• Correlation Analysis
©The National Graduate School of Quality Management v7 • 2
Day 2 Agenda
• Team Report Outs on Day 1 Material
• Process Capability
• MultiVari Analysis
• Confidence Intervals
• Control Charts
• Hypothesis Testing
• Contingency Tables
• Design of Experiments
• Grading
Class Preparation 30%
Team Classroom Exercises30%
Team Presentations 40%
10 Minute Daily Presentation (Day 2 and 3) on Application of previous days work
20 minute final Practicum application (Last day)
Copy on Floppy as well as hard copy
Powerpoint preferred
Rotate Presenters
Q&A from the class
➪ A Quality Level
✳ 6σ = 3.4 Defects per Million Opportunities
➪ A Structured ProblemSolving Approach
✳ Phased Project: Measure, Analyze, Improve, Control
➪ A Program
Sweet Fruit
Design for Manufacturability
Process Entitlement
Bulk of Fruit
Process Characterization
and Optimization
Ground Fruit
Logic and Intuition
VALUE
STREAM TO WASTE DUE TO
THE INCAPABLE
CUSTOMER PROCESSES
Six
Six Sigma
Sigma Traditional
Traditional
Focus on Prevention Focus on Firefighting
Low cost/high throughput High cost/low throughput
Poka Yoke Control Strategies Reliance on Test and Inspection
Stable/Predictable Processes Processes based on Random Probability
Proactive Reactive
Low Failure Rates High Failure Rates
Focus on Long Term Focus on Short Term
Efficient Wasteful
Manage by Metrics and Analysis Manage by “Seat of the pants”
LSL T USL
Sales Receipts
On Time Delivery
Process Capacity
Order Fulfillment Time
Reduction of Waste
Product Development Time
Process Yields
Scrap Reduction
Inventory Reduction
Floor Space Utilization
HIRING NEEDS
BEADS ARE OUR BUSINESS
PRODUCTION SUPERVISOR
4 PRODUCTION WORKERS
2 INSPECTORS
1 INSPECTION SUPERVISOR
1 TALLY KEEPER
©The National Graduate School of Quality Management v7 • 25
STANDING ORDERS
• Follow the process exactly.
An
An understanding
understanding of of
Probability
Probability Distributions
Distributions is
is
necessary
necessary to:
to:
••Understand
Understand thethe concept
concept and
and
use
use of
of statistical
statistical tools.
tools.
••Understand
Understand the
the significance
significance of
of
random
random variation
variation in
in everyday
everyday
measures.
measures.
••Understand
Understand thethe impact
impact of
of
significance
significance on
on the
the successful
successful
resolution
resolution of
of aa project.
project.
Project Uses
Phase 1: •Establish baseline data
Measurement
characteristics.
Characterization
Phase 4:
•Use the concept of shift &
Control drift to establish project
expectations.
Cumulative Frequency
500
400
Cumulative Frequency
Frequency
Cumulative Percent
300 5000
200
100
0 0
70 80 90 100 110 120 130 70 80 90 100 110 120 130
"Head Count" Results from 10,000 people doing a coin toss 200 times.
Cumulative Cumulative Percent
Cumulativecount
countisissimply
simply the
thetotal
total frequency
frequency
count
countaccumulated
accumulated as as you
youmove
movefromfromleftleft to
to 100
right
rightuntil
until we
weaccount
accountfor forthe
the total
total population
populationof of Cumulative Percent
10,000
10,000people.
people.
50
Since
Since we
weknow
knowhowhowmany
many people
people were
were in
inthis
this
population
population (ie
(ie 10,000),
10,000),we
wecan
candivide
divide each
eachof of the
the
cumulative counts by 10,000 to give us a curve
cumulative counts by 10,000 to give us a curve
with
withthe
the cumulative
cumulativepercent
percentof
ofpopulation.
population. 0
70 80 90 100 110 120 130
"Head Count"
©The National Graduate School of Quality Management v7 • 41
COIN TOSS PROBABILITY EXAMPLE
based
based onon these
these percentages.
percentages.
50
Note
Note here
here that
that 50%
50% of of the
the
values
values are
are less
less than
than our
our
expected
expected value
value ofof 100.
100.
0
This
This means
means that
that in
in aa future
future
70 80 90 100 110 120 130
experiment
experiment setset up
up the
the same
same
way,
way, wewe would
would expect
expect 50% 50%
of
of the
the values
values to
to be
be less
less than
than
100.
100.
300
For
For example,
example, we we can
can see
see that
that the
the
200
occurrence
occurrence of of aa “Head
“Head count”
count” ofof less
less than
than
100
7474 or
or greater
greater than
than 124124 out
out of
of 200
200 tosses
tosses
0
70 80 90 100 110 120 130 isis so
so rare
rare that
that aa single
single occurrence
occurrence was was
"Head Count"
Results from 10,000 people doing a coin toss 200 times. not
not registered
registered outout of
of 10,000
10,000 tries.
tries.
Cumulative Percent
OnOn the
the other
other hand,
hand, wewe can
can see
see that
that the
the
100
chance
chance ofof getting
getting aa count
count near
near (or
(or at)
at) 100
100
isis much
much higher.
higher. With
With the
the data
data that
that we
Cumulative Percent
we
50
now
now have,
have, wewe can
can actually
actually predict
predict each
each ofof
these
these values.
values.
0
70 80 90 100 110 120 130
"Head Count"
©The National Graduate School of Quality Management v7 • 43
COIN TOSS PROBABILITY DISTRIBUTION
PROCESS
PROCESSCENTERED
CENTERED
ON
ON EXPECTEDVALUE
EXPECTED VALUE
% of population = probability of occurrence
600
CUM % OF POPULATION .003 .135 2.275 15.87 50.0 84.1 97.7 99.86 99.997
Common Occurrence
Rare Event
What
What are
are the
the chances
chances that
that this
this
“just
“just happened”
happened” IfIf they
they are
are small,
small,
chances
chances are
are that
that an
an external
external
influence
influence isis at
at work
work that
that can
can bebe
used
used to
to our
our benefit….
benefit….
sample
population
➣ Since the sample does not contain all the possible values,
there is some uncertainty about the population. Hence any
statistics, such as mean and standard deviation, are just
estimates of the true population parameters.
Inferential Statistics
©The National Graduate School of Quality Management v7 • 48
Inferential Statistics
Chances
Chances are are very
very 600 And
And the
the first
first 50
50
good
good that
that the
the trials
trials showed
showed
process
process distribution
distribution 500 “Head
“Head Counts”
Counts”
has
has changed.
changed. In In greater
greater than
than 130?
130?
400
Frequency
fact,
fact, there
there isis aa
probability
probability greater
greater 300
than
than 99.999%
99.999% that that
200
itit has
has changed.
changed.
σ
100
0
70 80 90 100 110 120 130
NUMBER OF HEADS 58 65 72 79 86 93 100 107 114 121 128 135 142
CUM % OF POPULATION .003 .135 2.275 15.87 50.0 84.1 97.7 99.86 99.997
Primarily
Primarily these
these distributions
distributions are
are used
used to
to test
test for
for significant
significant differences
differences in
in data
data sets.
sets.
To
To bebe classified
classified asas significant,
significant, the
the actual
actual measured
measured valuevalue must
must exceed
exceed aa critical
critical
value.
value. TheThe critical
critical value
value isis tabular
tabular value
value determined
determined by by the
the probability
probability distribution
distribution
and
and the
the risk
risk of
of error.
error. This
This risk
risk of
of error called αα risk
error isis called risk and
and indicates
indicates thethe probability
probability
of
of this
this value
value occurring
occurring naturally.
naturally. So, an αα risk
So, an risk of
of .05
.05 (5%)
(5%) means
means that
that this
this critical
critical
value
value will
will be
be exceeded
exceeded by by aa random
random occurrence
occurrence less less than
than 5%
5% ofof the
the time.
time.
Critical Critical
Value Value
DISPERSION
DISPERSION
How
How wide
wide aa population
population is
is spread.
spread.
DISTRIBUTION
DISTRIBUTION FUNCTION
FUNCTION
The
The mathematical
mathematical formula
formula that
that
best
best describes
describes the the data
data (we
(we will
will
cover
cover this
this in
in detail
detail in
in the
the next
next
module).
module).
600
500
400
300
200
100
0
70 80 90 1 00 110 1 20 130
What are some of the ways that we can easily indicate
the centering characteristic of the population?
mean = x = ∑x i
=
−2
= −.17 1
n 12 0
0
n=12 0
0
0
1
3
6 5 4 3 2 1 0 1 2 3 4 5 6 4
Mean ∑x = i −2
If we find the value half way (50%) through the data points, we 3
have another way of representing central tendency. This is 1 50% of data
points
called the median value. 1
0
0
Median
0
0
0
1
3
6 5 4 3 2 1 0 1 2 3 4 5 6 4
Median
Value
©The National Graduate School of Quality Management v7 • 55
WHAT IS THE MODE?
If we rank order (descending or ascending) the data set for
this distribution we find several ways we can represent central ORDERED DATA SET
tendency. 5
We find that a single value occurs more often than any other. 3
Since we know that there is a higher chance of this occurrence 1
in the middle of the distribution, we can use this feature as an 1
indicator of central tendency. This is called the mode. 0
0
Mode Mode
0
0
0
1
3
6 5 4 3 2 1 0 1 2 3 4 5 6 4
X −2
0
0
i
0
X = = = .17
6 5 4 3 2 1 0 1 2 3 4 5 6
1
3
4
n 12
ORDERED DATA SET
5
3
MEDIAN
1
n/2=6
1
0 (50 percentile data point)
0
n=12
0 Median
0 Here the median value falls between two zero
0
n/2=6
1
3
values and therefore is zero. If the values were
6 5 4 3 2 1 0 1 2 3 4 5 6 4
ORDERED DATA SET say 2 and 3 instead, the median would be 2.5.
Mode = 0
5
3
1
MODE
1
0 (Most common value in the data set)
}
0
0
Mode = 0
0
0
The mode in this case is 0 with 5 occurrences
1
3 within this data.
6 5 4 3 2 1 0 1 2 3 4 5 6 4
©The National Graduate School of Quality Management v7 • 57
SO WHAT’S THE REAL DIFFERENCE?
MEAN
MEAN
The
The mean
mean is is the
the most
most
consistently
consistently accurate
accurate measure
measure of
of
central
central tendency,
tendency, but but is
is more
more
difficult
difficult to
to calculate
calculate than
than thethe
other
other measures.
measures.
MEDIAN
MEDIAN AND
AND MODE
MODE
The
The median
median and
and mode
mode are
are both
both
very
very easy
easy to
to determine.
determine. That’s
That’s
the
the good
good news….The
news….The bad bad news
news
is
is that
that both
both are
are more
more susceptible
susceptible
to
to bias
bias than
than the
the mean.
mean.
MEAN
MEAN
Use
Use on
on all
all occasions
occasions unless
unless aa
circumstance
circumstance prohibits
prohibits its
its use.
use.
MEDIAN
MEDIAN AND
AND MODE
MODE
Only
Only use
use ifif you
you cannot
cannot use
use
mean.
mean.
600
500
400
300
200
100
0
70 80 90 1 00 110 1 20 130
What are some of the ways that we can easily indicate the dispersion
(spread) characteristic of the population?
Three measures have historically been used; the range, the standard
deviation and the variance.
©The National Graduate School of Quality Management v7 • 60
WHAT IS THE RANGE?
ORDERED DATA SET
The range is a very common metric which is easily
5
determined from any ordered sample. To calculate the range
simply subtract the minimum value in the sample from the 3
maximum value. 1
1
Range = x MAX − x MIN = 4 − ( −5) = 9 0
0
0 Range
0
0
1
3
6 5 4 3 2 1 0 1 2 3 4 5 6 4
Range
Min Max
X =
∑ Xi
=
−2
= .17 5 5(.17)=4.83
i
(4.83)2=23.32
n 12 3 3(.17)=2.83 (2.83)2=8.01
1
1(.17)=.83 (.83)2=.69
∑ (X − X)
2 1
i 61.67 0 1(.17)=.83 (.83)2=.69
s =
2
= = 5.6 0 0(.17)=.17 (.17)2=.03
n −1 12 −1 0 0(.17)=.17 (.17)2=.03
0 0(.17)=.17 (.17)2=.03
0 0(.17)=.17 (.17)2=.03
1
0(.17)=.17 (.17)2=.03
3
6 5 4 3 2 1 0 1 2 3 4 5 6 4 1(.17)=1.17 (1.17)2=1.37
3(.17)=3.17 (3.17)2=10.05
4(.17)=4.17 (4.17)2=17.39
61.67
©The National Graduate School of Quality Management v7 • 62
MEASURES OF DISPERSION
ORDERED DATA SET
5
3
Min=5 RANGE (R)
1
1
0 (The maximum data value minus the minimum)
R = X max − X min = 4 − ( −6) = 10
0
0
0
0
1
3
6 5 4 3 2 1 0 1 2 3 4 5 6 4 Max=4
DATA SET X i − X ( X −X) 2
VARIANCE (s2)
i
X =
∑ Xi
=
−2
= .17
5 5(.17)=4.83
3 3(.17)=2.83
(4.83)2=23.32
(2.83)2=8.01
n 12 1 1(.17)=.83 (.83)2=.69
1
0
1(.17)=.83 (.83)2=.69
(Squared deviations around the center point)
∑ (X − X)
0
0(.17)=.17 (.17)2=.03
2
0 0(.17)=.17
61.67
(.17)2=.03
i
s =
2
= = 5.6
0 0(.17)=.17 (.17)2=.03
0 0(.17)=.17 (.17)2=.03
1 0(.17)=.17
3 1(.17)=1.17
(.17)2=.03
(1.17)2=1.37
n −1 12 −1
6 5 4 3 2 1 0 1 2 3 4 5 6 4 (3.17)2=10.05
ORDERED3(.17)=3.17
DATA SET
5 4(.17)=4.17
3
1
(4.17)2=17.39
61.67 STANDARD DEVIATION (s)
1
0
0
(Absolute deviation around the center point)
s= s2 = 5.6 = 2.37
0
0
0
1
3
6 5 4 3 2 1 0 1 2 3 4 5 6 4
µ = X =
∑X i
1
2
10
15
N 3 12
4 14
(X )
2 5 10
−X 9
σ = s = ∑
2 2 i 6
11
n −1 7
12
8 10
9 12
10
Σ Xi
X
s2
©The National Graduate School of Quality Management v7 • 64
SO WHAT’S THE REAL DIFFERENCE?
VARIANCE/
VARIANCE/ STANDARD
STANDARD DEVIATION
DEVIATION
The
The standard
standard deviation
deviation is is the
the most
most
consistently
consistently accurate
accurate measure
measure of of
central
central tendency
tendency for
for aa single
single
population.
population. The
The variance
variance has has the
the
added
added benefit
benefit of
of being
being additive
additive over
over
multiple
multiple populations.
populations. BothBoth are
are difficult
difficult
and
and time
time consuming
consuming to to calculate.
calculate.
RANGE
RANGE
The
The range
range is is very
very easy
easy to
to determine.
determine.
That’s
That’s the the good
good news….The
news….The bad bad news
news
is
is that
that itit is
is very
very susceptible
susceptible toto bias.
bias.
VARIANCE/
VARIANCE/ STANDARD
STANDARD
DEVIATION
DEVIATION
Best
Best used
used when
when you
you have
have
enough
enough samples
samples (>10).
(>10).
RANGE
RANGE
Good
Good for
for small
small samples
samples (10
(10 or
or
less).
less).
LSL USL
12 10 8 6 4 2 0 2 4 6 8 10 12
m e
T i
Historically,
Historically, this
this shift
shift andand drift
drift
Original Study primarily
primarily impacts
impacts thethe position
position ofof
the
the mean
mean and and shifts 1.5 σσ
shifts itit 1.5
from
from it’s
it’s original
original position.
position.
©The National Graduate School of Quality Management v7 • 68
VARIATION FAMILIES
Sources of
Variation
To
To compensate
compensate for for these
these long
long
term
term variations,
variations, we
we must
must
consider
consider two
two sets
sets of
of metrics.
metrics.
Short
Short term
term metrics
metrics are
are those
those
which
which typically
typically are
are associated
associated
with
with our
our work.
work. Long
Long term
term metrics
metrics
take
take the
the short
short term
term metric
metric data
data
and
and degrade
degrade itit by
by an
an average
average of
of
1.5σ
1.5σ ..
Probability
Probability distributions
distributions are
are
necessary
necessary to:
to:
••determine
determine whether
whether anan event
event is
is
significant
significant or
or due
due to
to random
random
chance.
chance.
••predict
predict the
the probability
probability of
of specific
specific
performance
performance given
given historical
historical
characteristics.
characteristics.
Common Uses
Phase 1:
Measurement
•Baselining Processes
Characterization
Phase 2:
Analysis
Breakthrough
Strategy
Phase 3:
Improvement
X X X
Data points vary, but as the data accumulates, it forms a distribution which occurs naturally.
X X
4
Original Population Continuous Distribution
3
0
0 1 2 3 4 5 6 7
1
2
3
4
4
2
Subgroup Normal Distribution
1
0
Average
0 1 2 3 4 5 6 7
1
2
4
4
0
0 1 2 3 4 5 6 7
Mean µ x
Variance σ2 s2
Standard Deviation σ s
Process Capability Cp Cp
Binomial Mean P P
χ 2
=
e
Significant = χ 2
CALC
≥ χ 2
CRIT
α ,df
fe
Note that in each case, a limit has been established to determine what is
random chance verses significant difference. This point is called the critical
value. If the calculated value exceeds this critical value, there is very low
probability (P<.05) that this is due to random chance.
+/ +/
1σ = 68% 2 tail = 32% 2σ = 95 2 tail = 4.6%
1 tail = 16% %
1 tail = 2.3%
68.26% 95.46%
3σ +3σ
+3σ
3σ
Common
Common Test
Test Values
Values
+/
Z(1.6)
Z(1.6) == 5.5%
5.5% (1 tail αα =.05)
(1 tail =.05) 3σ = 99.7
2 tail = 0.3%
Z(2.0)
Z(2.0) == 2.5%
2.5% (2 tail αα =.05)
(2 tail =.05) %
1 tail = .15%
99.73%
Characterize
Characterize
Population
Population
•t Stat (µ ,n<30) •Z Stat (p) •Z Stat (p) •F’ Stat (n<10) • χ 2 Stat
(n>5)
• χ 2 Stat •t Stat (n<30) •t Stat (n<30) • χ 2 Stat
(σ ,n<10) (n>5)
• τ Stat • τ Stat
• χ 2 Stat (Cp) (n<10) (n<10)
©The National Graduate School of Quality Management v7 • 86
HOW DO POPULATIONS INTERACT?
Means Add
Population means interact in a simple intuitive manner. µ 1 +µ 2 =
µ new
µ ne
w
σ new
σ σ
1 2
Variations Add
Population dispersions interact in an additive manner
σ 1 +2
σ 2 =
2
σ new 2
©The National Graduate School of Quality Management v7 • 88
HOW DO POPULATIONS INTERACT?
SUBTRACTING TWO POPULATIONS
µ µ
1 2
Means Subtract
Population means interact in a simple intuitive manner. µ 1 µ 2 =
µ new
µ ne
w
σ new
σ σ
1 2
Variations Add
Population dispersions interact in an additive manner
σ 1
2
+ σ 22 =
σ new 2
©The National Graduate School of Quality Management v7 • 89
TRANSACTIONAL EXAMPLE
X = $60,000/week
s = $5,000
• Assuming nothing changes, what percent of the
time will shipments exceed orders?
©The National Graduate School of Quality Management v7 • 90
TRANSACTIONAL EXAMPLE
Orders Shipments
X = $53,000 in orders/week X = $60,000 shipped/week
s = $8,000 s = $5,000
To solve this problem, we must create a new distribution to model the situation posed
in the problem. Since we are looking for shipments to exceed orders, the resulting
distribution is created as follows:
X shipments −orders = X shipments − X orders = $60,000 − $53,000 = $7,000
sshipments −orders = 2
sshipments + sorders
2
= (5000)2 + (8000)2 = $9434
$7000 The new distribution looks like this with a mean of $7000 and a
standard deviation of $9434. This distribution represents the
$0 occurrences of shipments exceeding orders. To answer the original
Shipments > orders
question (shipments>orders) we look for $0 on this new distribution.
Any occurrence to the right of this point will represent shipments >
orders. So, we need to calculate the percent of the curve that exists
to the right of $0.
©The National Graduate School of Quality Management v7 • 91
TRANSACTIONAL EXAMPLE, CONTINUED
X shipments −orders = X shipments − X orders = $60,000 − $53,000 = $7,000
sshipments −orders = 2
s
shipments +s2
orders = (5000) + (8000) = $9434
2 2
µ0 −X $0 − $7000
= = .74 s
Shipments > orders
s $9434
Look up .74s in the normal table and you will find .77. Therefore, the answer to the original
question is that 77% of the time, shipments will exceed orders.
Correlation
Correlation Analysis
Analysis is
is
necessary
necessary to:
to:
••show
show aa relationship
relationship between
between
two
two variables.
variables. This
This also
also sets
sets the
the
stage
stage for
for potential
potential cause
cause and
and
effect.
effect.
Common Uses
Phase 1:
Measurement
Characterization
•Determine and quantify
Phase 2:
Analysis the relationship between
Breakthrough
factors (x) and output
Strategy characteristics (Y)..
Phase 3:
Improvement
Optimization
Phase 4:
Control
Output or y
variable
(dependent)
Correlation
Correlation
Y=
Y= f(x)
f(x)
As
As the
the input
input variable
variable changes,
changes,
there
there is
is an
an influence
influence or
or bias
bias
on
on the
the output
output variable.
variable.
Input or x variable (independent)
Positive
x x x
Negative
∑ (x − x)(yi − y)
••Calculate
Calculate sample
sample covariance
covariance
s xy
(( ))
i
sxy =
n −1 •Calculate ssxx and
•Calculate and ssyyfor
for each
each data
data
set
set
••Use
Use the
the calculated
calculated values
values to
to
sxy compute
compute rrCCAALCLC ..
rCALC =
sx s y ••Add
Add aa ++ for
for positive
positive correlation
correlation
and
and  for
for aa negative
negative correlation.
correlation.
While this is the most precise method to calculate
Pearson’s r, there is an easier way to come up with a fairly
close approximation...
••Plot
Plot the
the data
data on
on orthogonal
orthogonal axis
axis
••Draw
Draw an
an Oval
Oval around
around the
the data
data
••Measure
Measure thethe length
length and
and width
width
of
of the
the Oval
Oval
W
••Calculate
Calculate the
the coefficient
coefficient of
of
linear
linear correlation
correlation (r)
(r) based
based on
on
Y=f(
L the
the formulas
formulas below
below
x)
W
r ≈ ± 1 −
x L
6.7
r ≈ − 1 − = −.47
12.6
L + = positive slope
   
W
          
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
6.7 12.6  = negative slope
©The National Graduate School of Quality Management v7 • 104
HOW DO I KNOW WHEN I HAVE CORRELATION?
Ordered r CRIT • The answer should strike a familiar cord at this point…
Pairs We have confidence (95%) that we have correlation
5 .88 when rCALC > rCRIT .
6 .81
•Since sample size is a key determinate of rCRIT we need
7 .75
to use a table to determine the correct rCRIT given the
8 .71
number of ordered pairs which comprise the complete
9 .67 data set.
10 .63
15 .51 •So, in the preceding example we had 60 ordered pairs
of data and we computed a rCALC of .47. Using the table
20 .44
at the left we determine that the rCRIT value for 60 is .26.
25 .40
30 .36 •Comparing rCALC > rCRIT we get .47 > .26. Therefore the
50 .28 calculated value exceeds the minimum critical value
80 .22 required for significance.
100 .20 • Conclusion: We are 95% confident that the observed
correlation is significant.
©The National Graduate School of Quality Management v7 • 105
Learning Objectives
The
The Central
Central Limit
Limit Theorem
Theorem is:
is:
••the
the key
key theoretical
theoretical link
link between
between
the
the normal
normal distribution
distribution and
and
sampling
sampling distributions.
distributions.
••the
the means
means by by which
which almost
almost any
any
sampling
sampling distribution,
distribution, no no matter
matter
how
how irregular,
irregular, cancan be
be
approximated
approximated by by aa normal
normal
distribution
distribution ifif the
the sample
sample size
size is
is
large
large enough.
enough.
Common Uses
Phase 1:
•The Central Limit
Measurement Theorem underlies all
statistic techniques which
Characterization
rely on normality as a
Phase 2: fundamental assumption
Analysis
Breakthrough
Strategy
Phase 3:
Improvement
Optimization
Phase 4:
Control
Normal
What
What this
this means
means isis that
that no
no matter
matter what
what kind
kind
of
of distribution
distribution we
we sample,
sample, ifif the
the sample
sample size
size
is
is big
big enough,
enough, the
the distribution
distribution for
for the
the mean
mean
is
is approximately
approximately normal.
normal.
This
This is
is the
the key
key link
link that
that allows
allows us
us to
to use
use
much
much of of the
the inferential
inferential statistics
statistics we
we have
have
been
been working
working withwith so
so far.
far.
This
This is
is the
the reason
reason that
that only
only aa few
few probability
probability
distributions
distributions (Z, and χχ )) have
(Z, tt and have such
such broad
22
broad
application.
application.
IfIf aa random
random event
event happens
happens aa great
great many
many times,
times,
the
the average
average results
results are
are likely
likely to
to be
be predictable.
predictable.
Jacob
Jacob Bernoulli
Bernoulli
©The National Graduate School of Quality Management v7 • 113
HOW DOES THIS WORK?
n=2
n=5
n=10
Process
Process Capability
Capability Analysis
Analysis is
is
necessary
necessary to:
to:
••determine
determine the
the area
area of
of focus
focus
which
which will
will ensure
ensure successful
successful
resolution
resolution of
of the
the project.
project.
••benchmark
benchmark aa process
process toto enable
enable
demonstrated
demonstrated levels
levels of
of
improvement
improvement after
after successful
successful
resolution
resolution of
of the
the project.
project.
••demonstrate
demonstrate improvement
improvement after
after
successful
successful resolution
resolution of
of the
the
project.
project.
Common Uses
Phase 1:
•Baselining a process
Measurement primary metric (Y) prior to
starting a project.
Characterization
Phase 2:
•Characterizing the
Analysis capability of causitive
Breakthrough factors (x).
Strategy
Phase 3:
•Characterizing a process
Improvement primary metric after
changes have been
Optimization
implemented to
Phase 4: demonstrate the level of
Control improvement.
Upper
UpperandandLower
LowerStandards
Standards Single
SingleStandard
Standard
(Specifications)
(Specifications) (Specification)
(Specification)
Spec Spec
Spec Spec
Spec Spec
(Lower) (Upper)
(Upper)
(Lower)
µ spec)
(σ )
Spec Spec Spec Spec
Spec
Spec µµ Spec
Spec
(LSL)
(LSL) (USL)
(USL)
Note:
LSL = Lower Spec Limit
USL = Upper Spec Limit
• Calculation Values:
Upper Spec value = $200,000 maximum
No Lower Spec
µ = historical average = $250,000
s = $20,000
• Calculation: CPK = MIN (µ − LSL , USL −µ ) ($200,000 − $250,000)
= = .83
3σ 3 *$20,000
C PK =
0.3 382,089 0.1 884,930
0.4 344,578 0.1 864,334
0.5
0.6
308,538
274,253
0.2
0.2
841,345
815,940
3σ
0.7 241,964 0.2 788,145
0.8
0.9
211,855
184,060
0.3
0.3
758,036
725,747
We find that it bears a striking resemblance to the
1.0
1.1
158,655
135,666
0.3
0.4
691,462
655,422
equation for Z which is:
with the value µ µ 0
µ −µ 0
1.2 115,070 0.4 617,911
1.3 96,801 0.4 579,260
1.4 80,757 0.5 539,828
Z CALC = substituted for
1.5
1.6
1.7
66,807
54,799
44,565
0.5
0.5
0.6
500,000
460,172
420,740
σ MIN(µ LSL,USLµ ).
1.8
1.9
35,930
28,716
0.6
0.6
382,089
344,578
Making this substitution, we get :
1 MIN (µ − LSL, USL −µ ) Z MIN (µ −LSL ,USL −µ )
2.0 22,750 0.7 308,538
2.1 17,864 0.7 274,253
2.2 13,903 0.7 241,964 C pk = * =
2.3
2.4
10,724
8,198
0.8
0.8
211,855
184,060
3 σ 3
2.5 6,210 0.8 158,655
2.6
2.7
4,661
3,467
0.9
0.9
135,666
115,070 We can now use a table similar to the one on the
2.8
2.9
2,555
1,866
0.9
1.0
96,801
80,757 left to transform either Z or the associated PPM to
3.0 1,350 1.0 66,807
3.1 968 1.0 54,799 an equivalent Cpk value.
3.2 687 1.1 44,565
3.3 483 1.1 35,930
3.4
3.5
337
233
1.1
1.2
28,716
22,750
So,
So, ifif we
we have
have aa process
process which
which has
has aa short
short term
term
3.6
3.7
159
108
1.2
1.2
17,864
13,903 PPM=136,666
PPM=136,666 we we find
find that
that the
the equivalent
equivalent Z=1.1
Z=1.1 and
and
3.8 72.4 1.3 10,724
3.9
4.0
48.1
31.7
1.3
1.3
8,198
6,210
Cpk=0.4
Cpk=0.4 fromfrom the
the table.
table.
©The National Graduate School of Quality Management v7 • 127
Learning Objectives
MultiVari
MultiVari charts
charts are
are a:
a:
••Simple,
Simple, yet
yet powerful
powerful way
way to
to
significantly
significantly reduce
reduce the
the number
number
of
of potential
potential factors
factors which
which could
could
be
be impacting
impacting your
your primary
primary
metric.
metric.
••Quick
Quick and
and efficient
efficient method
method to to
significantly
significantly reduce
reduce the
the time
time and
and
resources
resources required
required toto determine
determine
the
the primary
primary components
components of of
variation.
variation.
Common Uses
Phase 1:
Measurement
Characterization
Optimization
Phase 4:
Control
old machinery
tooling humidity
supplier hardness
operator error temperature technique
fatigue wrong spec.
tool wear lubrication
handling damage
pressure heat treat
Webster’s
Dictionary Webster’s
Dictionary
Sources of
Variation
Step
Step 11 Sample 1
Step
Step 33
Sample 3
Sample 2
Time 2
Time 1
Common
Common Patterns
Patterns of
of Variation
Variation
The
The correct
correct sample
sample size
size is
is
necessary
necessary to:
to:
••ensure
ensure any
any tests
tests you
you design
design
have
have aa high
high probability
probability of
of
success.
success.
••properly
properly utilize
utilize the
the type
type ofof data
data
you
you have
have chosen
chosen or or are
are limited
limited to
to
working
working with.
with.
Common Uses
Phase 1:
•Sample Size
Measurement considerations are used in
any situation where a
Characterization
sample is being used to
Phase 2: infer a population
Analysis
characteristic.
Breakthrough
Strategy
Phase 3:
Improvement
Optimization
Phase 4:
Control
The possibility of error exists in almost every system. This goes for point values as
well. While we report a specific value, that value only represents our best estimate
from the data at hand. The best way to think about this is to use the form:
true value = point estimate +/ error
The error around the point value follows one of several common probability
distributions. As you have seen so far, we can increase our confidence is to go
further and further out on the tails of this distribution.
Point
Value
This “error band”
which exists around
the point estimate is
called the confidence
+/ 1s = 67% Confidence Band
interval.
Reality
Not different (Ho) Different (H1)
Not different (Ho) • Type II Error
Test Decision
Correct Conclusion
• β risk
• Consumer Risk
Decision Point
β Risk α Risk
A person does not feel well and checks into a hospital for tests.
Reality
Not different (Ho) Different (H1)
Not different (Ho) • Type II Error
Test Decision
Correct Conclusion
• β risk
• Consumer Risk
Error Impact
Ho: Patient is not sick
Type I Error = Treating a patient who is not sick
H1: Patient is sick
Type II Error = Not treating a sick patient
©The National Graduate School of Quality Management v7 • 152
HOW ABOUT ANOTHER EXAMPLE?
AA change
change is
is made
made to to the
the sales
sales force
force to
to save
save costs.
costs. Did
Did itit adversely
adversely impact
impact
the
the order
order receipt
receipt rate?
rate?
Reality
Not different (Ho) Different (H1)
Not different (Ho) • Type II Error
Test Decision
Correct Conclusion
• β risk
• Consumer Risk
σ σ
Mean
Mean X − t a / 2 ,n −1 ≤ µ ≤ X + t a / 2 ,n −1
n n
n −1 n −1
Standard
StandardDeviation
Deviation s ≤ σ ≤ s
χ 12−a / 2 χ a2/ 2
Process
Process Capability
Capability χ 12−a / 2,n −1 χ a2/ 2 ,n−1
Cp ≤ Cp ≤ Cp
n −1 n −1
p(1 − p ) p(1 − p )
Percent
PercentDefective
Defective p − Za / 2 ≤ p ≤ p + Za / 2
n n
These
These individual
individual formulas
formulas are are not
not critical
critical at
at this
this point,
point, but
but notice
notice that
that the
the only
only
opportunity
opportunity for
for decreasing
decreasing the the error
error band
band (confidence
(confidence interval)
interval) without
without decreasing
decreasing the
the
confidence
confidence factor,
factor, isis to
to increase
increase the
the sample
sample size.
size.
©The National Graduate School of Quality Management v7 • 154
SAMPLE SIZE EQUATIONS
2
Za /2σ Allowable error = µ  X
Mean
Mean n=
µ − X (also known as δ )
2
σ 2 Allowable error = σ /s
Standard
Standard Deviation
Deviation n = χ a /2 + 1
s
2
Zα / 2
n = p (1 − p )
Allowable error = E
Percent
PercentDefective
Defective
E
We want to estimate the true average weight for a part within 2 pounds.
Historically, the part weight has had a standard deviation of 10 pounds.
We would like to be 95% confident in the results.
• Calculation Values:
Average tells you to use the mean formula
Significance: α = 5% (95% confident)
Zα /2 = Z.025 = 1.96
s=10 pounds
µ x = error allowed = 2 pounds
2 2
Zα / 2σ 196
. *10
n= = = 97
• Calculation: µ − X 2
• Answer: n=97 Samples
We want to estimate the true percent defective for a part within 1%.
Historically, the part percent defective has been 10%. We would like to
be 95% confident in the results.
• Calculation Values:
Percent defective tells you to use the percent defect formula
Significance: α = 5% (95% confident)
Zα /2 = Z.025 = 1.96
p = 10% = .1
E = 1% = .01
• Calculation:
2 2
Zα / 2 1.96
• Answer: n = p (1 −Samples
n=3458 p ) = .1(1 −.1) = 3458
E .01
Understanding
Understanding the the confidence
confidence
interval
interval is
is key
key to:
to:
••understanding
understanding thethe limitations
limitations of
of
quotes
quotes in
in point
point estimate
estimate data.
data.
••being
being able
able to
to quickly
quickly and
and
efficiently
efficiently screen
screen aa series
series of
of
point
point estimate
estimate data
data for
for
significance.
significance.
Common Uses
Phase 1: •Used in any situation
Measurement where data is being
Characterization evaluated for significance.
Phase 2:
Analysis
Breakthrough
Strategy
Phase 3:
Improvement
Optimization
Phase 4:
Control
The possibility of error exists in almost every system. This goes for point values as
well. While we report a specific value, that value only represents our best estimate
from the data at hand. The best way to think about this is to use the form:
true value = point estimate +/ error
The error around the point value follows one of several common probability
distributions. As you have seen so far, we can increase our confidence is to go
further and further out on the tails of this distribution.
Point
Value
This “error band”
which exists around
the point estimate is
called the confidence
+/ 1s = 67% Confidence Band
interval.
••The
The confidence
confidence interval
interval
establishes
establishes aa wayway to
to test
test whether
whether
or
or not
not aa significant
significant change
change hashas
occurred
occurred in in the
the sampled
sampled
population.
population. This This concept
concept is is
called
called significance
significance or or hypothesis
hypothesis
testing.
testing.
••Being
Being able
able to
to tell
tell when
when aa
significant
significant change
change has has occurred
occurred
helps
helps in
in preventing
preventing us us from
from
interpreting
interpreting aa significant
significant change
change
from
from aa random
random event
event and
and
responding
responding accordingly.
accordingly.
All
Allof
ofour
ourwork
work was
was focused
focused in
inaanarrow
narrow time
time frame.
frame.
Over
Overtime,
time,other
otherlong
longterm
terminfluences
influences come
comeand
andgo go
which move the population and change some
which move the population and change some of itsof its
characteristics.
characteristics.
e
Tim
Confidence
Confidence Intervals
Intervals give
give us
us
Original Study the
the tool
tool to
to allow
allow us
us to
to be
be able
able to
to
sort
sort the
the significant
significant changes
changes from
from
the
the insignificant.
insignificant.
©The National Graduate School of Quality Management v7 • 166
USING CONFIDENCE INTERVALS TO SCREEN DATA
TIME
2 3 4 5 6 7
WHAT
WHAT KIND
KIND OF
OF PROBLEM
PROBLEM DO
DO YOU
YOU HAVE?
HAVE?
••Analysis
Analysis for
for aa significant
significant change
change asks
asks the
the question
question “What
“What happened
happened to
to
make
make this
this significantly
significantly different
different from
from the
the rest?”
rest?”
••Analysis
Analysis for
for aa series
series ofof random
random events
events focuses
focuses on
on the
the process
process andand asks
asks
the
the question
question “What
“What isis designed
designed into
into this
this process
process which
which causes
causes itit to
to have
have
this
this characteristic?”.
characteristic?”.
©The National Graduate School of Quality Management v7 • 167
CONFIDENCE INTERVAL FORMULAS
s s
Mean
Mean X − t a / 2 ,n −1 ≤ µ ≤ X + t a / 2 ,n −1
n n
n −1 n −1
Standard
StandardDeviation
Deviation s ≤ σ ≤ s
χ 12−a / 2 χ a2/ 2
Process
Process Capability
Capability χ 12−a / 2,n −1 χ a2/ 2 ,n−1
Cp ≤ Cp ≤ Cp
n −1 n −1
p(1 − p ) p(1 − p )
Percent
PercentDefective
Defective p − Za / 2 ≤ p ≤ p + Za / 2
n n
These
These individual
individual formulas
formulas enable
enable us
us to
to calculate
calculate the
the confidence
confidence interval
interval for
for many
many of
of
the
the most
most common
common metrics.
metrics.
©The National Graduate School of Quality Management v7 • 168
CONFIDENCE INTERVAL EXAMPLE
Over the past 6 months, we have received 10,000 parts from a vendor with
an average defect rate of 14,000 dpm. The most recent batch of parts
proved to have 23,000 dpm. Should we be concerned? We would like to
be 95% confident in the results.
• Calculation Values:
Average defect rate of 14,000 ppm = 14,000/1,000,000 = .014
Significance: α = 5% (95% confident)
Zα /2 = Z.025 = 1.96
n=10,000
Comparison defect rate of 23,000 ppm = .023
• Calculation: p(1 − p) p(1 − p )
p − Za / 2 ≤ p ≤ p + Za / 2
n n
.014 ( 1 − .014 ) .014 ( 1 − .014 )
.014 − 1.96 ≤ p ≤ .014 + 1.96 .014 − .0023 ≤ p ≤ .014 + .0023
10 , 000 10 ,000
• Answer: Yes, .023 is significantly .outside
012 ≤ p ≤ .of
016the expected 95% confidence interval of .
012 to .016.
We are tracking the gas mileage of our late model ford and find that
historically, we have averaged 28 MPG. After a tune up at Billy Bob’s
auto repair we find that we only got 24 MPG average with a standard
deviation of 3 MPG in the next 16 fillups. Should we be concerned? We
would like to be 95% confident in the results.
Control
Control charts
charts are
are useful
useful to:
to:
••determine
determine the
the occurrence
occurrence ofof
“special
“special cause”
cause” situations.
situations.
••Utilize
Utilize the
the opportunities
opportunities
presented
presented by by “special
“special cause”
cause”
situations”
situations” toto identify
identify and
and correct
correct
the
the occurrence
occurrence of of the
the “special
“special
causes”
causes” ..
Common Uses
Phase 1:
•Control charts can be
Measurement effectively used to
determine “special cause”
Characterization
situations in the
Phase 2: Measurement and
Analysis
Analysis phases
Breakthrough
Strategy
Phase 3:
Improvement
Optimization
Phase 4:
Control
Remember our earlier work with confidence intervals? Any occurrence which falls
outside the confidence interval has a low probability of occurring by random chance
and therefore is “significantly different”. If we can identify and correct the cause, we
have an opportunity to significantly improve the stability of the process. Due to the
amount of data involved, control charts have historically used 99% confidence for
determining the occurrence of these “special causes”
Point
Value
Special cause
occurrence.
A control chart is simply a run chart with confidence intervals calculated and drawn in.
These “Statistical control limits” form the trip wires which enable us to determine
when a process characteristic is operating under the influence of a “Special cause”.
+/ 3s =
99% Confidence
Interval
First
First things
things first:
first:
••Select
Select the
the metric
metric to
to be
be evaluated
evaluated
••Select
Select the
the right
right control
control chart
chart for
for the
the
metric
metric
••Gather
Gather enough
enough data
data to
to calculate
calculate the
the
control
control limits
limits
••Plot
Plot the
the data
data on
on the
the chart
chart
••Draw
Draw the
the control
control limits
limits (UCL
(UCL && LCL)
LCL)
onto
onto the
the chart.
chart.
••Continue
Continue the
the run,
run, investigating
investigating and
and
correcting
correcting the
the cause
cause ofof any
any “out
“out of
of
control”
control” occurrence.
occurrence.
©The National Graduate School of Quality Management v7 • 179
How do I select the correct chart ?
LCL = X − A R
2 LCL = D R 3
n D4 D3 A2
2 3.27 0 1.88
3 2.57 0 1.02
4 2.28 0 0.73
5 2.11 0 0.58
6 2.00 0 0.48
7 1.92 0.08 0.42
8 1.86 0.14 0.37
9 1.82 0.18 0.34
P( 1 − P )
UCL p = P + 3 UCLnp = n P + 3 n P( 1 − P )
n
P( 1 − P )
LCLp = P − 3 LCLnp = n P − 3 n P( 1 − P )
n
Note: P charts have an individually calculated control limit for each point plotted
U
UCLu = U + 3 UCLC = C + 3 C
n
U
LCLu = U − 3 LCLC = C − 3 C
n
Note: U charts have an individually calculated control limit for each point plotted
Time
Time to
to Find
Find and
and Fix
Fix the
the cause
cause
••Look
Look for
for patterns
patterns in
in the
the data
data
••Analyze
Analyze the
the “out
“out of
of control”
control” occurrence
occurrence
••Fishbone
Fishbone diagrams
diagrams and
and Hypothesis
Hypothesis
tests
tests are
are valuable
valuable “discovery”
“discovery” tools.
tools.
Hypothesis
Hypothesis testing
testing is
is
necessary
necessary to:
to:
••determine
determine when
when there
there is
is aa
significant
significant difference
difference between
between
two
two sample
sample populations.
populations.
••determine
determine whether
whether there
there isis aa
significant
significant difference
difference between
between aa
sample
sample population
population and
and aa target
target
value.
value.
Common Uses
Phase 1:
Measurement
Characterization
We
We think
think we
we see
see something…well,
something…well,
we
we think…err
think…err maybe
maybe itit is…
is… could
could
be….
be….
But,
But, how
how do
do we
we know
know for
for sure?
sure?
Hypothesis
Hypothesis testing
testing is
is the
the key
key by
by
giving
giving us
us aa measure
measure of of how
how
confident
confident we
we can
can be
be inin our
our
decision.
decision.
©The National Graduate School of Quality Management v7 • 192
SO HOW DOES THIS HYPOTHESIS STUFF WORK?
Critical
Point
Statistic Value
We
We determine
determine aa critical
critical value
value from
from aa probability
probability table
table for
for the
the statistic.
statistic. ThisThis
value
value is
is compared
compared withwith the
the calculated
calculated value
value wewe get
get from
from ourour data.
data. IfIf the
the
calculated
calculated value
value exceeds
exceeds thethe critical
critical value,
value, thethe probability
probability of of this
this occurrence
occurrence
happening
happening duedue to
to random
random variation
variation isis less
less than
than our test αα ..
our test
©The National Graduate School of Quality Management v7 • 193
SO WHAT IS THIS ‘NULL” HYPOTHESIS?
Mathematicians are like Frenchmen, whatever you say to them they
translate into their own language and forth with it is something entirely
different. Goethe
Population
Population Population
Population
Average
Average Variance
Variance
Many test statistics use α and others use α /2 and often it is confusing
to tell which one to use. The answer is straightforward when you
consider what the test is trying to accomplish.
If there are two bounds (upper and lower), the α probability must be
split between each bound. This results in a test statistic of α /2.
If there is only one direction which is under consideration, the error
(probability) is concentrated in that direction. This results in a test
statistic of α . α α /
Examples
Critical
Critical
Value
Value
Critical
Critical
Value
2 Critical
Critical
Value
Examples Value Value
Examples
Examples
Ho:µµ 11>µ
Ho: >µ 22 Common
Rare Rare Common Rare
Ho:µµ 11=µ
Ho: =µ 22
Occurrence Occurrence Occurrence Occurrence
Occurrence
Ho:σσ 11<σ
Ho: <σ 22 Ho:σσ 11=σ
Ho: =σ 22
α Probability α /2 Probability 1−α Probability α /2
1−α Probability Probability CC
onofidenecneceInIn
tt
nfid
Test
TestFails
Failsin
inone direction==αα
onedirection Test
TestFails
Failsin
ineither direction==αα /2
eitherdirection /2
©The National Graduate School of Quality Management v7 • 197
Sample Average vs Sample Average
Coming up with the calculated statistic...
n1= n2 ≤ 20 n1 + n2 ≤ 30 n1 + n2 > 30
2 Sample t
2 Sample Tau 2 Sample Z
(DF: n1+n22)
X1 − X 2 X1 − X 2
τ =

2 X1 − X 2  t=
1
+
1 [( n − 1) s12 + ( n − 1) s22]
ZCALC =
σ 12 σ 212
dCALC
R1 + R 2 n1 n2 n1 + n2 − 2 +
n1 n2
• Ho: µ =µ
1 2
Receipts 1 Receipts 2
• Statistic Summary: 3067 3200
n1 = n2 = 5
DF=n1 + n2  2 = 8
2730 2777
Significance: α /2 = .025 (2 tail) 2840 2623
tcrit = 2.306 (From the table for α =.025 and 8 DF) 2913 3044
• Calculation: 2789 2834
s1=130, s2= 227
X1=2868, X2=2896
X1 − X 2
t=
tCALC =(28682896)/.63*185=.24 1 1 [( n − 1) s12 + ( n − 1) s22]
+
• Test: n1 n2 n1 + n2 − 2
Ho: tCALC < tCRIT
Ho: .24 < 2.306 = true? (yes, therefore we will fail to reject the null hypothesis).
• Conclusion: Fail to reject the null hypothesis (ie. The data does not support the conclusion that there is a
significant difference
©The National Graduate School of Quality Management v7 • 200
Sample Average vs Target (µ 0)
Coming up with the calculated statistic...
n ≤ 20 n ≤ 30 n > 30
Several changes were made to the sales organization. The number of receipts
was gathered both before and after the changes. Determine if the samples have
equal variance with 95% confidence.
• Ho: s12 = s22
• Statistic Summary: Receipts 1 Receipts 2
n1 = n2 = 5 3067 3200
Significance: α /2 = .025 (2 tail) 2730 2777
F’crit = 3.25 (From the table for n1, n2=5) 2840 2623
• Calculation: 2913 3044
R1=337, R2= 577 R MAX ,n1 2789 2834
F’CALC =577/337=1.7 ′ =
FCALC
R MIN ,n2
• Test:
Ho: F’CALC < F’CRIT
Ho: 1.7 < 3.25 = true? (yes, therefore we will fail to reject the null
hypothesis).
• Conclusion: Fail to reject the null hypothesis (ie. can’t say there is a significant difference)
p1 − p0 p1 − p2
ZCALC = ZCALC =
p0 (1 − p0 ) n1 p1 _ n2 p2 n p _ n p 1 1
1 − 1 1 2 2 +
n n1 + n2 n1 + n2 n1 n2
ANOVA
Anova
Anova is
is aa powerful
powerful method
method for
for
analyzing
analyzing process
process variation:
variation:
••Used
Used when
when comparing
comparing two
two or
or
more
more process
process means.
means.
••Estimate
Estimate the
the relative
relative effect
effect of
of the
the
input
input variables
variables on
on the
the output
output
variable.
variable.
Common Uses
Phase 1:
Measurement
Characterization
Phase 2:
Analysis • Hypothesis Testing
Breakthrough
Strategy
Phase 3:
Improvement
• Design of Experiments
Optimization
Phase 4:
(DOE)
Control
Use validated data & your process experts to identify key variables.
L ite r s P e r H r B y F o r m u la t i o n
G r o u p M e a n s a r e In d i c a t e d b y L i n e s
2 5
3 Groups = 3 Treatments
{
L i te r s P e r H r
2 0
µ
Ove ra ll M e a n
τ i
n
1 5
1 2 3
Let’s say we run a study where weF ohave three groups which we are evaluating for a
r m u la t i o n
significant difference in the means. Each one of these groups is called a “treatment”
and represents one unique set of experimental conditions. Within each treatments, we
have seven values which are called repetitions.
©The National Graduate School of Quality Management v7 • 212
Analysis Of Variance ANOVA
Linear Model
L ite r s P e r H r B y F o r m u la t i o n
G r o u p M e a n s a r e In d i c a t e d b y L i n e s
2 5
3 Groups = 3 Treatments
{
L i te r s P e r H r
2 0
µ
Ove ra ll M e a n
τ i
n
1 5
1 2 3
The first step in being able to perform this analysis is to compute the “Sum
of the Squares” to determine the variation between treatments and within
treatments.
∑∑( y ) ∑∑( y )
a n a a n
ij  y = n∑ ( yi  y ) + ij  yi
i =1 j =1 i =1 i =1 j =1
Since we will be comparing two sources of variation, we will use the F test
to determine whether there is a significant difference. To use the F test, we
need to convert the “Sum of the Squares” to “Mean Squares” by dividing by
the Degrees of Freedom (DF).
The Degrees of Freedom is different for each of our sources of variation.
DFbetween = # of treatments  1
DFwithin = (# of treatments)(# of repetitions of each treatment  1)
DFtotal = (# of treatments)(# of repetitions of each treatment)  1
Note that DFtotal = DFbetween + DFwithin
SS
MeanSquar e =
Degrees of Freedom
F Critical
Value
Statistic Value
We determine a critical value from a probability table. This value is compared with
the calculated value. If the calculated value exceeds the critical value, we will reject
the null hypothesis (concluding that a significant difference exists between 2 or more
of the treatment means.
©The National Graduate School of Quality Management v7 • 216
Class Example  Evaluate 3 Fuel Formulations
Is there a difference?
Here we have an example of three different fuel formulations that are expected to
show a significant difference in average fuel consumption. Is there a significant
difference? Let’s use ANOVA to test our hypothesis……..
The first step is to come up with the Mean Squares Between. To accomplish this
we:
•find the average of each treatment and the overall average
•find the difference between the two values for each treatment and square it
•sum the squares and multiply by the number of replicates in each treatment
•divide this resulting value by the degrees of freedom
©The National Graduate School of Quality Management v7 • 218
Class Example  Evaluate 3 Fuel Formulations
Is there a difference?
Step 2: Calculating the Mean Squares Within
Step #2 = Calculate the Mean Squares Within Value
2
Treatment # Data (Xi) Treatment Average (Ao) Difference (XiAo) Difference (XiAo)
1
1
19
18
18.0
18.0
1.0
0.0
1.0
0.0
The next step is to come up with
1
1
21
16
18.0
18.0
3.0
2.0
9.0
4.0
the Mean Squares Within. To
1
1
18
20
18.0
18.0
0.0
2.0
0.0
4.0
accomplish this we:
1
1
14
0
18.0
18.0
4.0
0.0
16.0
0.0
•square the difference
1
1
0
0
18.0
18.0
0.0
0.0
0.0
0.0
between each individual
2
2
20
15
19.0
19.0
1.0
4.0
1.0
16.0 within a treatment and the
2
2
21
19
19.0
19.0
2.0
0.0
4.0
0.0 average of that treatment
2 17 19.0 2.0 4.0
•sum the squares for each
2
2
22
19
19.0
19.0
3.0
0.0
9.0
0.0 treatment
2
2
0
0
19.0
19.0
0.0
0.0
0.0
0.0 •divide this resulting value by
2 0 19.0 0.0 0.0
3 23 22.0 1.0 1.0 the degrees of freedom
3 20 22.0 2.0 4.0
3 25 22.0 3.0 9.0
3 22 22.0 0.0 0.0
3 18 22.0 4.0 16.0
3 24 22.0 2.0 4.0
3 22 22.0 0.0 0.0
3 0 22.0 0.0 0.0
3 0 22.0 0.0 0.0
3 0 22.0 0.0 0.0
Sum of Squares Within (SSw) 102.0
Degrees of Freedom Within (DFw = (# of treatments)(samples 1) 18
Mean Squares Within (MSw=SSw/DFw) 5.7
Contingency
Contingency tables
tables are
are helpful
helpful
to:
to:
••Perform
Perform statistical
statistical significance
significance
testing
testing on
on count
count oror attribute
attribute data.
data.
••Allow
Allow comparison
comparison of of more
more than
than
one
one subset
subset of
of data
data to
to help
help
localize
localize KPIV
KPIV factors.
factors.
Common Uses
Phase 1:
Measurement
Characterization
Conduct “ad hoc” training for your team prior to using the tool
Gather data, use the tool to test and then move on.
Keep it simple
AA contingency
contingency table table isis just
just another
another way way of of hypothesis
hypothesis testing.
testing. Just
Just like
like the
the
hypothesis
hypothesis testing
testing we we have
have learned
learned so so far,
far, we
we will
will obtain
obtain aa “critical
“critical value”
value” from
from
aa table
table (χ
(χ in
22
in this
this case)
case) and and use
use itit asas aa tripwire
tripwire for
for significance.
significance. We We then
then use
use
the
the sample
sample datadata to calculate aa χχ CCAALCLC value.
to calculate value. Comparing
Comparing this this calculated
calculated value
22
value
to
to our
our critical
critical value,
value, tells
tells us
us whether
whether the the data
data groups
groups exhibit
exhibit no
no significant
significant
difference
difference (null
(null hypothesis
hypothesis or or HHoo)) or
or whether
whether oneone oror more
more significant
significant differences
differences
exist
exist (alternate
(alternate hypothesis
hypothesis or or HH11).).
Ho:
Ho:PP1=P 2=P3…
1=P2=P3…
H1:
H1:One
Oneor
ormore
morepopulation
population(P)
(P)isissignificantly
significantlydifferent
different
χ
µ
2
CRI
No significant T
One or more significant
differences differences LSL T USL
Ho H1
χ 2
CALC
•Define the hypothesis you want to test. In this example we have 3 vendors
from which we have historically purchased parts. We want to eliminate one
and would like to see if there is a significant difference in delivery performance.
This is stated mathematically as Ho: Pa=Pb=Pc. We want to be 95% confident in
this result.
•We then construct a table which looks like this example. In this table we have
order performance in rows and vendors in columns. Orders On Time Vendor A Vendor B Vendor C
Orders Late
•We ensure that we include both good and bad situations so that the sum of
each column is the total opportunities which have been grouped into good and
bad. We then gather enough data to ensure that each cell will have a
count of at least 5.
Vendor A Vendor B Vendor C
Orders On Time 25 58 12
• Fill in the data. Orders Late 7 9 5
•Add up the totals for each column and row. Orders On Time
Vendor A
25
Vendor B
58
Vendor C
12
Total
95
Orders Late 7 9 5 21
Total 32 67 17 116
Collapse
Collapsethe
the table
table
••IfIfwe
wewere
wereplacing
placingside
sidebets
betsininaanumber
numberof ofbars
barsand
and
wondered if there were any nonrandom factors
wondered if there were any nonrandom factors at play.at play.
WeWegather
gatherdata
dataand
andconstruct
constructthe
thefollowing
followingtable:
table:
Bar #1 Bar #2 Bar #3
Won Money 5 7 2
Lost Money 7 9 4
••Since
Sincebarbar#3
#3does
doesnot
notmeet
meetthethe“5“5orormore”
more”criteria,
criteria,
we do not have enough data to evaluate that
we do not have enough data to evaluate that particular particular
cell
cellfor
forBar
Bar#3.
#3. This
Thismeans
meansthat
thatwe wemust
mustcombine
combine the
Bar #3 Collapsed
the
into Bar #2
data with that of another bar to ensure that
data with that of another bar to ensure that we havewe have
significance.
significance. This
Thisisisreferred
referredto toas
as“collapsing”
“collapsing”the the
table. The resulting collapsed table looks like
table. The resulting collapsed table looks like the the
following:
following:
Bar #1 Bar #2&3
Won Money 5 9
Lost Money 7 13
••We
Wecan
cannow
nowproceed
proceedtotoevaluate
evaluateour
ourhypothesis.
hypothesis.
Note
Notethat
thatthe
thedata
databetween
betweenBar
Bar#2#2and
andBar
Bar#3
#3will
willbe
be
aliased and therefore can not be evaluated separately.
aliased and therefore can not be evaluated separately.
•Calculate the percentage of the total contained in each row by dividing the row
total by the total for all the categories. For example, the Orders on Time row
has a total of 95. The overall total for all the categories is 116. The percentage
for this row will be row/total or 95/116 = .82.
Vendor A Vendor B Vendor C Total Portion
Orders On Time 25 58 12 95 0.82
Orders Late 7 9 5 21 0.18
Total 32 67 17 116 1.00
•.The row percentage times the column total gives us the expected occurrence
for each cell based on the percentages. For example, for the Orders on time
row, .82 x32=26 for the first cell and .82 x 67= 55 for the second cell.
Actual Occurrences
Vendor A Vendor B Vendor C Total Portion
Orders On Time 25 58 12 95 0.82
Orders Late 7 9 5 21 0.18
Total 32 67 17 116 1.00
Expected Occurrences
Vendor A Vendor B Vendor C
Orders On Time 26 55
Orders Late
•Now we need to calculate the χ 2 value for the data. This is done using the
formula (ae)2/e (where a is the actual count and e is the expected count) for
each cell. So, the χ 2 value for the first column would be (2526)2/26=.04.
Filling in the remaining χ 2 values we get:
Actual Occurrences Expected Occurrences
Vendor A Vendor B Vendor C Total Portion Vendor A Vendor B Vendor C
Orders On Time 25 58 12 95 0.82 Orders On Time 26 55 14
Orders Late 7 9 5 21 0.18 Orders Late 6 12 3
Column Total 32 67 17 116 1.00
2 2 2
Calculations ( χ =(ae) /e) Calculated χ Values
Vendor A Vendor B Vendor C Vendor A Vendor B Vendor C
Orders On Time 2
(2526) /26
2
(5855) /55
2
(1214) /14 Orders On Time 0.04 0.18 0.27
Orders Late 2
(76) /6
2
(912) /12
2
(53) /3 Orders Late 0.25 0.81 1.20
©The National Graduate School of Quality Management v7 • 231
Now what do I do?
Performing
Performingthe
theAnalysis
Analysis
••Determine
Determinethe thecritical
criticalvalue
valuefrom
fromthetheχχ 2table.
table. To Toget
2
get
the
thevalue
valueyouyouneed
need33pieces
piecesof ofdata.
data. TheThedegrees
degreesof of
freedom are obtained by the following
freedom are obtained by the following equation; equation;
DF=(r1)x(c1).
DF=(r1)x(c1). In Inour
ourcase,
case,we wehave
have33columns
columns(c) (c)
and 2 rows (r) so our DF = (21)x(31)=1x2=2.
and 2 rows (r) so our DF = (21)x(31)=1x2=2.
••The
Thesecond
secondpiecepieceof ofdata
dataisisthe
therisk.
risk. Since
Sincewe weare are
looking
looking for .95 (95%) confidence (and α risk = 1
for .95 (95%) confidence (and α risk = 1
confidence)
confidence)we weknow
knowthe theαα risk
riskwill
willbebe.05.
.05.
••In theχχ 2table
Inthe table,,wewefind
findthat
thatthethecritical
criticalvalue
valuefor forαα
2
==.05
.05and
and22DF DFto tobe
be5.99.
5.99. Therefore,
Therefore,our ourχχ CRCITRIT ==
22
5.99
5.99
••Our calculatedχχ 2value
Ourcalculated valueisisthe
thesum
sumof ofthe
theindividual
2
individual
cell χ 22
values. For our example this
cell χ values. For our example this is . is .
04+.18+.27+.25+.81+1.20=2.75.
04+.18+.27+.25+.81+1.20=2.75. Therefore,Therefore,ourour
χχ CALC ==2.75.
22
2.75.
CALC
••We
Wenownowhave
haveall allthe
thepieces
piecesto toperform
performour ourtest.
test. Our
Our
Ho: is χ 22
< χ 22
. Is this true? Our data
Ho: is χ CACLACLC < χ CRCITRIT . Is this true? Our data shows shows
2.75<5.99,
2.75<5.99,therefore
thereforewe wefail
failto
toreject
rejectthethenull
null
hypothesis that there is no significant
hypothesis that there is no significant difference difference
between
between
©The the
Nationalthe vendor
vendor
Graduate performance
performance
School ininthis
thisarea.
of Quality Management area.
v7 • 232
Contingency Table Exercise
We have a part which is experiencing high scrap. Your team thinks that since it is
manufactured over 3 shifts and on 3 different machines, that the scrap could be
caused (Y=f(x)) by an off shift workmanship issue or machine capability. Verify with
95% confidence whether either of these hypothesis is supported by the data.
Construct
Construct aa
contingency
contingency table
table of
of
the
the data
data and
and Actual Occurrences
interpret
interpret the
the results
results Machine 1 Machine 2 Machine 3
for
for each
each data
data set.
set. Good Parts 100 350 900
Scrap 15 18 23
Actual Occurrences
Shift 1 Shift 2 Shift 3
Good Parts 500 400 450
Scrap 20 19 17
FUNDAMENTALS
Design
Design of
of Experiments
Experiments isis
particularly
particularly useful
useful to:
to:
••evaluate
evaluate interactions
interactions between
between 22
or
or more
more KPIVs
KPIVs and
and their
their impact
impact
on
on one
one or
or more
more KPOV’s.
KPOV’s.
••optimize
optimize values
values for
for KPIVs
KPIVs to
to
determine
determine the
the optimum
optimum output
output
from
from aa process.
process.
Phase 1:
Measurement
Characterization
Phase 2:
Analysis
Breakthrough
Strategy
•Verify the relationship
Phase 3:
Improvement
between KPIV’s and
KPOV’s by manipulating
Optimization the KPIV and observing
Phase 4: the corresponding KPOV
Control change.
•Determine the best KPIV
settings to optimize the
KPOV output.
AA design
design of
of experiment
experiment introduces
introduces
purposeful
purposeful changes
changes inin KPIV’s,
KPIV’s, so
so
that
that we
we can
can methodically
methodically observe
observe
the
the corresponding
corresponding response
response inin the
the
associated
associated KPOV’s.
KPOV’s.
x f(x) Y=f(x)
Variables
How
How do
do you
you know
know how
how
•Input, Controllable (KPIV) much
much aa suspected
suspected KPIV
KPIV
•Input, Uncontrollable (Noise) actually
actually influences
influences aa
•Output, Controllable (KPOV) KPOV?
KPOV? You You test
test it!
it!
©The National Graduate School of Quality Management v7 • 241
Design of Experiments, Terminology
4 factors Mathematical
Mathematical objects
objects are
are sometimes
sometimes as
2
as
4 evaluated peculiar
peculiar asas the
the most
most exotic
exotic beast
beast or
or bird,
bird,
IV and
and the
be
the time
be well
time spent
well employed.
spent in
employed.
in examining
examining them
them may
may
For
For the
the purposes
purposes of of this
this training
training wewe will
will teach
teach only
only full
full factorial
factorial (2
(2k)) designs.
k
designs.
This
This will
will enable
enable you
you to to get
get aa basic
basic understanding
understanding of of application
application and and use
use the
the
tool.
tool. In
In addition,
addition, the
the vast
vast majority
majority ofof problems
problems commonly
commonly encountered
encountered in in
improvement
improvement projects
projects cancan bebe addressed
addressed with with this
this design.
design. IfIf you
you have
have any
any
question
question on on whether
whether the the design
design isis adequate,
adequate, consult
consult aa statistical
statistical expert...
expert...
©The National Graduate School of Quality Management v7 • 243
The Yates Algorithm
Determining the number of Treatments
We work for a major “Donut & Coffee” chain. We have been tasked to
determine what are the most significant factors in making “the most
delicious coffee in the world”. In our work we have identified three
factors we consider to be significant. These factors are coffee brand
(maxwell house vs chock full o nuts), water (spring vs tap) and coffee
amount (# of scoops).
Use
Use the
the Yates
Yates algorithm
algorithm to
to design
design the
the experiment.
experiment.
• Draw upTreatment
the design
A Busing
C the
AB Yates
AC algorithm.
BC ABC
1 + + + + + + +
2 + +  +   
3 +  +  +  
4 +     + +
5  + +   + 
6  +   +  +
7   + +   +
8    + + + 
©The National Graduate School of Quality Management v7 • 247
So, How do I Conduct a DOE?
We
We can
can see
see the
the impact,
impact,
but
but how
how dodo we
we know
know ifif
these
these results
results are
are
significant
significant or
or just
just
random
random variation?
variation?
Ranked Degree of
Impact
A (6.75)
AB (5.25)
BC (3.25)
B (2.25) What
What tool
tool do
do you
you
AC (1.75)
ABC (1.25)
think
think would
would bebe
C (0.25) good
good toto use
use in
in this
this
situation?
situation?
©The National Graduate School of Quality Management v7 • 252
Confidence Interval for DOE results
Confidence
Confidence Interval
Interval
Ranked Degree of
Impact == Effect
Effect +/
+/ Error
Error
A (6.75)
Some of these factors
AB (5.25) do not seem to have
BC (3.25) much impact. We can
B (2.25) use them to estimate
AC (1.75) our error.
ABC (1.25)
C (0.25) We can be relatively
safe using the ABC and
the C factors since they
offer the greatest
chance of being
insignificant.
Confidence = ± tα / 2 , DF ∑ ( ABC 2
+ C 2
)
Ranked Degree of DF
Impact
DF=# of groups used
6σ IMPROVEMENT PHASE
Vital Few Variables
GUTTER!
Establish Operating Tolerances
7 8
soft ball
OR
158 141
hard ball
no wristband Oily Lane Soft Ball
154 159
soft ball
The Ball Type depends on the Lane Condition....
This is called an Interaction!
©The National Graduate School of Quality Management v7 • 260
Where do we go from here?
With Wristband
and