Causal Decision Making and Causal Effect Estimation Are Not the Same…and Why It Matters

Causal decision making (CDM) at scale has become a routine part of business, and increasingly, CDM is based on statistical models machine learning algorithms. Businesses algorithmically target offers, incentives, recommendations to affect consumer behavior. Recently, we have seen an acceleration research related causal effect estimation (CEE) using machine-learned models. This article highlights important perspective: not the same as CEE, counterintuitively, accurate CEE necessary for CDM. Our experience that this well understood by practitioners or most researchers. Technically, estimand interest different, implications both modeling use We draw recent highlight three implications. (1) should carefully consider objective function learning, if possible, optimize “treatment assignment” rather than effect-size estimation. (2) Confounding affects differently. The upshot here supporting it may be just good even better learn with confounded data unconfounded data. (3) all support because proxy might do better. third observation helps explain least one broad common practice seems “wrong” first blush—the widespread noncausal targeting interventions. last two are particularly in practice, acquiring (unconfounded) “sides” counterfactual can quite costly often impracticable. These observations open substantial ground. hope facilitate area pointing articles from multiple contributing fields, them written five years.