We propose to apply and further develop analytic methods for estimating the effect of public health interventions using complex longitudinal data from observational studies. Our focus will be the estimation of the effects and attribute risks of various interventions on the risk of coronary heart disease in the Nurses' Health Study. Examples of hypothetical interventions are *a random half of the smokers quit smoking at baseline' and 'inactive individuals increase their activity to at least 30 minutes brisk walking per day1. In contrast to standard statistical methods, the methods we propose allow us to estimate consistent attributable risks, appropriately adjust for time-dependent confounding due to intermediate variables, estimate the effect of joint interventions on multiple risk factors, and estimate the effect of dynamic interventions. Our research will produce state-of-the-art estimates of the effects of public health interventions on risk factors for coronary heart disease that use (i) one of the largest observational cohorts available, and (ii) the most powerful methods available for causal inference from complex longitudinal data. We will estimate the effect of hypothetical interventions using the parametric g-formula (based on regression and Monte Carlo simulation), marginal structural models (based on inverse-probability weighting), and nested structural models (based on g-estimation). These methods have shown their superiority over standard analytic methods in some areas, but their practical impact in cardiovascular disease epidemiology is largely unknown because they have not been systematically applied to large cohorts with longitudinal data on time-varying risk factors and health outcomes. The main reasons for the lack of widespread used of these methods is their perceived technical complexity, the scarcity of research on how to translate them into practical applications, and the lack of standard software. We will adapt and further develop these methods to address realistic questions in cardiovascular epidemiology and produce user-friendly and computationally-efficient software to apply them.
Showing the most recent 10 out of 52 publications
