1

I read everywhere that repeated measures ANOVA is inferior to mixed modelling (since it doesn't handle missing data as well and relies on sphericity assumption). G*Power doesn't tell you how to compute sample sizes for linear mixed models. Should I just do the calculation for RM-ANOVA and go with that number?

My study is: one group takes drug A, another drug B (blinded), surveys with a few continuous questions are given at 5 follow up appointments.

StatsNTats
  • 111
  • 1
  • 2
  • 11

1 Answers1

3

In general, because designs in which linear mixed models are to be used can be complex, it is better to work with a simulation-based approach rather than to rely on specific formulas. If you happen to work in R, you can have a look at the powerlmm package that facilitates this.

Dimitris Rizopoulos
  • 17,519
  • 2
  • 16
  • 37
  • Wow thanks! Not sure why I couldn't find that on google before. Also I just realized I have your book next to me on my desk :) – StatsNTats Oct 24 '18 at 19:38
  • I just realized there are a few parameters that it's very difficult (basically impossible) for me to make guesses for (variance of the random slopes, within subject variance) when running the simulation. What do people usually do in this situation? I'm thinking of telling them to choose a time point that they are most interested in and then doing a more simple calculation for a t.test of means or simple regression (response ~ baseline value + group). Or just trying to find some similar study. – StatsNTats Oct 25 '18 at 20:28
  • Yes, it typically helps if you have some pilot data from which you can estimate all these parameters. If not, then either you can do as you suggested but then the sample size calculation is not for the longitudinal effect or you could present the results under different logical values for these parameters. – Dimitris Rizopoulos Oct 26 '18 at 08:37