Heterogeneous Treatment Effect Instrumental Variables (HTE-IV)
Instrumental Variables Estimation with Heterogeneous Treatment Effects · Also known as: HTE-IV, LATE estimator, IV with effect heterogeneity, local average treatment effect IV
Heterogeneous treatment effect IV applies instrumental variables estimation while explicitly acknowledging and modelling that the treatment effect differs across units. Rather than recovering a single average effect, it focuses on the Local Average Treatment Effect (LATE) — the causal effect for compliers, the subpopulation whose treatment status is actually shifted by the instrument — and extends analysis to variation in that effect across observed subgroups.
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 HTE-IV when you have a valid instrument that shifts treatment for some units but not others, and you suspect or want to test whether the treatment effect varies across observed subgroups. It is appropriate with observational data where treatment assignment is endogenous and randomisation is infeasible, and when the policy question involves targeting or distributional effects. The instrument must satisfy relevance (strong first stage, ideally F > 10), exclusion restriction, and monotonicity. Do not use it when the instrument is weak, when you have no theoretical basis for effect heterogeneity, when subgroup sample sizes are too small for reliable inference, or when you want to recover the Average Treatment Effect for the full population rather than for compliers.
Strengths & limitations
- Identifies a well-defined causal effect (LATE) for compliers under weaker assumptions than OLS, without requiring full randomisation.
- Explicitly models who benefits from treatment, enabling more targeted policy recommendations than a single average effect.
- Robust to selection bias and endogeneity of treatment, provided the instrument is valid.
- Can combine with pre-specified subgroup analysis to test theoretically motivated hypotheses about effect modifiers.
- The LATE framework is transparent about the population to which the estimate applies, avoiding overgeneralisation.
- The LATE is specific to compliers and may not generalise to always-takers, never-takers, or the full population, limiting external validity.
- Finding instruments that are both relevant and plausibly exogenous is difficult in practice; a weak or invalid instrument produces severely biased estimates.
- Subgroup LATE analysis multiplies the number of estimates and raises concerns about multiple testing and reduced statistical power within strata.
- The monotonicity assumption (no defiers) is untestable and may be implausible in some settings.
- Interpretation of effect heterogeneity is conditional on complier composition, which may differ across subgroups.
Frequently asked
What is the difference between LATE and ATE in this context?
The LATE (Local Average Treatment Effect) is the causal effect of treatment for compliers — units whose treatment status changes because of the instrument. The ATE is the average effect for the entire population. IV with binary instruments recovers the LATE, not the ATE, because it only uses variation generated by the instrument, which moves only compliers.
How do I know whether my instrument is strong enough?
Report the first-stage F-statistic. A rule of thumb from Staiger and Stock (1997) is F > 10 for a single instrument; more recent guidance by Lee et al. (2022) suggests higher thresholds for reliable inference. If F is below 10, consider using weak-instrument-robust inference methods such as the Anderson-Rubin test.
Can I estimate heterogeneous LATEs with machine learning?
Yes. Methods such as causal forests (Wager and Athey, 2018) and double/debiased machine learning (Chernozhukov et al., 2018) extend IV to estimate conditional average treatment effects flexibly. These approaches regularise the estimation while maintaining valid inference, and are especially useful when the number of potential effect modifiers is large.
What if heterogeneity is driven by differences in complier composition across subgroups?
Complier composition can differ across subgroups: for example, women and men who comply with an instrument may have different baseline characteristics. Marginal treatment effect (MTE) methods, developed by Heckman and Vytlacil, model how effects vary with the propensity to comply, helping separate genuine effect heterogeneity from compositional differences.
When should I prefer HTE-IV over heterogeneous-effect propensity score matching?
Use HTE-IV when treatment is endogenous and you have a valid instrument; propensity score matching assumes treatment is as good as random conditional on observables (unconfoundedness). If unconfoundedness is plausible, matching can recover effects for a broader population. If a valid instrument exists and endogeneity is the main concern, IV is preferred even though it restricts inference to compliers.
Sources
- Imbens, G. W., & Angrist, J. D. (1994). Identification and Estimation of Local Average Treatment Effects. Econometrica, 62(2), 467-475. DOI: 10.2307/2951620 ↗
- Angrist, J. D., & Pischke, J.-S. (2009). Mostly Harmless Econometrics: An Empiricist's Companion. Princeton University Press. ISBN: 978-0691120355
How to cite this page
ScholarGate. (2026, June 3). Instrumental Variables Estimation with Heterogeneous Treatment Effects. ScholarGate. https://scholargate.app/en/causal-inference/heterogeneous-treatment-effect-instrumental-variables
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.
- 2SLS RegressionEconometrics↔ compare
- Heterogeneous Treatment Effect Propensity Score MatchingCausal inference↔ compare
- Instrumental Variables in Health ResearchHealth Economics↔ compare
- Local Average Treatment EffectCausal inference↔ compare