On 25 Nov 2002 16:23:00 +0200 Laaksonen Seppo <[email protected]> wrote:
> The general formula for the estimand based on MI is not rational for > categorical variables. The average of multiply imputed categories provides > usually a decimal, not any particular category (if gender is missing, the > result may be between female and male). Hence, you have again to decide which > of these several imputed values to use as the finally imputed value. Or > otherwise, you have to build a new extended file with all imputed data sets > (pooled data set so that each sampling weight has been divided by the number > of > imputations), and to estimate your final distribution from this data set. Or > are there other solutions? I really want to look various parameters (or their > estimates) including maximum values, minimum values, ... frequency > distributions, ..., not totals or means or regression coefficients. > Seppo > > That is not how you do multiple imputation. You get the appropriate draws from the appropriate distribution of the multinomial variable, fit multiple models, and average the MODELS. Possibilities for getting multiple imputations of multinomials include predictive mean matching (used in the R/S-Plus function aregImpute (http://hesweb1.med.virginia.edu/biostat/s/Hmisc.html) and using a polytomous logistic model as in the MICE library for R and S-Plus (see http://hesweb1.med.virginia.edu/biostat/rms for more information). You can also consider single imputation in which dummy variables are replaced by probabilities of category membership (see work by Vach, Schemper, and others). For binary predictors, such single imputation can work surprisingly well when the response variable's model is a binary logistic model, but this may not generalize to other response models. -- Frank E Harrell Jr Prof. of Biostatistics & Statistics Div. of Biostatistics & Epidem. Dept. of Health Evaluation Sciences U. Virginia School of Medicine http://hesweb1.med.virginia.edu/biostat
