Marginal structural models can be a useful tool for analyzing longitudinal data in which variables are observed repeatedly over time that both confound future treatments and are also on the causal pathway from past treatments to the outcome of interest. I explain what is meant by counterfactual events, give the definition of causality, and introduce marginal structural models. A simple simulation study is presented in which both classical statistical methods and marginal structural models are valid and give virtually identical results. The key difference is that classical methods account for confounding by matching or adjusting for the confounders in the model, while marginal structural models create a weighted pseudo-population in which the causal links between the confounders and the treatment have been broken. This permits the direct effect of the treatments on the outcome to be estimated from this pseudo-population.
A slightly more complicated example is given in which classical methods fail. Here, adjusting for the confounder is necessary to obtain an unbiased estimate of the effect of future treatment on the outcome. However, this confounder is on the causal pathway between past treatment and the outcome, and adjusting for it blocks its effect on the outcome through this pathway.
The application of marginal structural models to real data is discussed. These models are less powerful and are harder to interpret than classical methods. In my opinion they should be avoided whenever classical methods are appropriate. They can be very useful in longitudinal studies when classical methods break down.
Copies of the slides can be found here: Introduction_to_Marginal_Structural_Models.pdf
I | Attachment | Action | Size | Date | Who | Comment |
---|---|---|---|---|---|---|
Introduction_to_Marginal_Structural_Models.pdf | manage | 135.6 K | 09 May 2016 - 10:07 | WilliamDupont |