Identifying the Split-plot and Constructing an Analysis George A. Milliken Department of Statistics Kansas State University
2 Complex Split-plot Designs 2.Often used but Not Recognized Designs 3. Often Miss or Inappropriately Analyzed Could Spend several Hours Describing and Discussing Complex Split Plot Designs I will use an Example to Demonstrate some of the Ideas Involved 1.Very Useful Efficient Designs
3 Hydrothermal Processing of Wheat Gluten Slurry at 3 concentrations---10% 14% 18% Path --- long or short (time in cooker) Temp F of cooker Drying methods -- Air (room temp), Hot (heated) Measure solubility--put sample of the part into a flask of water and measure Time to dissolve IN SECONDS; Four Replications of 36 Treatment Combinations
4 Time in Seconds for product to dissolve for SHORT path. PATH=SHORT TEMP=250TEMP=275TEMP=300 REPCONCHOTAIRHOTAIRHOTAIR
5 Time in Seconds for product to dissolve for Long path. PATH=LONG TEMP=250TEMP=275TEMP=300 REPCONCHOTAIRHOTAIRHOTAIR
6 Analysis of Variance Results
7 Conclusions from AOV Significant Concentration by Temperature Interaction Estimate of Variance is Compare the Conc*Temp Cell Means
8
9
10
11 Response Surface Model Since Levels of Concentration and Temperature are Quantitative, fit RESPONSE SURFACE type model using Path and Dry as Categorical variables
12 Final Response Surface Model
13 Conditions with Maximum Response PATHDRYCONCTEMP EST MAX RESPONSE SHORTHOT SHORTAIR LONGHOT LONGAIR GRAPHICS FOLLOW WITH 95% CI CONTAIN MAX
14
15
16
17
18 How was the experiment executed? Part 1 Slurry at 3 concentrations---slurry tank 10% 14% 18% Make a tank of Slurry using one of the concentrations Do this in Random Order – Obtain four Replications of each concentration---- Completely Randomized Design Tank is the Experimental Unit for levels of Slurry—the entity to which levels of Slurry are Randomly Assigned
19 Graphical Representation of The Experiment – Tank as EU Slurry Concentration 10%14%18% Tank 1Tank 2Tank 3Tank 4Tank 5Tank 6 Completely Randomized Design
20 Tank Level of Analysis SourcedfDivisor Concentration3Error(Tank) Error(Tank)9
21 How was the experiment executed? Part 2 TANK is BLOCK of Six BATCHES Take Six BATCHES from TANK--apply the Six Combinations of PATH*TEMP to the BATCHES RANDOMLY assign Combinations of PATH*TEMP to the Six BATCHES from each TANK BATCH is EXPERIMENTAL UNIT for combinations of PATH*TEMP BATCH Design is Randomized Complete Block where TANK is the Blocking Factor
22 Graphical Representation of The Experiment – Batch as EU SHORTLONG Path by Temperature Combinations Batches TANK Batches TANK 1 … Each Tank is a Block of Six Batches for levels of Path by Temperature Combinations
23 BATCH Level of Analysis SourcedfDivisor Blocks=Tanks11 Path1Error(BATCH) Temp2Error(BATCH) Path*Temp2Error(BATCH) Conc*Path2Error(BATCH) Conc*Temp4Error(BATCH) Conc*Temp*Path4Error(BATCH) Error(BATCH)45
24 Graphical Representation of The Experiment – Part as EU Batch Batch(Tank) is Block of Two Parts – for levels of DRY AIRHOT DRY METHOD TANK PART
25 PART Analysis SourceDfDivisor Blocks=Batches71Error(Part) Dry1Error(Part) Conc*Dry2Error(Part) Path*Dry1Error(Part) Temp*Dry2Error(Part) Path*Temp*Dry2Error(Part) Conc*Path*Dry2Error(Part) Conc*Temp*Dry4Error(Part) Conc*Temp*Path*Dry4Error(Part) Error(Part)54
26 Appropriate Model Includes Factorial Effects for Levels of Conc x Path x Temp x Dry Three Sizes of Experimental Units, each with an ERROR TERM 1TANK 2BATCH 3PART
27 Analysis of Variance for Split-plot ns
28 Estimates of the Variance Components for Split-plot Sum of Variance Component Estimates = Same as CR Estimate of Variance
29 Comparisons of Split-plot and CRD analyses Using Split-plot Error Structure Discovered Conc*Temp*Path*Dry interaction Exists in the Data Set CRD analysis found Conc*Temp interaction Significant while split-plot analysis didn’t CRD analysis pools the three error terms together and the resulting error is not appropriate for any of the comparisons
30 Response Surface Model with Split-plot Errors--AOV
31 Response Surface Model with Split-plot Errors
32 Conditions with Maximum Response PATHDRYCONCTEMP EST MAX RESPONSE SHORTHOT SHORTAIR LONGHOT LONGAIR GRAPHICS FOLLOW WITH 95% CI CONTAIN MAX
33
34
35
36
37 Comparisons of 95% Confidence Regions for Maximum Response Path=Short Dry=Hot
38 Comparisons of Split-plot and CRD Response Surface Models Split-plot Response Surface Model is more complex Many more relationships are occurring than discovered using CRD Predicted Response Surface Sweet spots are larger for Split-plot than for CRD
39Conclusions-1 Ignoring the error structure can provide a different response surface model Ignoring the error structure will provide the illusion that there is a smaller sweet spot in the surface Incorporating the split-plot error structure into the model provides appropriate tests, comparisons, resulting model and sweet spot
40 Conclusions -2 Failure to identify the appropriate Design Structure and use it in the modeling process CAN LEAD TO VERY MISLEADING RESULTS Acknowledgments: Departments of Grain Science and Agricultural and Biological Engineering for the experiment Version 8 of PROC MIXED of the SAS® System
41 SAS System Code for ANOVA proc mixed cl DATA=TIME ; class rep conc path temp dry; title 'Model using the split-split-plot error treated as aov with means'; model time=conc|path|temp|dry; random rep(conc) path*temp*rep(conc); lsmeans path*dry*temp conc*path*dry conc*temp/diff;
42 SAS System Code for RSM proc mixed cl data=time; class rep xconc xtemp path dry ;**xconc=conc and xtemp=temp; title 'Final regresson model using split-split-plot error structure'; model time=conc conc*conc temp conc*temp conc*conc*temp path dry conc*dry conc*temp*dry path*dry conc*path*dry conc*conc*path*dry temp*conc*path*dry temp*temp*conc*path*dry /solution SINGULAR=1e-11 ddfm=KR outpm=pred; random rep(xconc) path*xtemp*rep(xconc);
43 THE END THANK YOU FOR YOUR ATTENTION