Join 3,512 readers in helping fund MetaFilter (Hide)


How can I counterbalance my experimental research design?
April 21, 2010 7:51 AM   Subscribe

Can someone help me with counterbalancing an experiment that I will be conducting?

My experiment involves a between-subjects factor (group, 2 of them) and a within-subjects factor (facial expression @ 3 levels, happy, sad, and neutral). In other words, a 2x3 repeated measures factorial ANOVA.

I want participants to view four different people using each of these facial expressions. So, there would be a total of 12 excerpts (3 different facial expressions for each of the four people) that would be accompanied by music. Each person's facial expression (x3) would be accompanied by the same piece of music. In other words, 3 facial expressions and 1 excerpt of music for each of the four persons.

Therefore, I need to counterbalance the design in such a way that even though people are seeing different facial expressions, they are less likely to be "tipped off" that the music is the same or that they are seeing the same conductor's face (albeit different facial expression) next to one another. I think this goes beyond a Latin Square.

Any suggestions? I appreciate your thoughts.

Thank you.
posted by bengalsfan1 to Education (4 answers total)
 
You could perhaps change the treble/bass mix, or tempo, to make the music "appear" different even if it isn't.
posted by Hiker at 7:57 AM on April 21, 2010


Use at least three different musical selections. Shuffle these so that no participant sees the same face paired with the same selection, and each face + expression is paired with each selection an equal number of times. You'll end up with main effects for musical selection (which you'll ignore), facial expression, and person. You might be able to reduce the variability introduced by doing this considerably by choosing similar but not identical selections.

Do you really need each individual participant to see each person with each expression? (I don't think each individual needs to be exposed all twelve conditions, each group does. You need compare responses to each condition, and differences between the groups in these responses. You don't need to expose each individual member of each group to all twelve conditions to do that.)
posted by nangar at 8:49 AM on April 21, 2010


Actor and music are collinear, so you will not be able to separately determine actor effects and music effects or an actor-emotion interaction and a music-emotion interaction.

It's quite possible that order effects will exist in the worst way: their current response is influenced by their previous response (the residual, not just their mean) or their previous response to this actor/music. You can change up the order per-subject so that you get different actor/emotion orders and model the responses as correlated in time.

My recommendation would be to not have actor/music collinear; randomly or balanced distribute actors to different music.
posted by a robot made out of meat at 8:53 AM on April 21, 2010 [2 favorites]


If the concern is about tipping the subjects off (and you have the time), you could add some "filler" trials in which the expressions are also randomly paired with other music. This will create enough noise that they don't see the underlying pattern. Then just focus your analysis on the trials of interest. Key here, of course, is randomizing the order between each participant, so you don't get order effects.

I agree with robot, though: if you want to separately determine actor effects and music effects, or interactions, the current setup is inadequate; you would need to vary actor with music completely.
posted by forza at 3:50 PM on April 21, 2010


« Older How do I play guitar like this...   |  Tips for newbie to safely chop... Newer »
This thread is closed to new comments.