The results obtained from surveys are used to estimate characteristics of populations. Calibration estimators use auxiliary information to improve the efficiency of these survey estimates. Calibration is used in the U.S. in the Census Bureau's Current Population Survey (CPS), the Bureau of Labor Statistics' Consumer Expenditure Survey, and a variety of other Federal surveys as well as surveys administered by other nations. Some of the key uses of calibration are to (i) reduce variances, (ii) correct for nonresponse, and (iii) correct coverage deficiencies of a sample frame.
A primary assumption with calibration is that the sample estimates are calibrated to true population values known without error. Often, however, these population values are estimates obtained from other surveys which could possess sampling variance. These estimated controls add variability to survey estimates and may produce bias if the source of controls is inaccurate. The use of estimated controls in surveys is common but is virtually never accounted for in survey variance estimation. This research will investigate methods that will account for this extra imprecision in variance estimates and also study the effects of bias in the population vlues and in the analytic survey itself. The specific goals of this research will be to (1) analyze the effect of benchmarks and analytic survey estimates that are biased due to coverage errors with extensions done for general regression (GREG) estimators and other calibration estimators; (2) develop variance estimators for GREG and other calibration estimators that account for estimated benchmarks and undercoverage by sample frames; (3) develop some practical rules for determining whether estimated benchmarks are too variable to be worth using; and (4) illustrate theoretical findings via simulation studies based on real survey data sets.
