Following McGuire (1978) for the basic theoretical model, we assumed that the decision-maker is the combined state and local system. This approach allowed us to model state and local spending on both public and private goods in a consistent way. The budget constraint consists of state fiscal capacity augmented by federal grants. The state and local governments then make spending and taxation choices subject to this budget constraint. States that have preferences for public spending over private consumption tend to raise more from their citizens in taxes and allocate such tax revenue to public spending. The theoretical foundation for the linear expenditure model used in this report is described more fully in McGuire (1978) and Appendix A to this report.
i) Dependent Variable: Spending Per Capita
The dependent variable in the regression was total state and local spending per capita on public functions defined for five general areas: (1) cash assistance, (2) Medicaid,7 (3) other non-health social services (e.g.,foster care, child care, low-income energy assistance), (4) spending on public hospitals8; and (5) non-social welfare (e.g., education, transportation, law enforcement).9 The total state and local spending for particular categories includes the federal grants that the state and localities spend. However, such federal grants are not reported in the Census of Governments at the same level of detail as is the spending activity. Federal grants are reported only for social welfare spending as a whole, which fails to include public hospitals, and for non-social welfare spending. This practice makes measuring spending impossible from state and local sources at the higher levels of disaggregation for which we report types of social welfare spending. The price deflators used in the analysis depended on the type of spending. For cash assistance and other non-health social welfare spending, we used the general gross domestic product (GDP) price deflator. For Medicaid and public hospitals, we used the Consumer Price Index (CPI) for health care, which we believe better captures price trends in the health sector than does the overall deflator.
ii) Explanatory Variables
The explanatory, or independent, variables in the regression models included a measure of fiscal capacity, measures of the need for social welfare spending, state effect dummy variables, and year dummy variables. In addition, some but not all models attempted to capture price effects of federal grants using a McGuire-type analysis (see Appendix A).
We considered the possibility of creating a consistent data series incorporating some or all of the elements of the PCPI, RTS, and TTR approaches, as discussed in the Introduction. However, the PCPI model is the easiest and most reliable to implement because consistent data are available across states from the Bureau of Economic Analysis (BEA). Therefore, in our models, state and local resources (i.e., fiscal capacity) are measured by state per capita personal income, deflated by the implicit price deflators of the general GDP relative to 1996.
Federal grant amounts for social welfare and non-social welfare functions as measured by the Census data on intergovernmental revenues appear as explanatory variables in the model. Because both variables enter the "budget constraint" for the public decision-makers, the effects of federal grants might be thought comparable to the effects of personal income. However, research on public expenditures has identified what has been termed the "flypaper effect" (i.e., money sticks where it hits) through which federal grant money exerts a greater stimulatory effect on public spending than increases in private income on public spending (Gramlich, 1977; Hines & Thaler, 1995; Gamkhar & Oates, 1996). We accommodated the likelihood of a flypaper effect by introducing grants into the model as explanatory variables separately from our measure of state fiscal capacity. We also adjusted federal grants for inflation using the general GDP price deflator.
As measures of the (relative) need for spending on social welfare in the core model, we included:
- Number of poor persons per capita,
- Number of unemployed persons per capita, and
- Population density.
These variables are thought to capture need for social welfare spending for several reasons. First, families in poverty are more likely to qualify for cash assistance and also need social services. Second, unemployment captures economic downturns and also economic hardship associated with involuntary unemployment. Third, population density is a variable that has in other studies been correlated with government spending. It undoubtedly captures a number of effects, including urbanization and special resource costs associated with a high population density.
A number of additional factors might influence state spending. We have chosen the three variables identified above because poverty, unemployment, and population density are thought to proxy in slightly different ways for the need for social welfare spending. We considered adding poverty for various subgroups, such as children and elderly, but the data were unavailable over a sufficient time period. We also experimented with per capita number of persons in urban areas, but this variable was highly correlated with population density, and population density seemed more strongly associated with per capita spending than urbanization.
State Effect Variables
The state effect variables, as described above, were dummy variables defined as one for a particular state and zero otherwise. Inclusion of such variables in the regression allowed us to estimate separate intercepts for each state.10 These state effects were useful in assessing how much of the spending on certain categories of social welfare was due to explanatory variables, such as per capita personal income, per capita federal grants, or need variables, and how much was due to an inherent willingness of the state to spend on the particular function. The state effect variables were important in our analysis of spending patterns in the six states selected for more in-depth study and in comparing spending among rich and poor states in general.
We captured time trends in the models primarily by dummy variables for year. The magnitude of the coefficients on the year dummies indicated whether spending increased or decreased following certain seminal policy initiatives or shifts, such as welfare reform or expansion in Medicaid eligibility. We did not interact the year dummies with the state dummies in general because we had no specific hypotheses to test state by state. However, we estimated the quartile regressions with and without the year dummies, essentially estimating different time trends for rich and poor states.