The Current Employment Statistics survey is conducted by the U.S. Bureau of Labor Statistics and State Employment Security Agencies to produce monthly estimates of employment, hours and earnings by industry for the U.S., States, and areas. This establishment survey is currently undergoing a redesign. Sample design research indicates that a simple but well executed probability design could considerably reduce mean squared error compared to samples selected using the current realized sampling rates. What this research has not considered is that the realized sampling rates are not deviations from a more optimum design as much as they are the result of low participation rates when units are first solicited for the survey, particularly among units in the largest employment size classes. This research compares bias of estimators resulting from nonresponse adjustment using information available on these nonparticipants from administrative records from the State Unemployment Insurance (UI) programs for earlier months, with more traditional survey methods. In other words, unlike most other studies, in this research we do not assume nonrespondents to be missing at random. For this study, we can evaluate the effectiveness of various imputation procedures since responses for every month for every unit are available from the administrative records consisting of UI accounts.
The data used in this study are discussed in Section 2. This will include a brief discussion of the CES survey and our test population. Section 3 presents the methods used in the various imputation routines. Section 4 describes the evaluation criteria used to analyze the results. Section 5 contains our results and comparisons of the imputation methods. Conclusions from this paper are contained in Section 6.