ESTIMATING CAUSAL EFFECTS IN LONGITUDINAL STUDIES

点击翻译按钮获取中文摘要

(Adapted from Investigator's Abstract) Time-varying exposures and treatments are ubiquitous in epidemiology and medicine. The primary goal of many studies is to determine the joint causal effects of exposures received at different times on some outcome of interest. In recent years, Robins and others developed comprehensive approaches for defining and estimating such causal effects. This formal framework provides insight into the analysis and interpretation of experimental and observational data in general and into dealing with confounding by a variable affected by the study exposure in particular, a problem which can manifest itself in both observational and randomized studies. The application seeks to extend these insights and simplify their application. Regression, broadly understood, is the most common way to analyze data on time-varying exposures in observational studies, and includes methods for dealing with the time varying covariates and correlated outcomes typical in longitudinal studies. In the presence of confounding by a variable affected by exposure regression will sometimes produce misleading or biased estimates on the joint causal effects of exposure received at different times. Nonetheless, there has been little attempt to quantify the extent of the bias. The application's first aim is to quantify the bias in these familiar methods and to learn when it is valid to assign causal interpretation to regression estimates. In the presence of confounding by a variable affected by exposure, Robins' G-estimation approach provides an alternative to familiar methods for analyzing randomized and observational studies which adopts attractive features of both. Nonetheless, it has been little used in applications. The application's second goal is to simplify the application of this approach by developing easy-to-use estimators that are reasonably efficient in common situations and making available software to perform these analyses. Causal interpretation of exposure-outcome associations in observational studies depends on untestable assumptions. The usual conduct of observational studies of time-varying exposures makes some of the commonly made assumptions suspect. The G-estimation approach allows valid estimation of exposure effect under less restrictive and more flexible assumptions. The application's third goal is to extend earlier work showing how to use G-estimation to obtain valid estimates under these more flexible and better justified assumptions.