I’ll be talking about pragmatic trials & per-protocol effects.
One benefit over existing approaches: unlike weighted quartile sums regression, this method doesn’t assume all components work in the same direction.
Using this framework helps us define the causal estimands of interest for the agent-based model *and* ensures we’ve been explicit about our target population.
I should have made a slide gif to tweet now but ... hindsight is 20/20 & all that 🤷🏼♀️🤷🏼♀️
To paraphrase: “With doubly robust methods & relatively mild conditions, we can estimate the average treatment effect with machine learning!!”
Second important caveat: even with doubly-robust methods, to get good confidence interval coverage you need to use sample splitting (and even then it might not be super great 😬)