Distributionally balanced sampling designs

We propose Distributionally Balanced Designs (DBD), a new class of probability sampling designs that target representativeness at the level of the full auxiliary distribution rather than selected moments. In disciplines such as ecology, forestry, and environmental sciences, where field data collection is expensive, maximizing the information extracted from a limited sample is critical. More precisely, DBD can be viewed as minimum discrepancy designs that minimize the expected discrepancy between the sample and population auxiliary distributions. The key idea is to construct samples whose empirical auxiliary distribution closely matches that of the population. We present a first implementation of DBD based on an optimized circular ordering of the population, combined with random selection of a contiguous block of units. The ordering is chosen to minimize the design-expected energy distance, a discrepancy measure that captures differences between distributions beyond low-order moments. This criterion promotes strong spatial spread, and yields low variance for Horvitz-Thompson estimators of totals of functions that vary smoothly with respect to auxiliaries. Simulation results show that approximate DBD achieves better distributional fit than state-of-the-art methods such as the local pivotal and local cube designs. Hence, DBD can improve the reliability of estimates from costly field data, making distributional balancing effective for constructing representative surveys in resource-constrained applications.

💡 Research Summary

The paper introduces a novel class of probability sampling designs called Distributionally Balanced Designs (DBD), which aim to make the empirical distribution of auxiliary variables in the sample as close as possible to the full population distribution. Traditional balanced sampling methods, such as the cube method, focus on matching only the means of auxiliary variables and therefore provide variance reduction primarily for linear relationships. Spatially balanced designs (e.g., GRTS, Local Pivotal Method) improve geographic spread but do not guarantee optimal distributional fit. DBD addresses both shortcomings by minimizing a global discrepancy measure – the energy distance – between the sample distribution and the population distribution.

Energy distance, a member of the Maximum Mean Discrepancy (MMD) family, captures differences in all moments and geometric shape. The authors prove that the expected energy distance directly controls an upper bound on the mean‑square error of the Horvitz–Thompson estimator for any target variable that is a smooth function of the auxiliaries (Proposition 1). Consequently, reducing the expected energy distance simultaneously reduces variance for linear trends, non‑linear relationships, and spatial patterns.

To make the optimization tractable, the population is arranged in a circular order (u). A sample of fixed size (n) is obtained by selecting a contiguous block of length (n) starting from a uniformly random position. The design class therefore consists of all possible circular permutations of the population indices, each yielding exactly (N) possible samples with equal inclusion probabilities (π_i=n/N). The objective is to find the permutation (u^*) that minimizes the expected energy distance (\bar E(u;n)). Because the search space grows factorially, the authors employ simulated annealing with a simple swap move (exchange of two positions) and an O((n)) update formula for (\bar E). The algorithm iteratively cools the temperature, accepts uphill moves with a probability that depends on the increase in (\bar E), and records the best permutation found. An efficient implementation is provided in the R package rsamplr.

Standard variance estimators based on second‑order inclusion probabilities become unstable under DBD because many pairwise inclusion probabilities are near zero (the design spreads units far apart). The authors therefore propose a local‑mean variance estimator. For each sampled unit, the k‑nearest neighbours in the auxiliary space are identified, and the within‑neighbourhood variance is computed. This estimator automatically adapts: if the target variable is smooth in the auxiliary space, local neighbourhoods capture the remaining variation; if there is no relationship, the estimator reduces to the usual population variance. Values of (k) between 2 and 4 work well in practice, while (k=n) recovers the classical independent‑observation variance formula.

Three simulation studies illustrate the method. The first examines convergence of the expected energy distance as a function of annealing iterations and demonstrates robustness across multiple runs. The second compares DBD with existing designs (Local Pivotal, Local Cube, GRTS) on synthetic populations of varying dimensionality and non‑linear auxiliary‑target relationships. DBD consistently achieves lower mean energy distance, better spatial spread, and reduced Horvitz–Thompson variance. The third applies DBD to a real forest‑inventory dataset, showing a 15–30 % reduction in mean‑square error of total estimates relative to competing designs.

The paper also discusses computational scalability. Each swap update requires O((n)) operations, making the algorithm feasible for populations of several thousand units on standard hardware. Because the design class is based on systematic sampling with a circular ordering, it integrates easily with existing field protocols.

In conclusion, Distributionally Balanced Designs provide a unified framework that simultaneously ensures global distributional balance and spatial spread. By targeting the full auxiliary distribution rather than selected moments, DBD delivers variance reduction for a broad class of target variables while remaining computationally practical. Future work may extend the approach to unequal inclusion probabilities, multi‑stage sampling, and alternative kernel‑based discrepancy measures.