Randomized complete block design

Recall 1-way AOV: “completely randomized design with

single factor”. There are 8T experimental units (book is
now using notation R ), randomly assigned among >
treatments; objective of study is to obtain inferences about
the treatment means ." , .# , ..., .> . Reparameterization will
be helpful for extending the AOV model to more
complicated designs:

!" œ ." ' .

!# œ .# ' .
ã
!> œ .> ' .

The !3 's are the treatment effects; here . is the overall

mean (the mean of the .3 's) and !" * !# * â * !> œ 0.

The statistical model becomes

]34 œ . * !3 * %34

where %34 µ normala0ß 5 # b, or equivalently,

]34 µ normalˆ. * !3 , 5 # ‰
The RCBD design incorporates an additional categorical
variable, called a blocking variable or block, to account
for possible heterogeneity in experimental circumstances.
For instance, a typical block in an agricultural experiment
is a field— fields differ substantially in soil quality, etc.,
and the same experimental treatment might produce
different means in different fields.

Formally, the design is as follows: within each of , blocks,

assign 1 experimental unit at random to each of >
treatments. Thus, all treatments appear within each block,
and each block-treatment combination receives 1
experimental unit, which produces the observed response
C34 :

block
treatment 1 2 ,
1 C"" C"# â C", '
C "†
2 C#" C## â C#, '
C #†
ã ã ã
> C>" C># â C>, '
C >†

mean '
C †" '
C †# '
C †,

The statistical model is

]34 œ . * !3 * "4 * %34

where D !3 œ 0, D"4 œ 0, and %34 µ normala0, 5 # b. Here

3 4
"4 is the effect of the 4th block on the mean of ]34 .
Note that "4 is the same for every treatment. It is not
possible to estimate an interaction between block and
treatment, because each block/trt combination is assigned
only 1 experimental unit.

ML estimates of model parameters:

s œ'
. C †† œ DDC34 ÎR
3 4
s3 œ '
! C 3† ' '
C ††
s œ'
" C '' C
4 †4 ††

Š a,'"ba>'"b ‹!! ’C34 ' Š. s ‹“

> , #
# 1
s œ
5 s*!
s3 * " 4
3œ"4œ"

R ' a, ' 1b ' Ð> ' 1Ñ ' 1 œ a, ' 1ba> ' 1b

Note: R œ ,>, and

Sums of squares:
SSablockb œ >! ˆ' C †† ‰ œ > D "
,
C †4 ' '
# s#
4 4
4œ"

SSatreatmentb œ ,! a' C †† b# œ , D !
>
C 3† ' ' s 3#
3
3œ"

SSaerrorb œ ! ! ’C34 ' Š. s ‹“ œ D/ #

> , #
s*!
s3 * " 4 34 3,4
3œ" 4œ"

And as always, SSatotalb œ ! ! aC34 ' '

C †† b#
> ,

3œ" 4œ"
The AOV table for the RCBD design:

Source SS df MS J

block SSablockb ,'" SSablockb MSablockb

MSaerrorb
SSatrtb SSatrtb MSatrtb
,'1
trt >'" MSaerrorb
SSaerrorb SSaerrorb
>'1
a, ' "ba> ' "b
error a,'1ba>'1b

total SSatotalb ,> ' 1

Hypothesis tests

Block effect

H! : "" œ "# œ â œ ", œ 0

Ha : Á

ablockb
Test statistic is 0 œ MS
MSaerrorb using
Fa, ' ", a, ' "ba> ' "bb distribution.

Treatment effect

H! : !" œ !# œ â œ !> œ 0
Ha : Á

Test statistic is 0 œ MSatrtb

MSaerrorb using Fa> ' ", a, ' "ba> ' "bb
distribution.
RCBD remarks

1. Missing observation(s) produces unbalanced design;

block and treatment are no longer orthogonal (& hand-
calculating formulas are no longer available). Parameters
can still be estimated, and effects tested, in PROC GLM.
Use type III sums of squares:

MODEL Y=BLOCK TRT / SS3;

Type III SS tests the effect last, after the other variables are
entered in the model.

2. If blocks have been randomly chosed from some

defined population of blocks (block might be a person in a
study, say, or a plot of land picked from farms in southern
Idaho), one might consider a model in which
"4 µ normalˆ0, 5"# ‰

This is called a random effects model (or, more precisely,

a mixed effects model) & the experiment is called a
random blocks experiment (rather than randomized
block). We will look at a few of these types of models
toward the end of the course. The ordinary RCBD model
is a fixed effects model.