Testing Treatment Effect Heterogeneity with Functional Outcomes
Abstract
Assessing treatment-effect heterogeneity is important in causal inference, particularly for functional outcomes where effects may vary over time and the common effect is unknown. We propose a nonparametric procedure that assesses treatment-effect homogeneity through a testable marginal implication. By comparing within-arm difference distributions, the construction removes the unknown common functional effect without estimating it. We measure departures between these distributions through an integrated squared discrepancy of their characteristic functionals. A full-support Gaussian integrating measure yields an explicit bounded-kernel representation and identifies equality of the difference laws without requiring moment conditions on the outcomes. We construct an unbiased generalized U-statistic that accounts for dependence among overlapping differences, derive its null, fixed-alternative, and local-alternative limits, and develop a centered multiplier bootstrap for inference. The framework extends to multiple treatment arms with a shared control. For computation, a random-direction approximation combined with a nine-state recursion reduces the cost to linear order in the sample size for fixed numerical settings. Simulation studies assess the finite-sample performance of the proposed procedure relative to several benchmark methods. An application to the Irish gas smart-metering trial illustrates its use for assessing heterogeneity in functional treatment responses.
Related
- Shape-Preserving Treatment Effects for Functional Outcomes
- Testing for Unobserved Heterogeneity in Censored Duration Models: EM Approach
- Identifying and Measuring Treatment-Effect Heterogeneity across Sites through Within-Site Randomization
- Splitting the Difference: Interpretable Causal Forests for Treatment Effect Heterogeneity and Bias
- Robust Global Testing for Mean Functions in Sparse and Dense Functional and Longitudinal Data
BibTeX
@misc{feng2026testing,
title = {{Testing Treatment Effect Heterogeneity with Functional Outcomes}},
author = {Yongzhen Feng and Jie Li and Xiaojun Song},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.23397},
eprint = {2610.23397},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.23397},
}
RIS
TY - UNPB
TI - Testing Treatment Effect Heterogeneity with Functional Outcomes
AU - Yongzhen Feng
AU - Jie Li
AU - Xiaojun Song
PY - 2026
DA - 2026/10/16
PB - arXiv
DO - 10.48550/arxiv.2610.23397
UR - https://arxiv.org/abs/2610.23397
AB - Assessing treatment-effect heterogeneity is important in causal inference, particularly for functional outcomes where effects may vary over time and the common effect is unknown. We propose a nonparametric procedure that assesses treatment-effect homogeneity through a testable marginal implication. By comparing within-arm difference distributions, the construction removes the unknown common functional effect without estimating it. We measure departures between these distributions through an integrated squared discrepancy of their characteristic functionals. A full-support Gaussian integrating measure yields an explicit bounded-kernel representation and identifies equality of the difference laws without requiring moment conditions on the outcomes. We construct an unbiased generalized U-statistic that accounts for dependence among overlapping differences, derive its null, fixed-alternative, and local-alternative limits, and develop a centered multiplier bootstrap for inference. The framework extends to multiple treatment arms with a shared control. For computation, a random-direction approximation combined with a nine-state recursion reduces the cost to linear order in the sample size for fixed numerical settings. Simulation studies assess the finite-sample performance of the proposed procedure relative to several benchmark methods. An application to the Irish gas smart-metering trial illustrates its use for assessing heterogeneity in functional treatment responses.
ER -