How to choose a causal method

No single method wins every problem. Start with the causal question and data-generating process, not with the flashiest algorithm.

Quick guide

• Randomized experiment → simple difference in means or regression with covariates.
• Binary treatment, rich covariates, cross-section → matching, IPW, or doubly robust ML.
• Policy timing, treated vs control regions → difference-in-differences (check parallel trends).
• Eligibility cutoff or sharp rule → regression discontinuity.
• Endogenous treatment with valid instrument → IV.
• Many covariates, heterogeneous effects → causal forests or DML.
• Unknown graph, exploratory phase → discovery plus expert review.

Run more than one

Credible practice compares multiple methods that rely on different assumptions. If PSM, IPW, and a causal forest agree, confidence increases. If they diverge, investigate overlap, functional form, and unmeasured confounding before recommending action.