View
217
Download
1
Embed Size (px)
Citation preview
1
Design & Analysis of Multi-Stratum Randomized Experiments
Ching-Shui Cheng
June 5, 2008
National SunYat-sen University
2
Randomization model, Null ANOVA, Orthogonal designs
3
4
Randomization model for designs with simple block structures
where are the treatment effects.
Nelder (1965a) gives simple rules for determining .
5
6
Block designs (b/k)
The covariance matrix has three different entries: a common variance and two covariances depending on whether the corresponding pair of observations are in the same block or not.
7
A model of this form also arises from the following mixed-
effects model:
Random block effects
8
9
has spectral form
where is the orthogonal projection matrix onto theeigenspace of with eigenvalue
The eigenspaces are independent of the values of thevariances and covarinces: co-spectral.
Each of these eignespaces is called a stratum.
10
Completely randomized design
11
Block design (b/k)
12
Row-column design ( )
13
Total sum of squares
14
Total variability
15
Suppose (in which case does not contain treatment effects, and therefore measures variability among the experimental units)It can be shown that
th stratum variance
Null ANOVA
16
One error term
Complete randomization
17
b/k: s=2
18
19
Two error terms
20
21
22
Three error terms
23
24
Nelder (1965a) gave simple rules for determiningthe degrees of freedom, projections onto the strataand the sums of squares in the null ANOVA for anysimple block structure.
25
Null ANOVA
26
27
28
McLeod and Brewster (2004) Technometrics
29
From the d.f. identity, we can write down a yield identity which gives projections to all the strata. For convenience, we index each unit by multi-subscripts, and as before, dot notation is used for averaging. The following is the rule given by Nelder:
Expand each term in the d.f. identity as a function of the n's; then to each term in the expansion corresponds a mean of the y's with the same sign and averaged over the subscripts for which the corresponding n's are absent.
30
31
32
Miller (1997) Technometrics
33
34
There are 8 nontrivial strata
35
36
37
38
Projections of the data vector onto different
strata are uncorrelated between and
homoscedastic within.
39
Estimates computed in different strata are uncorrelated.
Estimate each treatment contrast in each of the strata in which it is estimable, and combine the uncorrelated estimates from different strata.
Simple analysis results when the treatment contrasts are estimable in only one stratum.
40
Designs such that falls entirely in one stratum are
called orthogonal designs.
Examples:
Completely randomized designs
Randomized complete block designs
Latin squares
41
42
Completely randomized design
43
44
Complete block designs
The two factors T and B satisfy the condition of proportional frequencies.
45
Under a row-column design such that each treatment appears the
same number of times in each row and the same number of times in
each column (such as a Latin square),
46
47
ANOVA for an orthogonal design
48
49
50
51
52
53
54
55
56
Three replications of 23
Treatment structure: 2*2*2
Block structure: 3/8
57
58
Incomplete blocks
The condition of proportional frequencies cannot besatisfied. has nontrivial projections in both the inter- and intra-block strata. Hence there is treatmentinformation in both strata.
Interblock estimateIntrablock estimate Recovery of interblock information
59
60
61
62
63
64
In general, there may be information for treatment contrasts
in more than one stratum.
Analysis is still simple if the space of treatment contrasts
can be decomposed as , where each ,
consisting of treatment contrasts of interest, is entirely in
one stratum.
Orthogonal designs
65
Treatment structure: 22 Block structure: 6/2
Information for the main effects is in the intrablock stratum andinformation for the interaction is in the interblock stratum
66
ANOVAdf
InterblockAB 1Residual 4
IntrablockA 1B 1Residual 4
Total 11
67
ANOVA for 3/2/2df
Replicates 2
InterblockAB 1Residual 2
IntrablockA 1B 1Residual 4
Total 11
68
69
Treatment structure: 3*4Block structure: 3/3/4
70
71
72
Miller (1997)
The experiment is run in 2 blocks and employs
4 washers and 4 driers. Sets of cloth samples
are run through the washers and the samples
are divided into groups such that each group
contains exactly one sample from each washer.
Each group of samples is then assigned to one
of the driers. Once dried, the extent of wrinkling
on each sample is evaluated.
73
Treatment structure:
A, B, C, D, E, F: configurations of washers
a,b,c,d: configurations of dryers
Block structure:2 blocks/(4 washers* 4 dryers)
74
75
Source of variation d.f. block stratum
AD=BE=CF=ab=cd 1 block.wash stratum
A=BC=EF 1B=AC=DF 1 C=AB=DF 1D=BF=CE 1E=AF=CD 1F=BD=AE 1
76
block.dryer stratum
a 1b 1c 1d 1ac=bd 1bc=ad 1
77
block.wash.dryer stratum
Aa=Db 1Ba=Eb 1Ca=Fb 1Da=Ab 1Ea=Bb 1Fa=Cb 1Ac=Dd 1Bc=Ed 1Cc=Fd 1Dc=Ad 1Ec=Bd 1Fc=Cd 1Residual 6Total 31
78
Blocks Block/((Sv/St/Sr)*Sw) 4/((3/2/3)*7) Treatments Var*Time*Rate*Weed 3*2*3*7
79
Factor [nvalues=504;levels=4] Block & [levels=3] Sv, Sr, Var, Rate & [levels=2] St, Time & [levels=7] Sw, WeedGenerate Block, Sv, St, Sr, SwMatrix [rows=4;columns=6; \values="b1 b2 Col St Sr Row"\1, 0, 1, 0, 0, 0,\0, 0, 1, 1, 0, 0,\0, 0, 1, 1, 1, 0,\1, 1, 0, 0, 0, 1] CkeyAkey [blockfactor=Block,Sv,St,Sr,Sw; \Colprimes=!(2,2,3,2,3,7);Colmappings=!(1,1,2,3,4,5);Key=Ckey] Var, Time, Rate, WeedBlocks Block/((Sv/St/Sr)*Sw)Treatments Var*Time*Rate*WeedANOVA
80
Block stratum 3 Block.Sv stratumVar 2Residual 6 Block.Sw stratumWeed 6Residual 18
Block.Sv.St stratumTime 1Var.Time 2Residual 9
81
Block.Sv.Sw stratumVar.Weed 12Residual 36
Block.Sv.St.Sr stratumRate 2Var.Rate 4Time.Rate 2Var.Time.Rate 4Residual 36
82
From the archive of S-news email list:
“I'm having trouble with what I'm sure is a very
simple problem, i.e., specifying the correct
syntax for a simple split plot design. To check
my understanding I'm using the cake data set
from Table 7.5 of Cochran & Cox. .....
I can't figure out what error term I need to
specify .....”
83
“Speaking as an absolutely committed devotee of Splus, I think that the Splus syntax for analyzing experimental designs containing random effects absolutely SUCKS!!! …..I am firmly of the opinion that talking about ‘splitplots’ as such is counter-productive, confusing, and antiquated. The ***right*** way to think about such problems is simply to decide which effects are random and which are fixed. Then think about expectedmean squares, and choose the denominator of your F-statistic so that under H_0 the ratio of the expected mean squares is 1 ….”
84
“It's a cross-Atlantic culture clash. …..This divide-and-conquer idea of separating thedata using linear functions with homogeneous error variances, that is, the spectral decomposition of the variance matrix, I find immensely simplifyingand natural. …..The contrary notion of putting everything togetherin one big anova table with interactions between random and fixed effects standing in forwhat really are error terms and pondering which one goes on the bottom of which other is to lose the plot entirely. ……” ---- Bill Venables
85
“The great advantage of the western Pacific view of specifying an experiment is that the commands follow from the design of the experiment. There is no need to know anything about fixed or random effects, or denominators of F ratios. Experimenters know what their treatments are, and they know how they randomised their experiment. That’s all they need to specify the model, and the correct analysis pops out whether it was a split plot, confounded factorial, Latin square, or whatever. …..”
86
Bailey (1981) JRSS, Ser. A
“Although Nelder (1965a, b) gave a unified treatment of what he called ‘simple’ block structures over ten years ago, his ideas do not seem to have gained wide acceptance. It is a pity, because they are useful and, I believe, simplifying. However, there seems to be a widespread belief that his ideas are too difficult to be understood or used by practical statisticians or students.”