harrelfe
commented
6 years ago This is uncharted territory with not much literature to go on. It's likely if you do the most cumbersome analysis (full separate analysis on each completed dataset, run through robcov before combining for multiple imputation) will work. This is a good advertisement for a full Bayesian approach which doesn't require all this ad hockery. In the frequentist world we're making this stuff up as we go along. For Bayesian approaches, in which it's very easy to add random effects for cluster effects and get exact analyses, see https://github.com/paul-buerkner/brms/blob/master/vignettes/brms_missings.Rmd
nds109
peterzaren
Thanks to pointing me to this site, Frank. I would like to run an ordinal regression model on imputed data. I can do this using aregImpute and fit.mult.impute but the problem is I would like to correct for clusters using e.g. robcov or bootcov. As far as I understand, we cannot do this in fit.mult.impute. I would have thought we need to correct the variance covariance matrix for each fitted model before pooling? I wonder if you have a work around for this? I have tried various other R packages but would like to stay in the R Hmisc and rms modelling suite if possible as I am most comfortable there. In addition, it seems that anova etc. are available from rms on the pooled model after multiply imputation. I have not seen this in other R packages (I am unsure why – is this tricky to do?) and I would like the anova type table for hypothesis testing, contrasts etc. If you have any suggestions of how to account for the clustering that would be much appreciated. Other things I thought of was to run a mixed effects model with ID as a random effect through fit.mutl.impute but I couldn’t get that to work. Another thing I thought of was to find a regression function which computes the cluster robust SEs inside the regression function (as an argument) but I couldn’t find a suitable method for that.