A simple pooling method for variable selection in multiply imputed datasets outperformed complex methods.

Area of Science:

• Statistics
• Biostatistics
• Epidemiology

Background:

• Variable selection is crucial for prognostic model development after multiple imputation.
• Pooling methods are essential for integrating results from multiply imputed datasets.
• Evaluating different pooling methods is necessary to optimize prognostic modeling.

Purpose of the Study:

• To compare the performance of four pooling methods (D1, D2, D3, and Median-P-Rule) for variable selection in multiple imputed datasets.
• To assess these methods using both simulated and real-world data for logistic regression models.
• To determine the most effective pooling strategy for different variable types (categorical, dichotomous, continuous).

Main Methods:

• Simulated datasets with varying sample sizes and correlations, including missing at random data.
• Application of Multiple Imputation (m=5) to generated and real-world NHANES datasets.
• Comparison of four pooling methods (D1, D2, D3, MPR) against selection from a full model.

Main Results:

• The Median-P-Rule (MPR) demonstrated superior performance in pooling and selecting categorical variables, and improved prognostic model stability.
• MPR performed comparably to other methods for continuous and dichotomous variables.
• In real-world data, MPR was identified as the most sensitive, simple, and easy-to-apply method.

Conclusions:

• The Median-P-Rule (MPR) is recommended for pooling categorical variables (>2 levels) after multiple imputation, especially with Backward Selection.
• MPR is also advised for continuous and dichotomous variables due to its consistent performance.
• The MPR method offers a simple and effective approach for variable selection in prognostic modeling.