Dear list members, my task is to analyze mixed effect models (nested structure: time, small groups and/or class) which contain NAs. After reading the literature, I have some technical questions I hope to get answers for and which I was not able to solve fully to be sure to proceed in the right way. I use R for statistics.
(1) Which MI model to choose for lme's? PAN or treating every time-point (6 time points) as a distinct variable and then using another approach of MI (norm, mice, aregImpute...)? The proposal to treat different time-points as distinct variables was mentioned in literature but not deeply discussed. What is the better approach? (2) How to pool and compare models with different parameters and to test for effects? It seems that lme() or lmer() will be used for analysis. For the fixed parameters, I assume the common Rubins-rules are standard procedure and should not be a problem in multilevel models. Is this right? But how to proceed with the random effects? Is this standard procedure as well? With my naive understanding, these are no "real" estimates but just covar-structures. So how to pool them? I also read that lme() does not really provides S.E.s for the random effects. Is there a pragmatic way of obtaining them ? (3) For this work, anova tables have to be made. Reviewing the postings on this list, it seems that for F-statistics and R-square -> a log() transformation is useful before averaging them. But does this mean (I am not a statistician) that then the F-statistic is asymptotic ~N(0,1) after log() is used? (for large samples of course...) Thus, the necessary variance estimate to apply Rubins-rules is reduced to u <- 1 ubar = 1/m * sum(u) ?? or what is the variance of a F-statistic? I thought a F-statistic has no S.E. For further estimates, the same procedure can be applied to pool other elements of anova tables like ssq/msq??? etc. which have to be reported. (4) Lastly, is it appropriate to use the known likelihood ratio test to compare "normal" (i.e. not glm's) mixed effect models as long es they are nested (full model versus reduced model)? If this is so, has anyone coded this LRT in R/SPLUS (or maybe SAS) so that I can adopt it to my needs? I feel quite uncomfortable to code it myself without being sure that the resulting code is alright and I am not really capable of calculating matrices. Thank you very much, with best regards Leo G?rtler Germany (Berg) -- Psychologist --
