Hájek Estimator in Clustered & Stratified Designs

• Hájek estimator is a ratio estimator for unbiased estimation of sample average treatment effects in clustered and stratified experimental designs.
• It employs design-based variance estimation using large-stratum and hybrid approaches to account for heterogeneity and small sample sizes.
• Extensions include covariate-adjusted regression, generalized U-statistics, and jackknife variance methods ensuring robust, reproducible inference.

The Hájek estimator encompasses a broad class of ratio estimators prominently used for unbiased estimation of sample average treatment effects (SATE) in clustered and stratified experimental designs. It is pivotal in design-based inference frameworks, extending to generalized U-statistics and supporting theoretically robust variance estimation, including the development of jackknife and hybrid variance estimators. The estimator retains Neyman-style conservativeness and delivers valid inference under varied strata sizes and covariate structures, particularly when heterogeneity and small sample sizes challenge standard variance estimation strategies (Wang et al., 2024, Juergens, 15 Sep 2025).

1. Mathematical Formulation and Properties

Consider a stratified, clustered randomized experiment with $B$ strata (blocks). Within stratum $b$, there are $n_b$ clusters, each cluster $i \in b$ containing $m_{bi}$ units. The normalized cluster weights are defined by

$$w_{bi} = \frac{m_{bi}}{\sum_{b'=1}^B \sum_{i' \in b'} m_{b'i'}}$$

and the stratum weights by $\omega_b = \sum_{i \in b} w_{bi}$, summing to one across all strata.

The observed, cluster-weighted means in each stratum (for treatment $z \in \{0,1\}$) are

$$\bar y_{bz}^{\rm obs} = \frac{\sum_{i\in b:\,Z_{bi}=z} w_{bi} Y_{bi}}{\sum_{i\in b:\,Z_{bi}=z} w_{bi}}$$

with $Z_{bi}$ indicating cluster assignment. The Hájek estimator for SATE is

$b$0

which is unbiased for the finite-population SATE: $b$1 whenever each stratum's randomization is complete (Wang et al., 2024).

2. Design-Based Variance Estimation

Variance estimation for the Hájek estimator in clustered, stratified settings relies on weighting and stratum-specific variance calculations. For each cluster and arm, define

$b$2

with $b$3 typically replaced by an estimate of the marginal mean. Within each stratum, the sample variance is

$b$4

The large-stratum variance estimator is

$b$5

where $b$6 are the weighted stratum marginal means. This estimator is asymptotically conservative—its expectation exceeds the true sampling variance unless treatment effects are constant within clusters (Wang et al., 2024).

To maintain finite-sample validity, especially for small strata (e.g., matched pairs), a hybrid variance estimator is recommended:

• Use the large-stratum estimator if $b$7.
• Use a pairwise estimator if $b$8. Combine by stratum size: $b$9 with $n_b$0 (large/pairwise) (Wang et al., 2024).

3. Asymptotics, Conservativeness, and Theoretical Guarantees

Key regularity conditions for consistency include: bounded cluster sizes, non-degenerate assignment probabilities within strata, increasing number or size of strata, and control on extreme values to ensure central limit theorems apply. Both large-stratum and hybrid estimators guarantee

$n_b$1

so that Wald-type confidence intervals achieve at least nominal asymptotic coverage (Wang et al., 2024).

In hypothesis testing, score-style tests based on inversion of the Hájek estimating equations yield exact levels under the sharp null and are asymptotically valid under the weak null.

4. Extensions: Covariate Adjustment and Regression Approaches

Precision gains are available via regression adjustment within the Hájek framework. Using weighted least squares (WLS), regress outcomes on treatment indicators and baseline covariates with weights $n_b$2. The resulting coefficient on $n_b$3 is a covariate-adjusted Hájek estimator $n_b$4, which remains unbiased for SATE by design-based arguments. Variance estimation on regression residuals, optionally employing cluster-robust adjustments (e.g., HC2 or CR2), ensures conservative inference even with model misspecification (Wang et al., 2024).

5. Generalized U-Statistics, Hájek Projection, and Jackknife Variance Estimation

For generalized U-statistics, the asymptotic behavior of the estimator’s variance is governed by the Hájek projection, the leading term in the Hoeffding decomposition. The Hájek-domination condition requires

$n_b$5

where $n_b$6 is the U-statistic, and $n_b$7 its order-1 projection. Under this condition and $n_b$8, nonparametric jackknife and delete-$n_b$9 jackknife variance estimators achieve ratio-consistency: $i \in b$0 This framework encompasses classical fixed-order results as special cases and generalizes to settings where the kernel order grows with $i \in b$1 or sampling is incomplete. It places the regular jackknife on equal theoretical footing with the infinitesimal jackknife (Juergens, 15 Sep 2025).

6. Empirical Performance and Practical Recommendations

Simulation studies indicate that the Hájek estimator is unbiased under heterogeneous treatment effects, even when other estimators (e.g., Imai–King–Nall, regression with fixed effects) become biased, especially when heterogeneity correlates with cluster or stratum size. The large-strata variance estimator is conservative in small strata, leading to over-coverage, while the hybrid estimator restores nominal coverage levels without excessive interval width.

Applied examples—nutritional studies in paired clusters, educational trials with cluster-level covariates—demonstrate that the hybrid variance estimator yields coverage near nominal, while traditional estimators can over-cover by a large margin. Covariate adjustment reduces standard error without jeopardizing coverage (Wang et al., 2024).

7. Significance and Connections to Broader Methodological Literature

The Hájek estimator and its associated variance estimators constitute a robust, design-based approach to impact estimation in randomized experiments with complex assignment structures (clustering, stratification, fine blocking). They require only randomization (not model specification) for validity. The connection to U-statistics and jackknife variance estimation aligns with longstanding principles in nonparametric inference and further supports its extensibility to modern high-dimensional and adaptive designs (Juergens, 15 Sep 2025).

Under the minimal variance-dominance condition, validity of asymptotic normal approximations and jackknife variance estimates is unified across a wide array of applications, from classical fixed-order statistics to contemporary nonparametric regression estimators. This approach ensures rigorous, reproducible inference in both large-scale and small-sample, highly stratified designs.