Using Causal Machine Learning Methods to Inform Tobacco Regulatory Science

点击翻译按钮获取中文摘要

PROJCT SUMMARY/ABSTRACT Inconsistent findings regarding whether and how E-cigarette (EC) use influences subsequent tobacco use behaviors complicate evidence-based tobacco regulation. Among youth, EC use is associated with greater risk of transitioning to combustible cigarette (CC) smoking, but estimated effect sizes of EC exposure vary substantially across studies. Among adults who currently smoke CCs, ECs show potential to help quit CC smoking in some studies but not in others. These inconsistent findings may be due in part to a preponderance of observational studies, use of small size cross-sectional data, and inadequate control for covariates, Further, despite considerable heterogeneity in the size of estimated EC exposure effects, whether specific characteristics modify the EC exposure effects has been largely ignored in literature. Understanding how ECs influence subsequent CC smoking, particularly among vulnerable subgroups (e.g., age, gender), and their intersectionality, will help inform regulatory activities that address tobacco-related health disparities. Lastly, it is unclear whether estimated EC exposure effects from a certain population subgroup or at a certain time can be generalized to different subgroups or times. Generalizable EC exposure effects could provide critical evidence for tobacco regulators. To address these knowledge gaps, this study aims to use causal machine learning methods to determine the influence of ECs on subsequent CC smoking, in overall US youth and adult populations and in vulnerable subgroups, and to explore methods for estimating generalizable EC exposure effects. A secondary analysis of the longitudinal Population Assessment of Tobacco and Health study will be conducted to address the following specific aims. Aim 1: Determine average exposure effects of EC use on subsequent CC smoking in youth and adults. Aim 2: Determine heterogeneous EC exposure effects among vulnerable subgroups (age, gender, poverty, race/ethnicity). Aim 3: Evaluate the performance of causal machine learning methods to generalize EC exposure effects using both simulated and PATH Study data. To successfully accomplish these aims and develop into an independent methodologist in tobacco regulator science (TRS), I will obtain training in the following areas: 1) TRS theories and measures, especially health disparities in TRS; 2) Causal inference methods for evaluating exposure effects; and 3) Machine learning skills for high-dimensional data analysis. During the award period, I will be supported in my research and training goals by my institution and interdisciplinary mentoring team, which consists of experts in the fields of TRS, causal inference, machine learning, and health disparities. The K01 research and training experience will result in an R01 with the overarching goal of extending causal machine learning methods for generalization to address more complex real-world questions. In the long term, I will bring TRS, machine learning methods, and causal inference together to address pressing issues in TRS. This effort will put causal machine learning methods in the hands of tobacco researchers and facilitate the use of complex data to inform FDA regulations.