2
$\begingroup$

I am currently using the tmle function in R to perform targeted maximum likelihood estimation as a doubly-robust alternative to propensity score methods in identifying an average treatment effect. This is an ideal method, as I have a large number of covariates, and I'm not worried in estimating the effect of specific covariates on the outcome. However, with propensity score methods, I generally report on covariate balance. I've been trying to find some similar diagnostic to show that the estimation of the probability of treatment resulted in a sufficiently low level of bias. Anyways, just wondering if anyone knew of a way to pull such diagnostic info?

$\endgroup$

1 Answer 1

1
$\begingroup$

See my answer here for a related question. TMLE does not balance the covariates or adjust the sample in anyway; it estimates potential outcomes for each individual using an outcome model and adjusts the difference in the estimated potential outcome means using a function of the propensity score (called the "clever covariate"). You cannot report balance after TMLE, which is why it is unclear how well it is doing at removing systematic bias in your sample. This is the basis of my critique of it here. You basically have to just trust that it is doing its job based on its asymptotic properties. You can evaluate the quality of some of its predictions using cross-validation statistics, but those do not correspond exactly to how well bias is being removed.

$\endgroup$
1
  • $\begingroup$ Thanks, Noah! This is really helpful. $\endgroup$ Jul 28, 2022 at 20:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.