There are two changes with the Current Population Survey (CPS) that affect both the state and county poverty estimates. (1) All CPS data used in the 2000 estimates had weights controlled to demographic population estimates that were "2000-based," i.e., that used Census 2000 results as the base population from which updated population estimates were constructed. Although the CPS weights were from Census 2000, the initial sample design was developed from the 1990 census. (2) The 2001 and 2002 CPS Annual Social and Economic Supplements (ASEC) significantly expanded the sample of the previous CPS March supplement by also interviewing part of the sample in February and April, yielding direct CPS estimates with more precision (lower variance) than would have been obtained from the traditional March supplement sample. We use the estimates based on the expanded sample for income years 2000 and 2001 (from the 2001 and 2002 CPS ASECs, respectively) whenever ASEC estimates for these years are used in the estimation procedures. Further information on the sample expansion can be found on the CPS website.
State Model Changes
The most important change to the state models for 2000 was the use of estimates from the expanded ASEC sample. The expanded sample yielded direct estimates with lower variances than would otherwise have been obtained, which resulted in model-based estimates for 2000 with lower variances as well.
The use of the ASEC data from the expanded sample also necessitated a re-examination of the sampling error models used to smooth the direct estimates of the ASEC sampling error variances. We had available ASEC estimates and corresponding direct variance estimates for income years (IYs) 2000 and 2001. In previous years, sampling error models for the ASEC variance estimates included state random effects that allowed for systematic state-to-state variations from the fitted generalized variance functions (GVFs). These models, however, were fitted to at least three successive years of ASEC direct estimates of sampling variances and covariances. With the switch to using the expanded sample we felt we needed to refit the sampling error models using only the variances and covariances estimated from the expanded sample, which limited us to two years of this data. We were concerned about trying to estimate the state random effects using only two years of data, and so dropped the state random effects from the sampling error model and used the fitted GVFs without modifications for estimated state random effects. (For further discussion of the sampling error models see "Sampling Error Modeling of Poverty and Income Statistics for States," (Otto and Bell 1995) on the Published Papers section of this web site.)
As was the case for last year's estimates for IY 1999, in the IY 2000 models for age group poverty ratios and median income we used the corresponding Census 2000 estimates for IY 1999 as regression predictor variables. While this aspect of the models thus remained unchanged from last year, it was a departure from what we had planned. Our plan was to let "Census 2000 residuals" replace the Census 2000 estimates as regression predictors in this year's models, analogous to what we had done with 1990 census data in most models before the Census 2000 estimates became available. However, statistical comparisons of the models for IY 2000 favored the use of Census 2000 estimates rather than Census 2000 residuals, and so the Census 2000 estimates have been used in this year's models. We plan to continue to make such model comparisons in the future and will switch to use of Census 2000 residuals when this becomes the favored choice, something that becomes more and more likely the further we move beyond the census IY.
In last year's models for IY 1999, we used informative prior distributions with mean zero for the regression coefficients in the poverty ratio models. (See 1999 State-Level Estimation Details.) This was done because of the special situation that applied in the census year, namely, that Census 2000 and the 2000 ASEC were both estimating poverty for the same year (1999). In this situation, the administrative records regression variables were expected to be less relevant in the models, and empirical comparisons confirmed this (for the poverty ratio models, though not for median income.) The prior distributions were used to "shrink" the corresponding regression coefficient estimates towards zero, and to reduce the uncertainty about these regression coefficients in the model. This year the situation is different. For IY 2000 the administrative records regression variables provide updated information relative to the Census 2000 estimates, so such prior distributions were not felt to be appropriate and were therefore not used.
County Model Changes
The use of CPS ASEC direct estimates of poverty and median household income had a minimal effect in county model-based estimates. Although the sample size for national and state estimates increased significantly, the sample increase in most counties was generally small. Consequently, sample size for all but a handful of counties remained quite small. We observed no meaningful global effect on the SAIPE county estimates. We did observe a greater number of counties in sample and a slight increase in the direct estimate's contribution to the final SAIPE estimate.