Evaluation of Confidence Interval Methodology for the National Compensation Survey

Glenn Springer, Martha Walker, Steven P. Paben, and Alan H. Dorfman

Abstract

The National Compensation Survey (NCS) is a Bureau of Labor Statistics (BLS) program that provides data on occupational wages. An investigation by Casady, Dorfman, and Wang 1996 (CDW) suggested that the standard 95% confidence intervals (C.I.) for domain means or totals, when based on the standard normal distribution and standard methods of variance estimation, tend to yield less than the actual 95% coverage. The estimation of means or totals within an occupation is a case of domain estimation presented in Cochran (1977, pg. 34) since the observations in the sample falling within a specified occupation are not known prior to sampling. Even though the sample size is large enough to support standard normal estimations, the individual occupations can be represented by a small number of establishments. CDW presented new nonstandard methods that offer an improvement, giving intervals with more accurate coverage, typically at or close to the nominal 95% coverage. These intervals tend to be longer than the standard intervals and depend on the use of a t-statistic having degrees of freedom dependent on the available domain data. The increase in length will vary with domain, and will depend on the particular method for C.I. construction that is used. In Harpenau, Coleman, Lincoln (HCL 1995) this was shown to be true for data from the Occupational Compensation Survey Program (OCSP). We modified this method to make it suitable to the multi-stage design of the NCS. Using NCS data, an artificial sampling frame was created and simulated samples were selected. The standard normal confidence intervals were compared to confidence intervals using the t-distribution with weighted degrees of freedom for estimates of means and quantiles. Coverage properties for confidence intervals using the non-standard approach were found to be superior to the standard normal approach.