 |
 |
Dear Geraint, > I'm analysing some old event related data collected without a > null trial and wanted to pick your brain on current thinking about the > interpretability of the simple main effects of individual events. > > For example, if I have an SPM of trial type A>B and inspection of the > parameter estimates shows that this is primarily due to a negative > weighting on trial type B, then is it possible to make any meaningful > interpretation of that negative weighting in isolation(i.e. 'trial type B > causes a deactivation in area X')? I remember strong views being expressed > at methods meetings that this was inappropriate, but I wonder if you could > rehearse the reasons why this is so? On the one hand I can see that if > event A is (say) correlated differently with some covariate (e.g. > globals) then event B, then the sign of the parameter estimates may be > spurious (Geoff Aguirre's paper). But on the other, if all elements of the > model are correctly specified then won't the 'baseline' fitted by each > component of the model truly reflect the 'real' baseline? You are absolutely right. In the context of an event-related design with only too trial types (that are not separated by a long SOA) the efficiency of the estimates of the contrast of effects for any single event is exceedingly low (i.e. can be regarded as unestimable) and the meaning of the parameter estimates for one event dissappears (only the contrast of parameter estimates testing for differences can be interpreted). Think of it like this: Say each event evoked a box-car response that lasted for the SOA. The sum of the event-specific regressors is now a constant. This means that the solution for either is underdetermined because these regressors are collinear with the constant term. Only the difference is estimable. This is an extreme example but speaks to the effect you are dealing with. All the very best - Karl %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%