Dealing with Item-level Missingness in Multilevel Data Structure

Missing data is pervasive in large-scale survey research with multiple scale measurements and nested data structures. While there are some suggestions on how to handle item-level missing data, there are no methods proposed and studied on how to handle these missing values in clustered data structures. I studied and compared multiple imputation methods on item-level missing data using a real dataset. The results show that: (1) item-level imputation generally performs better compared to scale-level imputation; (2) JM-AM algorithm and MICE-LV algorithm with random effects in the imputation model are preferred with less bias, and (3) specifying item data as ordinal in imputation models produces less bias. The limitations and cautions to consider when using multiple imputation for item-level missing data have been discussed