Estimand frameworks for causal inference in clinical research and regulatory decision making
Abstract
Causal inference for clinical and regulatory decision making requires a precise definition of the causal estimand, that is, what we want to estimate. Three frameworks are commonly used to describe causal estimands: PICO and its extensions, the target trial framework, and the E9(R1) Addendum of the International Council for Harmonisation (ICH). Because these frameworks were developed by different scientific communities, they operate at different levels of detail, emphasize different aspects of the causal estimand, and use different terminology for closely related concepts. Here we compare the three frameworks and translate across their terminologies. PICO is a compact framework that helps jumpstart a conversation about the causal question, but it does not encode sufficient information to fully characterize a causal estimand. The target trial framework characterizes the causal estimand precisely by specifying the components of the protocol of the randomized trial that would answer the causal question: eligibility criteria, treatment strategies, assignment, outcome, start and end of follow-up, and causal contrast. E9(R1) starts from intercurrent events and lists strategies to address them, some of which describe estimators rather than estimands, or imply estimands that are not relevant to decision makers or cannot be estimated under reasonable assumptions. We argue that the target trial framework subsumes the other two, and we provide a map of the correspondences across the three frameworks.
Related
- Automated, Efficient, and Model-Free Inference for Randomized Clinical Trials via Data-Driven Covariate Adjustment
- When AI Generates Covariates: Causal Typing and Estimand Drift in Sequential Experiments
- Estimand or estimator? Comparing vowel overlap measures against a known ground truth
- Aggregation Distortion in Multilevel Mediation: Estimand Geometry, Design Dependence, and Diagnostics
- CausalSmith: A Formally Grounded, Self-Improving Agentic Framework for Automated Research in Causal Inference
BibTeX
@misc{hernan2026estimand,
title = {{Estimand frameworks for causal inference in clinical research and regulatory decision making}},
author = {Miguel A. Hernán and Mats J. Stensrud and Xabier García-Albéniz and Issa J. Dahabreh and Jonathan A. C. Sterne},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.28983},
eprint = {2610.28983},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.28983},
}
RIS
TY - UNPB
TI - Estimand frameworks for causal inference in clinical research and regulatory decision making
AU - Miguel A. Hernán
AU - Mats J. Stensrud
AU - Xabier García-Albéniz
AU - Issa J. Dahabreh
AU - Jonathan A. C. Sterne
PY - 2026
DA - 2026/10/20
PB - arXiv
DO - 10.48550/arxiv.2610.28983
UR - https://arxiv.org/abs/2610.28983
AB - Causal inference for clinical and regulatory decision making requires a precise definition of the causal estimand, that is, what we want to estimate. Three frameworks are commonly used to describe causal estimands: PICO and its extensions, the target trial framework, and the E9(R1) Addendum of the International Council for Harmonisation (ICH). Because these frameworks were developed by different scientific communities, they operate at different levels of detail, emphasize different aspects of the causal estimand, and use different terminology for closely related concepts. Here we compare the three frameworks and translate across their terminologies. PICO is a compact framework that helps jumpstart a conversation about the causal question, but it does not encode sufficient information to fully characterize a causal estimand. The target trial framework characterizes the causal estimand precisely by specifying the components of the protocol of the randomized trial that would answer the causal question: eligibility criteria, treatment strategies, assignment, outcome, start and end of follow-up, and causal contrast. E9(R1) starts from intercurrent events and lists strategies to address them, some of which describe estimators rather than estimands, or imply estimands that are not relevant to decision makers or cannot be estimated under reasonable assumptions. We argue that the target trial framework subsumes the other two, and we provide a map of the correspondences across the three frameworks.
ER -