Dirichlet process mixture models to impute missing predictor data in counterfactual prediction models: an application...

Area of Science:

• Statistics
• Machine Learning
• Biostatistics

Background:

• Handling missing data is a significant challenge in statistical modeling, especially for clinical prediction.
• Missing predictor information complicates the development and application of regression models in practice.

Purpose of the Study:

• To present a flexible Bayesian approach for managing missing predictor information in regression models.
• To provide practitioners with full posterior predictive distributions for missing data and outcomes.
• To apply this method to a counterfactual treatment selection model for type 2 diabetes therapies.

Main Methods:

• Utilized a Bayesian framework combining a regression model and a Dirichlet process mixture model (DPMM).
• DPMMs were employed to flexibly model the joint distribution of predictors.
• The approach was applied to a counterfactual treatment selection model for type 2 diabetes.

Main Results:

• DPMMs effectively model complex predictor relationships and handle incomplete data under missing-at-random assumptions.
• The framework incorporates uncertainty from missing data into parameter estimates and treatment effect predictions.
• Identified which variables, if collected, could offer the most additional information in scenarios with multiple missing predictors.

Conclusions:

• DPMMs provide a flexible method for modeling complex covariate structures and addressing missing predictor data in clinical prediction models.
• DPMM-based counterfactual prediction models support clinical decision-making by enabling predictions with appropriate uncertainty for individuals with incomplete data.