Rubin's imputation rules when datasets different sizes

I am performing a Cox Model on imputed datasets created by proc mi. I then analyse using the Cox Model (phreg) and combine the parameter estimtes using Rubin's rules to obtain an overall (log) hazard ratio with SEs and CIs (proc mianalyze). My query is that prior to mianalyze I do selection on one of the imputed variables x_imp (say) ie I filter so that I only use observations where x_imp < cutoff in each imputed dataset. Therefore the imputed data sets I actually use are filtered by that variable and differ in size depending on what the imputed values x_imp actually are and how may observations were filtered out. Can I just check that Rubin's rules implemented by proc mianalyze is still valid for imputed data sets with unequal sizes ie it is still correct to the method to obtain a pooled standard error derived from different components that reflect the within and between sampling variance of the mean difference in the multiple imputed datasets when those datasets are of difference sizes.

many thanks


Less is more. Stay pure. Stay poor.
So you had missing data in one of your inclusion variables. Then based on imputed values you excluded some obs,but also used that variable in your Phreg model?

Can you try to rephrase your question i want to make sure i am following. It might help us if you write out your models, e.g., y = x1 + x2, x2 is imputed variable > ???
thanks for your reply, yes you are quite right hlsmith, I had several variables with missing values, for one of the I excluded some obs based on imputed values so the imputed data sets I went on to use have diffferent numbers of obs


Less is more. Stay pure. Stay poor.
I would say the rubin method may be off given varying degrees of freedom. What are the differences between sample sizes? That may direct my next suggestion.