 |
 |
Michael Wolmetz wrote: > After scouring the listserv, I'm still left wondering whether it's > possible to do a hierarchical regression with > non-time-series-dependent continuous variables within spm2. Will Penny replied: > The most common examlpe of this is what neuroimaging people call > 'random effects analysis'. Since I had a similar question I've been sitting on, I thought I'd jump in here. I think Michael's asking about hierarchical regression, which is not the same thing as the tiered analysis used to do random effects analyses of subjects with repeated measures in fMRI. If I can venture a summary based on my limited knowledge of the subject, with hierarchical regression you specify the order in which parameters for each covariate are estimated. So in the case you have covariates that are not perfectly orthogonal, you get to pick which one gets first crack at explaining variance in your data. The idea being that when you're done, you can say that your second covariate is getting credit only for variance that couldn't have been explained by the previously entered covariates. This might be done at the single subject level for typical designs. I have the feeling SPM used to do this automatically for covariates of no interest, I don't know if that's the case currently. I thought two reasonable-sounding approaches would be to (a) do the regression iteratively, using the residuals as the data for each step after the first; and (b) pre-orthogonalize the covariates appropriately in advance, so that "later" covariates have variance explained by "earlier" covariates removed. I haven't thought it through, though, and I know "reasonable-sounding" isn't a good standard for statistical techniques. So any thoughts on this would be greatly appreciated (including the thought that I'm completely confused). dan