📊 statistics

More with Less - Bethel Allocation and Precision-Preserving Sample Size Reduction via Hierarchical Bayes Modelling

This paper proposes a practical two-stage strategy that combines multivariate constrained Bethel allocation with Hierarchical Bayes small area modelling to determine the minimum sample size required to simultaneously meet precision targets for multiple variables across all geographic domains, thereby enabling significant cost reductions for national statistical offices.

Original authors: Siu-Ming Tam

Published 2026-03-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Siu-Ming Tam

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the head chef of a massive catering company (a National Statistical Office). You have a strict budget, but your boss (the government and the public) is demanding a very specific menu: they want to know exactly how many people are employed, unemployed, and how many hours they work, broken down not just for the whole country, but for every single neighborhood and region.

The problem? You have a lot of ingredients (data) to gather, but your budget is shrinking.

Traditionally, chefs solve this by trying to guess the perfect amount of each ingredient needed for every single dish. If they need a tiny pinch of salt for the soup but a whole cup for the stew, they just buy a whole cup for everything to be safe. This wastes money and leaves you with too much salt.

This paper proposes a smarter, two-step recipe to cook the same delicious meal using 80% less ingredients without ruining the taste.

The Problem: The "Max of Everything" Mistake

Currently, many statistical offices use a clumsy method. They calculate the perfect sample size for "Employment," then for "Unemployment," then for "Hours Worked."

To get a good estimate for "Unemployment" (which is rare and hard to find), they might need to interview 68,000 people.
To get "Hours Worked," they might only need 200 people.

The old way says: "Let's just interview 68,000 people for everything to be safe."
The Flaw: This is like buying 68,000 cups of salt just because you need one cup for the stew. You are wasting a fortune. Worse, even with 68,000 people, you might still fail to get a precise answer for the small neighborhoods because the math wasn't set up correctly for all the dishes at once.

The Solution: A Two-Stage Strategy

The author, Siu-Ming Tam, suggests a clever two-step process to shrink the sample size while keeping the results accurate.

Stage 1: The "Perfect Puzzle" Solver (Bethel Allocation)

First, we stop guessing. We use a sophisticated mathematical tool called Bethel Allocation.

Think of this as a master puzzle solver. Instead of solving for "Soup" then "Stew" separately, it looks at the whole table at once. It asks: "What is the absolute smallest number of people we need to interview to get a precise answer for Employment, Unemployment, AND Hours Worked, for the whole country AND every single neighborhood?"

It finds the "Goldilocks" number. In the paper's example, this "perfect" number is 91,308 people. This is already better than the old wasteful method, but it's still expensive.

Stage 2: The "Smart Borrower" (Hierarchical Bayes)

This is the real magic trick. The paper asks: "Can we interview even fewer people if we use a little bit of magic?"

The magic is called Hierarchical Bayes (HB) Modelling.
Imagine you are trying to guess the average height of people in 10 different small towns.

The Old Way (Direct Estimate): You go to a tiny town with only 5 people. You measure them. If one person is a basketball player, you think the whole town is huge. If they are all children, you think the town is tiny. Your guess is shaky because you have so little data.
The HB Way (Borrowing Strength): You say, "I know these 5 people, but I also know the average height of the whole country and the neighboring towns. I will use that big picture to 'borrow' strength."

If the tiny town's data looks weird, the model gently pulls the estimate toward the national average, using the data from the other 99 towns to fill in the gaps. It's like using a telescope to see a distant star clearly, even if your eyes are weak.

Because this "borrowing" makes the estimates so much more precise, you don't need to interview as many people to get the same level of accuracy.

The Result: Cutting the Cost by 80%

The paper tested this on a fake population of 1 million people (a simulation).

The Old Wasteful Way: Needed 68,697 people (and still failed some accuracy checks).
The "Perfect Puzzle" Way (Stage 1): Needed 91,308 people (to be perfectly safe).
The New Smart Way (Stage 1 + Stage 2): Needed only 18,262 people.

That is an 80% reduction in the number of people you need to interview!

Is it Safe? (The Taste Test)

You might worry: "If we interview fewer people, won't the results be wrong?"

The author ran a massive test (1,000 simulations) to check.

Did the estimates hit the target? Yes. The "taste" was perfect.
Did the confidence intervals work? Yes. The "guarantees" held up 95% of the time, just like a standard survey.
What's the catch? The only difference is how we explain the uncertainty. Instead of saying "We are 95% confident based on how we picked the people" (Design-based), we say "Given the data and our smart model, there is a 95% chance the answer is here" (Model-based).

The Big Takeaway

Statistical agencies are often told they must cut costs. This paper says: "Don't just cut the budget; cut the waste."

By combining a smart mathematical allocation (Bethel) with a model that borrows strength from neighbors (Hierarchical Bayes), governments can get the same high-quality, detailed data for their citizens while interviewing 80% fewer people. It's like getting a gourmet meal for the price of a sandwich, without sacrificing the flavor.

1. Problem Statement

National Statistical Offices (NSOs) face a critical dilemma: the demand for granular, domain-level statistics is rising, while survey budgets are fixed or shrinking. Traditional multi-purpose surveys (surveying multiple variables across multiple geographic domains) suffer from inefficient sample allocation strategies.

The Flawed Status Quo: The common practice is to calculate a Neyman allocation separately for each target variable and then take the element-wise maximum of the resulting stratum sample sizes.
- Deficiency 1 (Budget Inflation): This approach massively over-samples for variables whose optimal allocation is dominated by another variable, wasting budget.
- Deficiency 2 (Precision Failure): Even with the inflated budget, this method often fails to guarantee precision for specific domains. The stratum weights optimized for national precision of one variable do not necessarily protect domain-level precision for another.
The Gap: While the Bethel allocation algorithm correctly solves the multivariate constrained optimization problem (finding the minimum sample to meet all precision targets), the resulting sample size can still be prohibitively large because it relies solely on direct survey estimates.

2. Methodology

The paper proposes a novel two-stage strategy to minimize sample size while rigorously maintaining pre-defined precision targets (Coefficient of Variation, CV) for all variables and domains.

Stage 1: Multivariate Constrained Optimisation (Bethel Allocation)

Instead of the "Neyman-max" approach, the authors use the Bethel algorithm (Bethel, 1989) to find the globally minimum sample size ( $n^*$ ) that simultaneously satisfies CV constraints for all $K$ variables across all $D$ domains.

Objective: Minimize total cost $\sum c_h n_h$ .
Constraints: Ensure $CV(\hat{Y}_{d,k}) \leq g_{d,k}$ for all domains $d$ and variables $k$ .
Nature: This is a design-based approach. Precision guarantees rely solely on the randomization distribution of the sampling design, requiring no statistical models for the target variables.

Stage 2: Hierarchical Bayes (HB) Sample Reduction

The core innovation is using HB small area estimation to "borrow strength" across domains, allowing for a further reduction in sample size ( $n_{HB} < n^*$ ) while maintaining precision.

Models Used:
- Binary Variables: Logit-normal binomial HB model (Rao & Molina, 2015).
- Continuous Variables: Gaussian Fay-Herriot area-level model.
Mechanism: By shrinking stratum estimates toward covariate-predicted means, HB models reduce variance. This "precision surplus" is exchanged for a smaller sample size.
The Reduction Algorithm:
1. Start with the Bethel master sample ( $n^*$ ).
2. Draw nested sub-samples at decreasing fractions ( $\alpha$ ).
3. Fit HB models and evaluate against four gates:
  - CV Gate: All domain/national CVs meet targets.
  - Convergence Gate: MCMC convergence ( $\hat{R}_{max} \leq 1.05$ ).
  - National Accuracy Gate: Relative bias within tolerance.
  - Domain Accuracy Gate: Mean Absolute Relative Error (MARE) and Max ARE within tolerance.
4. Select the maximum reduction fraction ( $\alpha^*$ ) where all gates pass for all variables simultaneously (Minimax rule).

Prior Calibration

A critical component is calibrating the HB prior (specifically the variance of random effects, $\sigma^2_v$ ). The authors propose a grid search using a surrogate for "truth" (direct estimates from the previous survey cycle) to select hyperparameters that maximize credible interval coverage while minimizing domain error.

3. Key Contributions

Reframing the Optimization Problem: Shifts the question from "How to allocate a fixed sample?" to "How small can the sample be made while meeting all precision targets?"
Correcting NSO Practices: Demonstrates mathematically and empirically that the "Neyman-max" approach is doubly flawed (wasteful and insufficient) and that Bethel allocation is the necessary baseline.
Integration of Design and Model-Based Inference: Successfully bridges the gap between design-based sample determination (Bethel) and model-based estimation (HB) to achieve cost reduction.
Operational Framework: Provides a concrete, four-gate algorithm for determining the optimal sample reduction fraction in a multi-variable, multi-domain context.

4. Results

The methodology was validated using a Monte Carlo study ( $B=1,000$ replications) on a synthetic labour-force population of 1 million individuals (100 strata, 10 domains, 3 variables: Employment, Unemployment, Hours Worked).

Sample Size Reduction:
- Neyman-max (Status Quo): 68,697 (Driven by the rare "Unemployment" variable; fails domain precision for unemployment).
- Bethel Allocation (Stage 1): 91,308 (Required to meet all constraints simultaneously using direct estimates).
- HB-Combined (Stage 2): 18,262.
- Outcome: An 80% reduction in sample size compared to the Bethel benchmark, and significantly smaller than the traditional Neyman-max approach.
Precision and Accuracy:
- CV Targets: At $n_{HB} = 18,262$ , all national (≤3%) and domain (≤8%) CV targets were met for all three variables.
- Coverage: Monte Carlo credible interval coverage was close to or exceeded the nominal 95% level (93.0% for Employment, 97.7% for Unemployment, 100.0% for Hours Worked).
- Bias: National relative bias was negligible (<1.54%). Domain MARE was low, though Unemployment showed higher error in small domains (4.10% MARE), reflecting the inherent difficulty of estimating rare events.
Robustness: The CV gate pass rate exceeded 95% across all replications, confirming the strategy works reliably under repeated sampling, not just for a single instance.

5. Significance and Implications

Cost Efficiency: Statistical agencies can achieve massive operational cost savings (up to 80% in the simulation) without compromising the quality of published statistics.
Paradigm Shift: The paper legitimizes the shift from purely design-based inference to model-based inference for sample size determination. It argues that credible intervals (model-based) offer more intuitive interpretations for decision-makers than confidence intervals (design-based), provided the model is well-specified.
Practical Applicability: The proposed method relies on standard tools (R packages R2BEAT and mcmcsae) and uses data proxies (previous survey cycles) for calibration, making it feasible for real-world NSO implementation where population truth is unknown.
Future Directions: The authors note the need for extensions to multi-stage cluster designs, rotational panel surveys, and robustness checks against model misspecification.

In conclusion, the paper demonstrates that by combining Bethel allocation (for optimal design) with Hierarchical Bayes modelling (for precision enhancement), statistical agencies can drastically reduce survey costs while maintaining rigorous precision standards across all required outputs.