Using aregImpute for Proportional Odds Model

[Jacquelyn Neal]

Hi Dr. Harell,

One of the objectives of my final project is using binary logistic regression models to predict conversion from normal cognitive function to either mild cognitive impairment (MCI) or Alzheimer's Disease for a baseline normal population. 

Currently, my outcome variable "conversion" is a binary factor. I was thinking "conversion" could potentially be an ordered factor for a normal population (whether they convert to MCI or directly to Alzheimer's Disease). I was reading the documentation for aregImpute and transcan, and the transcan documentation stated that fit.mult.impute will not work with a proportional odds model. How would you recommend imputing data if an ordinal variable is the outcome of interest?

Thanks!

[Frank Harrell]

I only see this mentioned in the help file for fit.mult.impute which says that the problem exists when doing regression imputation, not when using the default method of predictive mean matching.  But there is a general problem if Y has any NAs.

So it is safe to use aregImpute and fit.mult.impute if you use predictive mean matching and if Y has no NAs.