This study involves a task that measures arousal during a task with three trials of increasing cognitive complexity. In the last trial, all subjects score the same in terms of performance though demonstrate differing levels of arousal.
As far as I know, there are few strict rules on choosing a covariance structure in a multilevel model. I usually try several and choose the one with the lowest AIC or BIC (they usually agree; I don't have a strong preference). I usually test AR(1), unstructured and compound symmetry, as these seem to make sense for longitudinal data (although others do, as well).
what statistical package are you using, SPSS? you can put variables into the equation
I would say it depends on your statistical analysis. If you are using some form of random effect to account for repeated measurements, you can use AIC or BIC as Peter Flom mentionned above. If you are using a GEE approach, then I would say use an independence or unstructured.
These are all god answers but I am looking to specifically define the heterogeneity in such a way that variance decreases over the trials, eventually becoming zero for all subjects.
I use SPSS, SAS or MatLab.
Hi Peter, when would you use AR(2) or higher? Would it make sense to do this, and then test with the LR Test to determine which is best? I have an analysis where I tried to do that, but I get an unreasonable result, where AR(10) is preferred. This is with unbalanced panel data, with a very big T=100, although not every panel gets even close to 100. In the end, I think I'll just stick with AR(1) to keep it simple, but not sure if it is as "valid" as it could be. Thoughts? Thanks!
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression