86
1 Design & Analysis of Multi-Stratu m Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

  • View
    217

  • Download
    1

Embed Size (px)

Citation preview

Page 1: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

1

Design & Analysis of Multi-Stratum Randomized Experiments

Ching-Shui Cheng

June 5, 2008

National SunYat-sen University

Page 2: 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

Page 3: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

3

Page 4: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

4

Randomization model for designs with simple block structures

where are the treatment effects.

Nelder (1965a) gives simple rules for determining .

Page 5: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

5

Page 6: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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.

Page 7: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

7

A model of this form also arises from the following mixed-

effects model:

Random block effects

Page 8: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

8

Page 9: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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.

Page 10: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

10

Completely randomized design

Page 11: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

11

Block design (b/k)

Page 12: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

12

Row-column design ( )

Page 13: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

13

Total sum of squares

Page 14: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

14

Total variability

Page 15: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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

Page 16: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

16

One error term

Complete randomization

Page 17: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

17

b/k: s=2

Page 18: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

18

Page 19: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

19

Two error terms

Page 20: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

20

Page 21: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

21

Page 22: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

22

Three error terms

Page 23: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

23

Page 24: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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.

Page 25: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

25

Null ANOVA

Page 26: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

26

Page 27: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

27

Page 28: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

28

McLeod and Brewster (2004) Technometrics

Page 29: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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.

Page 30: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

30

Page 31: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

31

Page 32: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

32

Miller (1997) Technometrics

Page 33: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

33

Page 34: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

34

There are 8 nontrivial strata

Page 35: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

35

Page 36: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

36

Page 37: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

37

Page 38: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

38

Projections of the data vector onto different

strata are uncorrelated between and

homoscedastic within.

Page 39: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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.

Page 40: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

40

Designs such that falls entirely in one stratum are

called orthogonal designs.

Examples:

Completely randomized designs

Randomized complete block designs

Latin squares

Page 41: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

41

Page 42: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

42

Completely randomized design

Page 43: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

43

Page 44: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

44

Complete block designs

The two factors T and B satisfy the condition of proportional frequencies.

Page 45: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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),

Page 46: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

46

Page 47: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

47

ANOVA for an orthogonal design

Page 48: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

48

Page 49: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

49

Page 50: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

50

Page 51: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

51

Page 52: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

52

Page 53: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

53

Page 54: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

54

Page 55: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

55

Page 56: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

56

Three replications of 23

Treatment structure: 2*2*2

Block structure: 3/8

Page 57: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

57

Page 58: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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

Page 59: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

59

Page 60: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

60

Page 61: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

61

Page 62: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

62

Page 63: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

63

Page 64: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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

Page 65: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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

Page 66: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

66

ANOVAdf

InterblockAB 1Residual 4

IntrablockA 1B 1Residual 4

Total 11

Page 67: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

67

ANOVA for 3/2/2df

Replicates 2

InterblockAB 1Residual 2

IntrablockA 1B 1Residual 4

Total 11

Page 68: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

68

Page 69: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

69

Treatment structure: 3*4Block structure: 3/3/4

Page 70: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

70

Page 71: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

71

Page 72: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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.

Page 73: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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)

Page 74: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

74

Page 75: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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

Page 76: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

76

block.dryer stratum

a 1b 1c 1d 1ac=bd 1bc=ad 1

Page 77: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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

Page 78: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

78

Blocks Block/((Sv/St/Sr)*Sw) 4/((3/2/3)*7) Treatments Var*Time*Rate*Weed 3*2*3*7

Page 79: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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

Page 80: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

80

 

Block stratum 3 Block.Sv stratumVar 2Residual 6 Block.Sw stratumWeed 6Residual 18

Block.Sv.St stratumTime 1Var.Time 2Residual 9  

Page 81: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

81

 Block.Sv.Sw stratumVar.Weed 12Residual 36

Block.Sv.St.Sr stratumRate 2Var.Rate 4Time.Rate 2Var.Time.Rate 4Residual 36

 

Page 82: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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 .....”

Page 83: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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 ….”

Page 84: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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

Page 85: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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. …..”

Page 86: 1 Design & Analysis of Multi-Stratum Randomized Experiments Ching-Shui Cheng June 5, 2008 National SunYat-sen University

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.”