Research Reports

An Empirical Study on Using CPS and ACS Survey Data in Bivariate State Poverty Models

Elizabeth T. Huang and William R. Bell

KEY WORDS: Small area estimation, Fay-Herriot model, American Community Survey, Current Population Survey

ABSTRACT

An empirical study was carried out to examine the potential benefits to state age group poverty ratio models of using data from both the CPS (Current Population Survey) and ACS (American Community Survey) via bivariate models. We use a Bayesian approach to assess the potential benefits of using both data sources by comparing prediction error (posterior) variances from the bivariate models with those from the corresponding univariate models applied to the CPS or ACS data separately. We focus mostly on the potential of the bivariate models to reduce posterior variances for the CPS equation, but present results for the ACS equation as well. We examined alternative models for doing this using CPS estimates for income years 2000-2002 and 2004, and ACS data from the demonstration period of survey years 2000-2003, as well as the full production ACS 2005 data. The models include regression variables constructed from administrative records data and Census 2000 state poverty ratios, along with state random effects and sampling error components. Our main conclusions are the following. (1) Use of an unrestricted bivariate model can be expected to yield at most small average improvements in posterior variance for the CPS equation, and these small improvements would come at the expense of occasional large posterior variance increases. (2) A restricted bivariate model (one that assumes that the regression coefficients, except intercept, in both equations are the same) is not rejected by the data, and yields larger average improvements in posterior variance for the CPS equation, but still produces occasional large posterior variance increases. (3) The occasional large posterior variance increases in the CPS equation of the bivariate models occur for states with large regression residuals in the ACS equation. (4) Use of either the unrestricted or restricted bivariate models makes no improvements of substance in posterior variances for the ACS equations compared to the ACS univariate models. There is too much sampling error in the CPS direct estimates for them to convey much useful information for improving estimation in the ACS equation. (5) Conclusions (1)–(4) hold whether we are using ACS demonstration survey or full production data. (This conclusion is a little tentative because we were limited to analyzing only one year of ACS full production data.) However, posterior variances in the ACS equations (from univariate or bivariate models) are substantially lower when using ACS full production data compared to ACS demonstration survey data, due to the larger sample size and lower sampling error of the full production data.

CITATION:

Source: U.S. Census Bureau, Statistical Research Division

Created: November 16, 2007
Last revised: November 16, 2007