate this knowledge into the estimation. fixed these two priors as part of their surveillance
Consequently he felt comfortable with taking protocol for any future drug-adverse event
n 0 =2.5 prior quarters examined and r0 =2.5 combination they studied.
prior events per quarter. This corresponded to They were ready to assume (step 2, Table VI) that
V=2.5*2.5 = 6.25 prior expected events in the counts of quarterly adverse events follow a Poisson
W = n 0 = 2.5 prior quarters. His prior density, distribution (1,2) with a mean adverse event rate
D=GAMMADIST( ,6.25, 1/ 2.5,FALSE), is also (number per quarter), . Now they were willing
graphed in Figure 11. His prior gives what he felt to begin their surveillance of drugadverse event
was a proper weighting of prior probability above combinations.

PROCESS VALIDATION Process Design 83

David LeBlond

Table VI: Knowledge building for a Poisson defects per unit rate, , using the
Gamma distribution.
Figure 2 Step Use of the Gamma distribution functions in Excel
1. Summarize prior knowl- Express knowledge about as a gamma distribution based on prior experience, data,
edge or expert opinion. Let n be the prior number of units of material examined and d be the
0 0

prior number of defects observed in the n units. Then take parameters of the gamma
0

distribution to be V=n r and W=n . In cases where little or nothing is previously known
0 0 0

about , or in cases where decisions are to be made on data alone, let V=1 and choose
W near 0.
2. State process model Assume defects are distributed randomly throughout the material according to a Poisson model
with unknown rate parameter, .
3. Acquire new data Take a random sample of n units of material and observe d defects in these n units.
4. Combine prior knowl- Express posterior knowledge about as a gamma distribution with parameters V = n0r0+d and
edge with new data to ob- W=n0+n. Obtain point, interval, and distributional estimates of as described in the text.
tain posterior knowledge
5. Use posterior knowl- The predictive posterior defect rate per unit is = (n0r0+d)/(n0+n).
edge to make predictions
of future samples
6&7. Acquire confirma- Refine model if necessary. Incorporate the posterior knowledge (step 4) as the new prior
tory data and compare knowledge.
it with prediction. Re-
turn to step 1.

After some examination (step 3, Table VI), they
discovered one particular drug for which d=26
headaches were reported over a two-year period
(n=8 quarters). Using the equations for V and W in
step 4 of Table VI, they obtained the following point
estimates for :
Alinas point estimate (3.4 events/quarter)
was a little higher than Craigs (3.1 events/
quarter). This was consistent with Alinas use
of an uninformative prior that gives substantial
probability to high rates. The posterior
distributional estimates were obtained using
the gamma density function from Table V, D
=GAMMADIST( ,V,1/W,FALSE), and these are
shown in Figure 11. Despite the very different
appearance of the prior distributions, the posterior
distributions appeared similar. Craigs distribution
showed less spread than Alinas because of the
additional information contributed by Craigs
informative prior.
To estimate the 95% one-sided lower credible
bound of , they used the quantile equation
from Table V, = GAMMAINV(0.05, V, 1/W), and
obtained the following:
Alinas one-sided 95% credible lower bound on :
2.4 events per quarter
Craigs one-sided 95% credible lower bound on :
2.2 events per quarter.
Because the lower bound was above
two events per quarter, they launched an
investigation of this particular drug-adverse
event combination.
SUMMARY AND CONCLUSIONS
Process knowledge building and estimation
are important components of validation and
compliance. In this column, a methodology by
which estimation of random variables is guided by
their probability distribution models in an overall
knowledge-building process was illustrated. These
three simple examples show how subjective prior
knowledge can be quantitatively combined with
knowledge contained in data in a way that facilitates
communication and improved decision making.
The provided Excel functions allow decision makers
and knowledge experts to participate and become
familiar with this approach to the knowledge-
building process. It is hoped that readers will find
these examples instructive and useful.
In the three examples illustrated, the
mathematical operations are relatively simple
and can be performed by someone familiar with
using Excel statistical functions. For more complex
models, analytical expressions for the various
probability distribution functions may not be
available in closed form and computer simulation
using a package such as WinBUGS (11) is required.
In such cases, or in any case where aspects of
statistical modeling or use of prior knowledge
are unclear, the advice of a trained statistician is
recommended.
ACKNOWLEDGMENT
Thanks to Diane Wolden and Paul Pluta for their
careful reading of this text and for their many

84 PROCESS VALIDATION Process Design


suggestions and corrections that have made it more
readable than otherwise possible.
GLOSSARY
Bayes rule: A process for combining the information
in a data set with relevant prior information
(theory, past data, expert opinion and
knowledge) to obtain posterior information.
Prior and posterior information are expressed
in the form of prior and posterior probability
distributions, respectively, of the underlying
physical parameters, or of predictive posterior
distributions of future data.
Calibrated estimation method: An estimation method
whose reliability and accuracy are known from
theory or computer simulation.
Conditional probability: The probability of an event (say
A) occurring, given that some other relevant
event (say B) has occurred, symbolized P(A|B).
For example, let A = house is on fire and B =
smoke alarm is sounding.
Confidence interval: A random interval estimate of a
(conceptually) fixed quantity which is obtained
by a estimation method calibrated such that
the interval contains the fixed quantity with a
certain probability (the confidence level).
Credible interval: An interval estimate of a random
variable, based on its probability distribution,
which contains its value with a certain
probability (the credible probability level).
Data: Measured random variable values, assumed to be
generated by some hypothetical model, which
contain information about the parameters of
that model.
Distributional estimate: An estimate of a random
variable that is of a probability distribution that
summarizes complete knowledge about the its
possible values.
Estimation method: A statistical procedure that uses
data to produce estimates of quantities.
Expected value: A point estimate of a random variable that is
the weighted mean of its possible values, with the
weight being the probability density at each value.
Interval estimate: An estimate of a random variable
consisting of a range (or interval) of values
that provides a measure of both location and
spread. Often the range is associated with some
probability or confidence level.
Knowledge: As used here, knowledge about a random
variable is inversely related to the spread of its
probability distribution.
Median: A point estimate of a random variable that
is the 50th percentile (or 0.5 quantile) of its
probability distribution.
Originally published in the Summer 2008 issue of The Journal of Validation Technology
David LeBlond
Mode: A point estimate of a random variable that is
the value at which its probability density is
maximized.
Model: A description of a data generating process
that includes parameters (and possibly
other variables) whose values determine the
distribution of data produced.
Point estimate: An estimate of a random variable
consisting of a single number that gives some
idea of the location of the value of the random
variable along its measurement scale.
Prior distribution: A subjective distributional estimate
of a random variable, obtained prior to any
data collection, that consists of a probability
distribution.
Posterior distribution: A distributional estimate of a
random variable that updates information from
a prior distribution with new information from
data using Bayes rule.
Predictive posterior distribution: A distributional
estimate (or prediction) of a random variable
(typically future data) that combines posterior
knowledge about model parameters with
variability in data generated by the model.
REFERENCES
1.LeBlond, D., Data, Variation, Uncertainty, and Probability
Distributions, Journal of GxP Compliance, Vol. 12, No. 3, pp
30-41, 2008.
2.LeBlond, D., Using Probability Distributions to Make
Decisions, Journal of Validation Technology, Spring 2008,
pp 2 14, 2008.
3.International Conference on Harmonization, ICH
Harmonised Tripartite Guideline on Pharmaceutical
Development Q8, Current Step 4 version, November 10, 2005.
4.Price, R., (1763) An essay towards solving a problem in the
doctrine of chances, Dale, A Most Honourable Remembrance:
The Life and Work of Thomas Bayes, Springer, New York, 2003.
5.Deming, W. Edwards, Some Theory of Sampling, Dover
Publications, 1966.
6.Essentially, all models are wrong, but some are useful,
quoted from Box, George E. P.; Norman R. Draper, Empirical
Model-Building and Response Surfaces, p. 424, Wiley, 1987.
7.To endure uncertainty is difficult, but so are most of
the other great virtues, quoted from Bertrand Russell,
Philosophy for Laymen, Unpopular Essays, George Allen and
Unwin, London, 1950.
8.Box, G. and Tiao, G., Bayesian Inference in Statistical Analysis,
Addison-Wesley Pub. Co., Reading, MA, 1973.
9.Gelman, A., Carlin, J., Stern, H., and Rubin, D., Bayesian Data
Analysis, 2nd Edition, Chapman and Hall/ CRC, New York, 2004.
10. Bolstad, W., Introduction to Bayesian Statistics, 2nd Edition,
John Wiley & Sons, Hoboken, New Jersey, 2007.
11. Cowles, M., Review of WinBUGS 1.4, American
Statistician 58(4), 330-336, 2004. JVT
PROCESS VALIDATION Process Design 85
David LeBlond
Understanding Hypothesis
Testing Using Probability
Distributions
David LeBlond
Statistical Viewpoint addresses principles of
statistics useful to practitioners in compliance and
validation. We intend to present these concepts in a
meaningful way so as to enable their application in
daily work situations.
Reader comments, questions, and suggestions are
needed to help us fulfill our objective for this column.
Suggestions for future discussion topics or questions
to be addressed are invited. Readers are also invited
to participate and contribute manuscripts for this
column. Case studies sharing regulatory strategies
are most welcome. Please contact coordinating
editor Susan Haigney at shaigney@advanstar.com

PS
UXP
with comments, suggestions, or manuscripts for
publication.
KEY POINTS
The following key points are discussed:
t

4DJFOUJGJD
JOGFSFODF
SFRVJSFT
CPUI
EFEVDUJWF
BOE

inductive inference
t

5ISFF
NBJO
BQQSPBDIFT
UP
JOEVDUJWF
JOGFSFODF
BSF

used
t

'JTIFSJBO
JOEVDUJPO
VTFT
UIF
1WBMVF
BT
B
NFBTVSF

of evidence against the null hypothesis
t

/FZNBO1FBSTPO
JOEVDUJPO
DPOUSPMT
UIF
MPOHSVO

decision risk over repeated experiments
t

#BZFTJBO
JOEVDUJPO
PCUBJOT
EJSFDU
QSPCBCJMJTUJD

measures of evidence from the posterior
distribution
[ ABOUT THE AUTHOR
For more Author
information, David LeBlond, Ph.D., has 29 years experience in the pharmaceutical and medical diagnostics
go to fields. He is currently a principal research statistician supporting analytical and pharmaceutical
gxpandjvt.com/bios development at Abbott. David can be reached at David.LeBlond@abbott.com.
86 PROCESS VALIDATION Process Design
t

"
1WBMVF
JT
UIF
QSPCBCJMJUZ
PG
PCTFSWJOH
B
SFTVMU

as extreme or more extreme than that observed,
assuming the null hypothesis is true
t

5IF
5ZQF
*
FSSPS
SBUF
JT
UIF
QSPCBCJMJUZ
PG

incorrectly rejecting the null hypothesis
t

5IF
5ZQF
**
FSSPS
SBUF
JT
UIF
QSPCBCJMJUZ
PG

incorrectly failing to reject the null hypothesis
t

5IF
#BZFT
GBDUPS
JT
B
NFBTVSF
PG
FWJEFODF
JO
GBWPS

of the null hypothesis contained in the data
t

1WBMVF
FTUJNBUJPO
JT
CBTFE
PO
VOPCTFSWFE
SFTVMUT

more extreme than those observed, so it may over-
state the evidence against the null hypothesis
t

5IF
1WBMVF
GSPN
B
QPJOUOVMM
IZQPUIFTJT
sided test of equality, is very difficult to interpret. A
confidence or credible interval should be provided
in addition to the P-value.
t

"OBMZTJT
PG
WBSJBODF
JT
B
'JTIFSJBO
IZQPUIFTJT
UFTU

for the equality of means of two or more groups
t

5IF
#BZFTJBO
BQQSPBDI
PGGFST
BO
JOTJHIUGVM

re-interpretation of the analysis of variance in
terms of the joint posterior distribution of the
group means.
INTRODUCTION
The first issue of Statistical Viewpoint (1) presented
basic probability distribution concepts and Microsoft
Excel tools useful in statistical calculations. The
second (2) reinforced these concepts and tools
through eight scientific decision-making examples.

The third and fourth (3, 4) illustrated how probability
distributions aid in process knowledge building. This
issue shows how probability distributions are central
to the understanding of hypothesis testing. Some of
the concepts of probability distributions introduced in
the first four issues of this column will be helpful in
understanding what follows here.
Product, process, or method development can be
viewed as a series of decisions based on knowledge
building: What type of packaging should be employed?
What is the optimum granulation time? Can a
proportional model be used for calibration? The tools
we use to make such decisions are many. We rely on
prior theory and expert knowledge to conceptualize
the underlying mechanisms, select prototypes for
testing, design experiments, and prepare our minds
to interpret experimental results. We use dependable
measuring systems to acquire new data. Often, when
the results of our experimental trials are unclear (and
even sometimes when they are obvious), we employ
statistical and probabilistic methods to guide us in our
decision making.
We humans are reasonably good at exploring our
world, finding explanations for things, and making
predictions from our theories. Unfortunately, while we
all have a sense of rational intuition that (for the most
part) serves us well, the process we use (or should use)
to build understanding and make optimal decisions
from data has been the subject of heated debate by
philosophers, scientists, and mathematicians for
centuries (reference 5 gives a nice, readable overview;
see references 6 and 7 for more details). While the
debate shows no sign of concluding in our own time,
three noteworthy approaches have emerged. Here
we discuss some history and key concepts of each
approach and illustrate the central role probability
distributions play with two simple examples.
EXAMPLE 1: TABLET POTENCY
Consider the case of a development team concerned
with the true mean potency, averaged over batches,
produced by a tablet manufacturing process. In this
case, the tablet label claim (LC) and target for the
manufacturing process is 100%LC. Individual batches
may have a mean potency that deviates slightly from
100%, but batch means <90%LC are unacceptable.
While the team believes the process is adequate,
their objective is to provide evidence that the process
produces acceptable batches. How can the team
validate their belief that the process mean potency
is acceptable? We will use this example to illustrate
the three difference systems of scientific inference for
doing this.
David LeBlond
THE PROCESS OF SCIENTIFIC INFERENCE
Let us start by defining several important terms as
follows (Note that these and additional terms are
defined in the Glossary section at the end of this
article):
t

Parameter: In statistics, a parameter is a quantity
of interest whose true value is to be estimated.
Generally a parameter is some underlying variable
associated with a physical, chemical, or statistical
model. When a quantity is described here as
true or underlying the quantity discussed is a
parameter.
t

Hypothesis: A provisional statement about the
value of a model parameter or parameters whose
truth can be tested by experiment.
t

Null hypothesis (H0): A plausible hypothesis
that is presumed sufficient, given prior knowledge,
until experimental evidence in the form of a
hypothesis test indicates otherwise.
t

Alternative hypothesis (Ha): A hypothesis
considered as an alternative to the null hypothesis,
though possibly more complicated or less likely
given prior knowledge.
t

Inference: The act of drawing a conclusion
regarding some hypothesis based on facts or data.
t

Inductive inference: The act of drawing a
conclusion about some hypothesis based primarily
on data.
t

Deductive inference: The act of drawing a
conclusion about some hypothesis based entirely
on careful definitions, axioms, and logical
reasoning.
We can identify the following four types of activities
in the decision-making process which are illustrated in
Figure 1.
State Hypotheses About True Mean
State one or more hypotheses about the underlying
true mean potency parameter. In Figure 1, three
possible hypotheses (true process mean potency = 92,
96, and 102) are illustrated. Such hypotheses are called
point hypotheses because they specify a single fixed
value of an underlying parameter. Useful hypotheses
often specify a range of values and are referred to as
composite hypotheses. Note the following definitions:
t

Composite hypothesis: A statement that gives
a range of possible values to a model parameter.
For example, Ha: true mean > 0 is a composite
hypothesis.
t

Point (simple) hypothesis: A statement that
a model parameter is equal to a single specific
value. For example, H0: true mean = 0 is a simple
hypothesis.
PROCESS VALIDATION Process Design 87
David LeBlond
Figure 1: Scientific inference applied to the
mean potency of a tablet manufacturing
process.
1. State hypotheses about true mean
2. Make H: 92 H: 96 H: 102
deductive
Inferences
4. Make
80.5 89.8 96.2 102.7 108.3
inductive
inferences
3. Obtain a sample estimate of mean
For example 1, if low potency is a concern, the
team may be interested in testing a composite null
hypothesis such as
H0: true mean = or > 96%LC
against a composite alternative hypothesis such as
Ha: true mean < 96%LC.
The value of 96%LC might be considered the lowest
process mean potency consistent with an acceptable
process. That is, if the team guesses that the true
process standard deviation is about 2%LC, then 96%
would be safely 3-sigma above the unacceptable
lower limit of 90%LC.
There is an asymmetry to the hypotheses such that
the null hypothesis, H0, is considered a priori most
likely, requiring the fewest assumptions (i.e., the
process is performing acceptably). Following Ockhams
Razor (8), the simplest hypothesis is often the default
H0. The hypothesis chosen as H0 is usually the one
that requires the lower burden of proof in the minds of
decision makers.
If the team believes the process is adequate, the
alternative hypothesis, Ha, is considered less likely
than H0 and would require postulation of some
special cause, defined as follows:
t

Special cause: When the cause for variation
in data or statistics derived from data can be
identified and controlled, it is referred to as
a special cause. When the cause cannot be
identified, it is regarded as random noise and
referred to as a common cause.
Still the team needs supportive evidence because,
if the true process mean is < 96%LC, the process may
result in subpotent or failing batches.
Make Deductive Inferences
The team believes that batch mean potency
88 PROCESS VALIDATION Process Design
measurements will be normally distributed about
the true process mean potency. Thus they have a
mechanistic/probabilistic model to predict the likely
range of measured batch mean potency values to
expect. This kind of model is known as a likelihood
model, defined as follows:
t

Likelihood model: A description of a data
generating process that includes parameters (and
possibly other variables) whose values determine
the distribution of data produced. Specifically, the
likelihood is
likelihood = Probability of the observed data
if the hypothesis is true. [Equation 1]
Note that the predicted range of the observed data
depends on the hypothesized true mean potency. The
range will be different for H0 and Ha, with a mean
process potency of 96%LC being considered borderline
acceptable. The act of predicting (or simulating)
future data from such a likelihood model is purely
deductive. Such predictions are always true as long as
the underlying model and hypothesized value of the
underlying parameter are true.
Obtain Sample Estimate Of Mean
The team obtains potency measurements for 10
batches made using the process by testing composite
samples. This is the experimental part of the decision-
making process. The measured batch potencies
constitute raw data. For inferences a summary of
the data will be sufficient. In the present case, the
observed mean of 93%LC and standard deviation of
5%LC were obtained. The team noted that 93%LC is
below 96%LC. But is it far enough below 96%LC to
reject H0?
Make Inductive Inferences
From Figure 1 we see that induction is the opposite
of deduction: On the basis of the data, we evaluate
which hypothesis (H0 or Ha) is most likely. When
we reason inductively, we reflect back from the
observations to some underlying truth about
nature, in this case the true process mean potency.
Unlike deduction, even if our data are valid, there
is no guarantee that our conclusions about nature
are correct. What we hope to do is make optimal
scientific decisions and/or acquire evidence in
favor of one or more of the hypotheses we have
considered, with some appreciation for the decision
risks.
Given the larger than expected standard
deviation estimate (5 instead of the expected
2%LC), could the value of 93%LC have been due to
random variation? What inductive inference can the
team make about the true process mean potency?

To help the team with this decision we must
review some background on methods of inductive
inference.
THREE SYSTEMS OF INDUCTIVE
INFERENCE
In the following sections we describe the three
systems of inductive inference most commonly
employed today. Each description opens with
a brief historical perspective followed by an
application of the methodology to our Example 1.
Fisherian Induction
The term Fisherian seems appropriate because it
was R. A. Fisher who described the approach with
the greatest clarity and laid its statistical foundations.
Before 1900, inductive inference from data was
informal. The discipline of statistics, as we know it
today, was in its infancy. While many probability
distributions and models were known, workers
typically summarized their data using tables and
graphs and made visual comparisons with theoretical
predictions. In 1900 the British mathematician Karl
Pearson described what is now called the Chi-square
test (9). This innovation was followed in 1908 with a
t-test for means by an Irish brewer, William Gosset,
better known to us as Student (10), and in 1925 with
ANOVA and an associated F-test for comparing
groups of means by the English geneticist and
mathematician Ronald Fisher (11). The F-distribution
was so named in his honor (12).
The following are definitions of some terms that are
central to the Fisherian approach:
t

Statistic: A summary value (such as the mean or
standard deviation) that is calculated from data.
A statistic is often used because it provides a good
estimate of a parameter of interest
t

Sampling distribution: The distribution of data
or some decision statistic calculated from data
t

t-statistic: The decision statistic used in Students
t-test consisting of the ratio of a difference between
an observed and hypothesized mean divided by
its estimated standard error
t

F-statistic: The decision statistic used in Fishers
analysis of variance hypothesis test consisting of
the ratio of two independent observed variances
calculated from normally distributed data
t

Analysis of variance (ANOVA): A hypothesis
test that uses the F-statistic to detect differences
among the true means of data from two or more
groups.
The Chi-square, t-test, and ANOVA hypothesis tests
(and many others developed subsequently) rely on the
following ideas:
David LeBlond
t

3FEVDJOH
SBX
EBUB
UP
TPNF
EFDJTJPO
TUBUJTUJD

whose sampling distribution is known from the
likelihood, Equation 1
t

6TJOH
B
i1WBMVFw
GPS
EFDJEJOH
XIFUIFS
BO

observed set of data deviates from a hypothesized
probability distribution more than would be
expected from random error. P-value, is defined
as: The probability of obtaining a result at least
as extreme as the one that was actually observed,
given that the null hypothesis is true. The fact that
p-values are based on this assumption is crucial to
their correct interpretation.
Fisher offered his opinion about the proper P-value
criterion for decision making (11, page 80):
We shall not often be astray if we draw a
conventional line at 0.05.
Over 80 years later his conventional line is widely
used as a criterion for approval of new pharmaceutical
products.
In our tablet-manufacturing example, Fisherian
induction would have us summarize our data as a
t-statistic, which can be done by using Excel syntax as
follows:
t statistic = sqrt(Sample Size)*(Observed Mean
H0)/(Standard Deviation)
= sqrt(10)*(93 - 96)/5
= -1.8974 [Equation 2]
If the team were to repeat their experiment using 10
different (independent) batches, they would of course
not get the same t-statistic because of sampling and
measurement variation. Conceptually, if they repeated
the experiment many times, and if the true value of
the process mean was equal to 96%LC, the sampling
distribution of the t-statistics would be the probability
distribution given in Figure 2. This distribution is known
as the Students t-distribution.
Notice that the observed t-statistic (-1.8974) is
relatively far to the left side of the t-distribution. This
is because the observed mean of 93%LC is somewhat
below the hypothetical limit of 96%LC. Does this
mean that the team should reject H0? Fisherian
induction suggests that they should reject H0 (in favor
of Ha) if it is unlikely to obtain such a t-statistic or
one even more extreme by random chance alone. In
this case even more extreme would include all those
values equal to or less than -1.8974. The probability
of observing such extreme values by chance alone is
PROCESS VALIDATION Process Design 89
David LeBlond
Figure 2: Fisherian inductive inference model.
Observed value = 1.8974
Probability density
Sampling
distribution
for H0

UIF
1WBMVF
JOEFY
JODMVEFT
OPU

P-value = 0.045
0 also all those t-statistic values of lesser value that
Observed statistic (t)
equal to the area under the distribution curve to the
left of -1.8974. This probability is known as the P-value
and we can obtain it easily using the Excel cumulative
distribution function as follows:
P-value=TDIST(-observed t-statistic, sample size-1,1)
=TDIST(1.8974, 9, 1)
= 0.045.
Thus, such extreme (or even more extreme) values of
the t-statistic would occur by chance alone on average
less than 1 time in 20 repeats of this experiment,

UIF
1WBMVF
JT
OFJUIFS
PG
UIF

or less than 5% of the time. If we use Fishers
conventional line, the team should reject H0 and
conclude that the true process mean is below 96%LC
and thus unacceptable. The P-value is also called
the significance level of the hypothesis test and
when it is low (say < 0.05) it is considered a measure
of evidence against the null hypothesis, defined as
follows:
t

Measure of evidence: In scientific studies,
hypothesis testing is used to build evidence for
or against various hypotheses. The P-value and
Bayes factor (see below) are examples of measures
of evidence in Fisherian and Bayesian induction,
respectively.
While it offers an objective criterion, the P-value
may not be an optimal measure of evidence of the
validity (or not) of the null hypothesis (13), and must
be interpreted with care (14). The following cautions
apply to the Fisherian perspective:
t

"T
XJUI
BMM
TUBUJTUJDBM
QSPDFEVSFT

UIF
NFUIPEPMPHZ
As with Fisherian induction, Neyman-Pearson
is only applicable as long as all assumptions
of the likelihood model (the data generation
model) apply. For instance, in our example we
assume normality and independence of the
measurements.
t

/PUJDF
JO
FRVBUJPO

UIBU
UIF
NBHOJUVEF
PG
UIF

t-statistic depends directly on the square root
90 PROCESS VALIDATION Process Design
of the sample size. If our team had tested 1,000
instead of only 10 batches, it is likely that the
t-statistic would have been significant even if
the observed mean had been only slightly below
96%LC. Thus it is always wise to supplement a
hypothesis test with a confidence interval for
the mean potency (see references 3 and 4 for a
discussion of confidence intervals).
t

#Z
JUT
WFSZ
OBUVSF
only the observed t-statistic value, -1.8974, but
were not observed. In a sense, we are rejecting
H0, because it has failed to predict low t-statistic
values that were not in fact observed. The P-value
is similar to the index often used to classify the
relative performance of students: the observed
t-statistic is in the lower 5% of its class. Within
that category there are many poorer performers
with whom we are not concerned, but the P-value
disregards this information.
t

5IF
iDPOWFOUJPOBM
MJOFw
EFDJTJPO
QPJOU
GPS
B

P-value of 0.05 may not be appropriate in all cases.
The P-value per se does not take into account the
consequences of making an incorrect judgment
concerning H0. Further, it is difficult to integrate
the P-value index with measures of decision error
consequences.
t

.PTU
JNQPSUBOUMZ
following:
a. The probability that the initial experimental
result will repeat, or
b. The probability that H0 is true.
To obtain these probabilities we need to use one of
the other two systems of induction.
Neyman-Pearson Induction
Between 1927 and 1933, two of Fishers
contemporaries extended his groundbreaking ideas
about hypothesis testing. Egon Pearson was the son of
Karl Pearson (the developer of the Chi-square test) and
a colleague of Fisher in London. Jerzy Neyman was a
Polish mathematician in Warsaw who, among other
things, developed the idea of confidence intervals.
Their collaboration (15) led to a more general concept
of inductive inference. They developed an extremely
useful theory of optimal testing. Some key aspects of
their ideas are shown in Figure 3.
induction recognizes a decision statistic having a
known sampling distribution. The range of values of
the decision statistic regarded as unlikely (if H0 is true)
is called the rejection region. This region is in the tail
or tails of the sampling distribution and corresponds to
some fixed tail probability (e.g., 0.05) called the Type I
error. Type I error is defined as follows:

t

Type I error: A decision error that results in
falsely rejecting the null hypothesis when in fact it
is true.
They envisioned decision makers using this
decision statistic for all future hypotheses tests. In this
way, all future hypothesis tests would have a fixed
probability (e.g., 0.05) of incorrectly rejecting H0.
Different Type I errors could of course be chosen for
different situations, but the important point was to
ensure that the rate of incorrectly rejecting H0 would
be understood for each decision.
For instance, in the current example, our team may
agree that a Type I error rate of 0.05 (i.e., one error in
20 hypothesis tests) is appropriate in their situation.
Using the TINV EXCEL function, this error rate
corresponds to the following fixed t-statistic:
Fixed t-statistic =-TINV(2*(Type I error rate),Sample
size - 1)
= - TINV(2*0.05, 10-1)
= - 1.8331.
According to the Neyman-Pearson scheme, any
observed t-statistic less than -1.8331 would result in a
rejection of H0. For our tablet manufacturing example
we would reject H0 because -1.8974 < -1.8331. We
should note here that the calculation of the t-statistic
in the equation would be modified if larger (rather
than smaller) potencies, or if both larger and smaller
potencies were considered unacceptable.
In comparing the Neyman-Pearson paradigm
(Figure 3) with that of Fisher (Figure 2) notice that an
additional distribution (Ha) is added. Neyman and
Pearson recognized the need to consider the sampling
distribution of the statistic (the t-statistic in the present
example) for both a specific H0 (such as when the true
mean potency equals 96%LC) and at a specific Ha
(representing some arbitrary true mean potency that
might be considered unacceptable). The probability
of incorrectly accepting H0 when in fact Ha is true, is
called a Type II error, defined as follows:
t

Type II error: A decision error that results in
failing to reject the null hypothesis when in fact it
is false.
Of course this Type II error will depend on the
specific Ha being considered. One can make a plot of
Type II error as a function of the value of the parameter
(e.g., the true mean potency) associated with Ha.
Such a plot is called a Power Curve or an Operating
Characteristic curve for the hypothesis test, defined as
follows:
t

Power (or operating characteristic) curve:
David LeBlond
Figure 3: Neyman-Pearson inductive
inference model.
Decision point
Sampling Sampling
Probability density
distribution distribution
for Ha for H0
Type I error Type II error
t statistic
Power is equal to 1 minus the Type II error rate.
The power curve of a hypothesis test is a plot of
the Power versus the true value of the underlying
parameter of interest.
The calculation of such power curves is an
important part of experimental planning; however, it
involves the use of probability distribution functions
(such as the non-central t distribution) that are not
available in Excel.
One can picture these decision risks as a 2x2 table
such as in Figure 4. In adopting a Neyman-Pearson
paradigm, the team has moved from the objective
of developing evidence with respect to their specific
experiment, to the objective of using a methodology
with assured decision risks. The Neyman-Pearson
approach does not concern itself with whether
or not the true mean process potency is below or
above 96%LC. It only assures the team that the
many decisions they will make over their careers to
reject H0s will be incorrect only 1 time in 20 (i.e., a
probability of 0.05). Examples of practical situations
where this point of view is appropriate are listed as
follows:
t

Control charting. For monitoring critical
quality measures of a process or method, it may
be useful to know the probability of incorrectly
identifying an out of control situation (Type I
error) or of failing to detect a condition that is
unacceptable (Type II error).
t

Diagnostics screening. A diagnostic test is
analogous to a hypothesis test. The sensitivity
(1 - Type II error rate) and specificity (1 - Type I
error rate) of a diagnostic test are key measures that
determine the medical value of the reported test
result when the prevalence of disease in the tested
population is known.
t

Validation and product acceptance
testing. For judging production costs and
allocating resources, it may be desirable to fix the
manufacturers risk (Type I error, risk of incorrectly
PROCESS VALIDATION Process Design 91
David LeBlond
Figure 4: Neyman-Pearson hypothesis testing
error types.
True state of nature*
H0 Ha
Conclusion
from H0 no error Type II error
experiment
Ha Type I error no error
*Choosing the wrong H0 or Ha to study is
sometimes called a Type III error.
failing an acceptable batch) or consumers risk
(Type II error, risk of incorrectly passing a batch of

UIF
/FZNBO
some defined level of unacceptability).
t

New drug application regulatory
acceptance. For maintaining standards of risk, a
regulatory agency may find it desirable to require
all studies of a given type (e.g., clinical trials, shelf-
life estimation, bio-equivalence tests) to maintain
Type I error benchmarks. Requirements with
respect to Type II error can assure that sample sizes
are adequate (e.g., for safety studies).
The Neyman-Pearson approach permeates much of
todays scientific decision making. It represents a high
watermark in terms of objectivity and consistency in
inductive inference. When we use hypothesis testing
methodology whose Type I and II error risks are
known, we say that we are using a calibrated method,
and this can have many advantages. A calibrated
hypothesis test is defined as follows:
t

Calibrated hypothesis test: A hypothesis test
method whos Type I error, on repeated use, is
known from theory or computer simulation.
However, the Neyman-Pearson approach may not
be the appropriate paradigm for all situations. Some
important considerations are listed as follows:
t

8IJMF
UIF
BQQSPBDI
EPFT
GJY
EFDJTJPO
FSSPS
SBUFT

over a series of experiments, it does not by itself
provide a measure of evidence concerning H0 or
Ha in any specific experiment.
t

*O
NPTU
BDUVBM
TUVEJFT
UIFSF
XJMM
CF
NVMUJQMF

hypotheses tests, which may or may not be
independent. We refer to groups of hypothesis
tests that are associated with the same decision as
a family. Thus we must consider both the family-
wise error rates as well as the individual test rates
obtained from Neyman-Pearson methodology.
These family-wise rates suffer from a condition
called multiplicity in that they can be difficult to
predict.
t

5IF
/FZNBO1FBSTPO
5ZQF
*
FSSPS
SBUF
NVTU
OPU

be confused with the Fisherian P-value. The Type
I error rate is the rate of falsely rejecting H0 over
92 PROCESS VALIDATION Process Design
many experiments. The P-value is a measure of
evidence (albeit imperfect) against H0.
t

3 JHJE
BEIFSFODF
UP
B
5ZQF
*
FSSPS
SBUF
MFBET

to conceptual problems. Type I error rates of
0.0499 and 0.0501 are very close in any practical
situation, yet they could lead to very different
decisions.
t

*G
UIF
EFDJTJPO
TUBUJTUJD
EPFT
OPU
GBMM
JO
UIF

rejection zone, the Neyman-Pearson formulation
recommends that H0 be accepted. However,
from a scientific point of view it is more
appropriate to fail to reject H0. A larger
experiment would have a larger rejection zone that
might include the observed result.
t

"T
XJUI
UIF
'JTIFSJBO
TZTUFN
Pearson system relies solely on the likelihood
(probabilistic model of data generation) for
both deductive and inductive inferences.
However, developing and building evidence for
mechanistic, predictive models often requires
strong theory and experience. Finding and using
such models as a part of risk management control
strategies is the key to regulatory initiatives such
as quality by design (QbD) (16). To incorporate
such prior knowledge in a quantitative way, we
must use Bayesian induction that grew out of an
earlier age.
Bayesian Induction
In 1739, the Scottish empiricist philosopher David
Hume posed the following problem in inductive
inference (17):
tis only probable that the sun will rise tomorrow
we have no further assurance of this fact than what
experience affords us.
Knowing the underlying probabilities of events
was critical to the active 18th-century insurance and
finance industries whose profits depended on accurate
inductive inferences from available experience and
theory. It was also a problem of some interest to the
liberal theologians of that time. In 1763, the problem
was addressed quantitatively for the first time by two
non-conformist ministers, Thomas Bayes and Richard
Price (18). Their solution, Bayes rule, is to probability
theory what the Pythagorean theorem is to geometry.
The following definitions apply:
t

Bayes rule: A process for combining the
information in a data set with relevant prior
information (e.g., theory, past data, expert
opinion, and knowledge) to obtain posterior
information. Prior and posterior information
are expressed in the form of prior and posterior
probability distributions, respectively, of the
underlying physical parameters, or of predictive
posterior distributions of future data.

t

Prior distribution: A subjective distributional
estimate of a random variable, obtained prior to
any data collection, which consists of a probability
distribution.
t

Posterior distribution: A distributional
estimate of a random variable that updates
information from a prior distribution with new
information from data using Bayes rule.
t

Bayesian induction: A process for inductive
inference in which the P-value is replaced with
the posterior probability that the null hypothesis
is true. In Bayesian induction, the respective
prior distributions and data models (likelihoods)
constitute the null and alternative hypotheses. In
addition, one must specify the prior probability (or
odds) that the null hypothesis is true.
Two centuries later, Harold Jeffreys, a British
astronomer, greatly extended the utility of Bayes
rule (19). In terms of hypothesis testing, it may be
summarized as follows:
Probability that the hypothesis is true, given
observed data
= K*(Probability of the observed data if the
hypothesis is true)
x(Prior probability that the hypothesis is true),
[Equation 3]
and by reference to Equation 1 we see that
Probability that the hypothesis is true, given
observed data
= K*(Likelihood)x(Prior probability
that the hypothesis is true).
[Equation 4]
Thus we see that evidence for (or against) a given
hypothesis may be obtained directly from probability
theory, as long as we can supply the following:
t

5IF
MJLFMJIPPE

XIJDI
JT
BWBJMBCMF
GPS
NPTU
0.045 while that to the left of 96%LC (i.e., Ha range) is
practical problems. It is the same likelihood
required for Fisherian and Neyman-Pearson
approaches. However, in the Bayesian approach we
must also have the following:
t

5IF
WBMVF
,
JO
&RVBUJPO


XIJDI
TPNFUJNFT

requires computing technology and numerical
methods that have only recently become available.
In many common cases, however, such as those
we consider here, K can be easily evaluated in
Excel.
t

5IF
QSJPS
QSPCBCJMJUZ
PG
UIF
IZQPUIFTJT
5IJT
QSJPS

probability should be based on existing theory and
expert knowledge. It can be problematic because
experts will differ in their prior beliefs. On the other
hand, this Bayesian approach provides a quantitative
David LeBlond
tool to gauge the effects of different prior opinions
on a final conclusion. If prior knowledge is lacking,
it is logical to assign equal probability to each of the
hypotheses under consideration (e.g., 0.5 to H0 and
0.5 to Ha).
For our tablet-manufacturing example, it is
straightforward to apply a Bayesian approach. We have
shown previously how to obtain the prior and posterior
distributions of a normal mean (see reference 3, Table
IV and reference 20, pp. 78-80). As illustrated in Figure
5, the prior (or posterior) probabilities of H0 and Ha
are simply the areas under the prior (or posterior)
distribution of the mean over the respective ranges of the
mean (in the present example, below and above 96%LC
for Ha and H0, respectively).
Lets calculate the probability of truth of H0 and Ha
before and after the data are examined by the team. This
particular example is illustrated in Figure 6.
Before examining data: All knowledge about the
true process mean comes from the prior distribution. The
team used a noninformative prior for both the mean and
standard deviation. In Figure 6, the prior distribution for
the mean is essentially flat and indistinguishable from
the horizontal axis. This prior distribution was so broad
that about half the probability density (i.e., area under
the prior distribution of the mean) lies below 96%LC
and half above. Thus the prior probabilities of H0 and
Ha were each very close to 0.5. We can use short-hand to
state this:
PriorProbH0 = 0.5 and PriorProbHa = 0.5.
While the team actually felt that H0 was more likely,
they used this noninformative prior to provide a more
objective test.
After examining data: All knowledge about
the true process mean comes from the posterior
distribution. Notice in Figure 6 that the area under the
distribution to the right of 96%LC (i.e., H0 range) is
0.955, so that
PostProbH0 = 0.045 and PostProbHa = 0.955.
Thus the team can be 95.5% confident (in a true
probabilistic sense) that the null hypothesis, H0, is
false. It is also useful to consider such probabilities in
terms of odds as illustrated in Figure 7.
Odds is defined as follows:
t

Odds: The ratio of success to failure in probability
calculations. In the case of hypothesis testing
where only H0 or Ha are possible (but not both), if
the probability of truth of H0 is ProbH0, then the
odds of H0 equals ProbH0/(1-ProbH0).
PROCESS VALIDATION Process Design 93
David LeBlond

Figure 5: Bayesian induction inference model. that H0 is true to its prior odds. A B value of
1/10 means that Ha is supported by the data 10
times as much as H0. Because the Bayes factor
Prior or Ha H0 is normalized by the prior odds, it is measure of
Probability density

posterior
evidence that primarily reflects the observed data.
distribution

The Bayes factor is a measure of evidence supplied

by data in a hypothesis testing situation. Like the
Probability that Probability that P-value, various decision levels have been proposed
Ha is true H0 is true
(see reference 19, page 432).
Notice something important here: The Bayesian
Parameter
PostProbH0 and the Fisherian P-value for our
manufacturing example are both equal to 0.045.
Figure 6: Visualizing a one-sided hypothesis
This seems remarkable considering that these indices
test for a normal mean using its posterior dis-
are not measuring the same thing. The P-value is the
tribution.
probability of observing data at least as extreme as
was observed, while the PostProbH0 is the probability
0.3 that H0 is true. However, it can be shown that in many
Prior Ha H0
common situations (e.g., one-sided hypothesis tests
Probability density

Posterior
0.2 involving normally distributed data) the P-value will
Probability that Probability that
be equal to the PostProbH0 when an appropriate non-
Ha is true = 0.955 H0 is true = 0.045 informative prior is used to calculate PostProbH0 (see
0.1
examples in references 20 and 22).

0 PROBLEMS WITH THE TWO-SIDED

86 88
Before data are examined, the prior odds of H0 are
defined as
Prior Odds of H0 = PriorProbH0/PriorProbHa =
0.5/0.5 = 1/1.
So the prior odds that H0 is true are 1 to 1. After
information from data has been incorporated, the
posterior odds of H0 are
Posterior Odds of H0 = PostProbH0/PostProbHa =
0.045/0.955 = 45/955.
So the posterior odds that H0 is true are 45 to 955.
This reduction in the odds of H0 from 1/1 to 45/955
is due to the evidence about H0 contributed by the
data. As shown in Figure 7, we can form a measure of
evidence by taking the odds ratio as follows:
B = (Posterior Odds of H0)/(Prior Odds of H0) =
(45/955)/(1/1) = 45/955 = 0.0452
90 92 94 96 98 100 HYPOTHESIS TEST FOR EQUALITY
Mean In Example 1, the hypothesis tested by our team is
one sided because the ranges for the mean for Ha and
H0 were each completely on one side or the other of
the hypothesized value, 96%LC. A one-sided test is
defined as follows:
t

One-sided test: A null hypothesis stated in such
a way that observed values of the decision statistic
on one side (either large or small but not both)
constitutes evidence against it.
Our example uses composite hypotheses because
Ha and H0 both consist of ranges rather than single
points. It is also common to consider a two-sided
situation in which the null hypothesis, H0, consists of
a single point. Say for instance our team obtained the
following data from their testing of n=10 batches:
Mean of 10 measured batch potencies = 95%LC, and
Sample standard deviation = 5%LC,
They might have considered testing a point-null
hypothesis such as
H0: true mean = 100%LC
against an alternative hypothesis such as

B is referred to as the Bayes factor (21,22). Bayes Ha: true mean is not = 100%LC.
factor is defined as follows:
t

Bayes factor (B): In Bayesian induction, the A Fisherian test of this point-null H0 is easily
Bayes factor is the ratio of the posterior odds executed in Excel as follows:

94 PROCESS VALIDATION Process Design


t-value = SQRT(10)*ABS(95-100)/5 = 3.16
P-value = TDIST(3.16,10-1,2) = 0.012,
that would lead to rejection of H0 if the common
line of 0.05 is employed.
Testing of point-null hypotheses such as this is
very common. Unfortunately, as shown below, this is
almost never a realistic or meaningful test.
Most would agree that there is little practical
difference between a process mean potency of
99.999%LC and 100.001%LC. Yet it can be shown that
if the sample size is large enough, H0 will be rejected
with high probability, even if the true deviation from
100.000%LC is only 0.001%LC. This sensitivity to
sample size is well known to statisticians because
decision statistics, such as the mean, become very
precise (small standard errors)but not necessarily
more accuratewhen sample size is increased. An
essentially correct hypothesis can be rejected when
the summary statistics are too precise. This type of
counter-intuitive behavior in hypothesis testing is
often due to an incorrect statement of the problem
known as a Type III error (see Figure 4 footnote). Type
III error is defined as follows:
t

Type III error: A decision error that results
in choosing the incorrect null or alternative
hypothesis for use in a hypothesis test.
Rather than consider a point H0, it may be more
appropriate to specify a small interval for H0. Type III
errors are common in hypothesis tests for normality. In
very large samples, normality is almost always rejected
despite the fact that a histogram agrees visually with
the fitted normal curve. Sometimes the rejection is
caused by minor imperfections in the data, such as
rounding, that are not material to the objectives of the
hypothesis test.
One advantage of the Bayesian approach is that
it forces one to think carefully about the correct
formulation of a hypothesis testing problem. From
the Bayesian hypothesis testing perspective, the prior
distributions for H0 and Ha are the hypotheses being
tested. An appropriate Bayesian approach to test this
point-null H0 is given by Schervish (23, example
4.22 pp 224-5). Under the Bayesian paradigm, we
require prior distributions for the mean and standard
deviation for both H0 and Ha. This is because, in
general, neither the true mean or standard deviation
parameters are known. Instead of a single parameter
we have two parameters to consider and these will
have joint prior and posterior probability distributions.
Joint probability distribution is defined as follows:
t

Joint probability distribution: A probability
distribution in which the probability density or
David LeBlond
Figure 7: The Bayes factor (B) as a measure of
evidence for H0.
Probability density
Prior
Ha H0
ProbHa ProbHO
+ Data
Probability density
Posterior
ProbHa ProbH0
ProbH0/ProbHa
Mean B=
ProbH0/ProbHa
mass depends on the values of two (bivariate) or
more (multivariate) parameters simultaneously. A
bivariate probability distribution can be visualized
as a surface mesh or contour plot.
The H0 prior for mean and standard deviation is
shown in Figure 8. The prior for the mean (top panel)
is a single spike at 100%LC, consistent with the point
H0. The prior distribution for the standard deviation
is a mildly informative IRG distribution with C=1 and
G=50 (3, Table IV) and is shown on the lower panel of
Figure 8.
The Ha prior for mean and standard deviation is
shown in Figure 9. Because under Ha, the mean is
permitted to vary, the joint distribution of mean and
variance is displayed as a surface mesh plot. This is a
joint LSSt-IRG distribution with R=1, T=100, U=7.07,
C=0.5, and G=50 (3, Table IV).
This same Ha prior is displayed in Figure 10 as
a contour plot. From this it is easier to see the two-
dimensional shape and range that is identified as the
alternate hypothesis.
Unlike the one-sided case, the point-null situation
requires that we also specify our prior beliefs about
the truth of H0 and Ha. If the team had no prior
preference for either H0 or Ha, they would assign equal
prior probabilities to each. We can express that using
shorthand notation as
PriorProbH0 = 0.5 and PriorProbHa = 0.5.
The calculation of the Bayes factor for this test can
easily be done in Excel (23, equation 4.23) and a value
of
B = 0.3495
PROCESS VALIDATION Process Design 95
David LeBlond

Figure 8: Point-null hypothesis visualized as the P-value groups the actual observation obtained
probability distributions for the mean and with much more extreme observations that were not
standard deviation. actually obtained. In the case of point-null hypothesis
testing, Bayesian and Fisherian conclusions rarely
Prior null hypothesis for mean agree (24, see pp 151, table 4.2). When two sound
methodologies lead to different conclusions we must
1.5
Probability mass

wonder whether we have misidentified the problem

1 (i.e., our friend, the Type III error again?). It is best to
0.5 consider carefully what we mean by not equal. We do
0
this in the following section.
When applicable, the Bayesian approach to
80 90 100 110 120
hypothesis testing gives answers in terms of
Mean probability statements about the parameters and the
hypotheses themselves, which are impossible with the
Prior null hypothesis for standard deviation Fisherian and Neyman-Pearson approaches. This is
very advantageous for risk analysis. Bayesian inductive
Probability density

1.5
inference is also useful in data-mining applications. As
1
an example, the US Food and Drug Administrations
0.5
Center for Drug Evaluation and Research now employs
0
a Bayesian screening algorithm as part of their internal
2 7 12 17 22
drug safety surveillance program (25). However, the
Standard deviation Bayesian approach can be more demanding for the
following reasons:
t

8IJMF
NBOZ
#BZFTJBO
QSPCMFNT
DBO
CF
TPMWFE
JO

Figure 9: Alternative hypothesis visualized as Excel, more complicated situations may require
a surface mesh plot of the joint probability for advanced computing packages such as WinBugs
the mean and standard deviation. (26). Calculation of the K in Equation 4 or of the
Bayes factor can sometimes be challenging (27).
t

#BZFTJBO
BQQSPBDIFT
SFRVJSF
TQFDJGJDBUJPO
PG

prior distributions and prior probabilities of the
0.006 0.005-0.006 hypotheses under study. While noninformative
0.004-0.005
0.005
priors may be used for objectivity, this may result
0.003-0.004
0.002-0.003 in a loss of information as well as lost opportunity
0.004 0.001-0.002 to debate and thereby take advantage of the prior
Probability
density 0.003
0-0.001
knowledge of different members of a project team.
It is always critical to understand any effect that
0.002
the prior may have on conclusions.
0.001 18
t

8IFO
VTFE
GPS
DPOGJSNBUPSZ
PS
EFNPOTUSBUJPO

13 studies such as clinical trials, validations, quality
0
80 88
7 control, or data mining, Bayesian hypothesis
96
104 2 Standard testing methods must be calibrated (usually by
112
120 deviation
Mean computer simulation) so that the Type I and II
error risks are known.

is obtained. Given this information, we can invert A good, readable, basic introductory textbook for
the equation for B given in Figure 7 to solve for the readers interested in the theory and methodology of
posterior probability of H0. Noting that PostProbHa = Bayesian hypothesis testing is Bolstad (28).
1 PostProbH0, we find that
THE TOST HYPOTHESIS TEST FOR
PostProbH0 =0.259. EQUIVALENCY
Sometimes, as with method transfers or process
So the team would conclude that there is a 25.9% changes in existing products or validation of new
probability that H0 is true and likely would not reject products, the objective may be to establish evidence
it. This is a troubling result because the Fisherian for equivalency, rather than equality, with the usual
point-null test above yielded a P-value of 0.012, clearly burden of proof reversed. It is best to define a range
suggesting rejection of H0. However, as noted above, for the mean difference, say L to H, indicative of

96 PROCESS VALIDATION Process Design

 David LeBlond

Figure 10: Alternative hypothesis visualized as a contour

equivalence. Then the null hypothesis is
plot of the prior joint probability distribution for the mean
and standard deviation.
H0: true mean difference < L or true mean
difference > H 22
0 . 0 0 5 -0 . 0 0 6
against the alternative hypothesis 0 . 0 0 4 -0 . 0 0 5
0 . 0 0 3 -0 . 0 0 4 18
Ha: L < true mean difference < H. 0 . 0 0 2 -0 . 0 0 3
0 . 0 0 1 -0 . 0 0 2
In this way we treat nonequivalence as the default 0 -0 . 0 0 1 14
state of nature and require the data to provide Standard
evidence that allows us to reject H0. A two one-sided deviation
testing (TOST) procedure (29) is based on requiring 10
a confidence or credible interval to be completely
contained within the range of Ha in order to reject H0.
TOST is defined as followed: 6
t

Two one-sided hypothesis test (TOST): A
hypothesis test for equivalency that consists of
2
two one-sided tests conducted at the high and
low range of equivalence, each of which must be 80 88 96 104 112 120
rejected at the Type I error rate in order to reject the Mean
null hypothesis of non-equivalence.
Figure 11: Visualizing the ANOVA null hypothesis (H0)
EXAMPLE 2: A TEST THAT THE MEANS OF
relative to the posterior distribution of the group deltas.
THREE GROUPS ARE EQUAL
The ANOVA procedure developed by Fisher is a
Ratio 0 .2
procedure for testing the null hypothesis of equality for
30-32
group means. Fishers ANOVA calculation procedure is 28-30 0 .1
illustrated elsewhere in this issue with an example (30) 26-28
delta_A
in which there are three groups (A, B, and C). We will 24-26 0

illustrate the Bayesian approach to ANOVA here. 22-24

20-22 -0 . 1
Zelen (31, pp 306-9) shows how to perform multiple
18-20 H0
regressions from a Bayesian perspective. His procedure 16-18
-0 . 2
can easily be adapted to the simple ANOVA case and a 14-16 Posterior
-0 . 3 mode
Bayes factor for the ANOVA can be calculated in Excel. 12-14
However, as with the point-null hypothesis, we will 10-12 -0 . 4
8-10
find that the conventional P-value tends to overstate
6-8 -0 . 5
the case for rejection of H0. Again, it is wise to ask if 4-6 0.1 0.2 0.3 0.4 0.5
-0.5 -0.4 -0.3 -0.2 -0.1 0
the ANOVA hypothesis test correctly frames the real 2-4
questions we want to ask. We will assume here that the 0-2 delta_B
null hypothesis of equality

H0: meanA = meanB = meanC,

is appropriate. The corresponding alternative
hypothesis is
Ha: one or more of the three means are not equal.
An ANOVA F-test is often used when the real
objective is to make comparisons among the group
means. When an appropriate noninformative prior
(3, 4) is used, the Bayesian approach to group means
comparison agrees exactly with the standard ANOVA
F-test (32, pp 123-43). But the Bayesian approach offers
a useful advantage.
The usual Fisherian paradigm regards the true
group means as fixed entities. So one can only make
comparisons with respect to the measured statistics,
using such things as confidence intervals. However,
the Bayesian paradigm regards the true group means
as having a joint posterior distribution. It is possible
to calculate and plot the joint posterior distribution
using Excel and to visualize the single point within this
distribution that corresponds to all the means being
equal (i.e., the H0). Such a graph for the example (30)
is shown in Figure 11, but it requires some explanation.
First, the axes for the plot are not the true group
means, but the deviations (deltas) of these true means

PROCESS VALIDATION Process Design 97

David LeBlond
from their true grand mean, M, where
M = (mean_A + mean_B + mean_C)/3,
or
mean_C - M = - (mean_A - M) - (mean_B - M),
and denoting the differences of the true group
means from their true grand mean with a delta, we
have
delta_C = -delta_A - delta_B.
Because delta_C can always be obtained knowing
delta_A and delta_B, it is not an independent random
variable. Consequently, we only need to concern our
selves with two of the three deltas, say delta_A and
delta_B. With three groups, we can actually visualize
the joint distribution as a two-dimensional contour
plot.
Second, the point in Figure 11 corresponding to H0
(true means all equal) is the point delta_A = delta_B
= 0, because only under this condition can all three
group means be equal.
Third, the contour lines form ellipses about the
joint distribution mode (the point delta_A = -0.22
and delta_B = -0.26 indicated as a red dot in Figure
11). (Recall the mode is the point of a probability
distribution having the maximum probability density).
They do not correspond to a probability density as
would be true for an actual distribution such as that
in Figure 10. Instead the contour lines are critical F
values. In Excel, the probability that the true value
of delta_A, delta_B is beyond a given contour ellipse
corresponding to F is FDIST(F,3-1,15-3). By analogy
with a univariate distribution, the further any point is
from the distribution mode, the larger the F value and
the less likely that point is as a candidate for the true
delta_A, delta_B.
Finally the contour line in Figure 11 that equals
the critical F value of 3.885, obtained in Excel as
FINV(0.05,3-1,15-3) (30), forms a 95% credible ellipse.
This is a bivariate analogue of a univariate 95% credible
interval (3). Notice in Figure 11 that H0 corresponds
to an F value of 6.142 and is, therefore, beyond the
95% credible ellipse. As with the traditional ANOVA
calculation, we reject H0.
By restating the hypotheses in terms of joint
posterior distributions that can be displayed visually,
the Bayesian perspective offers deeper insight into the
mechanics and interpretation of ANOVA.
SUMMARY
We have seen that scientific inference requires both
deductive and inductive inferences. The three main
98 PROCESS VALIDATION Process Design
approaches to inductive inference are: Fisherian, which
uses the P-value as a measure of evidence against the
null hypothesis; Neyman-Pearson, which controls the
long-run decision risk over repeated experiments; and,
Bayesian, which obtains direct probabilistic measures
of evidence from the posterior distribution.
Understanding of the P-value as the probability of
observing a result as extreme or more extreme than
that observed, assuming the null hypothesis is true,
is critical to its proper interpretation. The P-value
is based on unobserved results more extreme than
those observed, so it may overstate the evidence
against the null hypothesis. The P-value from a point-
null hypothesis, or two-sided test of equality, is very
difficult to interpret. A confidence or credible interval
should be provided in addition to the P-value.
The Bayes factor, like the P-value, is a measure of
evidence in favor of the null hypothesis contained
in the data. It is based on an odds ratio and permits
direct probability statements about the hypothesis
tests being considered.
The Type II error rate is the probability of incorrectly
failing to reject the null hypothesis when in fact the
alternative hypothesis is true. The Type II error rate
is related to Power and allows the construction of
operating characteristic curves.
The analysis of variance is a Fisherian point-null
hypothesis test for the equality of means of two
or more groups. The Bayesian approach offers an
insightful reinterpretation of the analysis of variance
in terms of the joint posterior distribution of the group
means.
The three approaches to hypothesis testing represent
major advances in the way we use data to make
inductive inferences about the products, processes,
and test methods we develop in regulated industries.
These approaches, and others not discussed here, have
done much to improve our decision-making. They
help us build evidence about underlying mechanisms
and provide us a common language for objective
communication of the evidence supporting our
conclusions.
Still, humility and caution are in order. Despite
these impressive methodologies, no single, coherent,
generally-accepted approach to inductive inference
has yet emerged in our time. All experimenters
know that Nature does not divulge her secrets easily.
For the present, it is perhaps unwise to apply any
induction methodology by rote without first carefully
considering whether it is consistent with our objective
and problem.
ACKNOWLEDGMENT
The text before you would be poorer if not for the help
of others. I am most sincerely grateful to Paul Pluta for
ideas, encouragement, and expert feedback; to Diane

Wolden who tirelessly kept the reader in mind; and to
Susan Haigney for expertly laying the product to print.
GLOSSARY
Alternative hypothesis: A hypothesis considered as
an alternative to the null hypothesis, though
possibly more complicated or less likely.
ANOVA (analysis of variance): A hypothesis test that
uses the F-statistic described by Fisher used to
detect differences among the true means of
data from two or more groups.
Bayes factor (B): In Bayesian induction, the Bayes factor is
the ratio of the posterior odds that H0 is true to
its prior odds. A B value of 1/10 means that the
Ha is supported 10 times as much as H0. Since
the Bayes factor is normalized by the prior odds,
it is measure of evidence that primarily reflects
the observed data.
Bayes rule: A process for combining the information
in a data set with relevant prior information
(theory, past data, expert opinion, and
knowledge) to obtain posterior information.
Prior and posterior information are expressed
in the form of prior and posterior probability
distributions, respectively, of the underlying
physical parameters, or of predictive posterior
distributions of future data.
Bayesian induction: A process for inductive inference in
which the P-value is replaced with the posterior
probability that the null hypothesis is true.
In Bayesian induction, the respective prior
distributions and data models (likelihoods)
constitute the null and alternative hypotheses.
In addition, one must specify the prior
probability (or odds) that the null hypothesis is
true.
Calibrated hypothesis test: A hypothesis test method
whos Type I error, on repeated use, are known
from theory or computer simulation.
Composite hypothesis: A statement that gives a range
of possible values to a model parameter. For
example, Ha: true mean > 0 is a composite
hypothesis.
Confidence interval: A random interval estimate of a
(conceptually) fixed quantity, which is obtained
by an estimation method calibrated such that
the interval contains the fixed quantity with a
certain probability (the confidence level).
Control chart: A time ordered plot of observed data
values or statistics that is used as part of a
process control program. Various hypothesis
tests are employed with control charts to detect
the presence of trends or unusual values.
Credible interval: An interval estimate of a random
David LeBlond
variable, based on its probability distribution,
which contains its value with a certain
probability (the credible probability level).
Data: Measured random variable values, assumed to
be generated by some hypothetical likelihood
model, which contain information about the
parameters of that model.
Deduction: The act of drawing a conclusion about some
hypothesis based entirely on careful definitions,
axioms, and logical reasoning.
F-statistic: The decision statistic used in Fishers analysis
of variance hypothesis test consisting of the
ratio of two independent observed variances
calculated from normally distributed data.
Fisherian induction: A process for inductive inference,
described most clearly by Ronald Fisher that
uses the P-value as a criterion for rejecting a
hypothesis.
Hypothesis: A provisional statement about the value of
a model parameter or parameters whose truth
can be tested by experiment.
Induction: The act of drawing a conclusion about some
hypothesis based primarily on limited data.
Inference: The act of drawing a conclusion regarding
some hypothesis based on facts or data.
Joint probability distribution: A probability distribution
in which the probability density or mass
depends on the values of two (bivariate) or
more (multivariate) parameters simultaneously.
A bivariate probability distribution can be
visualized as a surface mesh or contour plot.
Likelihood model: A description of a data generating
process that includes parameters (and possibly
other variables) whose values determine the
distribution of data produced.
Measures of evidence: In scientific studies, hypothesis
testing is used build evidence for or against
various hypotheses. The P-value and Bayes
factor are examples of measures of evidence in
Fisherian and Bayesian induction, respectively.
Mode: A point estimate of a random variable that is
the value at which its probability density is
maximized.
Multiplicity: When multiple hypothesis tests (e.g., control
chart rules) are applied to different aspects (e.g.,
trending patterns) of a data set, the overall false
alarm rate of any one test failing may be greater
than that of any single test when applied alone.
This statistical phenomenon is referred to as
multiplicity.
Neyman-Pearson induction: A methodology for
inductive inference, developed by Jerzy
Neyman and Egon Pearson, that considers
both a null and an alternative hypothesis.
PROCESS VALIDATION Process Design 99
David LeBlond
The null hypothesis is rejected in favor of the
alternative hypothesis if the observed value of
some statistic lies in its rejection region. The
statistic and the associated rejection region are
identified from statistical theory and are chosen
to provide desired Type I or II decision error rates
over repeated applications of the methodology.
Null hypothesis (H0): A plausible hypothesis that is
presumed sufficient to explain a set of data
unless statistical evidence in the form of a
hypothesis test indicates otherwise.
Ockhams Razor: The doctrine of parsimony that
advocates provisionally adopting the simplest
possible explanation for observed data.
Odds: The ratio of success to failure in probability
calculations. In the case of hypothesis testing
where only H0 or Ha are possible (but not both),
if the probability of truth of H0 is ProbH0, then
the odds of H0 equals ProbH0/(1-ProbH0).
One-sided test: A null hypothesis stated in such a way
that observed values of the decision statistic
on one side (either large or small but not both)
constitutes evidence against it.
P-value: The probability of obtaining a result at least as
extreme as the one that was actually observed,
given that the null hypothesis is true. The fact
that p-values are based on this assumption is
crucial to their correct interpretation.
Parameter: In statistics, a parameter is a quantity of
interest whose true value is to be estimated.
Generally a parameter is some underlying
variable associated with a physical, chemical, or
statistical model.
Point (simple) hypothesis: A statement that a model
parameter is equal to a single specific value.
For example, H0: true mean = 0 is a simple
hypothesis.
Posterior distribution: A distributional estimate of a
random variable that updates information from
a prior distribution with new information from
data using Bayes rule.
Power (or operating characteristic) curve: Power is
equal to 1 minus the Type II error rate. The
power curve of a hypothesis test is a plot of the
Power versus the true value of the underlying
parameter of interest.
Prior distribution: A subjective distributional estimate
of a random variable, obtained prior to any
data collection, which consists of a probability
distribution.
Probability density contour plot: A rendering of a
bivariate distribution in which the distribution
appears in the bivariate parameter space as
contours of equal probability density.
100 PROCESS VALIDATION Process Design
Probability density surface mesh plot: A three-
dimensional analogue of a two-dimensional
probability density plot in which there are two,
rather than only one, model parameters. The
bivariate distribution therefore appears as a
surface rather than as a curve.
Sampling distribution: The distribution of data or some
summary statistic calculated from data.
Special cause: When the cause for variation in data, or
statistics derived from data, can be identified
and controlled, it is referred to as a special
cause. When the cause cannot be identified, it is
regarded as random noise and referred to as a
common cause.
Statistic: A summary value (such as the mean or standard
deviation) that is calculated from data. A
statistic is often used because it provides a good
estimate of a parameter of interest.
t-statistic: The decision statistic used in Students t-test
consisting of the ratio of a difference between
an observed and hypothesized mean divided by
its estimated standard error.
Two-sided test: A null hypothesis stated in such a way
that either large or small observed values of the
decision statistic constitute evidence against it.
Two one-sided hypothesis test (TOST): A hypothesis test
for equivalency that consists of two one-sided
tests conducted at the high and low range of
equivalence, each of which must be rejected at
the Type I error rate in order to reject the null
hypothesis of non-equivalence.
Type I error: A decision error that results in falsely
rejecting the null hypothesis when in fact it is
true. It is sometimes referred to as the alpha-risk
or manufacturers risk.
Type II error: A decision error that results in failing to
reject the null hypothesis when in fact it is false.
It is sometimes referred to as the beta-risk or
consumers risks.
Type III error: A decision error that results in choosing
the incorrect null or alternative hypothesis for
use in a hypothesis test.
REFERENCES
1. LeBlond, D, Data, Variation, Uncertainty, and Probability
Distributions, Journal of GXP Compliance, Vol. 12, No. 3, pp
3041, 2008.
2. LeBlond, D, Using Probability Distributions to Make
Decisions, Journal of Validation Technology, Spring 2008,
pp 214, 2008.
3. LeBlond, D, Estimation: Knowledge Building with
Probability Distributions, Journal of GXP Compliance, Vol.
12 (4), 42-59, 2008. See Journal of Validation Technology, Vol.
14, No. 5, 2008 for correction to Table IV.

4. LeBlond, D, Estimation: Knowledge Building with
Probability DistributionsReader Q&A, Journal of Validation
Technology, Vol. 14(5), 50-64, 2008.
5. Marden, J, Hypothesis Testing: From p Values to Bayes
Factors, Journal of the American Statistical Association, 95(452)
1316-1320, 2000.
6. Stigler, S, The History of Statistics. The measurement of
uncertainty before 1900, Belknap Press, Cambridge, 1986.
7. Daston, L, Classical Probability in the Enlightenment, Princeton
University Press, Princeton, 1988.
8. William of Ockham, of the 14th century, advocated
adopting the simplest possible explanation for physical
phenomena, see Jeffreys, H (1961) Theory of Probability, 3rd
edition, Oxford University Press, NY, page 342.
9. Pearson, K, On the criterion that a given system of
deviations from the probable in the case of correlated
system of variables is such that it can be reasonably
supposed to have arisen from random sampling.
Philosophical Magazine 50, 157175, 1900.
10. Gosset, W, aka Student, The probable error of a mean,
Biometrika VI (1), 1-25, 1908.
11. Fisher, R, Statistical Methods for Research Workers, Oliver &
Boyd, Edinburgh, 1925.
12. Snedecor, G and Cochran, W, Statistical Methods, 6th
edition, Iowa State University Press, Ames, page 98, 1967.
13. Goodman, S, Toward Evidence-based Medical Statistics.
1. The P value fallacy, Annals of Internal Medicine, 130, 995-
1004, 1999.
14. Gibbons, J and Pratt, J, P-values: Interpretation and
Methodology, The American Statistician, 29(1) 20-25, 1975.
15. Neymen, J and Pearson E, On the Problem of the Most
Efficient Tests of Statistical Hypotheses, Philosophical
Transactions of the Royal Society, series A, volume 231, 289-
337, 1933.
16. International Conference on Harmonisation, ICH
Harmonised Tripartite Guideline on Pharmaceutical Development
Q8, Current Step 4 version, dated 10 November 2005.
17. Dale, A , A History of Inverse Probability, Springer, New York,
page 545, note 38, 1999.
Originally published in the Winter 2009 issue of The Journal of Validation Technology
David LeBlond
18. Price, R, (1763) An essay towards solving a problem
in the doctrine of chances, in Dale, A Most Honourable
Remembrance: The Life and Work of Thomas Bayes, Springer,
New York, 2003.
19. Jeffreys, H, Theory of Probability, 3rd edition, Oxford
University Press, Cambridge, 1961.
20. Gelman, A, Carlin, J, Stern, H, and Rubin, D, Bayesian Data
Analysis, 2nd Edition, Chapman and Hall/ CRC, New York,
2004.
21. Goodman, S (1999) Toward evidence-based medical
statistics. 2. The Bayes factor, Annals of Internal Medicine,
130, 1005-1013.
22. Casella, G and Berger, R, Reconciling Bayesian and
Freqentist Evidence in the One-Sided Testing Problem,
Journal of the American Statistical Association 82(397) 106-
111, 1987.
23. Schervish, M, Theory of Statistics, Springer-Verlag, New York,
1995.
24. Berger, J, Statistical Decision Theory and Bayesian Analysis, 2nd
edition, Springer-Verlag, New York, 1985.
25. Lincoln Technologies, WebVDME in Production at
the FDA, WebDVMA News, volume 2(2), page 1, 2005.
Available at HTTP://WWW.LINCOLNTECHNOLOGIES.
COM
26. Cowles, M, Review of WinBUGS 1.4, American Statistician
58(4), 330-336, 2004.
27. Kass, R and Raftery, A, Bayes Factors, Journal of the
American Statistical Association, 90(430) 773-795, 1995.
28. Bolstad, W, Introduction to Bayesian Statistics 2nd Edition, John
Wiley & Sons, Hoboken, New Jersey, 2007.
29. Schuirmann, D, On Hypothesis Testing to Determine
if the Mean of a Normal Distribution is Contained in a
Known Interval, Biometrics 37 617, 1981.
30. Vijayvargiya, A., One Way Analysis of Variance, Journal of
Validation Technology, Vol. 15, No. 1, 2009.
31. Zellner, A, An Introduction to Bayesian Inference in
Econometrics, John Wiley, New York, 1971.
32. Box, G and Tiao, G, Bayesian Inference in Statistical Analysis,
Addison-Wesley Pub. Co., Reading, MA, 1973. JVT
PROCESS VALIDATION Process Design 101
Paul L. Pluta

PQ=Confirmation
Paul L. Pluta

PQ Forum provides a forum for validation
practitioners to share information about stage 2
process qualification (PQ) in validation lifecycle.
Information about supporting activities such as
equipment and analytical validation is shared. The
information provided is intended to be helpful and
practical so as to enable application in actual work
situations. Our objective: Useful information.
Comments from readers are needed to help us
fulfill our objective for this column. Suggestions for
future discussion topics are invited. Please contact
column coordinator Paul Pluta at paul.pluta@
comcast.net or journal coordinating editor Susan
Haigney at shaigney@advanstar.com with comments,
suggestions, or topics for discussion.
t
The site validation policy should clearly state
the site strategy and approach to validation and
qualification. Appropriate personnel should be
trained on this policy.
t
If there is not complete confidence in the future
success of the validation PQ, protocols must
not be approved and validation should not be
initiated.
t
The site validation approval committee (VAC)
should have significant input into development
of the site policy and must then uphold the stated
validation standards and expectations.
t
The VAC should function as a surrogate
regulatory auditor when reviewing and approving
validation documents.

KEY POINTS INTRODUCTION

The following key points are discussed:
t
Validation performance qualification is expected
to be confirmatory (i.e., the validation is
expected to confirm the design, development,
and other support work associated with the
respective validation).
t
Personnel involved in validation may not know
or understand current validation expectations.
t
Validation documentation may also be
substandard due to incomplete development
work.
t
A strategy to change erroneous understanding
of validation and define current validation
expectations is proposed.
Validation has evolved over the years; what was
common and accepted practice in validation
performance qualification (PQ) years ago is not
acceptable today. The November 2008 FDA process
validation draft guidance (1) clearly states the current
expectation for PQ as follows: Stage 2Process
Qualification: During this stage, the process design
is confirmed as being capable of reproducible
commercial manufacture.
The draft guidance also states the following:
During the process qualification stage of process
validation, the process design is confirmed as being
capable of reproducible commercial manufacturing.
This stage has two elements: [1] design of the facility
and qualification of the equipment and utilities, and
[2] performance qualification (PQ). During this
stage, CGMP-compliant procedures must be followed

[
ABOUT THE AUTHOR
Paul L. Pluta, Ph.D., is a pharmaceutical scientist with extensive industrial development, manu-
For more Author facturing, and management experience. Dr. Pluta is also Adjunct Associate Professor at the Univer-
information, sity of Illinois, Chicago College of Pharmacy. He is also editor-in-chief of the Journal of Validation
go to Technology and the Journal of GXP Compliance. Dr. Pluta has written several chapters and edited
gxpandjvt.com/bios Cleaning and Cleaning Validation, Volume 1, Basics, Expectations, and Principles, published by PDA
and DHI Publishing. He may be contacted by e-mail at paul.pluta@comcast.net.

102 PROCESS VALIDATION Process Design


and successful completion of this stage is necessary
before commercial distribution.
A successful PQ will confirm the process design
and demonstrate that the commercial manufacturing
process performs as expected.
The level of monitoring and testing should
be sufficient to confirm uniform product quality
throughout the batch during processing.
After establishing and confirming the process,
manufacturers must maintain the process in a state of
control over the life of the process, even as materials,
equipment, production environment, personnel, and
manufacturing procedures change.
The key word that is restated in these sections of
the guidance is confirm.
There is clear and repeated evidence that
the US Food and Drug Administration expects
manufacturers to thoroughly understand the process
and confirm its acceptability in the PQ. The PQ
is expected to confirm the design, development,
and other preliminary work associated with the
validation. The confirmatory PQ is a decision point
that heavily contributes to the judgment to transition
responsibility for the process from development to
manufacturing for routine use. The FDA guidance
(1) states, Success at this (PQ) stage signals an
important milestone in the product lifecycle and
needs to be completed before the manufacturer
commences commercial distribution of the drug
product.
A successfully completed PQ is a strong statement
that development of the subject operation has
been completed and is ready for manufacture of
commercial product. The most effective way to
demonstrate readiness for routine use is through
a successfully executed PQproblem-free and
mistake-freethat clearly shows that the design and
development objective has been achieved.
The FDA guidance was written specifically for
manufacturing processes. However, validation
thought leaders are now applying the validation
lifecycle approach to equipment, facilities, utilities,
and other applications (i.e., any item or system to be
validated). The PQ for each of these items should
confirm acceptable performance. When the PQ
for any of these items is successfully completed,
responsibility for the object of the PQ (i.e., processes,
equipment, computer systems, etc.) is transferred for
routine use.
Why Dont We Confirm?
There are several reasons why the validation PQ
is not always confirmatory. The primary reason
is that people involved in the work of validation
(i.e., research and development [R&D] scientists,
Paul L. Pluta
engineers, and other technical people) are not aware
of current regulatory expectations. Validation is
performed in different areas on different things
related to manufacturing (e.g., manufacturing
processes, cleaning processes, analytical methods,
equipment, HVAC systems, water systems, computer
systems, and so on). Each of these is performed
by people with different expertise, knowledge,
and experience. These people have different
levels of awareness of regulations and regulatory
expectations for validation. Other factors may also
influence the validation approach. Project timelines
may cause an accelerated initiation of validation.
Commercial demand requirements may also cause
validation to be prematurely initiated. New active
pharmaceutical ingredients (APIs) or other materials
may not be available to perform experimental
runs prior to validation. Costs may be prohibitive,
especially when used material will be destroyed and
not commercially distributed.
All personnel involved in validation must clearly
know that validation is intended to confirm that
the process, equipment, facilities, etc. are well
understood, under good control, and can be expected
to perform reliably and consistently throughout their
respective lifecycle.
Again, the key message: Validation is
confirmation.
This discussion addresses the following:
t
Obsolete attitudes and beliefs. Some specific
examples of misunderstanding validation,
obsolete attitudes, and beliefs about validation.
t
Validation equals confirmation. A successful
PQ confirms that the subject of the validation
can be expected to reliably perform.
t
Defining current validation expectations.
A strategy to change erroneous understanding of
validation and clearly define current expectations
is presented. Topics addressed include validation
policy, training, experimental work, management
commitment, and associated topics.
OBSOLETE ATTITUDES AND BELIEFS
Validation managers that attended the IVT Validation
Week Europe conference in Dublin, Ireland in
March 2010 discussed obstacles they face in their
jobs. The first subject to be raised by these managers
was the general lack of understanding of current
validation expectations by people at their respective
manufacturing sites. People do not understand
that validation should confirm process design and
development. They learned validation at different
times over the last 20 to 30 years and have not stayed
current with regulatory expectations and industry
practice.
PROCESS VALIDATION Process Design 103
Paul L. Pluta
Incomplete Preparation For Validation
Several managers at the conference complained that
R&D, engineering, or technical people never totally
complete a project before initiating validation. These
people often do not understand that validation
should confirm their technical work. Their following
statements describe their attitudes and beliefs:
t
Validation is the final step in the development
process
t
We may have to do final process optimization in
validation
t
We will determine the final operating
parameters in validation
t
The project should be completely finished before
we start validation, but if we see something to
improve during validation, we will do it during
the validation.
These statements demonstrate a fundamental
misunderstanding of validation expectations.
Development work must be completed before
validation is initiated. Process parameters must
be completely defined and clearly stated in
manufacturing process directions. Processes using
finalized defined parameters are then confirmed in
validation.
Documentation Problems
Several managers also commented about validation
documentation problems caused by incomplete
technical work. For example, statements included the
following:
t
If there are major process problems during
validation, we explain them in the protocol. If
the explanation is good, then the process is
validated.
t
It is OK to make mistakes in paperwork because
we will explain them in the write-up.
t
As long as you sign and date changes in the
protocol, QA will approve them.
Validation documents are fundamental documents
frequently requested by regulatory auditors.
Regulatory expectations for validation documents
are high. Documents should demonstrate that
design and development of process, equipment,
or the system being validated has been completely
finished. A successful validation then confirms that
the process, equipment, etc., are well understood,
appropriately controlled, and able to be used
to reliably manufacture commercial product.
Validation documents with process problems,
errors, corrections, and other mistakes suggest
the oppositeand the more problems seen in the
documentation, the worse the impression. When
104 PROCESS VALIDATION Process Design
one validation manager complained to engineering
management about the quality of validation
documentation, she was told that FDA expects to see
mistakes and failures in validation. Mistakes show
that we really performed the validation, and didnt
drylab (falsify) the validation.
Validation professionals must persist in changing
the obsolete attitudes described above and help
their colleagues to know current requirements and
expectations.
VALIDATION EQUALS CONFIRMATION
Validation performance is expected to be
confirmatory (i.e., process, equipment, etc.
requirements are expected to be confirmed in
validation). There must be a good understanding
of whatever system is being validated. Testing
and results should demonstrate acceptable system
performance. The PQ is expected to confirm the
design, development, and other preliminary work
associated with the respective validation. There must
be an expectation of successful validation based on
technical knowledge and experience obtained before
validation is initiated. If necessary, development
work, research studies, and even scale-up runs may
need to be conducted prior to initiating the validation
protocol.
When the validation protocol is written, we
must know what to test and how to do the testing.
Sampling must be clearly defined, appropriate for use,
and justified. Acceptance criteria must be rational
and justified. There should be complete confidence
that validation tests will pass acceptance criteria.
We must not be learning the performance of our
equipment, or determining the limits of process
capability in validation.
We must not be testing something for the first
time when we perform validation. If we need to
do something for the first time, we should do it in
development or in an engineering study, and then
confirm its acceptability in validation. We must
not add a new requirement or specification to the
validation that has not been previously evaluated.
Laboratories must not be using a new test method
that has not been previously validated. We must not
be trying software for the first time in validation.
To again summarize, validation is confirmation of
what is already known.
Validation Should Be Boring
Consultant Michael Anisfeld, in his program
Fundamentals and Essentials of Validation (2),
teaches that validation should be boringmeaning
that there should be no surprises in the execution
of a validation protocol, and that the execution of
the protocol must yield expected successful results.

Those responsible for validation should have no
anxiety regarding validation performance. They
should be simply confirming what they already
knowthat all testing in the validation will meet
acceptance criteria. Professionals responsible for
processes, equipment, or systems being validated as
well as the personnel conducting the validations must
approach their work with a confident expectation
of complete success. They should not hope the
validation will pass. They should be bored waiting These documents must be in compliance with the
for expected results.
DEFINING CURRENT VALIDATION
EXPECTATIONS
Validation managers at the Dublin meeting
discussed key elements of a strategy to change
erroneous or obsolete understanding of
validation and clearly define current validation
expectations. These included validation policy,
training, validation approval committee (VAC)
responsibilities, experimental trials ahead of actual
PQ runs, personal and organizational reputations,
and senior management support.
Validation Policy
The site validation policy should clearly state
the site strategy and approach to validation and
qualification. These expectations should require
that validation protocols contain requirements that
are based on scientific and technical work done in
advance of validation. Design and development
work must be completed before validation is
initiated. The validation should then confirm the
design and development work. No optimization,
fine-tuning, or other development work is allowed
to be done in validation. The FDA 2008 process
validation draft guidance (1) describes this
approach in the lifecycle approach to validation.
Validation Training
After the policy is written and approved, training
of appropriate associated personnel must be
conducted. These included manufacturing, quality
assurance, technical support, engineering, and
other groups. Personnel involved in support work
that is done preliminary to validation are key
participants in this training. They must clearly
understand that their work must be completed
before starting validation, and that validation must
confirm their work. This training will be especially
useful to personnel who had learned validation
many years ago but had not remained current with
new developments or regulatory expectations. In
addition to the aforementioned site personnel,
Paul L. Pluta
their respective management must be part of this
training and support compliance to the policy.
Validation Approval Committee
The site VAC has an important role in the site
validation program. The VAC should have significant
input into development of the site policy. They have
responsibility for approving validation protocols,
results packages, and other associated documents.
site policy. Standards upheld by the site VAC sends
a strong message to those who are responsible for
development work and writing protocols. Successful
validation that confirms design and development
work must be an expectation that is demanded by the
VAC.
Pre-Work, Experimental Studies, And
Engineering Studies
Personnel initiating validation must be completely
confident of the future success of their protocols.
When personnel submitting protocols are not highly
confident of future success, protocols must not be
approved and validation should not be initiated.
Additional development work may be required
to address unanswered questions. Laboratory
experimental studies, pilot-scale work, or even full-
scale trials may be warranted depending on the
information available or risk of failure.
Personal And Organizational Reputations
Site personnel, their respective functional
organizations, and site management must support
and be committed to confirmation in PQ validation
and its ramificationscomplete and technically
based protocols and impeccable documentation.
Without commitments from all the aforementioned
individuals and groups, the validation program will
not succeed. Good technically-based validation
documentation provides a strong statement about the
organization and its management.
Good validation documentation also conveys a
message about personnel in the organization. One
validation manager suggested that authors who had
written substandard validation documents be asked
to evaluate the document from a reader perspective
what impression would such substandard documents
convey to an auditor? What judgments would
be made about the authors competency? How
significantly would the authors reputation be
damaged? Substandard documents convey an
unwritten message.
Validation documentation also reflects on the
site validation approval committee. The integrity
and competency of the site VAC is also in question
PROCESS VALIDATION Process Design 105
Paul L. Pluta
when a substandard document is approved. The
VAC should serve as a surrogate regulatory auditor
when reviewing validation documents and must
require that substandard documents be corrected or
rewritten before being approved. Again, substandard
documents convey an unwritten message.
Senior Management Support
Senior management support of policies that affect
multiple areas in the organization is vital to any
cross-function policy. Topics discussed above may
also affect costs and timelines. However, failures
in validation are often even more costly, required
extensive time and effort for correction, and are
an embarrassment to the organization and all
personnel involved. Without site management
support, these problems will not be corrected.
CONCLUSIONS
Validation performance is expected to be
confirmatory. The PQ is expected to confirm the
design, development, and other preliminary work
associated with the respective validation.
Personnel may not understand current validation
expectations. R&D or technical people may not
understand that development work should be
completed before validation is initiated. This often
leads to substandard documentation with process
problems, errors, corrections, and other mistakes.
The validation PQ is expected to be confirmatory.
There must be a good understanding of whatever
system is being validated. There must be an
expectation of success in validation based on
technical knowledge and experience obtained before
validation is initiated. New techniques, sampling,
testing, and associated activities must have been
tried before initiating validation. New acceptance
criteria should not be tried without prior evaluation.
Originally published in the Autumn 2010 issue of The Journal of Validation Technology
106 PROCESS VALIDATION Process Design
New software should not be tried for the first time
in validation. Validation is confirmation of what
we already know. Validation should be boring,
meaning that there should be no surprises in the
execution of a validation protocol.
A strategy for changing erroneous or obsolete
understanding of validation and clearly defining
current validation expectations is proposed. This
includes a validation policy that clearly requires
validation to confirm design and development work.
No optimization, fine-tuning, or other development
work must be allowed in validation. If there is not
complete confidence in the future success of the
validation PQ, protocols must not be approved
and validation should not be initiated. Validation
training of appropriate personnel and their
management is recommended. The site validation
approval committee should function as a regulatory
auditor to uphold site validation standards.
Substandard validation documents convey a
negatively impression of the site, the document
author, and the VAC to regulatory auditors and other
reviewers.
In brief, validation PQ equals confirmation.
REFERENCES
1. FDA, Guidance for Industry, Process Validation: General
Principles and Practices, Draft Guidance, November 2008.
2. Anisfeld, Michael, Fundamentals and Essentials
of Validation, www.Globepharm.org, Deerfield, IL,
USA. JVT
ARTICLE ACRONYM LISTING
APIs Active Pharmaceutical Ingredients
FDA US Food and Drug Administration
PQ Process Validation
R&D Research and Development
VAC Validation Approval Committee