Calibration plots for multistate risk predictions models.

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Lack of established methods for assessing calibration in multistate risk prediction models.
• Need for reliable calibration assessment in the presence of censoring.

Purpose of the Study:

• Introduce and evaluate techniques for generating calibration plots for multistate model transition probabilities.
• Assess the performance of these techniques under random and independent censoring.

Main Methods:

• Utilized pseudo-values (Aalen-Johansen estimator), binary logistic regression with inverse probability of censoring weights (BLR-IPCW), and multinomial logistic regression with inverse probability of censoring weights (MLR-IPCW).
• Simulated data with varying censoring levels to evaluate calibration curve estimation.
• Applied methods to a real-world cohort for predicting multiple chronic diseases.

Main Results:

• Pseudo-value, BLR-IPCW, and MLR-IPCW methods yielded unbiased calibration curve estimates under random censoring.
• These methods demonstrated robustness against independent censoring, with minimal bias in high-density prediction regions.
• MLR-IPCW provided additional insight through a calibration scatter plot.

Conclusions:

• Recommend pseudo-value or BLR-IPCW for calibration curves and MLR-IPCW for scatter plots.
• The developed methods are integrated into the "calibmsm" R package.
• These tools enhance the reliability of multistate models in risk prediction.