I am imputing the interaction terms in a three way interaction using mi impute chained; I use the Royston approach in which I compute the interaction terms prior to running the imputation and then enter them in the imputation model. The interaction terms consist of a dummy variable (college attainment) and two continuous variables (father's SES and year of birth with the latter variable having no missing values). I have a relatively large N (roughly 7500). For the college dummy, I use (logit, augment), and for the remaining terms, I use (pmm, knn(5)). Even with this specification, I obtain the error "error occurred during imputation of ..." for the logit model of the college dymmy and, using the noisily option, I can see that it is due to perfect prediction from the interaction terms involving the college dummy. I hope this is sufficient for understanding the problem I have. Otherwise, I can provide more detailed information.