Honesty in Causal Forests: When It Helps and When It Hurts
Information Systems Research · 14.10.2026
DOI: 10.1287/isre.2025.2343 ↗Kurz gefasst
Fachartikel in der Zeitschrift Information Systems Research, online erschienen am 14.10.2026. Verfasst von Yanfang Hou und 1 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 13 Literaturangaben.
Abstract (Originalsprache)
Causal forests are increasingly used to estimate who benefits most from an intervention, informing targeting and personalization in business and public policy. Widely used software often applies a safeguard called “honest estimation” by default. It splits the data: one part is used to find groups of people who respond differently to the intervention, and another to estimate the effect within each group. Our key finding is that splitting the data reduces the risk of mistaking chance patterns for real ones, but it also makes real differences harder to detect, which can make estimates less accurate. Across more than 7,000 benchmark data sets, we found that honest estimation can require up to 27% more data to reach the same accuracy as causal forests that did not use it. The takeaway is that honesty should be a choice, not a default. Whether it helps depends on the data and what the estimates are needed for. For organizations and policymakers planning experiments, its potential data cost should also be part of the design conversation.
Zitieren
APA
Hou, Y., & Fernández-Loría, C. (2026). Honesty in Causal Forests: When It Helps and When It Hurts. Information Systems Research. https://doi.org/10.1287/isre.2025.2343
Vancouver
Hou Y, Fernández-Loría C. Honesty in Causal Forests: When It Helps and When It Hurts. Information Systems Research. 2026. doi:10.1287/isre.2025.2343
BibTeX
@article{Hou2026,
author = {Hou, Yanfang and Fernández-Loría, Carlos},
title = {Honesty in Causal Forests: When It Helps and When It Hurts},
journal = {Information Systems Research},
year = {2026},
doi = {10.1287/isre.2025.2343}
}
EUVIMED