How To: Analyze a Split-Plot Design

Using STATGRAPHICS Centurion

by

Dr. Neil W. Polhemus

August 13, 2005

Introduction
When performing an experiment involving several factors, it is best to randomize the order in
which the experiments are performed. This reduces the chance that any unexpected effects, such
as a gradual change in some uncontrolled variable over time, will bias the results of the
experiment. Randomization protects the experimenter against the effects of “lurking variables”,
which are factors that affect the experiment but are not recognized until after the experiment is
completed (if they ever are). Randomization also insures that all experimental factors are
subjected to the same level of experimental error, which simplifies the type of analysis that needs
to be performed on the results.

In some cases, pure randomization is not practical, since certain factors may be hard to change.
For example, when experimenting with a chemical process, it may be very easy to change the
initial temperature by turning a dial, but changing a factor such as the type of catalyst may be
considerably more involved. In such cases, experimental designs such as the split-plot design
provide economical alternatives to full randomization. Unfortunately, the restricted
randomization complicates the statistical analysis that must be performed on the resulting data.

This “How To” guide shows how STATGRAPHICS Centurion can be used to analyze typical
split-plot designs. Two examples are considered, one involving categorical experimental factors
and the other involving quantitative factors.

Example #1
The first example we will consider comes from the latest edition of Statistics for Experimenters,
second edition (Wiley, 2005) by Box, Hunter and Hunter. It is an experiment designed to study
the corrosion resistance of steel bars that have been treated with four different coatings. The bars
have been randomly positioned in a furnace and baked at three different temperatures. Although
the position of the bars in the furnace could be randomized, multiple experiments involving a
particular temperature needed to be conducted at the same time, since it was impractical to
change the temperature of the furnace for each sample.

The following table shows the layout of the experiment. Six experimental runs were performed,
each at a selected temperature. During each run, four bars were baked, one bar with each coating.
The position of the bars in the oven was randomly determined for each run. However, the
temperature was changed in a systematic manner:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 1

 Run Temperature Coating of Bar by Position in Furnace
1 360 C2, C3, C1, C4
2 370 C1, C3, C4, C2
3 380 C3, C1, C2, C4
4 380 C4, C3, C2, C1
5 370 C4, C1, C3, C2
6 360 C1, C4, C2, C3
Figure 1: Experimental Design

The results of the experiment have been placed in a file called howto11.sf6, which has the
structure shown below:

Temperature Replicate Coating Corrosion

360 1 2 73
360 1 3 83
360 1 1 67
360 1 4 89
370 1 1 65
370 1 3 87
370 1 4 86
370 1 2 91
380 1 3 147
380 1 1 155
380 1 2 127
380 1 4 212
380 2 4 153
380 2 3 90
380 2 2 100
380 2 1 108
370 2 3 150
370 2 1 140
370 2 3 121
370 2 2 142
360 2 1 33
360 2 4 54
360 2 2 8
360 2 3 46
Figure 2: Experimental Results in File Howto11.sf6

The goal of the experiment is to determine the effect of the type of coating and the temperature
on corrosion.

Step 1: Plot the Data

The first step when analyzing any new data set is to plot it. In this case, a coded scatterplot is
very useful.

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 2

Procedure: X-Y Scatterplot

To plot the experimental data, let’s begin by pushing the X-Y Scatterplot button on the main
toolbar. On the data input dialog box, indicate the variables to be plotted on each axis as shown
below:

Figure 3: Data Input Dialog Box for X-Y Scatterplot

The resulting plot shows a general increase in corrosion with increasing temperature:

Plot of Corrosion vs Temperature

240
200
Corrosion

160
120
80
40
0
355 360 365 370 375 380 385
Temperature
Figure 4: X-Y Scatterplot for Chlorine Data

To introduce the type of coating into the plot, double-click on the graph to enlarge it and press
the Pane Options button on the analysis toolbar. This will display the following dialog box:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 3

 Figure 5: X-Y Plot Options Dialog Box

Enter Coating in the Point Codes field to generate a coded scatterplot:

Plot of Corrosion vs Temperature

240 Coating
1
200
2
Corrosion

160 3
4
120
80
40
0
355 360 365 370 375 380 385
Temperature
Figure 6: Coded X-Y Scatterplot

Each type of point symbol represents a different coating. At each of the 3 temperatures, there are
2 replicates for each type of coating. Based on this plot, one would be hard pressed to select the
best type of coating.

Step 2: Analyze the Data

The corrosion experiment was conducted to determine the effect of two factors, temperature and
type of coating, on a single response variable, corrosion. As in a standard factorial experiment,
data has been collected at all combinations of the levels of the two experimental factors. What
© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 4
makes this experiment special is that the size of the experimental unit for one factor is different
than for the other, causing the experimental error affecting the estimates of the factor effects to
be different. In particular, the furnace temperature was set only six times, corresponding to the
six rows in Figure 1. Each of these rows is called a “whole-plot”, with analogies to early
agricultural experiments in which one type of fertilizer might be applied to an entire field. Within
each whole-plot, the coatings are randomly allocated to positions within the furnace. Position is
referred to as the “subplot”.

There are two sources of variability in this experiment: variability in bringing the furnace to a
particular temperature, and variability amongst positions within the furnace. The variance of the
first source will be labeled σ W2 , while the variance of the second will be labeled σ S2 . Since
coatings were randomized across the subplots, estimates of differences between the coatings are
subject only to the subplot error, while estimates of the temperature effects involve both the
subplot and whole-plot error.

The secret to analyzing this split-plot experiment is to view it as two experiments, one contained
within the other. The whole-plot experiment involves changes in temperature. It can be viewed
as a one-way comparison amongst the three levels of temperatures, with two replicates at each
level:

T1: 360ºC T2: 370ºC T3: 380ºC

R1,R2 R1,R2 R1,R2

Each replicate represents a “whole-plot” or run of the furnace. A standard ANOVA table for the
whole-plot design would look like:

Source of Variation Degrees of Freedom

Temperature 2
Error (Replicates) 3

The second experiment crosses coating with the whole-plots:

T1 R1 T2 R1 T3 R1 T1 R2 T2 R2 T3 R2
C1 Y Y Y Y Y Y
C2 Y Y Y Y Y Y
C3 Y Y Y Y Y Y
C4 Y Y Y Y Y Y

One bar was baked at each of the 4×6 combinations of coating and whole-plot. The terms in the
ANOVA table for this experiment look like:

Source of Variation Degrees of

Freedom
Whole-Plots 5
Coating 3
Temperature×Coating 6
Error 9

Tests of significance for the terms in each experiment are done by comparing them to the
corresponding error component.
© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 5
Procedure: General Linear Models

A fully randomized factorial design, in which there is only one source of experimental error, is
usually analyzed using the Multifactor ANOVA procedure. When there is more than one source
of error, the General Linear Models procedure must be used instead. This is accessed from the
main STATGRAPHICS Centurion menu by selecting:

• If using the Classic menu: Compare – Analysis of Variance – General Linear Models.
• If using the Six Sigma menu: Improve – Analysis of Variance – General Linear Models.

The data input dialog box is shown below:

Figure 7: Data Input Dialog Box for General Linear Models

Although not strictly necessary, it is helpful to enter the factors in the order shown:

1. Whole-plot factors before subplot factors.

2. Replicate after Temperature, since the replicates are performed at each temperature level.

After completing the first dialog box, a second dialog box is displayed on which to specify the
statistical model. It should be completed as shown below:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 6

Figure 8: Model Specification Dialog Box for General Linear Models

Note the following:

1. The main effect of the whole-plot factor Temperature is specified by placing the single
letter A in the Effects field. Since specific values of temperature were used (360, 370,
380), it is a fixed rather than a random factor.

2. Replicate is entered using the notation B(A). This indicates that replicate (Factor B) is
nested within temperature (Factor A). Specifying the factors as nested indicates that the
experimental unit for the first replicate of T1 is not the same as the experimental unit for
the first replicate of temperature T2 or T3. Indeed, each combination of T and R forms a
separate whole-plot. Factor B is also specified to be a random factor, since the two
replicates are but a small random sample of all replicates that could have been performed.

3. After entering the whole-plot terms, the subplot effects are specified. First, main effects
of Coating are requested using the symbol C. Then the Temperature×Coating interaction
is entered as A*C. Since four specific coatings are being tested, Coating is also a fixed
rather than a random factor.

Pressing OK causes the specified model to be fit. The Analysis Summary pane summarizes the
fitted model. The top section of that summary is shown below:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 7

 General Linear Models
Number of dependent variables: 1
Number of categorical factors: 3
A=Temperature
B=Replicate
C=Coating
Number of quantitative factors: 0

Analysis of Variance for Corrosion

Source Sum of Squares Df Mean Square F-Ratio P-Value
Model 48517.8 14 3465.55 27.83 0.0000
Residual 1120.88 9 124.542
Total (Corr.) 49638.6 23

Type III Sums of Squares

Source Sum of Squares Df Mean Square F-Ratio P-Value
Temperature 26519.3 2 13259.6 2.75 0.2093
Replicate(Temperature) 14439.6 3 4813.21 38.65 0.0000
Coating 4289.13 3 1429.71 11.48 0.0020
Temperature*Coating 3269.75 6 544.958 4.38 0.0241
Residual 1120.88 9 124.542
Total (corrected) 49638.6 23
Figure 9: GLM Analysis Summary – Top Section

The most important information in the above table is in the section labeled Type III Sums of
Squares. The rightmost column of that table contains a P-Value for each term in the model. P-
Values less than 0.05 indicate effects that are statistically significant at the 5% significance level.
Both the main effect of Coating and the Temperature×Coating interaction are significant.

Also included in the Analysis Summary is the table shown below:

F-Test Denominators
Source Df Mean Square Denominator
Temperature 3.00 4813.21 (2)
Replicate(Temperature) 9.00 124.542 (5)
Coating 9.00 124.542 (5)
Temperature*Coating 9.00 124.542 (5)

Variance Components
Source Estimate
Replicate(Temperature) 1172.17
Residual 124.542
Figure 10: GLM Analysis Summary – Bottom Section

The F-Test denominators indicate which line in the ANOVA table has been used to test the
significance of each effect. For Temperature, the (2) indicates that it has been compared against
the whole-plot error on the second line of the ANOVA table (labeled Replicate(Temperature)).
The other factors have been compared to the Residual or subplot error in line 5.

Also shown are the estimates of the error components:

Whole-plot error variance: σˆ W2 = 1172 .17

Subplot error variance: σˆ S2 = 124.542

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 8

Notice that the subplot error, which represents differences between positions in the furnace, is
much smaller than the whole-plot error, which represents differences between runs of the
furnace. From the discussion in Box, Hunter and Hunter, it seems that this is evidence of
difficulties in maintaining the desired temperature when an experiment is run.

Step 3: Display the Results

Once the important factors have been identified, it is useful to display the estimated effects
graphically. To plot the main effects of each factor, press the Graphs button on the analysis
toolbar. If you select Means Plot, the mean response will be plotted at each level of a selected
factor. Using Pane Options, you can also choose an uncertainty interval to place around each
mean. In the plot below, the mean value of Corrosion is shown for each of the three
temperatures, together with Tukey’s HSD intervals:

Means and 95.0 Percent Tukey HSD Intervals

220
180
Corrosion

140
100
60
20
-20
360 370 380
Temperature
Figure 11: Means Plot for Temperature

Since all of the intervals overlap, we can not declare any means to be significantly different from
any other means, which matches the insignificant result for the Temperature main effects in the
ANOVA table. Using Pane Options to switch factors displays the following:

Means and 95.0 Percent Tukey HSD Intervals

140
130
Corrosion

120
110
100
90
80
1 2 3 4

Coating
Figure 12: Means Plot for Coating

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 9

In this plot, the intervals for the first three coatings all overlap, indicating no significant
differences amongst them. However, the uncertainty interval for Coating 4 does not overlap the
other intervals, indicating a significant difference between Coating 4 and the other 3 coatings.

We can also select Interaction Plot from the Graphs dialog box to display the interaction
between Temperature and Coating:

Interaction Plot
200 Coating
1
160 2
Corrosion

3
120
4
80

40

0
360 370 380
Temperature
Figure 13: Interaction Plot for Temperature×Coating

Each point on the interaction plot shows the mean value for a specific combination of
Temperature and Coating. Note that the difference between Coating 4 and the others is more
pronounced at 380º than it is at the other temperatures. Also, Coating 2, which gives lowest
Corrosion at 360º and 380º, is not the lowest at 370º.

We can also use Pane Options to add uncertainty intervals to the interaction plot:

Interactions and 95.0 Percent Tukey HSD Intervals

240 Coating
1
200
2
Corrosion

160 3
4
120
80
40
0
360 370 380
Temperature
Figure 14: Interaction Plot with Uncertainty Intervals

It appears that the only temperature at which the coatings are significantly different is 380º, since
all of the intervals overlap at the other temperatures.

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 10

Example #2
The second example comes from Design and Analysis of Experiments, sixth edition by Douglas
Montgomery (Wiley, 2005). It is an example of an experiment performed with a single wafer
plasma etching process. The goal of the experiment was to determine how Uniformity changed as
a function of 5 factors:

A = electrode gap
B = gas flow
C = pressure
D = time
E = RF power

The first 3 factors were relatively hard to change, while the last two could be changed easily.

An excellent design for studying the effects of 5 factors is the 25-1 fractional factorial design,
consisting of 16 runs at different combinations of two levels of each factor. The design is
resolution V, meaning that it can estimate clearly all main effects and two-factor interactions. In
running such a design, the 16 runs are ordinarily done in random order. Unfortunately, this
means that all of the factors will be changed frequently, which could be expensive and time-
consuming.

Montgomery describes an approach that reduces the number of times the hard-to-change factors
need to be changed. He suggests arranging the 16 runs in the 25-1 design as a split-plot design. In
his approach, each of the 8 combinations of the 3 hard-to-change factors defines a whole-plot.
Each time a whole-plot is created, two different combinations of factors D and E are tested. The
layout of the design is shown below:

Whole- A B C D E Uniformity
Plots
1 - - - - + 40.85
- - - + - 41.07
2 + - - - - 35.67
+ - - + + 51.15
3 - + - - - 41.80
- + - + + 37.01
4 + + - - + 91.09
+ + - + - 48.67
5 - - + - - 40.32
- - + + + 43.34
6 + - + - + 62.46
+ - + + - 38.08
7 - + + - + 31.99
- + + + - 41.03
8 + + + - - 70.31
+ + + + + 81.03
Figure 15: 25-1 Split-Plot Experiment

Each of the 16 rows of the table represents an experiment that was performed. The – and + signs
indicate whether a factor was run at its low or high level during a particular run. For example,
© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 11
during the first run, factors A, B, C and D were run at their low levels, while factor E was run at
its high level. The measured Uniformity for that run was 40.85. The eight whole-plots were run
in random order. Once the whole-plot conditions were set, the two runs in the whole-plot were
performed, also in random order. With this approach, the hard-to-change factors A, B and C
were changed less frequently then the easy-to-change factors D and E.

Step 1: Construct the Design

To construct the above experiment using STATGRAPHICS Centurion, the Design of
Experiments section can be used. From the main menu, begin by selecting:

• If using the Classic menu: DOE – Design Creation – Create New Design.
• If using the Six Sigma menu: Improve – Experimental Design Creation – Create New
Design.

On the first dialog box, select the Screening option (this includes the two-level factorials) and
indicate that 5 experimental factors are to be varied:

Figure 16: First Design Creation Dialog Box

On the second dialog box, specify the names of the 5 experimental factors:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 12

Figure 17: Second Design Creation Dialog Box

Select each button A through E and enter the following names:

A: electrode gap
B: gas flow
C: pressure
D: time
E: RF power

Since the low and high levels of each factor were not stated in Montgomery’s example, leave the
levels at -1.0 and 1.0.

On the third dialog box, specify a label for the response variable:

Figure 18: Third Design Creation Dialog Box

The fourth dialog box shows a list of screening designs for 5 factors:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 13

Figure 19: Fourth Design Creation Dialog Box

The design desired is the Half fraction. It has 16 runs and is Resolution V. Note that 0 degrees of
freedom are available to estimate the experimental error. This means that all 16 runs will be used
to estimate the average, the main effects, and the two-factor interactions. To obtain degrees of
freedom for estimating the experimental error, additional runs could be added on the final dialog
box:

Figure 20: Fifth Design Creation Dialog Box

Most commonly, centerpoints would be added at values of each factor positioned halfway
between the lows and the highs. In Montgomery’s example, however, no centerpoints were used,
so no formal statistical tests of significance will be possible.

After the final dialog box, the design will be created and placed into the STATGRAPHICS
datasheet:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 14

Figure 21: Datasheet Containing Generated Design

The order of the design differs somewhat from that shown in Figure 15. In particular, the 8
desired whole-plots consist of the following rows:

Whole-plot #1: runs 1 and 9

Whole-plot #2: runs 2 and 10
Whole-plot #3: runs 3 and 11
Whole-plot #4: runs 4 and 12
Whole-plot #5: runs 5 and 13
Whole-plot #6: runs 6 and 14
Whole-plot #7: runs 7 and 15
Whole-plot #8: runs 8 and 16

The two runs in each whole-plot have identical values for factors A, B and C, but the levels of
factors D and E are different. As stated earlier, the 8 whole-plots would be run in random order.
After selecting a whole-plot, the two runs would also be done in random order.

Step 2: Analyze the Results

To analyze the results of the experiment, enter the Uniformity values in the rightmost column as
shown in the above datasheet. Then select from the main STATGRAPHICS Centurion menu:

• If using the Classic menu: DOE – Design Analysis – Analyze Design.

• If using the Six Sigma menu: Improve – Experimental Design Analysis – Analyze Design.

Specify the response variable on the data input dialog box:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 15

Figure 22: Analyze Design Data Input Dialog Box

An analysis window will be created, containing several tables and graphs. Of particular interest
is the Pareto Chart, which plots the magnitude of each main effect and interaction in decreasing
order:

Pareto Chart for Uniformity

A:electrode gap
AB +
AE
B:gas flow -
E:RF power
AD
D:time
CD
AC
CE
BD
C:pressure
BC
DE
BE
0 4 8 12 16 20 24
Effect
Figure 23: Pareto Chart of Estimated Effects

The largest effects appear to involve factors A, B and E and some of their interactions.

Since the whole-plot factors are subject to a different experimental error than the subplot factors,
the results need to be analyzed in two groups. To analyze the whole-plot experiment:

1. Press the Analysis Options button on the analysis toolbar.

2. On the Analysis Options dialog box, press the Exclude button.
3. On the Exclude Effects Options dialog box, double-click on any effect involving either
subplot factor D or E to exclude it from the model:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 16

 Figure 24: Excluding Subplot Effects

4. Press OK twice to refit the model using only factors A, B and C.

5. Press the Graphs button on the analysis toolbar and select Normal Probability Plot
of Effects.
6. When the normal probability plot appears, double-click on it to maximize its pane and
then press the Pane Options button on the analysis toolbar. Complete the options
dialog box as shown below to request a Half-Normal Plot and to Label Effects:

Figure 25: Normal Probability Plot Options

The above steps will create a normal probability plot that can be used to help determine which
effects are statistically significant:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 17

 Half-Normal Plot for Uniformity
1.8

Standard deviations
A:electrode gap
1.5
1.2 AB
0.9 B:gas flow
0.6 AC
0.3 C:pressure
BC
0
0 1 2 3 4
Standardized effects
Figure 26: Half-Normal Plot for the Whole-Plot Experiment

The estimated effects that are merely manifestations of noise will appear at the bottom left and
should lie approximately along a straight line. The estimated effects corresponding to real signals
will lie off the line toward the upper right. In this case, it appears that both electrode gap and gas
flow are statistically significant and that they interact. Pressure does not seem to have a
significant effect.

NOTE: the normal probability plot is the only means of determining statistical
significance for the whole-plot factors in this experiment. Although the ANOVA table
shows P-Values for each effect, they are based on the subplot error which does not apply
to the whole-plot factors. Had degrees of freedom been available to estimate the whole-
plot error, perhaps from the inclusion of several runs with replicated centerpoints, then
the average result in each whole-plot could have been calculated and an analysis of
variance performed on the whole-plot averages.

It is also useful in this case to display an interaction plot for factors A and B. This may be done
by:

1. Press the Graphs button on the analysis toolbar and select Interaction Plots.
2. When the plot appears, double-click on it to maximize its pane.
3. Press the Pane Options button and select only electrode gap and gas flow.

The resulting plot is shown below:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 18

 Interaction Plot for Uniformity
77
gas flow=1.0
67

Uniformity
57

47 gas flow=-1.0
gas flow=-1.0
37 gas flow=1.0
-1.0 1.0
electrode gap

Figure 27: Interaction Plot for Electrode Gap and Gas Flow

At the low level of electrode gap, gas flow makes little difference. However, at the high level of
electrode gap, gas flow makes a big difference.

It is also useful to plot the estimated model for uniformity as a function of the two important
factors. The plot below was created by selecting Response Plots from the Graphs dialog box. It
plots the estimated corrosion as a function of electrode gap and gas flow, with the other factors
held halfway between their low and high values:

Estimated Response Surface

pressure=0.0,time=0.0,RF power=0.0

Uniformity
77 35.0
Uniformity

45.0
67
55.0
57 65.0
47 75.0
1
0.5
37 0
-1 -0.5
-0.6 -0.2 0.2 -1
0.6 1 gas flow
electrode gap

Figure 28: Uniformity versus Electrode Gap and Gas Flow

The strong interaction between the factors causes the surface to twist dramatically, resulting in
unusually high Uniformity when both electrode gap and gas flow are at their high levels.

The same analysis can now be repeated on the subplot experiment. Returning to the Exclude
Effects Options dialog box, we now create a model with any term (main effect or interaction) that
involves a subplot factor:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 19

Figure 29: Excluding Whole-Plot Effects

The resulting normal probability plot shows that factor E (RF power) is significant, as is its
interaction with factor A (electrode gap):

Half-Normal Plot for Uniformity

2
Standard deviations

AE
1.6
E:RF power
1.2
AD
D:time
0.8
CD
CE
0.4 BD
DE
BE
0
0 0.2 0.4 0.6 0.8 1 1.2
Standardized effects
Figure 30: Half-Normal Plot for Subplot Effects

Again, an interaction plot displays the essential information:

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 20

 Interaction Plot for Uniformity
78
RF power=1.0
68

Uniformity
58

48 RF power=-1.0

RF power=-1.0
38 RF power=1.0
-1.0 1.0
electrode gap

Figure 31: Interaction Plot for Electrode Gap and RF Power

As with gas flow, RF Power appears to have an effect only at the high level of electrode gap.

Conclusion
When the order of experiments cannot be fully randomized, a split-plot design is often useful.
Those factors that cannot be changed as easily are varied across large experimental units called
whole-plots, while the easily changed factors are varied across the subplots. Usually, the
experimental error amongst subplots is considerably less than that amongst whole-plots, so that
subplot factors and their interactions can be estimated with greater precision than the main
effects of the whole-plot factors.

As pointed out by the authors of the two books referenced in this guide, it is easy to slip into a
split-plot setup without realizing it. Often, the “logical” way to conduct an experiment prevents
one or more factors from being fully randomized. When randomization is restricted on one or
more factors, the usual tests of significance generated assuming full randomization may make
those factors appear more significant than they really are. For experimenters that face the
problem of not being able to fully randomize the order of their experiments, careful study of
split-plot designs is a must.

Note: The author welcomes comments about this guide. Please address your responses to
neil@statgraphics.com.

© 2005 by StatPoint, Inc. How to Analyze a Split-Plot Design - 21