Hierarchical Models for State-level AK Estimators in the Current Population Survey
Yuan Li
The George Washington University
Michael D. Larsen
George Washington University
The Current Population Survey is a multistage household probability sample produces monthly labor force estimates in the U.S. Adults in a household are interviewed for four months in a row, left out for eight months, and then included for four more months. This 4-8-4 rotation design produces overlap in the sample. Several weighting steps are used to adjust the ultimate sample to be representative of the population. In order to produce efficient estimates of labor force levels, an estimator, called the AK composite estimator, combines current estimates from 8 rotation panels and the previous month's estimate is applied. Finding the optimal values for the parameters, A and K, of AK estimator can be very helpful for estimating the employment and unemployment counts. Our method is to build a hierarchical model for each state. We contrast a univariate model with a bivariate model that includes correlation between A and K. The Gibbs sampler with multiple independent sequences is used for computations. Under the model the 51 state-level values of A and K experience shrinkage toward the overall mean values. Final unemployment estimates will use the modeled A and K values.