Vous êtes sur la page 1sur 51

Split-Plot Experimental Designs

1/21/2010
Copyright 2010 Kun Liang

Split-Plot Experimental Designs


Field
Block 1

Plot Genotype B
150 100 50 0 Split Plot or Sub Plot

Genotype C
0 100 150 50

Genotype A
50 100 150 0

Block 2

Genotype B
150 100 50 0

Genotype A
0

Genotype C
50 150 0

50 150 100 100

Genotype A
Block 3
100 50 0 150

Genotype B
0 100 150 50

Genotype C
50 100 150 0

Genotype B
Block 4 0

Genotype C
50 0

Genotype A
50 150 100 0
2

50 100 150 150 100

Split-Plot Experimental Designs


This experiment has two factors: genotype and fertilizer amount. Genotype has levels A, B, and C. Fertilizer has levels 0, 50, 100, 150 kg N/ha.

Genotype is called the whole-plot factor because its levels are randomly assigned to whole plots.
Fertilizer is called the split-plot factor because its levels are randomly assigned to split plots within each whole plot.
3

Definition of Experimental Units in Split-Plot Designs


Plots are the whole-plot experimental units because the levels of the whole-plot factor (genotype) are randomly assigned to plots. The split-plots are the split-plot experimental units because the levels of the split-plot factor (amount of fertilizer) are randomly assigned to split plots within each whole plot.

Thus we have two different sizes of experimental units in split-plot experimental designs.
4

Same Treatment Structure in a RCBD


Field
Block 1

B B A C B C A A C B C A 100 0 0 100 150 50 50 150 150 50 0 100


A A C A B B C C A C B B 150 0 50 50 100 50 100 0 100 150 150 0 C 0 A A B B B A C A C C B 0 100 100 50 0 150 50 50 150 100 150

Block 2

Block 3

Block 4

B C B A C A B C B C A A 0 150 50 150 100 0 150 50 100 0 100 50


5

Same Treatment Structure in a CRD


Field

B B A B A C A B A C C C 50 0 150 100 100 150 50 0 50 100 0 100


A A C B B B A 50 0 50 50 150 50 0 C 0 A A A A 0 100 150 0 C A C B B 0 100 50 150 0

B A B A B C A 0 150 150 50 150100 100

B B B C C C A C C C C B 50 100 100 150 100 50 150 50 150 0 150 100


6

Which Design to Use?


Consider a CRD if

1. There is equal interest in all treatment comparisons. 2. There is no natural way to group similar experimental units together in blocks.
3. No logistical constraints make complete randomization impractical.
7

Which Design to Use?


Consider a RCBD if

1. There is equal interest in all treatment comparisons.


2. There are potentially important sources of variation that can be controlled by blocking.

3. No logistical constraints make complete randomization within blocks impractical.


8

Which Design to Use?


Consider a split-plot design if

1. Comparisons among the levels of the split-plot factor are greater interest than comparisons among the levels of the whole-plot factor.
2. Logistical constraints make a CRD or RCBD impractical.
9

Split-Plot Experimental Design


Field
Block 1

Genotype C
0 100 150 50

Genotype A
50 100 150 0

Genotype B
150 100 50 0

Block 2

Genotype B
150 100 50 0

Genotype A
0

Genotype C
50 150 0

50 150 100 100

Genotype A
Block 3
100 50 0 150

Genotype B
0 100 150 50

Genotype C
50 100 150 0

Genotype B
Block 4 0

Genotype C
50 0

Genotype A
50 150 100 0
10

50 100 150 150 100

How would you design the microarray portion of the experiment?


Suppose the researcher in the split-plot experiment is willing to use 48 two-color microarray slides to measure gene expression in plants from the field. The researcher is primarily interested in understanding how gene expression changes in response to fertilizer rate within each genotype.
11

One Possible Design


Use a loop design to compare split-plot experimental units assigned adjacent fertilizer amounts within each whole-plot experimental unit.
A0 A50 A100 A150

Block 1 B0

B50

B100

B150

C0

C50

C100

C150

Consider the same loop designs for blocks 2, 3, and 4 perhaps reversing loop directions for two of the four blocks.
12

Split-plot design structures dont necessarily involve plots in the usual sense. 1) Suppose a total of 8 horses are assigned to two diets (A and B) using a completely randomized design (CRD) with 4 horses per diet. 2) After 4 weeks on the assigned diet, each horse is sacrificed and RNA samples are taken from the stomach and small intestine of each horse. 3) Affymetrix GeneChips are used to measure gene expression in each RNA sample with one GeneChip for each of the samples.
13

Conceptual Picture of the Experiment


Diet A stomach small intestine Diet A stomach small intestine Diet B stomach small intestine Diet B stomach small intestine

Diet A stomach small intestine

Diet B stomach small intestine

Diet B stomach small intestine

Diet A stomach small intestine

Horse is the whole-plot experimental unit

organ RNA sample is playing the role of the split-plot experimental unit
14

Conceptual Picture of the Experiment


Diet A stomach small intestine Diet A stomach small intestine Diet B stomach small intestine Diet B stomach small intestine

Diet A stomach small intestine

Diet B stomach small intestine

Diet B stomach small intestine

Diet A stomach small intestine

Diet is the whole-plot factor with levels A and B. Organ type is the split-plot factor with levels stomach and small intestine.
15

Conceptual Picture of the Experiment


Diet A stomach small intestine Diet A stomach small intestine Diet B stomach small intestine Diet B stomach small intestine

Diet A stomach small intestine

Diet B stomach small intestine

Diet B stomach small intestine

Diet A stomach small intestine

Note that the levels of the factor diet are randomly assigned to experimental units. That is not the case for the levels of the factor organ type, but we will analyze it and refer to it just as we would have for an experimental factors whose levels are randomly assigned. 16

How would we design the microarray portion of the experiment?


1) Suppose the researcher in this split-plot experiment can afford to use 8 two-color microarray slides to measure gene expression in the RNA samples.
2) The researcher is primarily interested in understanding how diets affect gene expression within each organ type.
17

One Possible Design


Diet A small stomach intestine Diet A stomach small intestine Diet A stomach small intestine Diet A stomach small intestine

stomach

small intestine

stomach

small intestine

stomach

small intestine

stomach

small intestine

Diet B

Diet B

Diet B

Diet B

Connect samples of the same organ type between horses treated with different diets because this will give us the tightest comparison between diets for each organ type (and this is the comparison of primary interest). 18

Overview of lecture
1) 2) 3) 4) 5) 6) What is a split plot design Partitioning the variability Pre-analysis checks Example ANOVA Reporting the Results Summary of ANOVA

Split Plot Designs


When there is more than one factor, we can have a split plot design: One or more repeated measures One or more between subjects measures The Split Plot design combines features of both between groups and within subjects designs. That is each level of factor A contains a different groups of randomly assigned subjects. On the other hand, each level of factor B at any given level of factor A contains the same subjects No such thing as a one factor mixed design.

Partitioning the variance


Partitioning the variance is done as for a standard ANOVA with small variations For the effects of interest A between groups effect is estimated A within subjects effect is estimated An interaction effect is estimated For the error terms A between subjects error is used for the between groups effect Within subjects error term used for the within subjects effect The within subjects error term is also used for the interaction effect since this includes a within subject component.

F ratios
For the main effects and the interaction there are separate F ratios calculated
FA MSA FB MSB MSS / A with (a 1) & (a )( s 1) df

MSBxS / A

with (b 1) & (a)(b 1)( s 1) df

FAB MS AB

MSBxS / A

with ( a 1)(b 1) & (a )(b 1)( s 1) df

Complete Example
Assessing a drug treatment to reduce systolic blood pressure
180 170 160 150 140 130 120 1 hour 5 hours 9 hours Drug Placebo

Pre-analysis checks
Homogeneity of Variance
For a mixed design it is necessary to check in a variety of different ways
Overall - Boxs M Within subjects - Mauchlys W Between groups - Levenes test

Normality
This can also be tested in a number of ways

SPSS conducts these tests.

Box's Test of Equality of Covariance Matrices

Homogeneity of variance
Box's M F df1 df2 Sig. 4.189 .466 6 724.528 .834

TIME

Mauchly's W .535

Approx. Chi-Square 5.633

df 2

Sig. .060

Levene's Test Blood Pressure (1 hr) Blood Pressure (5 hrs) Blood Pressure (9 hrs)

F .000 4.675 .034

df1 1 1 1

df2 10 10 10

Sig. 1.000 .056 .858

Normality

1 hour Drug 5 hours 9 hours 1 hour Placebo 5 hours 9 hours

Skew -1.369 0.857 0.711 -0.313 -0.668 0.456

S.E. z p 0.845 -1.620 0.105 0.845 1.014 0.310 0.845 0.841 0.400 0.845 -0.370 0.711 0.845 -0.791 0.429 0.845 0.540 0.589

Anova Summary
Mixed Design (alias Split Plot)

Source of Variation

Sum of df Mean Squares Squares

A (Treatment) 4761.000 1 4761.000 83.820 0.0000 B (Time) 2198.000 2 1099.000 32.324 0.0000 AB 1554.000 2 777.000 22.853 0.0000

Between Error 568.000 10 56.800 (Error BxS) 680.000 20 34.000

Summary of omnibus F results


A significant main effect of treatment
No further analyses required

A significant main effect of time


Post hoc tests required (no a priori prediction)

A significant interaction between treatment and time


Simple main effect analysis required

Main effect of time


Post hoc tests - Tukey
Comparisons Between Means for Selected Factor(s) * = p<0.05 ** = p<0.01 *** = p<0.001 **** = p<0.0001 Tukey test Comparison between levels of Time 1 hour 1 hour 5 hours vs 5 hours vs 9 hours vs 9 hours q = 2.97 q = 10.99 q = 8.02

*** ***

Simple main effects of Drug x Time interaction


Source of Sum of df Mean F p Variation Squares Squares Treatment at 1 hour 2352.000 1 2352.000 56.538 0.0000 5 hours 3888.000 1 3888.000 93.462 0.0000 9 hours 75.000 1 75.000 1.803 0.1894 Error Term 1248.000 30 41.600 Time at Drug 3484.000 2 1742.000 51.235 0.0000 Placebo 268.000 2 134.000 3.941 0.0361 Error Term 680.000 20 34.000

Summary of simple main effects


A significant simple main effect of Treatment at 1 hour The blood pressure for the drug and placebo groups are significantly different at 1 hour A significant simple main effect of Treatment at 5 hours The blood pressure for the drug and placebo groups are significantly different at 5 hours A non-significant simple main effect of Treatment at 9 hours The blood pressure for the drug and placebo groups are not significantly different at 9 hours A significant simple main effect of Time at Drug Post hoc tests required A significant simple main effect of Time at Placebo Post hoc tests required

Simple main effects of Treatment

180 170 160 150 140 130 120 1 hour 5 hours 9 hours Drug Placebo

Post hoc analysis of simple main effects


Comparison between levels of Time at level Drug 1 hour vs 5 hours 1 hour vs 9 hours 5 hours vs 9 hours at level Placebo 1 hour vs 5 hours 1 hour vs 9 hours 5 hours vs 9 hours q= q= q= 3.78 2.94 0.84 * q = 0.42 q = 12.60 q = 12.18

*** ***

Simple main effects of Time

180 170 160 150 140 130 120 1 hour 5 hours 9 hours Drug Placebo

Reporting the results


A two-way (2x3) mixed analysis of variance was conducted on systolic blood pressure. The independent variables included one between groups variable, treatment, with two levels (drug, placebo) and one within subject variable, time, with three levels (1 hour, 5 hours and 9 hours)
1 hour 132 (6.573) 160 (4.900) Time 5 hours 133 (8.832) 169 (4.517) 9 hours 162 (7.014) 167 (5.900)

Drug Placebo

Table 1: Means (and standard deviations) of blood pressure for the drug and placebo treatment groups for three times post adminstration

Reporting the results


Report the main effects after the descriptive statistics There was a significant main effect of treatment (F1,10=83.820, MSE=56.800, p<0.001). Overall the systolic blood pressure of the drug group (142.33) was less than that of the placebo group (165.33). There was a significant main effect of time (F2,10=32.324, MSE=34.000, p<0.001). Post hoc tukey tests (at p0.05) were conducted to examine further the effect of time. The average systolic blood pressure at 9 hours (164.50) after adminstration of the treatment was significantly greater than at 5 hours (151.00) and at 1 hour (146.00). The latter were not significantly different.

Reporting the results


There was a significant interaction between treatment and time (F2,20=22.853, MSE=34.000, p<0.001; see Table 1). Simple main effects analysis demonstrated that the drug and placebo groups systolic blood pressure was significantly different at 1 hour (F1,30=56.538, MSE= 41.600, p<0.001) and 5 hours (F1,30=93.462, MSE=41.600, p<0.001) but not at 9 hours (F1,30=1.803, MSE= 41.600, p=.189). At both 1 hour and 5 hours the systolic blood pressure was greater for the placebo group than drug group

The simple main effect of time was significant for both the drug group (F2,20=51.235, MSE=34.000, p<0.001) and the placebo group (F2,20=3.941, MSE=34.000, p<0.05). Post hoc tukey tests (at p0.05) were conducted to explore further these effects. For the simple main effect of time for the drug group blood pressure was different between 1 hour and 9 hours and between 5 hours and 9 hours. However, blood pressure at 1 hour and 5 hours were not significantly different. Blood pressure was lower at 1 hour and 5 hours than at 9 hours for the drug group. For the placebo group, there was one significant difference between 1 hour and 5 hours with blood pressure at 5 hours being significantly greater

Summary of ANOVA
ANOVA is a parametric statistical technique for testing the differences between means. ANOVA can be used to analyse both single factor and multifactorial designs. Anova can be used to analyse differences in between ANOVA , within subjects and mixed designs.

Summary of ANOVA
A number of assumptions are made by ANOVA that should be tested prior to analysis Significant results often require further analysis Both planned and unplanned comparisons can be conducted Interactions nearly always require further analysis A failure to find a significant result may be due to lack of statistical power

Split-Plot Experiment
Top Shrinkage by Wool Fiber Treatment and Number of Drying Revolutions
J. Lindberg (1953). Relationship Between Various Surface Properties of Wool Fibers: Part II: Frictional Properties, Textile Research Journal, Vol. 23, pp. 225-237

Data Description
Experiment to Compare 4 Wool Fiber Treatments at 7 Dry Cycle Lengths over 4 Experimental Runs (Blocks) Response: Top Shrinkage of Fiber Restriction on Randomization: Within Each block, each treatment is assigned to whole plot, then measurements made at each of 7 dry cycle times (split plots) Whole Plot Treatments: Untreated, Alcoholic Potash (15 Sec, 4Min, 15Min) Subplot Treatments: Dry Cycle Revolutions (200 to 1400 by 200) Blocks: 4 Experimental Runs (possibly different days)

Block Layout

Within each block, randomize the 4 treatments to the 4 whole plots

Revs 200 400 600 800 1000 1200 1400 200 400 600 800 1000 1200 1400 200 400 600 800 1000 1200 1400 200 400 600 800 1000 1200 1400

Trt A A A A A A A B B B B B B B C C C C C C C D D D D D D D

Whole Plot

Subplot

Marginal Means
run 1 2 3 4 average 23.63 24.77 22.48 22.07
trt 1 2 3 4 average 31.30 23.42 21.18 17.05

revs 200 400 600 800 1000 1200 1400

average 6.45 13.43 19.56 25.13 28.96 32.84 36.30

No Clear Run Effects

As Alcoholic Potash increases (TRT), Shrinkage Decreases

As #Revs increases, Shrinkage Increases

Interaction Plot - Trt*Revs


50

45

40

35

30 trt 1 trt 2 trt 3 trt 4

Shrinkage

25

20

15

10

0 0 200 400 600 800 Dryer Revolutions 1000 1200 1400 1600

Analysis of Variance
Source Treatments Blocks BlkxTrt (Error 1) Revs TrtxRevs Error 2 Total df 3 3 9 6 18 72 111 SS 3012.53 124.29 114.64 11051.78 269.51 99.16 14671.90 MS 1004.18 41.43 12.74 1841.96 14.97 1.38 F 78.84 P-Value 0.0000

1337.46 10.87

0.0000 0.0000

Note that there is a significant interaction (as well as main effects). Thus the profile relating shrinkage to # of revolutions differs by treatment

Decomposing the Revolution and Trt/Rev Interaction Sum of Squares into Polynomial Effects Note that for Revolutions, we have thus far treated these as nominal categories, however, it is a continuous variable We can break down its sums of squares into orthogonal polynomials representing linear, quadratic, cubic, components (6 in all) Graph appears to show at least linear and quadratic terms. Similar breakdown can be done on Treatment/Revolution interaction

Partitioning of SSRevs and SSTrtxRevs


Source Revs Revs Linear Revs Quadratic Revs Cubic Revs Quartic Revs Quintic Revs Sextic TrtxRevs TxR Linear TxR Quadratic TxR Cubic TxR Quartic TxR Quintic TxR Sextic Error 2 Total df 6 1 1 1 1 1 1 18 3 3 3 3 3 3 72 111 SS 11051.78 10846.83 198.88 2.87 1.43 0.12 1.65 269.51 154.37 104.77 7.41 0.77 0.87 1.32 99.16 14671.90 MS 1841.96 10846.83 198.88 2.87 1.43 0.12 1.65 14.97 51.46 34.92 2.47 0.26 0.29 0.44 1.38 F 1337.46 7875.93 144.40 2.08 1.04 0.09 1.20 10.87 37.36 25.36 1.79 0.19 0.21 0.32 P-Value 0.0000 0.0000 0.0000 0.1532 0.3113 0.7669 0.2773 0.0000 0.0000 0.0000 0.1559 0.9055 0.8892 0.8110

The Revolution Main effect and the Treatment/Revolution Interaction is made up of significant linear and quadratic components

Observed/Fitted Values - Quadratic Model


50

45

40

35

30

25

20

trt 1 trt 2 trt 3 trt 4 trt1(fit) trt(fit) trt3(fit) trt4(fit)

Shrinkage

15

10

0 0 200 400 600 800 Revolutions 1000 1200 1400 1600

Procedure for Obtaining Polynomial SS 1. Obtain coefficients for orthogonal polynomials from stat design or math source
Revs 200 400 600 800 1000 1200 1400 Linear -3 -2 -1 0 1 2 3 Quadratic 5 0 -3 -4 -3 0 5 Cubic -1 1 1 0 -1 -1 1 Quartic 3 -7 1 6 1 -7 3 Quintic -1 4 -5 0 5 -4 1 Sextic 1 -6 15 -20 15 -6 1

2. Obtain Linear Combinatio n of Means Across Revs L1 3 y 200 2 y 400 1y 600 0 y 800 1y1000 2 y1200 3 y1400 ... L6 1y 200 6 y 400 15 y 600 20 y 800 15 y1000 6 y1200 1y1400

Procedure for Obtaining Polynomial SS


3. Obtain the sums of squares for each contrast
2 (4)( 4) L1 Linear : SS1 (3) 2 (2) 2 (1) 2 (0) 2 (1) 2 (2) 2 (3) 2 ...

(4)( 4) L2 6 Sextic : SS 6 (1) 2 (6) 2 (15) 2 (20) 2 (15) 2 (6) 2 (3) 2 Where the (4)(4) 16 in numerator represents the number of

observatio ns per rev level (4 blocks x 4 treatment s)

This process can also be extended to the TreatmentxRev interaction by: 1) apply it within treatments (there are only 4 samples per Trt/Rev combination) 2) Sum each polynomial component over treatments and subtract results from 3)

EXCEL Calculations of Polynomial SS


All Trts(Rev) All Trts(Rev) Trt1 Trt2 Trt3 Trt4 Trt1 Trt2 Trt3 Trt4 Sum(T1:T4) Sum(T1:T4)-Rev L SS L L L L SS SS SS SS SS SS linear quadratic 137.78 -32.31 10846.83 198.88 161.33 -72.93 141.65 -18.80 132.28 -19.48 115.85 -18.05 3717.97 253.24 2866.39 16.83 2499.53 18.06 1917.32 15.51 11001.20 303.65 154.37 104.77 cubic 1.04 2.87 3.90 -0.27 0.30 0.23 10.14 0.05 0.06 0.03 10.28 7.41 quartic 3.71 1.43 0.80 5.63 1.30 7.12 0.02 0.82 0.04 1.32 2.20 0.77 quintic -0.80 0.12 -2.92 1.60 -2.93 1.05 0.41 0.12 0.41 0.05 0.99 0.87 sextic -9.76 1.65 -22.93 -7.75 -9.93 1.55 2.28 0.26 0.43 0.01 2.97 1.32

Note the second and last rows (of the numeric values) give the polynomial sums of squares for the factors Rev, and Trt x Rev, respectively.