Adding propensity scores to pure prediction models fails to improve predictive performance.

Area of Science:

• Statistical modeling
• Medical prediction
• Causal inference

Background:

• Increasing use of propensity scores in prediction modeling lacks theoretical basis.
• Confusion exists between predictive and causal inference modeling goals.
• Propensity scores are often incorrectly applied in prediction tasks.

Purpose of the Study:

• To formally assess the impact of propensity scores on predictive model performance.
• To compare propensity score methods against full covariate adjustment in prediction.
• To test the hypothesis that full covariate adjustment outperforms propensity scores.

Main Methods:

• Investigated logistic and proportional hazards regression models.
• Employed 500-fold random cross-validation for optimism correction.
• Compared full covariate adjustment models with propensity score adjustment and inverse probability weighting models.

Main Results:

• Multivariable models with full covariate adjustment demonstrated superior predictive performance.
• Full covariate adjustment yielded higher concordance indices and better calibration.
• Propensity score methods did not improve predictive performance over full covariate adjustment.

Conclusions:

• Propensity score techniques do not enhance prediction performance beyond multivariable adjustment.
• Propensity scores are not recommended for pure prediction modeling.
• Full covariate adjustment is the preferred method for medical prediction.