Comparing Sampling and Estimation Strategies in Establishment Populations

  • Kimberly A Henry
  • Richard Valliant University of Michigan
Keywords: best linear unbiased predictor, deep stratification, general regression estimator, measure of heteroscedasticity, optimal sample, robust variance estimation, weighted balance


Population structure is a key determinant of the efficiency of sampling plans and estimators. Variables in many establishment populations have structures that can be described by simple linear models with a single auxiliary variable and a variance related to some power of that auxiliary. If a working model can be devised that is a good approximation to the population structure, then very efficient sample designs and estimators are possible. This study compares alternative strategies of (i) selecting a pilot study to estimate the variance power and using that estimate to select a main sample and (ii) selecting a only main sample based on an educated guess about the variance power. We also examine a number of sampling plans, including probability proportional to size, deep stratification based on a measure of size, and weighted balanced sampling. Population totals are estimated by best linear unbiased predictors, general regression estimators, and some other choices often used in practice.