Penalised logistic regression and dynamic prediction for discrete-time recurrent event data

Area of Science:

• Biostatistics
• Statistical Modeling
• Survival Analysis

Background:

• Analyzing discrete-time recurrent event data is crucial for prediction.
• Existing methods like Aalen additive models and standard logistic regression have limitations.
• Aalen models struggle with negative estimates, while logistic regression faces stability issues with low event frequency.

Purpose of the Study:

• To develop an improved prediction method for discrete-time recurrent event data.
• To address the limitations of existing statistical approaches in predictive modeling.
• To propose a pragmatic compromise between model stability and prediction bias.

Main Methods:

• Comparison of Aalen additive models, standard logistic regression, and Firth penalized likelihood.
• Introduction of a novel modified penalized likelihood approach.
• Evaluation of methods on two real-world datasets.

Main Results:

• Aalen models can yield negative estimates, hindering prediction.
• Standard logistic regression with maximum likelihood estimation can be unstable.
• Firth penalized likelihood improves stability but introduces bias in predicted probabilities.
• The proposed modified penalized likelihood offers a balance between stability and bias.

Conclusions:

• The modified penalized likelihood method provides a robust alternative for predicting discrete-time recurrent event data.
• This approach offers a practical solution to the trade-offs between stability and bias in predictive modeling.
• The proposed method enhances the reliability of predictions in scenarios with time-varying effects and covariates.