gms | German Medical Science

Marginal Structural Models (MSMs) are popular in epidemiology to estimate the long-term effects of a (sequence of) treatment(s) while adjusting for time-varying confounding. Despite their popularity,

they do not allow for the estimation of treatment effect modification by time-varying covariates. Other methods, such as g-estimation of Structural Nested Mean Models (SNMMs) can be used to answer this type of questions, but their adoption has been hindered by the perceived complexity, and the lack of user-friendly implementations in widely used software. G-estimation is also less sensitive to positivity violations.

In this talk, we consider settings with time-varying confounding under the no unobserved confounding assumption, and study methods for estimating treatment effects of a time-varying binary treatment on a continuous outcome. We focus on doubly robust estimators, as this property is key for being able to exploit data-adaptive (machine learning) methods. In particular, we consider g-estimation of SNMM and estimation of MSMs via longitudinal targeted maximum likelihood estimation. We implement both strategies with either parametric or data-adaptive estimation of the nuisance modes (outcome model and propensity score models) to attenuate biases due to model misspecification, via the ensemble algorithm known as Super Learner.

Using a simulation study, we compare the finite sample properties of the two methods under different scenarios, including model misspecification and practical positivity violations.

We also provide an empirical application to a large-scale Electronic Health Records data from type 2 diabetes cohort undergoing treatment intensification.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.