Small Area Estimation for the Tobacco Use Supplement to the Current Population Survey

Introduction: The Tobacco Use Supplement to the Current Population Survey (TUS-CPS), conducted by the Census Bureau and sponsored by the National Cancer Institute (NCI), is a key source of national and state-level data on smoking and other tobacco use in the U.S. household population. However, policy makers and cancer researchers often need county-level data to evaluate tobacco control programs, and the TUS-CPS does not have enough sample at the county level to support estimates with adequate precision. In such case, estimates derived through small-area estimation (SAE) techniques may be preferred. Through a collaboration between the Census Bureau and NCI, we present research on estimating several different smoking-related variables (e.g., rate of people currently smoking, rate of people formerly smoking, rate of workplace smoking bans) for all U.S. counties.

Methods: We propose model-based county-level estimates using the following methodology. We used stepwise linear regression to develop an initial model for the estimated variables, using county-level auxiliary data on measures including education and income, and state-level data on smoking bans and tobacco control funding. We applied an arcsine-transformation to the variables to stabilize the variance, and implemented a state-level design effect to account for clustering. Our regression estimates are then input to a Fay-Herriot small-area model, which we extend to a Bayesian framework through a Markov Chain Monte Carlo (MCMC) simulation.

Results:  Our small-area models generated a new set of estimates with improved precision over the survey-based estimates. Three measures were computed to assess the goodness of fit, and those measures show that the Fay-Herriot model fits the data well for all the different outcomes. We also demonstrate the accuracy of our model through data exhibits.

Conclusions:  These techniques represent an effective means of generating estimates where there is small (or zero) county sample. Small-area modeling shows improvement over a design-based methodology and provides a better guideline for directing prevention funding. At present, we are extending our model to more recent years and will release the current data through NCI.