What happens when causal overlap becomes weak?

Question: In observational causal estimation, estimated propensity scores are extremely close to zero for some treated observations and extremely close to one for some controls. Which interpretation is most accurate? A) Identification remains unaffected because extreme propensity scores indicate highly predictable treatment assignment B) Only outcome-regression estimators are affected; weighting estimators remain stable after normalisation C) Inverse-probability weights can become unstable, and trimming may reduce variance while changing the effective target population D) Cross-fitting automatically restores positivity because each propensity model is trained on fewer observations Correct: C Explanation: Weak overlap creates very large inverse-probability weights and can produce unstable estimates with high variance. Clipping or trimming may improve numerical behaviour, but it can introduce bias or redefine the population for which the effect is being estimated. Topic: advanced ML / causal inference / positivity