Bayesian Dependent Functional Mixture Estimation for Area and Time-Indexed Data: An Application for the Prediction of Monthly County Employment

The U.S. Bureau of Labor Statistics (BLS) publishes employment totals for all counties on a monthly basis. BLS use the Quarterly Census Employment and Wages, where responses are received 6–7 month lagged basis aggregated to county, apply time series forecast model each county project forward current month, which ignores dependence among counties. Our approach treats these by-county as collection area indexed noisy functions that we co-model. includes predictor, trend seasonality terms by county. This application is first in Federal Statistical System address joint modeling expressing heterogenous patterns between them. We demonstrate Fourier outperforms locally-adaptive, intrinsic conditional autoregressive construction our degree expressed varies. County-indexed parameters 3 drawn from dependent Dirichlet process (DDP) prior allow borrowing information. show both spatial industry concentration predictors into probabilities co-clustering produces better prediction accuracy. DDP accounts possibility nearby may express distinct underlying economic structures. A feature framework it computes efficiently support production cycle. compare performances alternative formulations monthly, data 2002–2016.