I'm trying to perform a power calculation with G*Power.
There are two quite important options, the meaning of which is not clear to me:
"Number of groups" – what is this? I have a 2×2 repeated measures factorial design. Does this mean there are four groups, or, as the rest of the internet suggests, is this box for between-subjects factors? If it's for between-subjects factors, I assume this is 1 for a repeated measures factorial.
"Number of measurements" – also not entirely clear. Is this the number of data points I collected per participant? What if they've been averaged before they go into the ANOVA? Or is this the number of within-subjects conditions? So 4 for my experiment?
Any guidance on the meaning of these settings much appreciated – as far as I can tell these are undocumented.
Best Answer
GPower is assuming you have your data set up so that a row is a case (often a person), and a column is a measure.
For example, if we measured Y on three occasions, we'd have Y1, Y2, Y3, and we'd have three measures.
The groups are when you have a between case predictor - for example gender or experimental group. So when you have a 2x2 repeated measures design, you have four measures.
However, GPower assumes that you want to do 1 test with 3 df, which you don't, you want to do 3 tests, with 1 df (2 main effects, 1 interaction). I suspect that you should therefore be entering two as the number of measures. (And then the other parameters depend upon which of the effects you want to base your power on). Power analysis for this type of design gets complicated rapidly, and it's not clear how to enter the appropriate parameters into GPower. I prefer one of two other approaches which allow you to enter the data in matrix format.
First, D'Amico, et al, showed how to do this in SPSS, using the (old) manova command: paper here: http://www.ncbi.nlm.nih.gov/pubmed/11816450
Second, I showed how to do this as a structural equation model, in a paper here: http://www.biomedcentral.com/1471-2288/3/27/
Both of those approaches are a bit trickier to start with, but they are a lot more flexible.
A third approach is to ignore the fact that it's repeated measures when estimating power. The higher the correlations between your measures, the more power you have. But you don't know what those will be. If you estimate power as if the measures were independent, you know that your power analysis is conservative. The problem is that if the correlations are high, the power analysis might be very, very conservative.