Multiple imputation for discrete data: Evaluation of the joint latent normal model.

Area of Science:

• Statistics
• Biostatistics
• Data Science

Background:

• Missing data are common in clinical and social research.
• Multiple imputation (MI) is a preferred method for handling missing data.
• Two main MI strategies are joint modelling (JM-MI) and full conditional specification (FCS-MI).

Purpose of the Study:

• To evaluate the performance of the latent normal model for joint modelling multiple imputation (JM-MI).
• To compare JM-MI using latent normal models against FCS-MI for various data types.
• To assess the applicability of JM-MI for both single and multilevel imputation.

Main Methods:

• Extensive simulation study comparing latent normal JM-MI with FCS-MI.
• Application of JM-MI to data from the German Breast Cancer Study Group.
• Investigation focused on binary, categorical, ordinal, and count data.

Main Results:

• JM-MI using the latent normal model performed very well across most tested scenarios.
• The latent normal JM-MI approach sometimes outperformed FCS-MI.
• Performance was evaluated using data simulated from both the latent normal and general location models.

Conclusions:

• The latent normal model, as implemented in the R package jomo, is a reliable method for multiple imputation.
• Researchers can confidently use this JM-MI approach for single and multilevel missing data problems.
• JM-MI offers a robust alternative to FCS-MI, particularly when an explicit imputation model is desired.