TL;DR: Wrote a blog post exploring TMLE's "doubly robust" property through simulations. Finally got my head around how it works - basically gives you two chances to get it right (either outcome OR treatment model). Tested it with logistic regression + XGBoost and compared bias/variance across different model specifications. XGBoost + TMLE did pretty well at capturing complex relationships without manual specification.
https://www.kenkoonwong.com/blog/tmle/
Still learning here - would love feedback! Got some great input from Frank Harrell suggesting I examine 95% CI coverage, so I'm rerunning the simulations now. More to come on that front.
Has anyone actually applied TMLE in real-world observational data (not just sims)? Curious how it holds up when you don't know the true DGP. Any gotchas or tips appreciated!
------------------------------
Ken Koon Wong MD
Faculty
Cleveland Clinic Akron General
Akron OH
------------------------------