exchangeability assumption in single subject analysis
alenarto
Posted on 08/03/09 11:39:58
Number of posts: 41
Hi all ~ Am I correct to assume that the permutation tests for single subject analyses (trials in rows) fail to meet the exchangeability assumption (due to temporal correlations) and are therefore not interpretable?
Many thanks.
Agatha
Untitled Post
rmcintosh
Posted on 08/04/09 09:15:32
Number of posts: 394
Permuting single subject data is tricky. Strictly speaking, yes, the exchangability assumption is violated and renders the results questionable. There is a lot of work in the resampling stats world to try and get around this issue by removing the dependency or changing the permutation algorithm to only deal with valid exchanges (e.g, estimate temporal autocorrelation and permute outside it). The bootstrap has the same issue, though it may be somewhat less of a problem than the permutation. All I can suggest at this point is that if your trial spacing is sparse, the permutation may be more robust than if trials of the same type are too close together. In this case, you can probably interpret the permutation/bootstrap results with appropriate caution.
Randy