State Model Changes
There were two important changes to the state models for 2004. First, a switch was made to using prior census data in the form of census residuals rather than census estimates for all but the model for 65 and over poverty ratios. (Census residuals come from the regression of the corresponding Census 2000 estimates on 1999 values of the other predictor variables in the model.) Second, we reintroduced the food stamp participation rate variable into the 0-4, 5-17, and 18-64 poverty ratio models. Following is discussion of these two changes, and of a less important change concerning estimation of sampling error variances.
The change noted to use prior census residuals as predictor variables in the models rather than prior census estimates was made because statistical comparisons of model fit for 2004 favored the models with census residuals, whereas for 2000-2003 the corresponding model fit comparisons had mostly favored models with prior census estimates. For the age 65 and over poverty ratios the model fit comparison for 2004 still favored use of prior census estimates, and so the change to census residuals was not made for 65 and over. Somewhat similar results were observed in the 1990s, although the model fit comparisons favored use of (1990) census residuals for ages 0-4, 5-17, and 18-64 earlier in the decade. (SAIPE did not start producing estimates until income year 1993, though model fits done for evaluation purposes with data for 1990-1992 show that the models with census residuals would have been favored even earlier.) The models for age 65 and over poverty ratios and for median household income were switched to using census residuals in income year 1998.
In general, it appears that "sufficiently close" to the census year models using prior census estimates tend to fit better while "sufficiently far" from the census year models using census residuals tend to fit better. However, the meanings of "sufficiently close" and "sufficiently far" from the census year have been different in this decade than in the 1990s.
Another change for the 0-4, 5-17, and 18-64 poverty ratio models was the reintroduction of the state food stamp participation rate as a predictor variable. This variable was previously used in these models until 1998, when it was dropped because its regression coefficients had become statistically insignificant. (For discussion, see the documentation of estimation procedure changes for 1998.) This remained the situation until this year, though in recent years this was partly due to the continued use of the models with previous census estimates as predictors. (The other predictors in the models are less important when census estimates, rather than census residuals, are used in the models.) But with the switch to use of census residuals for 2004, statistical model comparisons favored reintroduction of the food stamp participation rate variable in the 0-4, 5-17, and 18-64 poverty ratio models.
The estimation of sampling error variances of the direct CPS ASEC state estimates was complicated by the fact that for 2004 the estimates drew on samples from both the 1990 census based CPS design and the newly introduced census 2000 based design. Direct variance estimates were still produced, but the different nature of the sample for the 2004 estimates raised concerns about whether the design effect parameters in the sampling error models could be different in 2004 from other years. So, despite changing the sampling error models last year to use the same design effect parameters for all years (see documentation of estimation procedure changes for the 2003 estimates), this year we went back to using different design effect parameters for all years (though only those for 2004 get used in the 2004 state poverty ratio and median income models). Even so, the estimated design effect parameters for 2004 from the fitted sampling error models were not dramatically different from the corresponding design effect parameters for 2000-2003.