Evaluating uses of data mining techniques in propensity score estimation: a simulation study.

Area of Science:

• Epidemiology
• Biostatistics
• Data Mining

Background:

• Propensity score modeling is crucial for observational studies.
• Optimizing exposure prediction using covariates is standard practice.
• Evaluating data mining techniques for exposure propensity score (EPS) models is essential.

Purpose of the Study:

• To assess the performance of various EPS models.
• To compare data mining techniques (recursive partitioning, neural networks) against logistic regression.
• To determine conditions for unbiased and efficient EPS model results.

Main Methods:

• Simulated data for a cohort study (n=2000) with binary exposure/outcome and covariates.
• Compared logistic regression (LR), recursive partitioning (RP1, RP2), and neural networks (NN) for EPS modeling.
• Evaluated c-statistics (C), standard errors (SE), and bias of exposure-effect estimates.

Main Results:

• Neural networks (NN) achieved the highest prediction accuracy (mean C: 0.86).
• All models showed small overall bias; NN estimates were least biased.
• Higher prediction accuracy (C) correlated with increased standard errors (SE), but not bias magnitude.

Conclusions:

• Simple logistic regression (LR) models for EPS provide robust effect estimates.
• Neural networks (NN) generally yield the least biased estimates.
• Prediction accuracy (C) is linked to increased SE, not bias magnitude, in EPS models.