Dear Joshua,

The way that FLAME and the multi-level GLM works is such that it automatically takes this information into account.  If a contrast is based on a larger amount of information then its VARCOPE will be smaller.  This has the effect of weighting this information more highly when it reaches the next level.

So there is nothing that you need to do - it is already being done.  And there is no problem with any systematic differences like this between patients and controls, as it does not create any bias.  It will simply affect the associated variance and consequently the sensitivity of the method to detect differences, but will still provide valid, unbiased results.

All the best,

Mark

On 18 Jul 2012, at 00:49, Joshua Lee wrote:

Hi all,

In event related designs where the coding of an event is determined by a participant's behavior (e.g. memory contrasts: remembered or forgotten), it is the case that participants will have varying numbers of events within each (e.g. one may have 85 remembered and 5 forgotten, while another might end up with 10 remembered and 70 forgotten).

It seems silly for a group analysis to weight equally parameter estimates from a subject with just 10 trials of an EV with parameter estimates derived from a subject with 80 trials of that EV. Obviously, one should value estimates that used more trials to make those parameter estimates. However, as only basic knowledge of statistics, I am not sure how to go about resolving the issue other than weighting cases in post-functional analyses.

It then occurred to me that I once read that FSL's FLAME carries up information from lower level analyses (i.e. COPES, VARCOPES, and degrees of freedom). My question is whether FLAME helps address this design issue, or whether you know of a good technique to de-weight cases whose parameter estimates are probably less reliable.

Also, it seems that groups may differ systematically on how many events get associated with each EV, say in a patient vs control comparison study. Does this systematicy violate any assumptions of FLAMEs mixed models?

Thanks for any input,

Joshua