In the state model that estimates poverty for people age 65 and over we modified the Supplemental Security Income (SSI) variable. For the 1996 and 1997 estimates we used an average of the March, June, September, and December estimates of the number of people age 65 and over receiving SSI payments as reported by the Social Security Administration. For the 1998 model we used a 12-month average. Although we believe the new definition of SSI participants makes more sense for our models, the new and old variables are in fact very closely related, and comparisons of the 4-month and 12-month averages showed almost no difference.
In the county median household income model, the number of bins used to interpolate the direct CPS estimate of median household income was increased. Due to inflation and economic growth, household income over the decade of the 1990's steadily increased to the point that several counties were "top coded" into the top income bin. The inferred estimates for these counties were given an arbitrary high value instead of the "true" direct estimate. Consequently, the estimated regression line was tilted in the direction of the assigned values. The additional income bins eliminated the bias and flattened the regression line such that the median household income estimates for rich counties were lowered and vice versa. However, the overall magnitude of the effect, in comparison with estimates created with too few bins, was very small for almost all counties.
State Model Changes
The state poverty ratio model for people age 65 and over and the model for household median income were changed so that they use the prior (1990) census data in the form of census residuals (residuals from regressing the corresponding 1990 census estimates on 1989 values of the other predictor variables in the model). Previously, both these models included the prior census estimates, rather than the census residuals. These models now use the prior census data in the same manner as the poverty ratio models for the other age groups.
In the median income model, we are now using IRS median income as a variable rather than the prior census median income multiplied by the ratio of current income year to census income year (1989) IRS median income. In some earlier years a similar variable was used but with the prior census median income multiplied by the ratio of current to census year IRS mean adjusted gross income. The evolution of this variable in the median income model was driven somewhat by data availability: originally we did not have tabulations of IRS median income.
Comparisons of model fits for the revised and original 65 and over poverty ratio models showed the revised model fits the CPS data better in recent years (since income year 1995). The same conclusions were drawn from comparing fits of the revised and original median income models.
For the state poverty ratio models for children under 5, children 5-17 (both related and total poor), and persons age 18-64, the food stamp participation variable was dropped from the model. Examination of model fit results for these age groups showed that food stamp participation was insignificant as a predictor variable for income years 1997 and 1998, and also in preliminary results obtained for income year 1999.
This contrasts with its performance as a predictor in earlier years; in model fits from 1989 to 1996 the food stamp participation variable was nearly uniformly significant (in every year modeled for all these age groups). We will continue to evaluate the value of food stamp data for future models.
The recent deterioration in the predictive performance of the food stamp participation variable in our models may be related to the effects of welfare reform legislation which went into effect in 1997. If different states implemented welfare reform in ways that make the relation between food stamp participation and poverty differ from state to state, this could make food stamp participation ineffective as a predictor of poverty in a model fitted to data for all the states.
County Model Changes
The county median household income model was altered to accommodate the unique qualities of the ratio of the number of dependent 1998 tax returns to the total number of returns, henceforth referred to as the "dependent ratio." Specifically, the relationship between the dependent ratio and all other variables in the model for Alaska counties was different than that relationship for the remainder of the U.S. counties. We modified the model by adding an interaction variable between the dependent ratio and an Alaska dummy variable. This allows us to estimate different slope parameters for the relationship between median household income and the dependent ratio for Alaska counties and the remainder of U.S. counties.
This change was made after model diagnostics and scatterplots showed the estimates for Alaska counties tended to be underestimated with respect to all other counties. Further inspection revealed that the dependent ratio was (1) the likely cause of low Alaska median household income estimates and (2) the relationship between the dependent ratio and other predictors and CPS median household income was fundamentally different for Alaska counties.
Goodness-of-fit statistics suggested that the model with the interaction term represented an improvement compared to the previous model. Furthermore, median household income estimates for Alaska no longer appear to be outliers in diagnostic scatter plots.