Home Statistical Dictionary Balanced Repeated Replication

Balanced Repeated Replication

A replicate-weight variance estimation method designed for stratified designs with two primary sampling units per stratum, using an efficient, balanced (Hadamard-matrix-based) pattern of half-sample replicates to estimate an estimator's sampling variance.

Sampling Theory & Survey MethodsAdvancedFree

In Plain English

Balanced repeated replication (BRR) is a variance-estimation technique built specifically for a very common survey design structure: strata that each contain exactly two primary sampling units. The naive approach would be to try every possible way of picking one PSU from each pair to form a 'half-sample' and see how the estimate varies, but with many strata that's an astronomically large number of combinations. BRR's clever trick is to use a specially structured, balanced pattern (based on mathematical objects called Hadamard matrices) that captures nearly all the same variance information using only a modest, manageable number of half-sample replicates, dramatically cutting the computational cost while still producing a statistically sound variance estimate.

Definition

Balanced repeated replication (BRR) estimates sampling variance for designs stratified into $H$ strata, each containing exactly two primary sampling units (a common design for large-scale surveys), by forming half-sample replicates, each replicate includes one PSU from every stratum, systematically excluding or down-weighting the other, and reweighting accordingly. Rather than forming all $2^{H}$ possible half-samples, which becomes computationally infeasible for even a moderate number of strata, BRR uses a balanced set of replicates constructed from the rows of a Hadamard matrix (an orthogonal $\pm 1$ matrix), which requires only a number of replicates equal to the smallest multiple of 4 at least as large as $H$ , while still providing an unbiased variance estimate for linear statistics; the BRR variance estimate is $Var_{BRR} (\hat{θ}) = \frac{1}{R} \sum_{r = 1}^{R} (\hat{θ}_{(r)} - \hat{θ})^{2}$ , averaging the squared deviation of each half-sample replicate estimate from the full-sample estimate. Fay's method, a common modification, uses a partial (rather than complete) down-weighting of the excluded PSU in each replicate to improve stability for nonlinear statistics.

Formula

Var_{BRR} (\hat{θ}) = \frac{1}{R} r = 1 \sum R (\hat{θ}_{(r)} - \hat{θ})^{2}

Notation

\hat{θ}_{(r)}

—Estimate computed from the r-th balanced half-sample replicate

R

—Number of replicates, the smallest multiple of 4 at least H

H

—Number of strata, each with exactly two PSUs

Properties

BRR's design requirement of exactly two PSUs per stratum is quite specific, this is a common and deliberate design choice for many large government and health surveys precisely because it enables BRR variance estimation, but BRR is not directly applicable to designs with a different number of PSUs per stratum without modification.
The Hadamard-matrix construction is what makes BRR computationally efficient relative to the naive alternative of forming every possible half-sample, for a design with, say, 100 strata, BRR needs only about 100-104 replicates (rather than 2^100), while still providing a mathematically balanced and unbiased variance estimate for linear statistics.
BRR is one of three major replicate-weight variance estimation families alongside the jackknife and the bootstrap, the choice among them in practice is often dictated by which method the survey organization that produced the dataset chose to publish replicate weights for, rather than by an analyst's own preference, since replicate weights are typically constructed once by the data producer.

At a Glance

Requiresa design stratified into strata with exactly two PSUs each

Mechanismbalanced half-sample replicates built from a Hadamard matrix pattern

Efficiencyachieves near-full variance information with far fewer replicates than 2^H

Common variantFay's method, using partial rather than complete down-weighting

Last updated September 8, 2026← Back to the Dictionary