Decision-theoretic comparison of point-identified LATT and set-valued ATT

Characterize, using a decision-theoretic framework, the region in which the point-identified credible-subpopulation local average treatment effect on the treated (LATT) dominates the set-valued average treatment effect on the treated (ATT), thereby converting the documented scope condition into a formal recommendation.

Background

The paper proposes replacing the conventional ATT with a credible-subpopulation LATT when parallel trends is credible only for selected treatment cohorts. The LATT is point-identified under parallel trends for those selected cohorts, whereas sensitivity analysis for the full ATT produces a set that may be wide and uninformative when some cohorts violate parallel trends. The paper emphasizes that the LATT's advantage depends on the informativeness of pre-treatment trends about post-treatment violations: selecting a narrower subpopulation can improve credibility and precision but reduces scope. The unresolved problem is to formalize this trade-off decision-theoretically by determining when the point-identified LATT is preferable to the set-valued ATT.