Multiple imputation (MI) of outcome variables and linear predictors has recieved the most attention so far in the MI literature. Research regarding imputation of incomplete non-linear terms (like interaction terms or quadratic or even higher-order relationships) on the other hand is still scarce. The present paper examined two ad hoc MI strategies and two more theoretically sound MI solutions (i.e. substantive model compatible MI) regarding bias in statistical inferences in a random intercept model that includes an interaction term and a quadratic term by means of Monte Carlo simulation. The distribution of predictor variables was either normal, heavy-tailed or skewed. Results show that only if the imputation model is fully compatible to the subsequent analysis model (and to the original data generating process), i.e. only when the imputation model includes non-linear terms as well as information regarding cluster membership, then point estimates and standard errors are unbiased. Currently available MI methods therefore need to be adjusted for situations, where distributional assumptions are violated to some extent.