Split Plots - UMN Statistics

1 Split PlotsGary W. OehlertSchool of StatisticsUniversity of MinnesotaNovember 1, 2014 What is a Split plot ? Split Plots are designs for factorial treatment are useful when we want to vary one or more of the factorsless often than the other factors ( , expensive to change, timeconsuming to change, logistically challenging to change, can onlybe applied to large units, etc).There are several ways to think about Split Plots , each useful indifferent example, you are blowing glass art figures and we areinterested in factors that affect fragility. You can set the annealingoven to two different temperatures, and you can make threedifferent sizes of oven takes hours to come to temperature and hours to cooldown. We do not want to change that frequently. Figure size,however, can be changed at we do is randomly assign temperatures to days. Then,within each day, we randomly choose an order for the three sizes :2B:3B:1B:2A:2B:1B:3B:2A:1B:3B:1B:2A:2B: 2B:3B:1A:1B:1B:2B:3A:1B:2B:3B:1In this schematic, A is temperature, B is size, and the littlecolumns represent is assigned to days, and size is assigned to the taskswithin a is nicely balanced, but all tasks within a daymust have thesame oven StructureTerminology of Split Plots comes from in a Split plot have structure.

2 We have big units, called wholeplots. The whole Plots comprise smaller units, called Split a sense, Split Plots are nested in whole our example, days are the whole Plots , and tasks within a dayare the Split randomly assign the levels of one factor to the whole is the whole plot treatment plot treatment factors are the hard-to-vary factors. In ourexample, temperature is the WP treatment each whole plot , you randomly assign the levels of theother factor to Split Plots . This is the Split plot treatment plot treatment factors are the easy-to-vary factors. In ourexample, size is the SP treatment a randomization perspective, whole Plots act like units forthe whole plot treatment a randomization perspective, whole Plots act like blocks forthe Split plot treatment sizes of units (one nested in the other) and tworandomizations. That gives us a Split plot RandomizationA second view of a Split plot is through an equivalent view of assign the treatments (combinations of whole plot andsplit plot treatment factors) to the Split Plots subject to tworestrictions.

3 All Split Plots in the same whole plot get the same level of thewhole plot treatment levels of the Split plot treatment factor occur in eachwhole restricted randomization is equivalent to the tworandomizations of the unit structure view is correct, but often not as insightful as the unitstructure view is most helpful when the whole plot is not physicallyapparent and it s really only the restricted randomization that leadsus to recognize a Split BlocksA Split plot design can also be viewed as an incomplete Plots are the incomplete blocks, and differences between thelevels of the whole plot treatment factor are confounded with block(whole plot ) , the randomization at the whole plot level induces arandom effect at the whole plot level ( , random blocks).We get information about the whole plot treatment factor viainterblock model and analysis for a Split plot are not that that assumes that you know that you have a Split plotexperiment.

4 Deciding that you (or someone else) have a Split plotis probably the hardest a model perspective, we get a random effect for each size ofunit. In effect, the randomization to a unit is represented by arandom effect at that unit we have a random whole plot term and a random Split plotterm (which cannot be distinguished from ordinary error).We cannot distinguish (S) from (Error).Note that this Hasse diagram looks just like the one we saw forcheese designs can lead to the same model can just use lmer() or lme() with a random effect for the wholeplots and proceed as at whole plot level are less precise than those at splitplot level. Similarly, less power at whole plot than two factors. We can have multiple factors at whole plotlevel and/or Split plot design at the whole plot level could be any one of ourblocking designs. RCB is very common at WP do additional balancing at Split plot level.

5 , take a crossover design (replicated LS), then add a second factor at the wholeplot (subject) whole plot factors, one Split plot whole plot factor, two Split plot whole plot factor, two Split plot factors; RCB at the WP design in B (Latin sq), but A randomly applied design in B (Latin sq), but A randomly applied tosubjects in RCB books only talk about Split Plots with whole plot of these books use a model of random blocks that interactwith the whole plot factor and the Split plot factor. This isnotthesame as what I have books tend to have an engineering orientation, so I call thisthe industrial Split plot don t use this Split plot designsOnce you have the idea of splitting units into smaller units, youcan Split more than Split Split plot has three sizes of units: whole Plots that are madeup of Split Plots which are made up of Split Split levels of nesting in the unit structure: Split Split Plots nestinto Split Plots , and Split Plots nest into whole need at least three factors: a whole plot treatment factor, asplit plot treatment factor, and a Split Split plot treatment : 5 B: 2 {5:2:2 C: 25:2:3 C: 35:2:1 C: 1B: 1 {5:1:1 C: 15:1:2 C: 25:1:3 C: 3B: 3 {5:3:3 C: 35:3:2 C: 25:3:1 C: 17 by 3 by 3 Split Split plot .}}}

6 This whole plot received level 5 offactor A; the three Split Plots and nine Split Split Plots are assignedas three levels of randomization and three sizes of units, we getthree random terms: one for whole Plots , one for Split Plots , andone for Split Split Plots (indistinguishable from error).We can have various kinds of blocking at the whole plot can have more than one factor at each randomization the randomization! Counting factors is not a way todistinguish between Split plot designs and Split Split plot designs(or even CRD). Split Split plot with CRD at WP blocks/strip Plots you get the idea of splitting (nesting) units, you could go allthe way to a Split Split Split Split Split plot if you wanted. I don tthink I ve seen beyond Split Split plot in the , we now have unit structure. We have seen nesting unitscross? Yes, they can build designs with unit structures that have nesting,crossing, or both.

7 Then we layer the treatment structure on top ofthat!Randomly apply three different primersto three horizontal apply four different paints tofour vertical horizontal units cross the vertical units on the same apply three different varietiesto horizontal applytwo different fertilizers to the two hori-zontal apply four irrigation levels tofour vertical have a blocked Split plot in the horizontal units and an RCB inthe vertical units, and the vertical units cross the horizontal typically need replication in blocks for this to work , blocks were the walls or the large chunks of basic model is to have a random effect for each kind of unit(randomization) and wherever units block, also called strip plot . Ignore paint, it s an RCB onprimer; ignore primer, it s an RCB on plot crossing an RCB. Ignore irrigation, it s a Split plot in Vand F. Ignore F, it s a strip plot in I and MeasuresRepeated measures look like Split Plots , but there is norandomization at the Split plot the Split plot treatment factor is time, and withrepeated measures we just keep measuring the same unitrepeatedly over does not like to be randomized,1so it s not a Split version arises when we can measure the same thingmultiple ways.

8 We literally just get multiple generic Dr. Who the repeated measures terminology: Whole Plots are called subjects. Whole plot treatment factors are called grouping factors. Split plot treatment factors are called trial our example, we prepare emulsions using three differentemulsifiers. We then measure each separate emulsion over emulsion is the subject. The emulsifiers form the groupingfactor. Time is the trial of looks like a Split plot , but no is happening is that we have experimented at the subjectlevel, but we observe a vector of responses across the trial vector of responses is probably correlated, not kind of correlation is potentially present among units we usein experimentation, but randomization of treatments to unitsscrambles the correlation to the point it can usually be , no randomization, no scrambling; the correlation comesthrough unaltered and potentially affecting approaches:1 Full multivariate univariate the Full multivariate analysis.

9 This requires a lot of data to workwell and many techniques we have not discussed. Take Stat 5401if you are interested in this Univariate summaries. Here you create some kind of statisticfrom the trial data for each subject, for example, the rate ofchange over time. You then treat this as the response for a subjectand do standard analysis. By looking at different summaries youcan examine different aspects of trial factor summaries are a legitimate approach, but you need tochoose the right summary (or summaries), and you have to figureout the relationship if you have more than one Univariate analysis approach. This approach says assume thereis a random subject effect and that this effect interacts with everytrial factor. With just a single trial factor this is equivalent to thestandard Split plot nature has been very kind to you and the data at the trial levelhave a covariance that satisfies a special condition, then theunivariate approach is the trial factor has only two levels, then the univariate approachis always you were unlucky and didn t get the special form of covariance,then tests at the trial factor level tend to be special condition (the Huynh-Feldt condition) is that alldifferences of repeated measures have the same case that satisfies the HF condition is sphericity: all variancesare the same and all correlations between trial levels within asubject are the same (the correlations don t have to be zero).

10 For multiple trial factors there is a generalization of sphericitycalled compound is a Mauchly Test for the HF condition, but it isverydependent on Modified univariate analysis. The modifications are for the treat it like a Split plot approach with old school mixed effectsanalysis. The modifications adjust the tests in an attempt to makethem less liberal (but not conservative).There is a Greenhouse-Geisser adjustment and a Huynh-Feldtadjustment. Both of these reduce the error DF for trial level testsby some factor estimated from the Model the correlation. The approach is possible with REML computations; it models and estimates the correlation, and thentakes the correlation into generally anticipate positive autocorrelation over time betweenthe observations for a single subject (separate subjects still beingindependent). There are many potential models for this, butautoregressive of order one (AR1) is the simplest and mostcommon.