Factorial designs (4 hours)

↵ Back to module homepage

What we discussed in the previous task was technically a one-way (also called one-factor) ANOVA. It's an analysis where you have one dependent variable (e.g., height) and one independent variable (e.g., group; this independent variable might have two levels [babies vs. NBA players] or it might have many levels [English speakers vs. Mandarin speakers vs. French speakers vs. Bantu speakers]).

Things get more interesting, and ANOVA gets more useful, when we consider designs with more than one independentvariable. If the variables are completely crossed (i.e., all combinations of each variable do actually occur in the design), that's called a factorial design. For example, imagine you measure the height of basketball players vs. jockeys, and you do that in both Hong Kong and Canada. Then you've got a factorial design:

  Canada Hong Kong
Basketball players Canadian basketball players HK basketball players
Jockeys Canadian jockeys HK jockeys

This is the sweet spot. This is a good way to design research. In a factorial design like this, you can examine not only the boring pairwise comparisons between specific groups (e.g. Canadian jockeys vs. HK basketball players), but also more complicated comparisons, like "basketball players overall vs. jockeys overall", "is the difference between basketball players and jockeys in Canada bigger than the difference between basketball players and jockeys in HK", etc. This opens you up to test more meaningful and interesting hypotheses.

In a factorial design like this, there are multiple different kinds of effects we can examine, including main effects, interactions, and simple effects. Read at least two of the below readings to get familiar with what these kinds of effects are. (If you're already pretty confident that you get the idea, you might just need to read the shorter ones, like Politzer-Ahles or Gelman. If you think you need more detailed and wordy explanation before you fully get it, make sure to read at least one of the longer ones, like or Krantz.) When you're done, continue on to the questions to test your understanding.

If you understand what the various effects are, you should be able to draw what they look like. For example, the below drawing shows a situation in which there is a main effect of Test (pretest vs. posttest), no main effect of Group (Mandarin vs. Cantonese) and no interaction:

A 2x2 barplot showing the following situation: Mandarin pretest = 5, Cantonese pretest = 5, Mandarin posttest = 7, Cantonese posttest = 7

This is just one possible situation that could occur in a factorial ANOVA. There are many others (main effects in both factors, but no interaction; main effect in one factor, an interaction, but no main effect in the other factor; no main effects, but an interaction; no main effects and no interaction; etc.). In situations where there is an interaction, there are a lot of possible kinds of interactions that can appear (the simple effect of pre-post for Mandarin speakers might be bigger than it is for Cantonese speakers, or might be smaller, or might even be in the opposite direction).

Draw at least three possible patterns of effects, like what I drew above. Make sure your graphs have labels (you can use the same labels as mine, or you can make up another set of factors that is more related to your own research). If you aren't confident using software like R/Excel/etc., it's ok to just draw your graphs by hand and take photos of them. For each graph, you should put a title or caption describing what situation it is (e.g., my graph above would be captioned "Main effect of Test, no main effect of Group, and no interaction").

Remember that in the first task of this module, I said that I almost never use ANOVA, because I can accomplish the same things with other tests. That also applies to factorial ANOVA. If I'm interested in looking at a main effect, or an interaction, or whatever, there are other ways to do it.

But this also depends on the nature of the data; it will work differently if I have independent data or paired (dependent, repeated measures) data. (Review the "Introduction to inferential statistics and t-tests module" if you don't remember what these terms mean.)

Imagine that I have a factorial study design with four groups (conditions), and they're all paired. In other words, the four measurements from the four conditions come from the same person; I have no independent groups. For example, maybe I gave people a syntax pretest and posttest, and also a phonetics pretest and posttest, so each person took four tests. Here's some fake data from a situation like this:

Participant Syntax pretest Syntax posttest Phonetics pretest Phonetics posttest
ptp00170858279
ptp00253803750
ptp00391947877
ptp00460586142
ptp00569815562

If I wanted to, I could test these in an ANOVA. (Technically I'd need a repeated-measures ANOVA. Just like t-tests include independent t-tests and dependent t-tests, ANOVA has different versions—normal ANOVA vs. repeated-measures ANOVA—depending on whether your data come from independent groups or paired groups.)

But I find it easier to use t-tests. Recall that a design like this includes two main effects, an interaction, and a bunch of simple effects. Imagine I want to see the main effect of test (i.e., compare the overall posttest scores to the overall pretest scores). That just means I want to see if the average of the syntax and phonetics post-tests is higher than the average of the syntax and phonetics pre-tests. Person #1 got 85 points on their syntax post-test and 79 points on their phonetics post-test, so their average post-test score is 82 (halfway between 85 and 79). I could calculate the average scores for each person's post-test and each person's pre-test, to get a table like the following:

Participant Pretest Posttest
ptp0017682
ptp0024565
ptp00384.585.5
ptp00460.550
ptp0056271.5

Now that it's two columns of data, it would be easy to do as a t-test. And the result you get for this t-test will be mathematically equivalent to what you get for the main effect of pre vs. post in a repeated-measures ANOVA.

Likewise, you can do the exact same thing to examine the syntax vs. phonetics: I can calculate each person's average syntax score (the average of their syntax pre-test and their syntax post-test) and average phonetics score, and compare them in a t-test. I won't write out the table here, but you should be able to figure it out.

Finally, I can also test interactions like this. In this kind of design an interaction would just mean that the pre-vs.-post difference for syntax is not the same as the pre-vs.-post difference for phonetics. (Or, alternatively, the syntax-phonetics difference at the pretest is not the same as the syntax-phonetics difference at the posttest; these are different but mathematically equivalent ways of expressing the same interaction.) So, again, I can turn this into a t-test. For each person, I can calculate their syntax improvement score (syntax post-test minus syntax pre-test, i.e., how much they improved on syntax) and their phonetics improvement score. Participant #1 went from a syntax pre-test of 70 to a syntax post-test of 85, i.e., this participant's syntax score improved by 15 points. Following this logic, we can make a table like this:

Participant Syntax improvement Phonetics improvement
ptp00115-3
ptp0022713
ptp0033-1
ptp004-2-19
ptp005127

Once again, I can do a simple t-test to see if people improved in syntax more than they improved in phonetics. If their improvement in syntax is different from their improvement in phonetics, that's an interaction. (A rule of thumb to keep in mind: something that can be explained by just referring to one variable is a main effect, and something that can only be explained by referring to both variables is an interaction. "People scored higher on the post-test than the pre-test" is a main effect, because I only mentioned the pre-vs.-post variable. "The amount that people improved from pretest to posttest was bigger for syntax than it was for phonetics" is an interaction, because I had to mention both the pre-vs.-post variable and the syntax-vs.-phonetics variable.)


All the above discussion was illustrating how I could turn an ANOVA into simpler t-tests if the conditions are all paired/dependent/repeated-measure (i.e., if the measurements come from the same person).

This strategy is not possible if I have independent groups (e.g., HK jockeys vs. Canadian jockeys vs. HK basketball players vs. Canadian basketball players; these are four separate groups of people [assuming that there's no multitalented person in the sample who's both a basketball player and a jockey). It's still true that I can analyze this without an ANOVA (I could do it using regression, which is a more complicated method which will be discussed in another module), but I can't turn it into t-tests.

Why not?

When you have finished these activities, continue to the next section of the module: "Multiple comparisons".


by Stephen Politzer-Ahles. Last modified on 2021-05-15. CC-BY-4.0.