We have been working with completely randomized designs, whereg treat-ments are assigned at random toNunits. Up till now, the treatments have had
no structure; they were justgtreatments. Factorial treatment structure ex- Factorials combine the levels of two or more factors to create treatments
ists when thegtreatments are the combinations of the levels of two or more factors. We call these combination treatments factor-level combinations or factorial combinations to emphasize that each treatment is a combination of one level of each of the factors. We have not changed the randomization; we still have a completely randomized design. It is just that now we are con-sidering treatments that have a factorial structure. We will learn that there are compelling reasons for preferring a factorial experiment to a sequence of experiments investigating the factors separately.
8.1 Factorial Structure
It is best to start with some examples of factorial treatment structure. Lynch and Strain (1990) performed an experiment with six treatments studying how milk-based diets and copper supplements affect trace element levels in rat livers. The six treatments were the combinations of three milk-based diets (skim milk protein, whey, or casein) and two copper supplements (low and high levels). Whey itself was not a treatment, and low copper was not a treatment, but a low copper/whey diet was a treatment. Nelson, Kriby, and Johnson (1990) studied the effects of six dietary supplements on the occur-rence of leg abnormalities in young chickens. The six treatments were the combinations of two levels of phosphorus supplement and three levels of calcium supplement. Finally, Hunt and Larson (1990) studied the effects of
166 Factorial Treatment Structure
Table 8.1:Barley sprouting data.
Age of Seeds (weeks)
ml H2O 1 3 6 9 12
4 11
9 6
7 16 17
9 19 35
13 35 28
20 37 45 8
8 3 3
1 7 3
5 9 9
1 10
9
11 15 25
sixteen treatments on zinc retention in the bodies of rats. The treatments were the combinations of two levels of zinc in the usual diet, two levels of zinc in the final meal, and four levels of protein in the final meal. Again, it is the combination of factor levels that makes a factorial treatment.
We begin our study of factorial treatment structure by looking at two-factor designs. We may present the responses of a two-way two-factorial as a table
Two-factor
designs with rows corresponding to the levels of one factor (which we call factor A) and columns corresponding to the levels of the second factor (factor B). For example, Table 8.1 shows the results of an experiment on sprouting barley (these data reappear in Problem 8.1). Barley seeds are divided into 30 lots of 100 seeds each. The 30 lots are divided at random into ten groups of three lots each, with each group receiving a different treatment. The ten treatments are the factorial combinations of amount of water used for sprouting (factor A) with two levels, and age of the seeds (factor B) with five levels. The response measured is the number of seeds sprouting.
We use the notationyijkto indicate responses in the two-way factorial.
In this notation,yijkis thekth response in the treatment formed from theith
Multiple
subscripts denote factor levels and replication
level of factor A and thejth level of factor B. Thus in Table 8.1,y2,5,3 = 25.
For a four by three factorial design (factor A has four levels, factor B has three levels), we could tabulate the responses as in Table 8.2. This table is just a convenient representation that emphasizes the factorial structure; treatments were still assigned to units at random.
Notice in both Tables 8.1 and 8.2 that we have the same number of re-sponses in every factor-level combination. This is called balance. Balance
Balanced data have equal replication
turns out to be important for the standard analysis of factorial responses. We will assume for now that our data are balanced with n responses in every factor-level combination. Chapter 10 will consider analysis of unbalanced factorials.
8.2 Factorial Analysis: Main Effect and Interaction 167
Table 8.2:A two-way factorial treatment structure.
B1 B2 B3
A1
y111
... y11n
y121
... y12n
y131
... y13n
A2
y211 ... y21n
y221 ... y22n
y231 ... y23n
A3
y311 ... y31n
y321 ... y32n
y331 ... y33n
A4
y411 ... y41n
y421 ... y42n
y431 ... y43n
8.2 Factorial Analysis: Main Effect and Interaction
When our treatments have a factorial structure, we may also use a factorial analysis of the data. The major concepts of this factorial analysis are main effect and interaction.
Consider a two-way factorial where factor A has four levels and factor B has three levels, as in Table 8.2. There areg= 12treatments, with 11 degrees of freedom between the treatments. We usei andj to index the levels of factors A and B. The expected values in the twelve treatments may be denoted µij, coefficients for a contrast in the twelve means may be denotedwij(where as usualPijwij = 0), and the contrast sum isPijwijµij. Similarly,yij•
is the observed mean in the ij treatment group, andyi•• andy•j• are the Treatment, row, and column means
observed means for all responses having leveliof factor A or leveljof B, respectively. It is often convenient to visualize the expected values, means, and contrast coefficients in matrix form, as in Table 8.3.
For the moment, forget about factor B and consider the experiment to be a completely randomized design just in factor A (it is completely randomized in factor A). Analyzing this design with four “treatments,” we may compute
a sum of squares with 3 degrees of freedom. The variation summarized by Factor A ignoring factor B
this sum of squares is denotedSSAand depends on just the level of factor A.
The expected value for the mean of the responses in rowiisµ+αi, where we assume thatPiαi= 0.
168 Factorial Treatment Structure
Table 8.3:Matrix arrangement of (a) expected values, (b) means, and (c) contrast coefficients in a four by three factorial.
(a) µ11 µ12 µ13 µ21 µ22 µ23 µ31 µ32 µ33 µ41 µ42 µ43
(b)
y11• y12• y13•
y21• y22• y23•
y31• y32• y33•
y41• y42• y43•
(c) w11 w12 w13 w21 w22 w23 w31 w32 w33 w41 w42 w43
Now, reverse the roles of A and B. Ignore factor A and consider the periment to be a completely randomized design in factor B. We have an
ex-Factor B ignoring
factor A periment with three “treatments” and treatment sum of squaresSSBwith 2 degrees of freedom. The expected value for the mean of the responses in columnjisµ+βj, where we assume thatPjβj = 0.
The effects αi and βj are called the main effects of factors A and B, respectively. The main effect of factor A describes variation due solely to the
A main effect describes variation due to a single factor
level of factor A (row of the response matrix), and the main effect of factor B describes variation due solely to the level of factor B (column of the response matrix). We have analogously that SSAandSSBare main-effects sums of squares.
The variation described by the main effects is variation that occurs from row to row or column to column of the data matrix. The example has twelve treatments and 11 degrees of freedom between treatments. We have de-scribed 5 degrees of freedom using the A and B main effects, so there must
Interaction is variation not described by main effects
be 6 more degrees of freedom left to model. These 6 remaining degrees of freedom describe variation that arises from changing rows and columns si-multaneously. We call such variation interaction between factors A and B, or between the rows and columns, and denote it bySSAB.
Here is another way to think about main effect and interaction. The main effect of rows tells us how the response changes when we move from one row to another, averaged across all columns. The main effect of columns tells us how the response changes when we move from one column to an-other, averaged across all rows. The interaction tells us how the change in re-sponse depends on columns when moving between rows, or how the change in response depends on rows when moving between columns. Interaction be-tween factors A and B means that the change in mean response going from leveli1 of factor A to leveli2 of factor A depends on the level of factor B under consideration. We can’t simply say that changing the level of factor A changes the response by a given amount; we may need a different amount of change for each level of factor B.
8.2 Factorial Analysis: Main Effect and Interaction 169
Table 8.4:Sample main-effects and interaction contrast coefficients for a four by three factorial design.
A
-3 -3 -3 -1 -1 -1
1 1 1
3 3 3
1 1 1
-1 -1 -1 -1 -1 -1
1 1 1
-1 -1 -1
3 3 3
-3 -3 -3
1 1 1
B
-1 0 1
-1 0 1
-1 0 1
-1 0 1
1 -2 1
1 -2 1
1 -2 1
1 -2 1
AB
3 0 -3
1 0 -1
-1 0 1
-3 0 3
-1 0 1
1 0 -1
1 0 -1
-1 0 1
1 0 -1
-3 0 3
3 0 -3
-1 0 1
-3 6 -3
-1 2 -1
1 -2 1
3 -6 3
1 -2 1
-1 2 -1
-1 2 -1
1 -2 1
-1 2 -1
3 -6 3
-3 6 -3
1 -2 1
We can make our description of main-effect and interaction variation more precise by using contrasts. Any contrast in factor A (ignoring B) has four coefficientsw⋆i and observed valuew⋆({yi••}). This is a contrast in the four row means. We can make an equivalent contrast in the twelve treatment means by using the coefficientswij = w⋆i/3. This contrast just repeatsw⋆i across each row and then divides by the number of columns to match up with the division used when computing row means. Factor A has four levels,
so three orthogonal contrasts partition SSA. There are three analogous or- Main-effects contrasts
thogonalwij contrasts that partition the same variation. (See Question 8.1.) Table 8.4 shows one set of three orthogonal contrasts describing the factor A variation; many other sets would do as well.
The variation in SSB can be described by two orthogonal contrasts be-tween the three levels of factor B. Equivalently, we can describeSSB with orthogonal contrasts in the twelve treatment means, using a matrix of contrast coefficients that is constant on columns (that is,w1j = w2j = w3j = w4j for all columnsj). Table 8.4 also shows one set of orthogonal contrasts for factor B.
170 Factorial Treatment Structure
Inspection of Table 8.4 shows that not only are the factor A contrasts orthogonal to each other, and the factor B contrasts orthogonal to each other,
A contrasts orthogonal to B contrasts for balanced data
but the factor A contrasts are also orthogonal to the factor B contrasts. This orthogonality depends on balanced data and is the key reason why balanced data are easier to analyze.
There are 11 degrees of freedom between the twelve treatments, and the A and B contrasts describe 5 of those 11 degrees of freedom. The 6 addi-tional degrees of freedom are interaction degrees of freedom; sample inter-action contrasts are also shown in Table 8.4. Again, inspection shows that
Interaction
contrasts the interaction contrasts are orthogonal to both sets of main-effects contrasts.
Thus the 11 degrees of freedom between-treatment sum of squares can be partitioned using contrasts intoSSA,SSB, andSSAB.
Look once again at the form of the contrast coefficients in Table 8.4.
Row-main-effects contrast coefficients are constant along each row, and add to zero down each column. Column-main-effects contrasts are constant down
Contrast coefficients satisfy zero-sum restrictions
each column and add to zero along each row. Interaction contrasts add to zero down columns and along rows. This pattern of zero sums will occur again when we look at parameters in factorial models.
8.3 Advantages of Factorials
Before discussing advantages, let us first recall the difference between facto-rial treatment structure and factofacto-rial analysis. Factofacto-rial analysis is an option
Factorial structure
versus analysis we have when the treatments have factorial structure; we can always ignore main effects and interaction and just analyze thegtreatment groups.
It is easiest to see the advantages of factorial treatment structure by com-paring it to a design wherein we only vary the levels of a single factor. This second design is sometimes referred to as “one-at-a-time.” The sprouting
One-at-a-time
designs data in Table 8.1 were from a factorial experiment where the levels of sprout-ing water and seed age were varied. We might instead use two one-at-a-time designs. In the first, we fix the sprouting water at the lower level and vary the seed age across the five levels. In the second experiment, we fix the seed age at the middle level, and vary the sprouting water across two levels.
Factorial treatment structure has two advantages:
1. When the factors interact, factorial experiments can estimate the inter-action. One-at-at-time experiments cannot estimate interinter-action. Use of one-at-a-time experiments in the presence of interaction can lead to serious misunderstanding of how the response varies as a function of the factors.
8.4 Visualizing Interaction 171
2. When the factors do not interact, factorial experiments are more ef-ficient than one-at-a-time experiments, in that the units can be used to assess the (main) effects for both factors. Units in a one-at-a-time experiment can only be used to assess the effects of one factor.
There are thus two times when you should use factorial treatment structure— Use factorials!
when your factors interact, and when your factors do not interact. Factorial structure is a win, whether or not we have interaction.
The argument for factorial analysis is somewhat less compelling. We usually wish to have a model for the data that is as simple as possible. When there is no interaction, then main effects alone are sufficient to describe the
means of the responses. Such a model (or data) is said to be additive. Additive model has only main effects
An additive model is simpler (in particular, uses fewer degrees of freedom) than a model with a mean for every treatment. When interaction is moderate compared to main effects, the factorial analysis is still useful. However, in some experiments the interactions are so large that the idea of main effects as the primary actors and interaction as fine tuning becomes untenable. For such experiments it may be better to revert to an analysis ofg treatment groups, ignoring factorial structure.
Pure interactive response Example 8.1
Consider a chemistry experiment involving two catalysts where, unknown to us, both catalysts must be present for the reaction to proceed. The response is one or zero depending on whether or not the reaction occurs. The four treat-ments are the factorial combinations of Catalyst A present or absent, and Catalyst B present or absent. We will have a response of one for the com-bination of both catalysts, but the other three responses will be zero. While it is possible to break this down as main effect and interaction, it is clearly more comprehensible to say that the response is one when both catalysts are present and zero otherwise. Note here that the factorial treatment structure was still a good idea, just not the main-effects/interactions analysis.
8.4 Visualizing Interaction
An interaction plot, also called a profile plot, is a graphic for assessing the
rel-ative size of main effects and interaction; an example is shown in Figure 8.1. Interaction plots connect-the-dots between treatment means
Consider first a two-factor factorial design. We construct an interaction plot in a “connect-the-dots” fashion. Choose a factor, say A, to put on the hori-zontal axis. For each factor level combination, plot the pair(i, yij•). Then
“connect-the-dots” corresponding to the points with the same level of factor
172 Factorial Treatment Structure
Table 8.5: Iron levels in liver tissue, mg/g dry weight.
Diet Control Cu deficient
Skim milk protein .70 1.28
Whey .93 1.87
Casein 2.11 2.53
B; that is, connect(1, y1j•), (2, y2j•), up to (a, yaj•). In our four by three prototype factorial, the level of factor A will be a number between one and four; there will be three points plotted above one, three points plotted above two, and so on; and there will be three “connect-the-dots” lines, one for each level of factor B.
For additive data, the change in response moving between levels of factor A does not depend on the level of factor B. In an interaction plot, that simi-larity in change of level shows up as parallel line segments. Thus interaction
Interaction plot shows relative size of main effects and interaction
is small compared to the main effects when the connect-the-dots lines are parallel, or nearly so. Even with visible interaction, the degree of interaction may be sufficiently small that the main-effects-plus-interaction description is still useful. It is worth noting that we sometimes get visually different impressions of the interaction by reversing the roles of factors A and B.
Example 8.2 Rat liver iron
Table 8.5 gives the treatment means for liver tissue iron in the Lynch and Strain (1990) experiment. Figure 8.1 shows an interaction plot with milk diet factor on the horizontal axis and the copper treatments indicated by different lines. The lines seem fairly parallel, indicating little interaction.
Figure 8.1 points out a deficiency in the interaction plot as we have de-fined it. The observed means that we plot are subject to error, so the line
Interpret “parallel”
in light of variability
segments will not be exactly parallel—even if the true means are additive.
The degree to which the lines are not parallel must be interpreted in light of the likely size of the variation in the observed means. As the data become more variable, greater departures from parallel line segments become more likely, even for truly additive data.
Example 8.3 Rat liver iron, continued
The line segments are fairly parallel, so there is not much evidence of inter-action, though it appears that the effect of copper may be somewhat larger for milk diet 2. The mean square for error in the Lynch and Strain experiment was approximately .26, and each treatment had replicationn = 5. Thus the standard errors of a treatment mean, the difference of two treatment means,
8.4 Visualizing Interaction 173
0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4
1 2 3
Milk diet I
r o n
1
1
1
2
2
2
Figure 8.1:Interaction plot of liver iron data with diet factor on the horizontal axis, using MacAnova.
and the difference of two such differences are about .23, .32, and .46 respec-tively. The slope of a line segment in the interaction plot is the difference of two treatment means. The slopes from milk diet 1 to 2 are .23 and .59, and the slopes from milk diets 2 to 3 are 1.18 and .66; each of these slopes was calculated as the difference of two treatment means. The differences of the slopes (which have standard error .46 because they are differences of differences of means) are .36 and .48. Neither of these differences is large compared to its standard error, so there is still no evidence for interaction.
We finish this section with interaction plots for the other two nutrition experiments described in the first section.
Chick body weights Example 8.4
Figure 8.2 is an interaction plot of the chick body weights from the Nelson, Kriby, and Johnson (1990) data with the calcium factor on the horizontal axis and a separate line for each level of phosphorus. Here, interaction is clear. At the upper level of phosphorus, chick weight does not depend on calcium. At the lower level of phosphorus, weight decreases with increasing calcium. Thus the effect of changing calcium levels depends on the level of phosphorus.
174 Factorial Treatment Structure
1 2
3 2
1 600
550
500
450
Calcium
Phosphorus
Mean
Interaction plot --- Data means for weight
Figure 8.2:Interaction plot of chick body weights data with calcium on the horizontal axis, using Minitab.
Example 8.5 Zinc retention
Finally, let’s look at the zinc retention data of Hunt and Larson (1990). This is a three-factor factorial design (four by two by two), so we need to modify our approach a bit. Figure 8.3 is an interaction plot of percent zinc retention with final meal protein on the horizontal axis. The other four factor-level combinations are coded 1 (low meal zinc, low diet zinc), 2 (low meal zinc, high diet zinc), 3 (high meal zinc, low diet zinc), and 4 (high meal zinc, high diet zinc). Lines 1 and 2 are low meal zinc, and lines 3 and 4 are high meal zinc. The 1,2 pattern across protein is rather different from the 3,4 pattern across protein, so we conclude that meal zinc and meal protein interact.
On the other hand, the 1,3 pair of lines (low diet zinc) has the same basic pattern as the 2,4 pair of lines (high diet zinc), so the average of the 1,3 lines should look like the average of the 2,4 lines. This means that diet zinc and meal protein appear to be additive.
8.5 Models with Parameters 175
45 50 55 60 65 70 75 80 85
1 2 3 4
Meal protein Z
i n c r e t e n t i o
n 1
1
1
1 2
2 2 2
3
3
3
3 4
4
4
4
Figure 8.3:Interaction plot of percent zinc retention data with meal protein on the horizontal axis, using MacAnova.
8.5 Models with Parameters
Let us now look at the factorial analysis model for a two-way factorial
treat-ment structure. Factor A has alevels, factor B has b levels, and there are A hasalevels, B hasblevels,n replications
nexperimental units assigned to each factor-level combination. Thekth re-sponse at theith level of A andjth level of B isyijk. The model is
yijk=µ+αi+βj+αβij +ǫijk,
whereiruns from 1 toa,jruns from 1 tob,kruns from 1 ton, and theǫijk’s Factorial model
are independent and normally distributed with mean zero and varianceσ2. Theαi,βj, andαβij parameters in this model are fixed, unknown constants.
There is a total ofN =nabexperimental units.
Another way of viewing the model is that the table of responses is broken down into a set of tables which, when summed element by element, give the response. Display 8.1 is an example of this breakdown for a three by two factorial withn= 1.
The term µ is called the overall mean; it is the expected value for the responses averaged across all treatments. The term αi is called the main
effect of A at leveli. It is the average effect (averaged over levels of B) for Main effects
leveli of factor A. Since the average of all the row averages must be the overall average, these row effectsαi must sum to zero. The same is true for
176 Factorial Treatment Structure
responses overall mean row effects
y111 y121 y211 y221 y311 y321
=
µ µ µ µ µ µ
+
α1 α1 α2 α2 α3 α3
+
column effects interaction effects
β1 β2
β1 β2 β1 β2
+
αβ11 αβ12
αβ21 αβ22 αβ31 αβ32
+
random errors
ǫ111 ǫ121 ǫ211 ǫ221 ǫ311 ǫ321
Display 8.1:Breakdown of a three by two table into factorial effects.
βj, which is the main effect of factor B at levelj. The termαβij is called the interaction effect of A and B in the ij treatment. Do not confuseαβij with
Interaction effects
the product ofαi andβj; they are different ideas. The interaction effect is a measure of how far the treatment means differ from additivity. Because the average effect in theith row must beαi, the sum of the interaction effects in theith row must be zero. Similarly, the sum of the interaction effects in the jth column must be zero.
The expected value of the response for treatmentij is E yijk=µ+αi+βj +αβij .
There are ab different treatment means, but we have 1 + a+b +ab
pa-Expected value
rameters, so we have vastly overparameterized. Recall that in Chapter 3 we had to choose a set of restrictions to make treatment effects well defined; we must again choose some restrictions for factorial models. We will use the following set of restrictions on the parameters:
Zero-sum restrictions on parameters
0 = Xa i=1
αi= Xb j=1
βj = Xa i=1
αβij = Xb j=1
αβij .
This set of restrictions is standard and matches the description of the param-eters in the preceding paragraph. The αi values must sum to 0, so at most