Innovative Data Science and Statistical Methods for Small Domain Estimation and Dissemination

Unit-level models, which model survey responses directly, offer a number of advantages over area-level models, which model aggregated estimates. However, accounting for a complex survey design becomes more challenging in the unit-level setting. In particular, little work has been done to extend such models to capture longitudinal designs, where temporal correlation exists at both the response and domain level. We consider a Bayesian hierarchical unit-level, model-based approach that handles Gaussian, binary, and categorical data, incorporates longitudinal dependence and multiscale time series structure, and accounts for informative sampling. To handle computational scalability, we develop an efficient Gibbs sampler with appropriate data augmentation. An empirical simulation study is conducted to compare the proposed approach to models that do not account for unit-level longitudinal correlation. Finally, using public-use microdata, we provide an analysis of the Household Pulse Survey that compares both design-based and model-based estimators and demonstrates superior performance for the proposed approaches. 

Due to smaller sample sizes and/or a lack of statistical reliability, some estimates for small domains (subpopulations) cannot be disseminated at NCHS. As a result, estimates may be suppressed or data may be aggregated across domains or time to produce more reliable estimates, which can mask potential differences in outcomes for some groups. Many approaches are available to improve the reliability of estimates for small domains and subsequently increase the number of estimates that can be disseminated. This presentation will describe some approaches used at NCHS to produce more reliable estimates for small subgroups of interest. These approaches include a new tool for small domain estimation (the enhanced modified Kalman filter) and using statistical learning methods to incorporate data from nonprobability surveys that may include oversamples of specific subpopulations. These model-based estimates can fill important data gaps for subpopulations of interest and improve dissemination. However, there are various challenges and limitations that are important to acknowledge, including (but not limited to): the bias-variance tradeoff; potential correlations between the selection probabilities for a nonprobability sample and the outcome variable(s) of interest; and when there are limited shared covariates across data sources to use in various data integration models.