Reliability of statistical inference after tree-based imputation

Determine whether statistical inference—specifically the validity of hypothesis testing and parameter estimation—is reliable when missing values are imputed using tree-based multiple imputation methods in empirical social science datasets. The purpose is to ascertain if standard inferential procedures yield trustworthy Type I error control and power when the imputation models are tree-based rather than parametric.

Background

The paper contrasts the widely used MICE with Predictive Mean Matching (PMM) against emerging tree-based imputation methods (e.g., Random Forest via MICE RF and missRanger, and Extreme Gradient Boosting via mixgb). While tree-based approaches can better capture complex, nonlinear relationships and handle mixed data types, prior evidence (e.g., with missForest) indicated potential inflation of Type I error in certain designs.

Given the increasing adoption of methods like missRanger in empirical studies, the authors highlight the unresolved question of whether statistical inference remains reliable when such tree-based imputation is used. This motivates their simulation paper assessing coefficient bias, Type I error control, and power across multiple methods and missingness scenarios.