Abstract:
|
We offer an asymptotically efficient, non-parametric way to assess treatment reliability via the blip function, B(W), the average treatment effect for a randomly drawn strata, W, or conditional average treatment effect. We can ask the two main questions of any function of a random variable: What are its mean and variance? The blip mean gives the more easily estimable causal risk difference where as blip variance measures reliability of treatment or the extent of effect modification. With blip variance we introduce the concept of clinical effect heterogeneity, which enables doctors and analysts to assess not only the average effect of treatment but also what an individual can expect from treatment. We can also assess how much precision in treatment can be gained in treating patients based on their covariates. Through simulations we will verify some of the theoretical properties of our proposed estimator and also point out some of the challenges in estimating blip variance, which lacks double robustness, unlike causal risk difference. We therefore open up a new direction in estimating such parameters. We will provide a demonstration, featuring new software.
|