A fair comparison of tree-based and parametric methods in multiple imputation by chained equations.

Area of Science:

• Epidemiology
• Biostatistics
• Data Science

Background:

• Multiple Imputation by Chained Equations (MICE) is a key method for handling missing data in epidemiology.
• Imputation can utilize parametric or nonparametric approaches, with tree-based methods often favored for nonlinear effects.
• Previous comparisons may have been biased by not including interaction terms in parametric models.

Purpose of the Study:

• To conduct a fair comparison of parametric and tree-based imputation methods within MICE.
• To evaluate the impact of including interaction terms in parametric imputation models.
• To guide epidemiologists in selecting appropriate MICE imputation strategies.

Main Methods:

• Simulation study comparing parametric and random forest imputation within MICE.
• Parametric imputation models were correctly specified to include interaction effects.
• Performance was assessed based on bias, coverage, and confidence interval width for effect estimates.

Main Results:

• Correctly specified parametric imputation models perform comparably to random forest imputation when estimating interaction effects.
• Parametric imputation showed slightly higher coverage for interaction effects but wider confidence intervals.
• Random forest imputation offered narrower confidence intervals but requires careful consideration of model specification.

Conclusions:

• Accurate specification of parametric imputation models, including interactions, is vital for unbiased estimates in MICE.
• Both correctly specified parametric and random forest imputation are viable options for handling missing data with interactions.
• Epidemiologists should carefully consider model specification when choosing imputation methods in MICE.