Stratified Two-Stage Cluster Sampling (STWCS)

• STWCS is a compound probability sampling design that partitions populations into strata and samples clusters (PSUs) and subclusters (SSUs) to achieve unbiased estimation.
• The method employs PPSWOR for primary cluster selection and SRS or systematic sampling for subunits, with design weights ensuring efficient variance estimation.
• It is applied in survey sampling, cluster-randomized trials, and annotation quality, demonstrating significant efficiency gains and cost reductions in large-scale studies.

Stratified Two-Stage Cluster Sampling (STWCS) is a compound probability sampling design that partitions the population into strata and, within each stratum, samples clusters (primary sampling units, PSUs) and then elements or subclusters (secondary sampling units, SSUs) from the sampled clusters. The STWCS framework provides a comprehensive toolkit for unbiased estimation of means, totals, and causal effects, and enables efficient cost-constrained inference in large-scale settings, including survey sampling, cluster-randomized trials, and annotation-based corpus quality assessment (Martinelli et al., 10 Jan 2026, Xiong et al., 2020, Barcaroli et al., 2022, Liu, 2023, Nugent et al., 2022).

1. Theoretical Foundation and Key Notation

STWCS operates within the classical survey sampling framework but augments it with dual-layer randomization and stratification:

• Population Structure: The universe consists of $H$ strata. Each stratum $h$ contains $N_h$ clusters, with each cluster $i$ containing $M_{hi}$ units.
• Measures of Size: $M_{hi}$ represents the size (often the number of target units) of cluster $i$ in stratum $h$; $M_h = \sum_{i=1}^{N_h} M_{hi}$ and total $N = \sum_{h=1}^H M_h$.
• Two-Stage Sampling: In stage I, $s_h$ clusters are sampled within each stratum—commonly by probability proportional to size without replacement (PPSWOR). In stage II, $s_{hi}$ units are sampled within sampled cluster $(h,i)$, often by SRS or systematic sampling.
• Sampling Probabilities: Stage I inclusion probability for cluster $i$ is $\pi_{hi} = (M_{hi} / M_h) s_h$. At stage II, conditional unit-level selection probability is $\rho_{hij} = s_{hi} / M_{hi}$.
• Design Weights: Each sampled unit $(h,i,j)$ receives design weight $w_{hij} = 1/(\pi_{hi}\rho_{hij})$. These weights propagate through all inference steps, ensuring unbiasedness of the resulting estimators (Xiong et al., 2020, Barcaroli et al., 2022, Nugent et al., 2022).

2. Sampling Algorithms and Estimation Procedures

STWCS implementation requires precise orchestration of stratum and cluster selection, randomization, and estimation:

• Strata Assignment: Auxiliary information (e.g., entity types, geographical region) is used to partition the population. In multivariate or multi-domain contexts, each domain or variable can impose separate precision constraints (Martinelli et al., 10 Jan 2026, Barcaroli et al., 2022).
• Cluster Selection (PSU stage): PPSWOR is standard; the R2BEAT package chooses clusters by balancing design effect, stratum-level size, and domain-level precision targets via extended Neyman-Tschprow-Bethel allocation (Barcaroli et al., 2022).
• Subunit Selection (SSU stage): SRS or proportional allocation of SSUs within selected clusters (e.g., $m_{hi} = m_h \cdot M_{hi} / \sum_j M_{hj}$).
• Estimator Construction: For means, the Horvitz–Thompson (HT) estimator is universally applicable:

$\hat\mu = \sum_{h=1}^H (M_h/N) \hat\mu_h,$

with $\hat\mu_h$ built from the weighted sums in sampled clusters/units (Xiong et al., 2020, Martinelli et al., 10 Jan 2026).

• Variance Estimation: Variance is estimated via the Sen–Yates–Grundy (SYG) estimator or with sample variances across strata and clusters:

$$\mathrm{Var}(\hat\mu) = \sum_{h=1}^H (M_h/N)^2 \frac{S_h^2}{n_h} [1 + \rho_h(b_h-1)],$$

where $S_h^2$ is stratum variance, $\rho_h$ is intra-class correlation, and $b_h$ average SSUs per cluster (Barcaroli et al., 2022).

Procedural workflows have been demonstrated for application to NEL accuracy estimation—clusters based on surface form, strata on entity-type labels, with iterative draws until reaching a target margin of error, and annotation cost models that incorporate per-unit and context-switching penalties (Martinelli et al., 10 Jan 2026).

3. Properties of Estimators: Unbiasedness, Efficiency, and Allocation

STWCS yields estimators with established theoretical properties:

• Unbiasedness: HT estimators using true sampling weights yield unbiased estimators for the population mean and for causal estimands (e.g., PATE) under the specified design (Xiong et al., 2020, Liu, 2023).
• Location Invariance: Estimators remain invariant under additive shifts in outcomes, since weighting does not depend on the outcome's level (Xiong et al., 2020).
• Minimum Variance Allocation: Classical Neyman allocation minimizes variance for a given cost for a single variable; in practice, Bethel's convex multivariate allocation is deployed when multiple variables and domains are present, with optimization over sample allocation subject to constraints on variability (CV, standard error, or margin of error) and survey cost (Barcaroli et al., 2022).
• Cost-Aware Sampling: When per-unit cost varies (e.g., annotation time includes context switching), optimal allocation favors strata or clusters with high variance and low per-unit cost, adopting $n_h \propto W_h s_h/\sqrt{c_h}$ for draw allocation (Martinelli et al., 10 Jan 2026).
• Statistical Guarantees: Confidence intervals derive from estimated variances using sample means, with the margin of error set as $z_{\alpha/2} \sqrt{\sum_{h=1}^H W_h^2(s_h^2/n_h)}$ for (approximately) Gaussian distributions (Martinelli et al., 10 Jan 2026, Barcaroli et al., 2022).

4. Applications: Survey Sampling, Randomized Trials, and Information Extraction

STWCS underpins a range of large-scale statistical tasks:

• Survey Sampling: National and regional surveys deploy STWCS for precise, efficient sample allocation across multiple strata and PSUs, including with cost and domain-specific accuracy constraints (Barcaroli et al., 2022).
• Experimental Design in CRTs: In cluster-randomized trials and matched-pair or finely block-stratified experiments, STWCS is used for efficient estimation of primary and spillover effects under partial interference assumptions, enabling unbiased average-of-averages estimators and optimally powered trials (Liu, 2023).
• Causal Inference with Complex Sampling: TMLE and related targeted estimation methods adapt to the STWCS framework by integrating design weights at both levels, constructing clever covariates to ensure double robustness and bias reduction in the presence of sampling-induced missingness or complex treatment allocation (Nugent et al., 2022).
• Named Entity Linking and Annotation Efficiency: In large-scale NEL corpora, STWCS achieves substantial annotation cost reductions by clustering by context and stratifying by semantic entity type, thereby minimizing both statistical and logistical inefficiencies (e.g., reducing annotation time by 29% relative to SRS at identical sample size and precision) (Martinelli et al., 10 Jan 2026).

5. Algorithmic Implementation and Practical Recommendations

Tools such as the R2BEAT package operationalize the theoretical principles of STWCS and facilitate its use in production environments (Barcaroli et al., 2022):

• Input Preparation: Construction of stratum-level, PSU-level, and SSU-level data structures, setting of domain-specific CV/error constraints.
• Bethel Allocation: Iteratively solves for stratum sample sizes while accounting for design effects, estimator effects, and cost parameters.
• PSU/SSU Selection: Supports PPSWOR, Sampford, SRS, and systematic methods.
• Diagnostics and Sensitivity: Built-in reporting of achieved vs. planned precision, item sensitivity to CV targets, design effect tracing.
• Budgetary Controls: Explicit cost models can be incorporated via per-unit costs at each sampling stage to optimize for budget constraints.

Empirical applications confirm that STWCS yields both efficiency gains and robust statistical guarantees when compared against alternative designs—SRS, simple random assignment, or single-stage methods. For complex survey or annotation tasks, use of fine-grained stratification, cost modeling, and regression adjustment for strong predictors further reduces variance without risk of inflation (Liu, 2023, Martinelli et al., 10 Jan 2026, Barcaroli et al., 2022).

6. Extensions: Optimal Design, Covariate Adjustment, and Multi-Domain Inference

Advanced STWCS designs exploit modern theory and computation for further gains:

• Optimal Stratification: Asymptotic efficiency is achieved by matching clusters or units using predicted outcomes or index functions optimized for the estimand of interest, leading to minimized MSE in primary or spillover effect estimation (Liu, 2023).
• Covariate Adjustment: Incorporating post-assignment regression adjustment for additional cluster-level covariates cannot increase asymptotic variance; OLS adjustment on tuple-level means is both simple and efficient (Liu, 2023).
• Multi-Variable, Multi-Domain Allocation: The full Tschprow-Neyman-Bethel allocation supports surveys and experiments targeting simultaneous control over precision in multiple domains or for multiple response variables. Practical implementation iteratively tunes allocations, design effects, and estimation methods (Barcaroli et al., 2022).
• Robustness to Missingness and Non-IID Structures: STWCS is compatible with modern semiparametric estimators (e.g., TMLE), which can recover efficiency and robustness even under complex missing data and sub-sampling structures, provided sampling and missingness probabilities are known or estimable (Nugent et al., 2022).

7. Empirical Performance and Current Usage

Empirical evidence confirms the efficacy and efficiency of STWCS across domains:

Application Area	Observed Sample Size (%)	Time/Cost Reduction	Reference
Named Entity Linking Quality	24.6%	29% annotation time	(Martinelli et al., 10 Jan 2026)
Cluster-Randomized Experiments	n/a	Variance reduction	(Xiong et al., 2020, Liu, 2023)
National/Official Statistic Surveys	Varies	Controlled by R2BEAT	(Barcaroli et al., 2022)

STWCS is robust, extendable, and has been adopted across both academic and applied settings for large-scale, cost- and precision-constrained inference.

---

References

• (Martinelli et al., 10 Jan 2026) "Efficient and Reliable Estimation of Named Entity Linking Quality: A Case Study on GutBrainIE"
• (Xiong et al., 2020) "The Benefits of Probability-Proportional-to-Size Sampling in Cluster-Randomized Experiments"
• (Barcaroli et al., 2022) "Two-stage Sampling Design and Sample Selection with the R package R2BEAT"
• (Nugent et al., 2022) "Blurring cluster randomized trials and observational studies using Two-Stage TMLE ..."
• (Liu, 2023) "Inference for Two-stage Experiments under Covariate-Adaptive Randomization"