gms | German Medical Science

Causal inference is almost exclusively conducted with outcome models, i.e., the outcome modeled as a function of a treatment of interest and covariates. It does not have to be this way. Confounding, loosely defined, is the result of an imbalance across treatment levels of variables that are predictors of the outcome [1]. Rather than model away the relationship between confounders and the outcome as is done with outcome regression, another way of addressing confounding is to remove the relationship between confounders and treatment using weights [2]. Inverse probability of treatment weighting can achieve this by creating a new pseudo-population where the confounders are balanced across treatment levels.

In this session, I will introduce inverse probability of treatment weighting including estimation of propensity scores, constructions of stabilized and unstabilized weights, balance and weight distribution diagnostics and estimation of effects using marginal structural models. I will discuss the advantages and disadvantages of approaching causal inference in this way the differences in interpretation between causal effects estimated with outcome regression and estimated with inverse probability of treatment weighting. This will include new insight on extra homogeneity assumptions required to estimate average treatment effects with outcome models that are not required when using inverse probability of treatment weighting and a precise definition of the type of treatment effects estimated by outcome regression when the homogeneity assumptions are not met.

Lastly, I will argue that causal inference should use outcome regression, inverse probability of treatment weighting as well as doubly robust estimation simultaneously as a form of triangulation to learn more about the validity of modeling assumptions required by each method.

The authors declare that they have no competing interests.

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