Local Average Treatment Effects (LATE)

• Local Average Treatment Effects (LATE) is the average causal effect for compliers, defined using potential outcomes and instrumental variables.
• It is crucial in settings with noncompliance, aiding analyses in randomized trials, policy evaluation, and high-dimensional empirical research.
• Advanced estimation methods, including dual instruments and machine learning techniques, improve robustness against bias and model misspecification.

The local average treatment effect (LATE) is the principal causal estimand in instrumental variables (IV) settings under noncompliance, characterizing the average causal effect of treatment for the subpopulation whose treatment status is shifted by the instrument (i.e., “compliers”). Its role is foundational in econometrics and biostatistics, spanning randomized controlled trials (RCTs) with non-adherence, policy evaluation under partial identification, cluster-randomized designs, and more advanced dynamic and high-dimensional frameworks. Recent research extends its identification and inference to a broad spectrum of empirical and theoretical scenarios.

1. Definition and Core Identification of LATE

The canonical LATE framework involves potential outcomes and treatments for each unit, with the instrument $Z$ influencing the actual treatment received, $D$, but not directly the outcome $Y$. Let $D_0$, $D_1$ be potential treatment statuses under $Z=0,1$, and $Y_0$, $Y_1$ the corresponding potential outcomes. The “complier” subpopulation is $\{i:D_1>D_0\}$. The LATE parameter is

$\text{LATE} = \mathbb{E}[Y(1)-Y(0)\mid D_1>D_0].$

Under the standard IV assumptions—(i) independence of $Z$, (ii) exclusion, (iii) monotonicity, and (iv) instrument relevance—LATE is identified via the Wald estimator: $$\text{LATE} = \frac{\mathbb{E}[Y\mid Z=1]-\mathbb{E}[Y\mid Z=0]}{\mathbb{E}[D\mid Z=1]-\mathbb{E}[D\mid Z=0]}$$ as originally established by Imbens and Angrist (1994). This estimator generalizes straightforwardly with covariate adjustment and possesses a clear probabilistic interpretation: it delivers the average causal effect for compliers even when treatment is not randomly assigned (Wang, 2023, Han et al., 2020).

2. Extensions: Imperfect Instruments and Partial Identification

Single instrumental variable strategies may fail when exclusion or monotonicity is violated. A contemporary advance is identification of LATE using two binary instruments, each potentially imperfect:

• The first instrument, $Z_1$, may have direct effects on $Y$ but satisfies independence and monotonicity.
• The second, $Z_2$, is exclusion-valid but may not satisfy monotonicity.

When direct effects of $Z_1$ are homogeneous across strata, all parameters (direct effect, subgroup proportions, and LATE) can be identified from the joint distribution of $(Y, D, Z_1, Z_2)$ via the "difference-in-differences" ratio: $$\text{LATE} = \frac{[E[Y|Z_1=1,Z_2=0]-E[Y|Z_1=0,Z_2=0]] - [E[Y|Z_1=1,Z_2=1]-E[Y|Z_1=0,Z_2=1]]} {[E[D|Z_1=1,Z_2=0]-E[D|Z_1=0,Z_2=0]] - [E[D|Z_1=1,Z_2=1]-E[D|Z_1=0,Z_2=1]]}$$ with sample analogues for estimation (Wang, 2023). This approach robustly mitigates exclusion failures in classical IV setups.

Simulations confirm unbiasedness of the two-instrument estimator where the standard IV estimator is biased if exclusion violations exist (see Table 1 in (Wang, 2023)).

3. Estimation and Inference Across Designs

a. Design-Based and Regression Approaches

In randomized or cluster-randomized experiments, the design-based Wald-IV estimator for LATE is computed via regression: $$\hat\tau_{LATE} = \frac{\text{ITT}_Y}{\text{ITT}_D}$$ where $\text{ITT}_Y$ and $\text{ITT}_D$ are intention-to-treat effects for outcome and treatment, possibly adjusted for covariates. Central limit theorems guarantee asymptotic normality under complete or blocked randomization, and consistent variance estimators are available for clustered/blocked designs (Schochet, 2024, Agbla et al., 2018).

b. Advanced Estimators

Doubly robust estimators utilize quasilikelihood methods weighted by the inverse instrument propensity score (IPWRA), yielding consistency if either the propensity-score or outcome models are correctly specified. Analytical or bootstrapped inference is valid, with robustness to moderate misspecification and sample size (Słoczyński et al., 2022).

Mixture models using substantive model compatible multiple imputation (SMC MIC) efficiently impute latent compliance classes, substantially outperforming conventional two-stage IV methods in binary or missing outcome settings and effectively utilizing auxiliary predictive information (DiazOrdaz et al., 2018).

c. High-Dimensional and Dynamic Settings

High-dimensional covariate setups leverage orthogonal (Neyman-orthogonal) scores and double/debiased machine learning (DML). Conditional quasi-LR analogues of the Anderson-Rubin statistic provide size-robust inference in weak identification regimes and with $p\gg n$ (Ma, 2023).

Dynamic extensions define period-specific and path-dependent LATEs (dynamic LATEs), incorporating time-varying instruments and treatments and exploiting sequential one-sided noncompliance or staggered adoption assumptions. Cross-period identification typically requires added effect-compliance independence conditions, with estimator construction paralleling modern DML frameworks (Sojitra et al., 2024, Casini et al., 16 Sep 2025).

4. LATE Under Covariates, Grouped Instruments, and Heterogeneous Effects

Saturated two-stage least squares (TSLS) in covariate-rich settings targets a weighted average of covariate-specific LATEs, but incurs substantial finite-sample bias (order $G/n$) as the number of instruments grows. The recommended alternative is the saturated IV estimator (SIVE), which employs jackknife bias correction and delivers asymptotically valid inference, robust even under weak instruments and unobserved heterogeneity (Boot et al., 2024).

Complex scenarios with multiple instruments and heterogeneous effects motivate instrument clubbing and plurality-based selection. Agglomerative clustering on estimated propensity scores, followed by reduced-form clustering, identifies sets of instruments with homogeneous complier groups and true-validity via plurality. This procedure yields consistent and asymptotically normal estimators, operationalizing valid LATE estimation even when many IVs are invalid (Apfel et al., 2022).

5. LATE Under Partial Identification and Sensitivity Analysis

When key assumptions (exclusion, monotonicity, instrument independence) are only partially met or when treatment is misclassified, the LATE and related marginal treatment effects (MTE) are only partially identified. Techniques include LP-based sharp bound computation nonparametrically, accommodating additional mean-independence, monotone treatment selection, and shape restrictions. These bounds—often dramatically narrowed by shape constraints—enable extrapolation to counterfactual environments and inform external validity, though the "intrinsic locality" of LATE remains a fundamental limitation (Han et al., 2020, Acerenza et al., 2021).

Sensitivity analysis methods address scenarios with censoring by death, where outcomes are only observed conditional on survival. A principal-strata mixing assumption, solved via GMM-type covariate balancing, enables identification and consistent estimation of the LATE among "complier-always-survivors." This approach is robust to misspecification of sensitivity parameters and provides efficiency gains over classic IPW estimators (Lee et al., 2018).

6. Power Analysis, Weighting Properties, and Distributional Effects

Analytical power calculations for LATE estimation must account for compliance rates, sample size, and effect size, as nuisance variances and strata means are not directly estimable. A distribution-free power analysis provides sharp lower bounds on power and conservative estimates for minimum detectable effects, relying on compliance rate alone (with optional bound tightening via ordered means) (Bansak, 2016).

Abadie's kappa theorem underpins weighting estimators for LATE; normalized weighting estimators (e.g., Abadie-Uysal) are recommended for their scale and translation invariance, denominator stability under one-sided noncompliance, and superior finite-sample properties compared to unnormalized alternatives (Słoczyński et al., 2022).

Distributional LATEs generalize standard mean-based LATE to the entire shift in outcome distributions among compliers. Orthogonal score-based, cross-fitted estimators (e.g., DML with Random Forests or Kolmogorov-Arnold Networks) robustly estimate quantile and CDF effects, with estimator performance and substantive conclusions depending strongly on the chosen nuisance function estimator (Shaw, 15 Jun 2025).

7. Practical Guidance and Implications

• Robust identification of LATE extends to designs with two imperfect instruments, multiple periods, staggered adoption, clustered samples, high-dimensional controls, and dynamic time series.
• In finite samples and under model uncertainty, practitioner attention should focus on estimator normalization, covariate balance, instrument strength, partial identification bounds, and cross-validation of inference methods.
• Empirical implementation is routinely feasible with off-the-shelf statistical software (e.g., R, Stata), but careful sensitivity and diagnostic analysis is essential.
• LATE remains fundamentally "local": its external validity hinges on access to richer instruments, auxiliary restrictions, or informative extrapolation bounds.
• Instrument selection and model choice for nuisance function estimation can substantively alter the empirical estimates of LATE and its heterogeneity.

---

Key references: (Wang, 2023, Han et al., 2020, DiazOrdaz et al., 2018, Słoczyński et al., 2022, Bansak, 2016, Miyaji, 2024, Boot et al., 2024, Zhong et al., 2024, Acerenza et al., 2021, Słoczyński et al., 2022, Shinoda et al., 2021, Apfel et al., 2022, Ma, 2023, Lee et al., 2018, Schochet, 2024, Sojitra et al., 2024, Agbla et al., 2018, Casini et al., 16 Sep 2025, Shaw, 15 Jun 2025).