A Choice of Performance Metrics for Evaluating Predictive Accuracy of Survival Models

This research critically assessed the predictive accuracy of parametric survival models (Weibull, Exponential, Log-logistic, and Gompertz) against penalized Cox PH models (Ridge, Lasso, and Elastic Net) using both simulated data (sample sizes of 100, 200, and 1000) and real-world data from the Nigerian Demographic and Health Survey (NDHS). The findings showed that parametric models, particularly the Weibull and Log-logistic models, consistently outperformed the others, achieving the highest Concordance Index (C-index) and the lowest Mean Absolute Error (MAE) and Mean Squared Error (MSE), indicating superior discrimination and calibration. In contrast, penalized Cox models underperformed, especially with a larger number of covariates, and the Gompertz model exhibited poor predictive performance under all conditions. Notably, parametric models remained stable and consistent even with smaller sample sizes and high-dimensional, complex data. These results highlighted the reliability of parametric models in survival analysis, particularly in small-sample and high-dimensional settings, offering key insights to inform future infant and child health research.