Heterogeneous Treatment Effect Sensitivity Analysis for Causality
Sensitivity Analysis for Causality under Heterogeneous Treatment Effects · Also known as: HTE sensitivity analysis, heterogeneous-effects sensitivity analysis, sensitivity analysis with effect heterogeneity, HTE robustness analysis
Heterogeneous Treatment Effect Sensitivity Analysis examines how robust subgroup-specific causal estimates are to unobserved confounding. Rather than testing a single average treatment effect, it asks whether the estimated variation in treatment effects across units or subgroups could be explained away by hidden bias, and at what level of hidden bias the causal conclusions for each subgroup would break down.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use this method after estimating heterogeneous treatment effects from observational data, when you want to assess whether the observed variation in effects across subgroups is robust to unobserved confounding. It is appropriate when treatment was not randomly assigned, when you suspect unmeasured differences between subgroups, and when the policy implication of heterogeneity is consequential. It is not a replacement for causal identification; it complements matching, weighting, or quasi-experimental designs. Do not use it as a substitute for randomisation or to manufacture credibility for poorly identified estimates — sensitivity analysis cannot rescue a fundamentally confounded design.
Strengths & limitations
- Provides an honest assessment of how robust heterogeneous effect estimates are to unmeasured confounding, going beyond simple significance tests.
- Allows researchers to communicate the fragility or credibility of subgroup findings in a principled, quantified way.
- Compatible with a wide range of HTE estimators — regression with interactions, causal forests, meta-learners, or matching-based estimators.
- Helps prioritise which subgroup conclusions are policy-relevant versus speculative.
- Does not require positing a specific unmeasured confounder — Gamma bounds agnostically cover all confounders of a given strength.
- Sensitivity bounds can be wide and uninformative when the sample in a subgroup is small, making it hard to distinguish robust from fragile conclusions.
- The Gamma framework assumes a specific model of how unmeasured confounding enters; violations of this model may over- or under-estimate sensitivity.
- Sensitivity analysis diagnoses vulnerability but does not fix it — if bounds are wide, the remedy is better identification, not more sensitivity analysis.
- Computationally intensive when combined with flexible ML-based HTE estimators and many subgroups.
Frequently asked
What does the sensitivity parameter Gamma mean?
Gamma is an upper bound on the odds ratio by which two similar units can differ in their probability of treatment due to an unmeasured covariate. Gamma = 1 means no hidden bias; Gamma = 2 means an unobserved variable could double the odds of treatment for one unit over another.
How is this different from standard sensitivity analysis?
Standard Rosenbaum sensitivity analysis tests whether the average treatment effect is robust. The heterogeneous-effects version applies the same logic separately to each subgroup or covariate stratum, so you learn which specific subgroup conclusions are fragile and which survive hidden bias.
Which estimators can I pair with this sensitivity framework?
Any estimator that produces subgroup-specific effect estimates: interacted OLS, propensity-matched subgroup comparisons, causal forests, meta-learners (X-learner, R-learner), or doubly robust CATE estimators. The sensitivity analysis is a post-estimation step applied to the resulting subgroup estimates.
What if my subgroup sample sizes are small?
Small subgroups yield wide sensitivity bounds, making it hard to draw firm conclusions. In this case, either pool subgroups into broader strata with adequate sample sizes, or interpret subgroup-specific findings as exploratory rather than confirmatory.
Does a high critical Gamma prove my heterogeneous effects are causal?
No. A high Gamma threshold means the findings are relatively robust to a broad class of hypothetical unmeasured confounders, which increases credibility. It does not rule out all possible confounders or establish causation with certainty.
Sources
- Rosenbaum, P. R. (2002). Observational Studies (2nd ed.). Springer. ISBN: 978-0387989679
- Crump, R. K., Hotz, V. J., Imbens, G. W., & Mitnik, O. A. (2008). Nonparametric tests for treatment effect heterogeneity. Review of Economics and Statistics, 90(3), 389-405. DOI: 10.1162/rest.90.3.389 ↗
How to cite this page
ScholarGate. (2026, June 3). Sensitivity Analysis for Causality under Heterogeneous Treatment Effects. ScholarGate. https://scholargate.app/en/causal-inference/heterogeneous-treatment-effect-sensitivity-analysis-for-causality
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Difference-in-DifferencesEconometrics↔ compare
- Doubly Robust EstimationCausal inference↔ compare
- Heterogeneous Treatment Effect Difference-in-DifferencesCausal inference↔ compare
- Propensity Score MatchingResearch Statistics↔ compare
- Sensitivity Analysis for CausalityCausal inference↔ compare