I am using Multiple Imputation to impute a continuous variable (X). I have a question regarding the generation of a new variable, starting from this imputed variable (what, in Stata jargon, would be called a passive variable).

In particular, my aim is to divide this variable X into, say, quartiles. Let's suppose that I imputed this variable $n$ times (X_1, ..., X_n). What am I supposed to do?

Should I create the quartiles "independently" for each of the $n$ imputations? In this case the cut-offs of the quartiles (i.e. the 25th, 50th and 75th percentiles) wouldn't be the same across the $n$ passive variables I created.

Should I pool the $n$ imputations, calculate the 25th, 50th and 75th percentiles and use those numbers to divide my $n$ imputed variables into quartiles?

  • 1
    $\begingroup$ You should pool the $n$ imputations but more important part of MI is the number of imputations, $n$. If I remember correctly, you need to impute at least 40 times with 50% missing information to guarantee less than 1% power falloff compared to the comparable full-information maximum likelihood analysis. This paper should be helpful. $\endgroup$
    – Ken
    Jun 21, 2012 at 15:06


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.