This estimator also belongs to a variety called calibration estimators, as the second term here “corrects” (or “calibrates”) the Horvitz-Thompson estimator for Y using known population totals for X.

Note that the estimator for Bd is based on sample data. When the sample size is small, this estimate may be unstable. To address this, one can pool the data over domains to produce a single image The resulting modified direct estimator, known as the survey regression estimator, is expressed as follows:

image

Gershunskaya illustrated how this estimator might be applied to NCSES R&D data. Let Xim be the known population payroll in industry-type i and state image be the Horvitz-Thompson estimate of payroll in industry-type i and state m, image be the Horvitz-Thompson estimate of payroll, national total for industry-type i, and image be the Horvitz-Thompson estimate of R&D funds, national total for industry-type i. Then, one can compute

image

using national data for industry i in the survey regression estimator to estimate

image

Gershunskaya pointed out that, although image in the survey regression estimator is based on a larger sample, the effective sample size still equals the domain sample size. To see why that is so, one can rewrite the survey regression estimator as

image

which shows that the survey regression estimator is a sum of the fitted values from a regression model based on predictors from the domain of interest, and it has a bias correction from weighting the residuals again from a regression using data only from that domain. Therefore, the eff-



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement