From Tokens to Policy: Minimal and Sufficient Heterogeneous Treatment Effects Identification

Heterogeneous Treatment Effect (HTE) characterization is crucial for understanding for whom an intervention works. Practical approaches trade expressivity for interpretability, but even full interpretability can still surface marginal effect modifiers that admit no causal reading, offering no guidance on what would happen intervening on them. We argue the right target is a minimal and sufficient HTE characterization given by the standalone exposure-interaction terms, i.e., the direct modifiers that d-separate the treatment effect from any other pre-treatment observation. This identification, previously out of reach, is now enabled by extensive multimodal pre-treatment measurements and representation learning pipelines. We recast the task as Markov-blanket discovery on a learned concept dictionary and introduce Neural EXposure Interaction Search (NEXIS), with provable asymptotic recall and finite-sample precision. We deploy NEXIS on two anti-poverty programs in Africa, augmenting each with satellite imagery capturing previously unmeasured environmental effect modifiers, leading to concrete and prescriptive hypotheses to maximize impact in their future iterations.