Stage-specific cancer incidence: an artificially mixed multinomial logit model

Area of Science:

• Biostatistics
• Epidemiology
• Oncology

Background:

• Prostate-specific antigen (PSA) testing has significantly impacted prostate cancer detection rates and patient prognoses.
• Analyzing the complex interplay between screening, disease progression, and patient outcomes presents a statistical challenge.
• Existing models struggle to capture the joint effects of diagnosis timing, tumor stage, and grade on survival.

Purpose of the Study:

• To develop a novel statistical modeling approach for analyzing stage-specific prostate cancer incidence.
• To establish a causal link between population-level screening policies and disease incidence patterns.
• To enable robust estimation for complex joint mixed models using large-scale data.

Main Methods:

• Developed a stable and structured Maximum Likelihood Estimation (MLE) approach for iterative parameter estimation.
• Utilized likelihood factorization to reduce computational complexity, processing smaller model dimensions.
• Employed generalized self-consistency and the quasi-Expectation-Maximization (EM) algorithm for mixed multinomial responses via Poisson likelihood.

Main Results:

• The proposed method allows for stable and iterative estimation in complex joint mixed models.
• Likelihood factorization significantly enhances computational efficiency for large datasets.
• The model successfully establishes a causal relationship between screening policies and stage-specific cancer incidence.

Conclusions:

• The developed MLE approach provides a powerful tool for analyzing complex survival-multinomial outcomes in prostate cancer.
• This methodology offers insights into the population-level impact of screening strategies on disease detection and progression.
• The findings facilitate a better understanding of the causal pathways linking screening interventions to cancer incidence.