What is Heterogeneous Treatment Effects?
Heterogeneous Treatment Effects are systematic differences in a treatment's causal effect across covariate-defined people, groups, settings, or time horizons rather than random variation around one population-average effect.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Separate prognostic risk from treatment-effect modification
A prognostic variable changes baseline outcome risk under both arms; an effect modifier changes the treatment contrast on the declared scale. High-risk patients can have larger absolute benefit even when relative effects are constant. The PATH Statement distinguishes risk-modeling and effect-modeling approaches and emphasizes patient-centered contrasts under both alternatives.
Discovery creates multiplicity and overfitting risk
One-variable-at-a-time subgroup tables lose power and invite contradictory stories when many attributes are tested. Flexible CATE learners can model interactions but create a larger search space. Prespecify confirmatory modifiers where possible; for exploration, use honest sample splitting or cross-fitting, retain null findings, correct multiplicity where relevant, and avoid interpreting variable importance as a causal mechanism.
Validate whether heterogeneity changes a decision
Useful HTE must generalize and cross a meaningful decision threshold after treatment cost, burden, harm, and uncertainty are considered. Validate calibration or grouped effects on held-out randomized data, inspect Overlap, compare policy value with treat-all and treat-none baselines, and test transport across sites and time. A visually varied CATE distribution alone is not evidence of actionable heterogeneity.
Key Characteristics
- Describes non-random variation in a declared causal contrast
- Can be hidden by an Average Treatment Effect near zero
- Differs from ordinary variation in baseline outcome risk
- Depends on effect scale, time horizon, and target population
- Can be studied through prespecified subgroups or flexible CATE models
- Requires out-of-sample validation before personalized decisions
Common Use Cases
- Identifying patient groups with different benefit-harm trade-offs
- Testing whether a product feature helps new and expert users differently
- Finding policy effects that vary across regions or baseline risk
- Designing follow-up experiments around plausible effect modifiers
- Evaluating whether personalization improves policy value over one rule
Example
Loading code...Frequently Asked Questions
How are HTE and CATE related?
HTE is the phenomenon or analysis goal: treatment effects vary across covariates or contexts. CATE is a conditional estimand used to quantify that variation. A CATE model can estimate a smooth surface, while HTE can also be examined through prespecified subgroup contrasts.
Does different baseline risk prove a Heterogeneous Treatment Effect?
No. A variable may predict outcomes equally under both treatments and leave the causal contrast unchanged. On an absolute scale, constant relative effects can still produce larger differences at higher baseline risk, so the effect scale must be declared.
Why are many subgroup analyses unreliable?
Treatment-by-subgroup interactions often have low power, while searching many cutoffs and attributes creates false discoveries. Small cells also worsen overlap and uncertainty. Prespecification, shrinkage, honest validation, multiplicity control, and replication reduce these risks.
Can a machine-learning model prove treatment-effect heterogeneity?
No. It can estimate patterns under a design and assumptions, but flexible models also fit noise. Evidence requires valid identification, held-out or honest estimation, uncertainty, calibration or grouped-effect checks, and preferably replication in another sample or experiment.
When is HTE actionable?
It is actionable when validated effect differences cross decision thresholds after benefit, harm, cost, capacity, fairness, and uncertainty are included. Ranking variation without stable policy-value improvement is not enough to justify personalized treatment.