Spending on Social Welfare Programs in Rich and Poor States. Final Report.. Chapter II: Approch

To answer the research questions identified in the Introduction, first we analyzed spending trends and patterns over 24 years for the 50 states plus the District of Columbia by using sample means from Census data and an econometric model estimated from the pooled time series and cross-section data. We supplemented this analysis with site visits and further analysis of qualitative and quantitative data from six states, selected as having high needs relative to their fiscal capacity. The econometric model estimates were used to identify states exhibiting a high propensity to spend on certain types of social welfare. Employing this information, we drew comparisons between rich states and poor states in general and also among the six states for which we had additional case study data.

We designed our field research in part to help interpret the estimated econometric models and extend the range of factors whose influences might be assessed. At the most general level, the analyses of the econometric data posed and sharpened the questions for the site visits and related analyses of the six states.

First, the site visits were used to weigh the credibility of different explanations of the estimated regression coefficients. For example, the econometric models found that population density exerted positive effects on cash assistance expenditures and negative effects on health-related spending. The site visits suggested hypotheses, consistent with a wide array of quantitative and qualitative data, as to why these differences might occur, at least in poor states.

Second, the econometric models estimated state effects for total as well as different types of social welfare spending. Intercepts estimated for each state,¹⁵ these coefficients represented an average level of spending for a particular state after controlling for the effects of all included variables, such as fiscal capacity, unemployment, and poverty. Because these state-effect estimates stripped off the linear effects of economic and demographic variables, they varied greatly among the six poor states and helped sharpen our analyses of institutional and political factors. For example, though Mississippi's spending on medical assistance could not be considered high in an absolute sense, it was sizeable after controlling for the state's fiscal capacity and other significant variables. Thus, the econometric analysis changed the question from why the state's spending on medical assistance was so low to why it was higher than we had expected, given the state's economy and demographics.

Third, the estimated state effects allowed us to examine with greater precision how states varied in the ways they combined, or failed to combine, different types of social welfare expenditures. For example, we found a fundamental division between poor states (i.e., between states that put enormous emphasis on medical assistance and other states whose long-run spending tendencies were more balanced between different functions). We estimated these different configurations, or packages, of spending through the econometric analysis and posed important questions for the site visits.

Fourth, the six state case studies allowed us to assess findings from the econometric analyses in light of state spending changes after fiscal year 2000, the last year for which Census Bureau spending data were available. For example, the models indicated that spending on cash assistance and Medicaid went up during recessions and down during economic booms, other things being equal, while non-health social services showed the opposite pattern. Because the states we studied were, for the most part, experiencing severe fiscal pressures after several years of economic growth, we could draw on quantitative and qualitative data in the case studies to tests these and other expectations.

We also estimated separate econometric models for each of the six states in the field research sample, and we thought these separate estimates would clarify other important differences and similarities among these states. However, with few exceptions, these separate models also turned out to be hard to interpret because of instability, we suspect, due to small degrees of freedom. Thus, we do not present these models in the current report.

⁷ The Census defines this category, which is primarily Medicaid spending, as "payments to medical vendors."

⁸ Census does not consider spending on state run hospitals social welfare spending because the patients at such hospitals might not be predominantly low-income. However, some part of the federal grants, which we measured by Census intergovernmental revenues, goes to support public hospitals primarily through the Medicare and Medicaid Disproportionate Share Hospital (DSH) program. Therefore, examining how spending on public hospitals varied across states was important. Unfortunately, Census considers grants for public hospitals to be grants for non-social welfare and fails to disaggregate grant amounts by detailed function. So, we were unable to identify federal grants for public hospitals and thus were unable to separate public hospital spending into a federal share component and a state and local share component as we could do for overall Census social welfare spending. Nonetheless, we used spending per capita on public hospitals as a dependent variable in most of our regression analyses where we did not have to identify separately the federal and state and local funding components.

⁹ Definitions for these five categories in the Census data used in our analysis appear in Exhibit II-1.

¹⁰ The coefficients on the explanatory variables were not allowed to vary across states in our general model, but we did estimate the regression separately for each quartile defined in terms of average per capita personal income. These quartile regressions estimated the coefficients separately for each quartile. However, the estimated state effects used in our cross-state analysis (see subsection III .B.4.) came from the regression estimated over all states. These state effects captured differences in state spending unexplained by the variables in the fixed coefficient model. Some part of these effects could be due to the fact that states had different responses (i.e., variable coefficients) to the explanatory variables.

¹¹ Public welfare expenditure includes all of the categories shown in Exhibit II-1, except Public Hospitals. The Census views spending on state-run public hospitals as outside its social welfare category. However, we included spending on public hospitals as a variable of interest, partly because state-run public hospitals receive Medicaid funding and also because low-income individuals might receive services in the public hospitals.

¹² The Federal Categorical Assistance category (E67) tracks federally funded programs and includes AFDC cash assistance, TANF cash assistance, or both, to the extent it passes through state accounts; federal Supplemental Security Income (SSI); plus state supplements. The only federal SSI included in E67 is retroactive federal payments to reimburse the state for payments made to individuals under state supplement programs. The Other Cash Assistance Programs category (E68) includes cash assistance programs not under federal categorical programs.

¹³ As noted, Vendor Payments for Medical Care is the largest category by far and consists mostly of Medicaid.

¹⁴ The Other Public Welfare category (E79) includes operational payments for administrative workers and payments for programs such as child care, foster care, low-income energy assistance, social services to the physically disabled, and programs funded by the Social Services Block Grant.

¹⁵ The "state effect" for each state was computed by adding the intercept to the coefficient of the dummy variable for the state.