Craig, I think it's best not to think of such variables as a series of dummies but as a single polytomous variable, when using multiple imputation. The S-Plus/R package MICE fits a polytomous logistic model to such variables and multiply imputes from the fitted multinomial distribution. The S-Plus/R function aregImpute in the Hmisc package essentially uses Fisher's optimum scoring method to score the categories to maximize R-squared, then uses predictive mean matching on predicted category scores to multiply impute. Both approaches result in "real" categories having realistic marginal distributions.
Frank Harrell On Tue, 28 May 2002 16:18:13 -0700 "Craig D. Newgard" <[EMAIL PROTECTED]> wrote: > Has anyone had experience with including a polychotomous categorical > variable, recoded as multiple dummy variables (each mutually exclusive), in > multiple imputation? I am using the Markov chain Monte Carlo method in SAS > proc MI, and am concerned about internally inconsistent coding of missing > values using this approach (and how to deal with these observations after > MI, and before the analysis). For example, we have a variable coding point > of maximum impact on the exterior of a vehicle involved in a motor vehicle > crash. There are 4 categories within the variable (frontal, right-lateral, > left-lateral, rear). I'd like to recode the variable as 3 dummy variables, > each with rear collision as the reference, in order to include the variable > in MI. Each of these 3 dummy variables would be mutually exclusive. That > is, there can only be one point of maximum impact per vehicle. I'm not sure > what to do with those observations imputed with a "1" for 2 or more of these > dummy variables (e.g. both frontal and right-lateral impacts) in MI, nor do > I know how much of a problem this may be. Any suggestions or similar > experience? Any thoughts are welcome. Thanks. > > Craig > > Craig D. Newgard, MD, MPH > Research Fellow > Department of Emergency Medicine > Harbor-UCLA Medical Center > 1000 West Carson Street, Box 21 > Torrance, CA 90509 > (310)222-3666 (Office) > (310)782-1763 (Fax) > [EMAIL PROTECTED] > > > -- 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