Data-Adaptive Estimation for Double-Robust Methods in Population-Based Cancer Epidemiology: Risk Differences for Lung...

Area of Science:

• Epidemiology
• Biostatistics
• Public Health

Background:

• Population-based cancer epidemiology requires robust statistical methods to assess mortality risks.
• Evaluating the performance of double-robust estimators is crucial for accurate analysis of binary exposures in cancer mortality studies.

Purpose of the Study:

• To propose a structural framework for population-based cancer epidemiology.
• To evaluate double-robust estimators for binary exposures in cancer mortality.
• To compare model selection strategies, including information criteria and machine learning algorithms.

Main Methods:

• Numerical analyses were conducted to assess bias and efficiency of double-robust estimators.
• Two model selection strategies were compared: Akaike's Information Criterion/Bayesian Information Criterion and machine learning algorithms.
• The performance of estimators was evaluated under various simulation scenarios, including model misspecification and near-positivity violations.

Main Results:

• In simulations, most estimators performed well with correct models, but the augmented inverse-probability-of-treatment weighting estimator showed significant bias.
• Under dual model misspecification and near-positivity violations, all double-robust estimators were biased; however, the targeted maximum likelihood estimator demonstrated the best bias-variance trade-off.
• Application to lung cancer patients in England revealed a 16% higher 1-year mortality risk for men and 18% for women diagnosed via emergency departments.

Conclusions:

• Data-adaptive model selection strategies based on machine learning algorithms are recommended for improved accuracy in cancer mortality studies.
• The findings highlight the importance of interventions for early detection of lung cancer symptoms.
• Double-robust estimators, particularly the targeted maximum likelihood estimator, show promise but require careful consideration of model selection and data conditions.