The increasing availability of complex longitudinal databases presents both challenges and opportunities for causal inference. A key challenge is the implementation of methods that can (1) compare clinically relevant dynamic regimes, and (2) appropriately adjust for measured time-varying confounding and selection bias. Semiparametric methods such as inverse probability weighting of marginal structural models and g-estimation of nested structural models can be used to achieve these goals. However, these methods do not easily accommodate some classes of dynamic regimes, and semiparametric estimates may have a large variance when obtained from many currently available databases. A fully-parametric alternative to these methods is the parametric g-formula. Though first proposed by Robins in 1986, the g-formula has been little used in practice, and thus its relevance remains largely unexplored. This presentation discusses recent refinements, software developments, and practical implementations of the parametric g-formula.
Keywords: Counterfactuals; G-formula; Time varying confounding
Biography: Miguels research is focused on methodology for Causal Inference from observational studies. In particular, he has focused on longitudinal studies, where both exposure, outcome, and covariates may vary over time. To the epidemiological audience he is well known for his impressive efforts to communicate difficult statistical concepts and methods. He is the co-author of a forthcoming book entitled Causal Inference.