I am using mice in R, a chained equations (sequential regression) algorithm, to impute a series of polytomous variables (e.g. scales from 1 to 5 or 1 to 3). Ultimately, my interest is in estimating statistics on simple proportions of categories on these variables (e.g. 1 and 2 versus other). There are two options for doing this recoding, i.e. before and after imputation.

Recoding after imputation would mean treating variables as polytomous in mice using a multinomial or ordinal regression model. On some categories (e.g. top or bottom, 1 or 5) the sample data is sparse, however. So I am not sure about the variance of imputing these categories by polytomous models. However, the model would be most adequate to the original data.

Recoding before imputation might solve this problem, because sparsely populated categories are summarized to indicator variables. In addition sequential regression would be limitted to less categories, in most cases thus 2 (e.g., 1 and 2 versus other). A logistic model with far less parameters would be used which may increase accuracy of imputations. But I am not sure whether recoding before imputation might change the data to an extend that imputations are not correct anymore.

So my questions are should I recode before or after imputation? Does this make any difference, is one or the other preferable?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.