Abstract
Design-based estimators for two-stage simple random sampling with regression can have a lack of precision (efficiency) when the primary sampling units (PSUs) vary in size, and PSU totals are approximately proportional to the size of a PSU. Precision and efficiency may deteriorate further for domain-specific estimators when PSUs contain elements from different domains. Design model-unbiased ratio-to-size estimators have been proposed as more efficient. This study introduces a variance estimator for a design model-unbiased ratio estimator. The estimator of variance is derived from a single-stage estimator of a variance of a ratio under two assumptions: the target variable (y) is equal to the sum of a model prediction and two error terms capturing residual errors and model estimation errors; and the existence of an unbiased estimator of model parameters. Extensive simulations confirmed the negative effects of unequal PSU sizes on the precision of the model-assisted estimators of variance and a superior performance of the proposed estimator of variance. These results were confirmed in the analysis of a regional two-stage survey of forest biomass. The proposed variance estimator was generally more efficient than an existing alternative and more stable across a large suite of design settings.