Could the problem be that randomise has trouble with repeated-measures models that have both within and between subject factors?
In April 2009 there was a list discussion between Kyle Kern and Tom Nichols. Kyle had a similar dataset to Miguel, with repeated measures in 2 groups (4 measures as opposed to Miguel's two), and wanted to examine group by time interactions (as part of a full factorial analysis)
Tom's response was that
" The reason that no one has chimed in is that randomise can't deal with this design in general. As you concisely state, you have 2-way repeated measures ANOVA, 1 within-subject factor (time) and 1 between subject factor (group) (and the repeated 'cluster' variable subject). In this advanced model, the appropriate way to do a permutation test depends on the question asked. If you are asking just about the within-subject factor (time), randomise *could* deal with this, just use the model as you have specified it *but* be sure to specify an exchangeability block file marking the subjects (e.g., transposed it would like like
1 1 1 1 2 2 2 2 3 3 3 3 ...
). Randomise will then know to only permute scans within subject (appropriate for a null hypothesis of no time effect, *and* implicitly assuming exchangeability within subject). But if you're asking about the between-subject factor, the appropriate way to permute is to shuffle subjects' data as a whole around. Or, equivalently to, permute the design ensuring that only permutations that move around sets of four rows together are used). That currently isn't implemented in randomise."
Could Tom's response to Kyle be the same reason Miguel's analysis won't work? I also have longitudinal data with more than one group, so I am keen to understand what analyses are possible, and which are not possible, in randomise.
thanks
Graham
|