Hi, I thought I’d post what appears to be the answer to this as it seems a lot of people are
unsure and there weren’t many responses, and if I post an incorrect ‘solution’ I’ll hopefully
get a response.
It seems that if you have different number of trials in different conditions, or different
durations of task-related activity in different conditions, you may have a problem. If the
there is not too big a difference, you don’t need to worry. If there is a big difference, this
is a problem, and your results could be flawed. What constitutes a 'big' difference is
apparently unclear. Examples would seem to include: if you use a subsequent memory
design; if you only use participants’ accurate trials; if you use a self-paced design; if you
have group differences in task duration/accurate trials etc.
This is apparently because your estimates of the beta values for different regressors have
different amounts of variance. A beta estimated using a lot of target-related events will be
better estimated than one using fewer target-related events. This can lead to biased results
in contrasts. You can deal with the problem by accounting for unequal variance at the first
level, but you can’t currently do this in SPM.
|