Multiple imputation for missing data via sequential regression trees.

Area of Science:

• Epidemiology
• Biostatistics
• Data Science

Background:

• Missing data is a common challenge in large epidemiologic studies, impacting analysis validity.
• Standard multiple imputation methods struggle with complex relationships (interactions, nonlinearities) among numerous variables.
• Accurate imputation is crucial for reliable inferences in studies with many data users and diverse analyses.

Purpose of the Study:

• To present a novel nonparametric approach for multiple imputation using sequential regression trees.
• To demonstrate the method's ability to capture complex variable relationships in large datasets.
• To improve the accuracy and reliability of imputations in complex epidemiologic settings.

Main Methods:

• Multiple imputation via chained equations (MICE) framework.
• Utilized sequential regression trees as nonparametric conditional models within MICE.
• Simulations were conducted to compare the new method against standard techniques.
• Applied the approach to a real-world dataset on adverse birth outcomes with over 100 variables.

Main Results:

• The nonparametric regression tree approach demonstrated superior performance in simulations for complex data structures.
• Imputations generated by the new method were more plausible, leading to more reliable inferences.
• Posterior predictive checks confirmed the validity of imputations in subsequent epidemiologic analyses.

Conclusions:

• Sequential regression trees offer a powerful, minimally tuned nonparametric alternative for multiple imputation in complex epidemiologic studies.
• This method enhances the handling of intricate variable interactions and nonlinearities, improving data integrity.
• The approach provides a more robust solution for missing data challenges in large-scale observational research.