Multiple imputation in the presence of non-normal data

Area of Science:

• Statistics
• Biostatistics
• Data Science

Background:

• Multiple imputation (MI) is a common method for handling missing data.
• Standard MI techniques often assume normally distributed variables, posing challenges for non-normal data.
• Guidance on imputing non-normally distributed continuous variables is limited.

Purpose of the Study:

• To compare different methods for imputing non-normally distributed continuous variables.
• To evaluate the impact of various transformations versus predictive mean matching (PMM) on statistical inferences.
• To provide evidence-based recommendations for handling non-normal missing data.

Main Methods:

• Simulated data from various non-normal distributions with 50% missingness (MCAR/MAR).
• Compared imputation on raw scale, log, Box-Cox, and non-parametric transformations, alongside PMM (type 1 and 2 matching).
• Assessed inferences on marginal means and associations with an outcome variable.

Main Results:

• Transformation can bias results if the relationship is linear on the untransformed scale.
• Transforming variables is important for accurately modeling non-linear relationships.
• Predictive mean matching (PMM) with type 1 matching offers a robust alternative.

Conclusions:

• The choice of imputation method for non-normal data depends on the variable's relationship with other variables.
• Avoid transformations if the relationship is linear on the raw scale; consider PMM type 1 matching.
• Accurate modeling of non-linear relationships necessitates appropriate variable transformation or specialized imputation techniques.