Weighted Averages of Conditional LATEs

• Weighted Averages of Conditional LATEs (WLATEs) is a framework that aggregates local treatment effects using user-defined weights to enhance causal inference in IV settings.
• It employs various weighting schemes—such as uniform, compliance-weighted, and policy-based—to target specific treatment effect subpopulations and improve external validity.
• Robust estimation methods, including TSLS, local polynomial regression, and kernel optimal matching, are applied to address effect heterogeneity and ensure stable inference.

Weighted averages of conditional local average treatment effects (WLATEs) generalize the LATE framework to allow estimation and interpretation of causal effects that are both local (i.e., identified for compliers or other principal strata) and averaged according to a user-specified or design-driven weighting scheme across covariates or latent characteristics. This class of estimands is essential for robust causal inference in instrumental variables (IV) settings with covariates, especially under effect heterogeneity, compliance heterogeneity, and when the estimand of interest is tied to real-world policies or actual experimental designs.

1. Formal Definition and Conceptual Foundation

Let $Y$ be the observed outcome, $D$ the binary treatment, $Z$ a binary instrument, and $X$ or $Z$ a set of observed covariates. The canonical LATE, as defined by Imbens and Angrist, is the average treatment effect among compliers, i.e., individuals whose treatment status genuinely responds to manipulation of the instrument. The conditional LATE, or CLATE, is defined as

$$\tau(z) = E[Y(1) - Y(0) | R = c, Z = z, \Delta \neq 0],$$

where $R$ is a running variable (as in RDD), and $\Delta$ encodes the compliance type at threshold $c$ (Caetano et al., 1 Feb 2026).

A weighted average of CLATEs (WLATE) takes the form

$\mathrm{WLATE}(w) = \int \tau(z) w(z) f_Z(z) dz,$

where $w(z) \geq 0$ is a weighting function and $f_Z(z)$ is the marginal density of covariates. This framework includes a broad class of estimands, including the LATE, compliance-weighted LATE (CWLATE), and bound-type averages in MTE approaches (Caetano et al., 1 Feb 2026, Han et al., 2020, Słoczyński et al., 2022, Boot et al., 2024).

2. Identification of WLATEs

Under standard IV or fuzzy RDD conditions—including continuity, local independence, overlap, and weak monotonicity—WLATEs are identified as a ratio of weighted reduced-form and first-stage discontinuities: $$\mathrm{WLATE}(w) = \frac{ \int w(z) \delta_Y(z) f_Z(z) dz }{ \int w(z) \delta_D(z) f_Z(z) dz },$$ where $$\delta_Y(z) = E[Y | R = c^+, Z = z] - E[Y | R = c^-, Z = z]$$ and similarly for the first-stage effect $\delta_D(z)$ (Caetano et al., 1 Feb 2026).

When IVs are discrete and/or only a single binary IV is observed, the WLATE can also be interpreted via the marginal treatment effect (MTE) function and its policy-relevant weighting scheme: $$\mathrm{WLATE}(\pi) = \iint w(x, u; \pi) \cdot \mathrm{MTE}(x, u)\, dF_{X,U}(x, u),$$ where $w(x, u; \pi)$ encodes the experimental or policy distribution, $u$ is the latent resistance to treatment, and $\pi$ parameterizes the policy or population of interest (Han et al., 2020).

3. Choice and Interpretation of Weighting Schemes

The choice of weights $w(z)$ determines the population over which treatment effects are aggregated. Several notable weighting schemes include:

Weighting Scheme	Definition	Interpretative Target
Uniform	$w(z) = 1$	Average CLATE (ATE if constant effects)
Compliance-weighted	$w(z) \propto [\delta_D(z)]^2$	Maximizes first-stage “strength” (CWLATE)
Policy/PRTE-based	$w(x, u; \pi)$ matches counterfactual policy exposure	Policy-relevant treatment effect under regime change
Kappa weighting	$w(x) = \Pr(D_1 > D_0 | X = x)$	Conditional average among principal strata (compliers)

The CWLATE maximizes correlation between the “instrument” and the first stage and is particularly robust in the presence of compliance heterogeneity (Caetano et al., 1 Feb 2026). Policy-relevant or external-validity effects are achieved by adapting $w(\cdot)$ to counterfactual exposure distributions, as in the PRTE/MTE framework (Han et al., 2020).

4. Estimation and Inference Procedures

When covariates are discrete, plug-in approaches estimate cell-specific reduced-form and first-stage jumps via local polynomial regression:

• For each $z_j$, estimate $\hat\delta_Y(j)$ and $\hat\delta_D(j)$,
• Plug into $$\hat\tau_Y = \sum_j \hat\pi_j \hat\delta_D(j) \hat\delta_Y(j)$$, $\hat\tau_D = \sum_j \hat\pi_j [\hat\delta_D(j)]^2$, yielding $\hat\beta_\mathrm{CW} = \hat\tau_Y / \hat\tau_D$ (Caetano et al., 1 Feb 2026).

Bias correction for local polynomial estimators follows robust bias-corrected inference, e.g., as developed by Calonico–Cattaneo–Titiunik, with explicit bandwidth-and-bias adjustment terms.

For general weighting types and continuous covariates, the kappa-weighting and semiparametrically efficient GMM/AIPW estimators generalize this approach (Słoczyński et al., 2022, Graham et al., 2018). Abadie’s kappa theorem establishes that for any measurable $g(Y, D, X)$,

$$E[g(Y,D,X) | D_1 > D_0] = \frac{E[\kappa \cdot g(Y,D,X)]}{\Pr(D_1 > D_0)},$$

where $\kappa$ is a function of the observed data and the propensity score $p(X)$ (Słoczyński et al., 2022). Normalized (Hajek-type) weighting estimators provide finite-sample stability and invariance properties. Under one-sided compliance, further normalization guarantees strictly positive denominators.

In saturated two-stage least squares (TSLS) settings with many covariate dummies, the “Saturated IV Estimator” (SIVE) provides median-unbiased inference for WLATEs even with weak instruments and high-dimensional controls. Fully robust variance estimation requires Hartley-Rubin-Kronmal (HRK) estimators and bias corrections (Boot et al., 2024).

Kernel optimal matching (KOM) offers a unified nonparametric estimation framework for arbitrary WLATEs by formulating a quadratic program that minimizes the conditional MSE of the weighted estimator over reproducing kernel Hilbert spaces (Kallus et al., 2019).

5. Sharp Bounds and External Validity

In settings where only a binary instrument is available or support is limited, most WLATEs are not point-identified without further shape restrictions. Nonparametric bounds, computed via linear programming over approximating sieves (e.g., Bernstein polynomials for MTE densities), are attainable under monotonicity, concavity, or rank invariance assumptions (Han et al., 2020). WLATEs thus enable credible counterfactual evaluation of policy-reform effects that depart from the observed instrument variation.

A key implication is that conventional LATE is a special case within the WLATE class—with a weight function selecting only the incremental compliers induced by the observed instrument variation. By contrast, WLATEs allow explicit connection to counterfactual distributions and thus external validity.

6. Empirical and Simulation Evidence

Simulation studies in fuzzy RDD and IV settings consistently show that the CWLATE estimator achieves improved mean squared error—often substantially so—relative to standard fuzzy RDD or covariate-adjusted estimators when compliance heterogeneity is present. For example, in calibrated Monte Carlo experiments, CWLATE estimators demonstrate superior stability as first-stage jumps vary (Caetano et al., 1 Feb 2026). Similarly, KOM estimators provide low bias and RMSE under positivity violations and model misspecification in both simulated and real datasets (Kallus et al., 2019).

Empirical applications—such as program evaluation in public health and education, or regression discontinuity analyses of cash transfers—reveal that WLATE-based inference is robust, precise, and often policy-relevant even under weak instrumental variation or substantial effect heterogeneity (Caetano et al., 1 Feb 2026, Boot et al., 2024).

7. Robustness, Implementation, and Practical Considerations

• WLATE estimation is robust to treatment effect heterogeneity and many-instrument settings when implemented with bias-correction and group-wise partitioning (Boot et al., 2024).
• Normalized weighting estimators enjoy desirable invariance properties (translation and scale equivariance), and are robust to outcome reparametrization (Słoczyński et al., 2022).
• Practical computation for WLATEs with complex weightings is feasible, with available software implementing kappa and CB-based estimators (e.g., “kappalate” and R’s “causalweight” (Słoczyński et al., 2022)).
• For continuous controls, discretization or kernel-based adjustment allows extension of the WLATE framework without loss of inferential validity (Graham et al., 2018).
• Under limited support or minimal overlap, sharp bounds can still be constructed if appropriate shape or monotonicity restrictions are credible (Han et al., 2020).

Taken together, the WLATE framework subsumes a broad class of estimands relevant for real-world causal inference: any weighted aggregation of local treatment effects conditional on observed or latent strata. WLATEs provide analytic and inferential flexibility for design-based and policy-driven effect estimation in the presence of treatment effect and compliance heterogeneity.