Causal Bandits Podcast Create Your Causal Inference Roadmap. Causal Inference, TMLE & Sensitivity | Mark van der Laan S2E6 | CausalBanditsPodcast.com
Sep 22, 2025
Mark van der Laan, a renowned professor at UC Berkeley and the mastermind behind Targeted Maximum Likelihood Estimation (TMLE), dives deep into causal inference. He differentiates between TMLE and double machine learning, emphasizing their unique applications. Mark shares insights on building a stepwise causal roadmap and the importance of uncertainty quantification. He discusses practical applications of his work and reflects on the role of large language models in research. His advice encourages diversity and rigor in the causal inference community.
AI Snips
Chapters
Books
Transcript
Episode notes
Formal Causal Roadmap
- A causal roadmap forces you to formalize the experiment, likelihood, causal model, identification, and estimation steps before analysis.
- This structure reveals separate sources of uncertainty: causal gap and statistical estimation error.
Quantify Sensitivity With A G-Value
- Plot how varying assumed unmeasured confounding (the causal gap) shifts your confidence intervals to test robustness.
- Report the maximal gap (G-value) that would overturn your conclusion to quantify plausibility of bias.
Use Negative Controls To Bound Bias
- Use negative controls and leave-one-out confounder checks to bound plausible unmeasured confounding.
- Translate those empirical shifts into interpretable units to judge how credible a causal gap is.
