Recovery and Reinvestment Act spending at the state level: Keynesian stimulus or distributive politics?

DOI 10.1007/s11127-011-9876-x Recovery and Reinvestment Act spending at the state level: Keynesian stimulus or distributive politics? Andrew T. Young Russell S. Sobel Received: 18 January 2011 / Accepted: 15 August 2011 Springer Science+Business Media, LLC 2011 Abstract We examine the US state-level pattern of American Recovery and Reinvestment Act (ARRA) spending. We relate spending to (1) Keynesian determinants of countercyclical policy, (2) congressional power and dominance, and (3) presidential electoral vote importance. We find that the ARRA is, in practice, poorly designed countercyclical stimulus. After controlling for political variables, coefficients on Keynesian variables are often statistically insignificant. When they are statistically significant they are often the incorrect sign. On the other hand, statistically significant effects are associated with majority party House of Representative appropriations subcommittee and authorization committee membership. One striking result is that the elasticity of ARRA spending with respect to the pre-arra ratio of federal grants and payments to federal taxes paid is estimated to be greater than unity in most specifications. States previously capturing large amounts of federal funds continue to do so under the ARRA stimulus. Keywords Fiscal stimulus Fiscal policy Political economy Public choice Congressional dominance model American Recovery and Reinvestment Act JEL Classification D7 H5 E6 1 Introduction On February 13th, 2009, the American Recovery and Reinvestment Act (ARRA) was passed by the US Congress. Four days later, President Barack Obama signed the bill into law. The ARRA provides some $787 billion in combined spending increases and tax cuts aimed at A.T. Young ( ) R.S. Sobel Department of Economics, West Virginia University, Morgantown, WV 26506-6025, USA e-mail: Andrew.Young@mail.wvu.edu R.S. Sobel e-mail: rsobel2@wvu.edu

Fig. 1 Stimulus effectiveness in the SRAS; AD model stimulating the national economy. 1 At the aggregate level, this would appear to be textbook Keynesian countercyclical policy. However, especially when it comes to how to allocate government spending geographically, macroeconomic policy is never as simple as increasing G in a model. The ARRA s spending component is $500 billion and projects must be defined; regions to locate the projects must be determined; individuals and firms to carry out the projects must be hired. These choices are made through the political process, and there is a large literature in public choice theory devoted to understanding this complex process. Ideally, to maximize the effectiveness of the stimulus spending from a Keynesian perspective, funds should be targeted at the areas or industries with the most unemployed resources, where the marginal propensity to consume (MPC) is high, or both. Figure 1 illustrates this graphically in the context of a short-run aggregate supply (SRAS) and aggregate demand (AD) model. In the upper panel, a given stimulus (in the form of an increase in AD) yields a larger increase in output/expenditures if there are slack resources (moving along the flatter portion of SRAS). Alternatively, starting from a higher initial output, a given stimulus 1 Basic facts, including expenditure and tax numbers, can be found at recovery.gov.

Fig. 2 Agency-reported Recovery and Reinvestment Act (log of) funds per capita (a) announced and (b) made available versus (log of) 2008 real per capita GDP by state. Note. Funds are on the horizontal axis shows up to a larger extent in nominal price increases (moving along the steeper portion of SRAS). In the bottom panel, an increase in aggregate demand is illustrated in terms of initial government stimulus spending (autonomous expenditures) and then multiplier effects (induced expenditures). A higher MPC will be associated with a larger elasticity of private expenditures with respect to each stimulus dollar. There will, all else equal, be larger multiplier effects and a greater induced shift in AD. In evaluating the ARRA, neither the Obama administration s Council of Economic Advisors (CEA) nor the Congressional Budget Office (CBO) consider, at least explicitly, the state-level spending patterns (e.g., CEA 2010; CBO 2010). The CEA (2010: 47 49) does try to assess state-level ARRA effects on employment. Its approach is to take an estimated aggregate employment creation number and calculate states shares of that aggregate according to (1) states shares of aggregate non-farm employment, (2) state-level shares of ARRA outlays, or (3) industry employment shares in each state. However, none of these approaches consider that multipliers may be different from state to state. If state-level ARRA spending patterns are not determined with these differences in mind, then policy-makers may be getting a smaller bang than they could for each stimulus buck. Have the Keynesian countercyclical policy criteria been important determinants of ARRA allocations across states? Figures 2 and 3 present the patterns of per-capita ARRA funds across US states in relation to per-capita income levels and unemployment rates, respectively. We consider two measures of ARRA funds as reported by federal agencies: (1) announced and (2) made available. 2 While there is some indication that states with higher unemployment rates are associated with more ARRA funds made available, there is also some indication that lower income levels correlate with fewer funds made available. 2 There are also recipient-reported measures of funding. We examined these as well as the agency-reported measures in a previous version of this paper and the primary empirical results reported below largely were the same. Also, since the ARRA stimulus was ongoing at the time our research was conducted, the agencyreported funds numbers are likely to be more indicative of the outcomes of political processes (rather than recipient-reported data which may simply reflect the date at which we gathered the data).

Fig. 3 Agency-reported Recovery and Reinvestment Act (log of) funds per capita (a) announced and (b) made available versus the January 2009 unemployment rate by state. Note. Funds are on the horizontal axis For ARRA funds announced, there appears to be no relationship with either unemployment or income. If countercyclical policy criteria are not the whole story, an alternative is that political factors have been important determinants of the ARRA allocations. Given a spending package such as the ARRA, it is possible that political incentives confronting both President Obama and members of the US Congress played crucial roles in the allocation of funds across states. States that are politically important in the next presidential reelection, or states with tenured and powerful congressional representation, may have been able to secure a larger funding shares than would have been warranted based on Keynesian criteria. To the extent that this is true, the effectiveness of the countercyclical policy may be compromised by both distributive politics and rent-seeking. The purpose of this paper is to examine the pattern of ARRA spending across states and ask what factors played a significant role in determining state-level funding. We relate statelevel funding to (1) Keynesian determinants of countercyclical policy, (2) congressional power and dominance, and (3) presidential electoral vote importance. In terms of congressional power and dominance our data cover both houses of Congress and include average legislator tenure and appropriations committee membership. We also estimate the effects of appropriations subcommittee and authorization committee membership by considering separately the state level funding for the four largest cabinet departments (Health and Human Services, Education, Energy, and Transportation). Furthermore, the party affiliation of committee and subcommittee membership is controlled for. In terms of electoral vote maximization our data include an electoral importance measure that is increasing in both a state s electoral votes and the closeness of the popular vote in the previous presidential election. We also control for the raw percentage of votes captured by Obama in the 2008 election. We find that the ARRA is, in practice, poorly designed countercyclical stimulus. After controlling for political variables, coefficients on Keynesian variables are often statistically insignificant. When they are statistically significant they often go in the opposite direction predicted by Keynesian theory. On the other hand, statistically significant and positive effects are associated with majority party membership on House of Representatives appropriations subcommittees and authorization committees.

Also, one very robust result is that the elasticity of ARRA spending with respect to the pre-arra ratio of federal grants and payments to federal taxes paid is statistically significant, positive, and estimated to be greater than unity in most specifications. We interpret this variable as a proxy for a state s accumulated ability to extract federal funds (extraction capital). States previously demonstrating the ability to capture large amounts of federal funds continued to demonstrate that ability with respect to the ARRA. Thus in practice, the ARRA distributed more money to the same states that were the most politically powerful in receiving federal money pre-stimulus. This paper is organized as follows. Section 2 discusses previous literature relevant to this paper, especially works that inform our selection of the political factors underlying the cross-state distribution of ARRA funding. The data included in our econometric analysis are discussed in Sect. 3 while the results of the analysis are presented in Sect. 4. Section 5 concludes. 2 Existing theory and evidence Our evaluation of the ARRA as countercyclical policy is informed largely by the standard, textbook fare of Keynesian macroeconomics. On the other hand, while politics matter may be an uncontroversial statement, choosing variables belonging in an analysis of distributive politics requires some care. What kind of political power is important; how do we measure it? In considering the determination of federal spending priorities, both the legislative and executive branches are potential sources of influence. Concerning the legislative branch, economists have elaborated a congressional dominance model (Weingast and Moran 1983; Weingast 1984; Moe 1987, 1997). In this model, congressional committee members hold monopolies of authorization and/or appropriation on legislation falling under their committee s jurisdiction. The seniority system guarantees a committee member property rights over his or her seat (Holcombe and Parker 1991; Weingast and Marshall 1988: 132). The seniority system also grants leadership positions based on tenure. Legislation is the result of political competition between self-interested representatives and senators where competitive advantages are functions of committee membership and seniority. Concerning the executive branch, the president wields influence by broadly setting agendas and, more directly, exercising (or threatening to exercise) the veto option to maximize electoral votes (e.g., Wright 1974). 3 Adding to the political competition, the executive branch favors legislation benefiting politically important states in the Electoral College. The underlying goal is reassuring reelection for the current president or a successor from the current president s party. We can refer to the above as the electoral vote maximization model. In this paper we analyze not only the empirical relevance of Keynesian versus political determinants of the ARRA, but also, along the political dimension, the relevance of the congressional dominance model relative to the electoral vote maximization model. As such, this paper is related to several previous studies presenting evidence of political rent-seeking in various contexts. 4 Most analogous to our study are the works by Wright (1974), Anderson and Tollison (1991), and Couch and Shughart (1997) linking New Deal spending across 3 Such an approximation of behavior is probably of increased relevance in the case of a first-term president. 4 For example, Faith et al. (1982) find that Federal Trade Commission (FTC) rulings tend to be more favorable in districts with congressional oversight committee membership. Young et al. (2001) present evidence that

states to measures of both congressional power and the importance of states electoral votes in the upcoming presidential election. The ARRA responded to the worst recession since the Great Depression. During the Great Depression, the New Deal was an analogous response. Anderson and Tollison (1991: 164) state: Traditional arguments imply that state-by-state allocation of [New Deal] spending was a direct function of the need [...] of the residents of various states for federal relief from economic distress. Their evidence, as well as that presented by Reading (1973) and Wright (1974), calls these traditional arguments into question. Anderson and Tollison (1991: 175) reach the pointed conclusion that the New Deal was not big government s Garden of Eden, but rather the more familiar stomping ground of Homo economicus. Our analysis is not designed to include the need for federal relief per se. However, our countercyclical determinants (e.g., per-capita income levels and unemployment rates) surely control for such. This being the case, our analysis can be interpreted as addressing, with respect to ARRA, the questions considered by the authors above. This paper is also related to series of reports by Brito and de Rugy (2010) and de Rugy (2010) examining recipient-reported data at the congressional district level, as well as papers by Lazarus and Reifler (2010) and Gimpel et al. (2010) that developed concurrently with our own. The last of these contributions (Gimpel et al. 2010: 4) concludes that the geographic distribution of funds allocated under the [ARRA] was poorly targeted to need. This echoes our own findings. Brito and de Rugy (2010) and Lazarus and Reifler (2010) focus on party affiliation at the level of congressional districts. Focusing on party affiliation can be viewed as an extension of the congressional dominance model, and we incorporate that extension into our own analysis. Our state-level analysis is more natural for the seniority and committee membership variables suggested by the congressional dominance model. (It also is more appropriate for estimating the political determinants of action in the US Senate.) We also, unlike these other authors, analyze authorization and subcommittee membership, as well as the previous level of federal grants and payments. 3 Data Our data on ARRA funding are obtained from the federal government s recovery.gov website. We collect both agency-reported funds measures for all 50 US states. Agencyreported data come from final federal agency reports, Federal Procurement Data System, and USASpending. 5 The funds measures that we analyze are funds announced and funds made available. The state-level ARRA funds measured are converted into per-capita values using 2008 population values from the US Census. The ARRA stimulus program remains a work in progress as of this writing. We collected data during a period when recovery.gov produced its last update as April 7th, 2010 (almost 14 months after the act s passing). As of that date, aggregate agency-reported funds announced were $324 billion (or just about 65% of the total legislated spending component). Table 1 reports the aggregate numbers for both of the funds measures. Table 1 also reports state IRS audit rates are sensitive to the political importance states in the upcoming presidential election. Kosnik (2005) reports that congressional representation affects the Federal Energy Regulatory Commission s (FERC s) relicensing of hydroelectric dams. As a final example, Coats et al. (2006) document the fact the Homeland Security grants following 9/11 were designed in a way consistent with electoral vote maximization. 5 http://www.recovery.gov/transparency/agency/pages/agencylanding.aspx.

Table 1 Recovery and reinvestment funds measures Agency-reported (Bil $) Funds announced Funds made available Total $324 $304 Notes. Source: www.recovery.gov. Data was collected when website stated its last update was April 7, 2010 Health and human services $56 $71 Education $82 $75 Transportation $34 $34 Energy $20 $18 Four dept. % of total 53 65 Table 2 Summary statistics for state-level per capita recovery and reinvestment total funds measures Agency-reported funds Announced per capita (FUNDAN) Made available per capita (FUNDAV) Mean $1,101 $1,038 Maximum $7,344 $1,916 Minimum $700 $717 Standard deviation $942 $204 Coefficient of variation 0.856 0.197 Stand. deviation (natural log) 0.371 0.177 Notes. Source:www.recovery.gov for funds and the US Census for 2008 state populations. Data was collected when website stated its last update was April 7, 2010. Coefficients of variation are computed as standard deviations divided by means separately funds administered by the Department of Health and Human Services, Department of Education, Department of Transportation, and Department of Energy. As the last row of Table 1 shows, these four departments together account for a substantial fraction of total ARRA funds 53% and 65% for funds announced and made available, respectively. The two state-level measures are positively correlated, but the correlation is not exceedingly strong (r = 0.543). While funds announced is a larger number publicized prior to funds actually made available, we believe that it is useful to examine both measures. Variables that are significant determinants based on variation in either funds measure will be deemed the more robust determinants of state-level ARRA funding. Summary statistics for the state-level per-capita ARRA funds measures are reported in Table 2. Standard deviations taken as percentages of the means (coefficients of variation) are between 0.856 and 0.197, for funds announced and made available, respectively. Fund announced, then, has notably more variation across US states. The distributions for both measures are highly skewed, the mean values being typically closer to the minimum values than the maximum values. We will be discussing regression results in terms of elasticities, and so the last row of Table 2 reports the standard deviations of the natural log of each per-capita measure. (Table 2 also uses the abbreviated terminology employed in subsequent tables: FUNDAN and FUNDAV for, respectively, funds announced and made available.) The ARRA funds measures serve as our dependent variables in our empirical analysis. We relate ARRA funds to independent variables associated with (i) Keynesian countercycli-

Table 3 Control variables: descriptions and sources Variable Description Sources Keynesian variables UNEMP January 2009, seasonally-adjusted state unemployment rate UNEMP Change in state unemployment rate from January 2008 BLS to January 2009 REV State tax revenue growth from 2007 to 2008 Census GDP Level of real state per-capita GDP in 2008 BEA GDP Growth in real state GDP from 2007 to 2008 BEA MPC Marginal propensity to consume in the state Luengo-Prado and Sørensen (2008) Demographic controls SENIORS Percent of a state s population over 65 years of age in 2008 BLS Census BLACK Percent of a state s population that was African Census American in 2008 LATINO Percent of a state s population that was Latino in 2008 Census Political variables FEDAID RFEDAID Federal grants & payments per-capita to state & local govts by state in 2008 FEDAID as a ratio of per-capita federal revenue collected by state in 2008 Census AVHOUSET Average tenure of state s US representatives Various AVSENATET Average tenure of state s US senators Various HAC (SAC) Number of members on the US House (Senate) Various Appropriations Committee HACD (SACD) Number of members on the US House (Senate) Various Appropriations Committee HACR (SACR) Number of members on the US House (Senate) Appropriations Committee Various EIM Electoral importance measure (Garrett and Sobel 2003) Electoral-vote.com OBAMA08 Dummy variable (1 if Obama Won State in 2008; Electoral-vote.com 0otherwise) IRS cal policy, (ii) the congressional dominance model and, (iii) the electoral vote maximization model. We also include (iv) various demographic controls. All of these variables, their definitions, and their sources, are reported in Table 3. Summary statistics for the (i) Keynesian variables, (ii) political variables, and (iii) individual states electoral importance to the president are reported in Tables 4 and 5, respectively. 3.1 Keynesian variables Countercyclical policy that aims at priming the pump in an economy would best be targeted at depressed states where economic activity is below potential and labor is idle. As such, we consider the levels of per-capita income and unemployment, as well as their changes, as potential determinants of ARRA funding.

Table 4 Summary statistics for Keynesian variables UNEMP UNEMP REV ln(gdp) GDP MPC Mean 7.158 2.622 0.043 10.487 0.001 0.330 Maximum 11.600 4.700 0.591 10.940 0.071 0.558 Minimum 3.700 0.900 0.068 10.102 0.055 0.012 Standard deviation 1.851 1.045 0.091 0.175 0.019 0.137 Notes. Sources: see Table 3 Table 5 Summary statistics for political variables FEDAID RFEDAID HAC SAC HACD SACD Mean $1,687.90 0.229 1.200 0.600 0.740 0.340 Maximum $4,381.80 0.589 7.000 1.000 5.000 1.000 Minimum $994.17 0.017 0.000 0.000 0.000 0.000 Standard deviation $651.67 0.136 1.400 0.495 1.006 0.479 HACR SACR AVHOUSET AVSENATET OBAMA08 EIM Mean 0.460 0.260 8.001 12.090 0.580 10.391 Maximum 3.000 1.000 36.000 37.000 1.000 52.340 Minimum 0.000 0.000 0.000 0.000 0.000 2.630 Standard deviation 0.706 0.443 5.711 8.614 0.499 9.295 Notes. Sources: see Table 3 The level of the state unemployment rate we include is from January of 2009, the last estimate of joblessness observable to policy-makers prior to the legislation s passing. Also, acknowledging that policy-makers focus more on the change in economic conditions immediately prior to voting on the stimulus package, we include the change in a state s unemployment rate from January 2008 through January 2009. State-level GDP is only available on an annual basis, so we include the 2008 level of real per-capita state GDP and the growth rate of real state GDP from 2007 through 2008. While less plausible as an actual consideration by policy-makers, we also consider that funds may be targeted to states where the marginal propensity to spend is high. This would be sensible if policy-makers wanted to maximize multiplier effects and generate the largest economic response to ARRA. We include the state-level estimates of marginal propensities to consume (MPCs) from Luengo-Prado and Sørensen (2008). While these MPC estimates are based on data from 1964 through 1998, we at least attempt to control for some relative spending tendencies across consumers at the state level that, perhaps, change only slowly over time. Another factor that we consider as potentially relevant for designing countercyclical stimulus is a state government s budgetary stance. The recent recession has been characterized by several state government budget crises where public spending and payments to public workers have been threatened. Since such decreases in state government expenditures may essentially offset federal stimulus, policy-makers may target funds where those decreases

appear likely. We include the growth in a state s own-source tax revenue from 2007 through 2008 to control for this possibility. 3.2 Congressional dominance variables The congressional dominance model suggests that the cross-state distribution of spending under an economic stimulus plan such as the ARRA is the result of political competition between self-interested legislators. Advantages enjoyed by a specific member of the US House of Senate in that competition are a function of committee membership and seniority. We account for seniority by including the average tenure of a state s senators and representatives, separately, at the time of the ARRA s enactment. In both the House and the Senate, appropriations committees deal with determining the specific funds allocated to federal agencies, departments, and for other purposes. The appropriations committees deal with drafting legislation specifying such funding allocations. We include the number of legislators for each state who are members of the House appropriations committee as an explanatory variable; likewise for the Senate appropriations committee. When we turn to analyzing a panel of four department-based ARRA measures, we extend the congressional dominance variables to include memberships in appropriations subcommittees and authorization committees that are associated with those specific executivebranch departments. A description of those variables is reserved for Sect. 4 (also see Table 8 below). Finally, we include 2008 federal grants and payments to state and local governments in a given state. We believe that it is sensible to evaluate cross-state ARRA allocations relative to initial federal government allocations. However, we also believe that this variable can be interpreted in terms of the congressional dominance model (broadly conceived) as a state s stock of extraction capital, i.e., accumulated political apparatus and skills aimed at capturing federal funds. In other words, a good measure of a state s political power in securing federal funds is how much they have been able to capture in the recent past. There may also be institutional stickiness associated with grant and payment flows; current committee assignments may be influenced partly by the characteristics, including party affiliation and representation of politically important special interests, of the members of a state s congressional delegation in the past. If this is true, then the federal grants and payments variable proxies for lagged values of congressional dominance variables. It could be that states with high levels of federal grants and payments also make large contributions to federal tax receipts. Therefore, initial federal grants and payments may indeed be a significant determinant of ARRA, but the appropriate measure is one net of federal taxes contributed. We consider this possibility by including in some regressions federal grants and payments as a percentage of federal revenues collected from the state in 2008. 3.3 Electoral vote maximization variables In contradistinction to what the model s name might naively suggest, we do not simply include the number electoral votes that a state commands. Rather, we follow Garrett and Sobel (2003) in constructing an electoral importance measure. First, we record the percentage of a state s popular vote won by Barack Obama in 2008 (call it X). We then calculate Y = 1 4 (X 0.5) 2. Our electoral importance measure is then a state s electoral votes multiplied by Y. The weight (Y ) placed on the electoral votes (X) isatamaximumofone for a popular vote share of 50%. The weight decreases symmetrically towards zero as the vote share goes to 0 or to 100%.

Intuitively, the electoral importance of a state increases in (1) the number of electoral votes the state has and (2) the expected closeness of the state in the next presidential election. For example, despite commanding 31 electoral votes, a Democratic president may be unlikely to allocate many resources to New York because a Democratic landslide is already highly probable. (Barack Obama won 63 percent of the popular vote in 2008.) On the other hand, North Carolina might be deemed worthy of a large amount of resources because, despite commanding only 15 electoral votes, it was a battleground state in the previous election and may be so again in the next presidential election. (Barack Obama won the state by just over 50% of the popular vote in 2008.) In addition to measuring a state s importance in Electoral College voting, we also include a dummy variable that takes the value of one if Barack Obama carried the state in the 2008 presidential election and zero otherwise. Including this variable is intended to control for the fact that, especially in the case of a first-term president, resources may be allocated to reward states for their past votes (perhaps in the hope of convincing voters that it behooves them to cast their votes for the incumbent or for his party s candidate if he is term-limited in the next election). 3.4 Demographic controls The most basic demographic for which we control is a state s population in 2008. In our regressions, dependent variables (i.e., ARRA funds measures) are always converted into per capita terms and logged. This means that our estimated coefficients on other regressors can be interpreted as elasticities of per-capita ARRA funds with respect to those regressors. We also include as right-hand-side controls the 2008 percentage of a state s population that is over 65 years of age; the percent that is African American and the percent that is Latino. As with all other variables, definitions and data sources are provided in Table 3. 4 Regression results Since we have a large number of regressors (17) relative to state observations (50) we first consider whether some of the Keynesian variables are irrelevant or redundant. Redundancy is a particular concern with the Keynesian variables since we consider real per-capita state GDP and unemployment in terms of both levels and changes. Also, we would expect state revenue changes to be highly correlated with economic conditions. 4.1 Keynesian variables Columns 2 and 3 in Table 6 report the results of ordinary least squares (OLS) regressions of the two (logged, per-capita) ARRA funds measures on (1) a constant, (2) the log of percapita federal grants and payments, and (3) the Keynesian variables. Included in the Keynesian variables, real GDP per-capita also enters regressions in logged form. For expositional purposes, in discussing results reported in Table 6 (and from all subsequent regressions), we will often refer to variables by the abbreviations listed and defined in Table 3. Table 6 reports the results of Breusch-Pagan (1979) tests for heteroscedasticity. The null hypothesis of homoscedasticity is rejected for the column 3 funds made available (FUNDAV) regression but not for the column 2 funds announced (FUNDAN) regression. Given this ambiguity and our prior that heteroscedasticity will be important for our US state cross-section, the standard errors reported here and for all subsequent regressions are White (1980) heteroscedasticity-consistent standard errors.

Table 6 Regressions of total funds measures on Keynesian controls Variable ln(fundan) ln(fundav) ln(fundan) ln(fundav) ln(fedaid) 0.229 0.343 (0.199) (0.076) RFEDAID 0.974 0.896 (0.334) (0.207) UNEMP 0.017 0.046 0.011 0.053 (0.045) (0.020) (0.043) (0.022) UNEMP 0.028 0.018 0.020 0.036 (0.088) (0.033) (0.076) (0.033) REV 0.910 0.420 0.579 0.289 (0.454) (0.175) (0.491) (0.196) ln(gdp) 0.129 0.317 0.267 0.641 (0.215) (0.124) (0.266) (0.127) GDP 0.494 1.279 1.051 1.021 (2.177) (1.004) (2.236) (0.839) MPC 0.292 0.011 1.888 0.077 (0.424) (0.129) (0.410) (0.125) Breusch-Pagan 0.508 2.485 0.415 1.617 F -stat R 2 0.130 0.651 0.366 0.620 Notes. White heteroscedasticity-consistent standard errors in parentheses. Statistical significance at the 10 percent, 5 percent, and 1 percent levels is denoted, respectively, by,, and. Regressions include constant (not reported) The unemployment level is statistically significant only for FUNDAV (at the 5% level). The coefficient point estimate (0.046) is of the correct sign (from a Keynesian perspective) and a one standard deviation increase in the unemployment rate (1.85%) is associated with about an eight percent increase in ARRA funding. 6 The change in the unemployment rate is not statistically significant in either the FUNDAN or the FUNDAV regression. The level of per-capita GDP level is also only statistically significant (at the 5% level) in the FUNDAN regression. The point estimate is 0.317 and a one standard deviation increase in per-capita GDP (about 17.5%) is associated with an increase in ARRA funds per-capita of about 5.5%. However, this effect is of the opposite sign predicted by Keynesian theory: all else equal, more ARRA funds are allocated to higher income states. Interestingly, for FUNDAV the estimated unemployment and GDP per-capita effects almost exactly offset one another given the variation in our state-level data. In the case of GDP per-capita growth, the estimated effects are never statistically significant. If anything, the levels of income and unemployment appear to be more relevant 6 The unemployment rate enters the regression as a percentage rate so if, for example, the unemployment rate for a state is 4% the corresponding observation is 4 rather than 0.04. We note this because other variables are ratios or in logged form where a 4% change indeed corresponds to 0.04.

than the changes in these variables. Also, the estimated effects on the Luengo-Prado and Sørensen (2008) MPC estimates are never statistically significant. Finally, the effect of state tax revenue growth on ARRA funds is significant at least at the 10% level in both the FUNDAN and FUNDAV regressions. The point estimates (0.910 and 0.420) are economically large given the about 10% standard deviation in REV. However, the sign of the effects in both cases implies that states where budgetary conditions were improving were allocated more ARRA funds. We are inclined to dismiss these estimated effects as spurious. While the sign on GDP is incorrect in the FUNDAV regression, from the perspective of countercyclical policy, one can imagine reasons why wealthier states (with, e.g., more educated populations and well-funded politicians) are able to extract disproportionate shares of 2009 s stimulus program. However, in the case of state tax revenue growth, no reasonable, direct, and positive relationship between tax revenue growth and ARRA funds seems plausible. For the regressions in columns 2 and 3, the log of FEDAID is only significant in the case of FUNDAV. However, columns 4 and 5 report the results of analogous regressions except for the replacement of logged FEDAID with the ratio RFEDAID. Now the (net) federal grants and payments variable is statistically significant at the one percent level for both funds announced and funds made available. Notably, the estimated elasticities are both close to unity (0.974 and 0.896 for, respectively). 7 Furthermore, the questionable REV significant effects disappear. For the FUNDAV regression, the correctly signed unemployment effect and the incorrectly signed per capita GDP effect remain, the latter of these more than doubling in size (as a point estimate). Table 6 suggests that the only reasonable evidence of ARRA funds being allocated according to Keynesian countercyclical criteria (in a way that would suggest improved policy) appears in the unemployment level coefficient for funds made available. On the other hand, when the income level and revenue growth coefficients are statistically significant they are always of the opposite sign suggested by Keynesian theory and, perversely, economically large. In the case of real per-capita GDP levels, the results may suggest that wealthier states are receiving a disproportionately large share of the stimulus. The remainder of the paper explores whether political determinants are better accounted for by political variables suggested by the congressional dominance and electoral vote maximization models. In the regressions that follow, we continue to include the per-capita GDP and unemployment levels as controls. (In some specifications these are replaced by their growth rates or first differences, respectively.) 4.2 Political variables Table 7 reports the results from OLS regressions of funds measures on the political variables while controlling for demographic factors (BLACK, LATINO, and SENIORS), real percapita GDP and unemployment rate levels, and the ratio of federal grants and payments to federal revenue paid (RFEDAID). Columns 2 and 3 report the results of our baseline specification for both FUNDAN and FUNDAV. The average tenure of a state s representatives has a statistically significant and positive estimated effect on both funds measures. (Significance is at the 10% level for FUNDAN and at the 1% level for FUNDAV.) These estimated effects are consistent with the congressional dominance model. However the point estimates (0.016 and 0.009, respectively) are 7 Neither can be distinguished from unity at the 10% significance level.

Table 7 Regressions of per capita funds measures on political controls, demographic controls, and selected Keynesian controls Variable ln(fundan) ln(fundav) ln(fundan) ln(fundav) ln(fundan) ln(fundav) RFEDAID 1.604 0.929 1.669 0.874 1.236 0.609 (0.369) (0.178) (0.392) (0.156) (0.342) (0.170) HAC 0.020 0.007 (0.062) (0.037) SAC 0.085 0.019 (0.063) (0.054) HACD 0.123 0.077 0.127 0.080 (0.130) (0.039) (0.135) (0.051) SACD 0.153 0.032 0.071 0.021 (0.093) (0.058) (0.091) (0.064) AVHOUSET 0.016 0.009 0.018 0.007 0.027 0.013 (0.008) (0.003) (0.010) (0.003) (0.013) (0.004) AVSENATET 0.008 0.003 0.008 0.002 0.010 0.001 (0.004) (0.002) (0.004) (0.002) (0.006) (0.003) OBAMA08 0.071 0.067 0.049 0.020 0.056 0.091 (0.124) (0.038) (0.098) (0.035) (0.082) (0.045) EIM 0.007 0.004 0.003 0.008 0.000 0.005 (0.010) (0.006) (0.011) (0.006) (0.011) (0.007) UNEMP 0.008 0.027 0.020 0.030 (0.025) (0.014) (0.032) (0.014) ln(gdp) 0.680 0.562 0.754 0.524 (0.304) (0.140) (0.345) (0.113) UNEMP 0.036 0.002 (0.043) (0.021) GDP 1.226 0.215 (2.114) (0.701) BLACK 0.009 0.001 0.008 0.002 0.007 0.001 (0.005) (0.001) (0.004) (0.001) (0.014) (0.002) LATINO 0.017 0.001 0.022 0.001 0.023 0.001 (0.011) (0.002) (0.014) (0.002) (0.014) (0.003) SENIORS 0.103 0.003 0.109 0.007 0.101 0.014 (0.083) (0.013) (0.087) (0.013) (0.090) (0.017) B-P F -stat 1.176 0.477 1.212 0.655 1.018 0.685 R 2 0.420 0.697 0.451 0.750 0.412 0.639 Notes. White heteroscedasticity-consistent standard errors in parentheses. Statistical significance at the 10 percent, 5 percent, and 1 percent levels is denoted, respectively, by,, and. Regressions include constant (not reported). Regressions reported in columns 3 and 4 include HACR and SACR (not reported); neither is ever statistically significant

somewhat modest. A one standard deviation increase in the average tenure of a state s representatives (5.711) is associated with only about a 10% increase in FUNDAN and a 5% increase in FUNDAV. (This is just about 28% of the standard deviation of logged funds in either case.) The average tenure of a state s senators is also statistically significant in the FUNDAN regression, but its sign is negative ( 0.008) in contrast to the prediction of the congressional dominance model. OBAMA08 is statistically significant at the 5% level in only the FUNDAV regression but is fairly small (0.067). A state having voted for Obama is associated with less than a 7% increase in ARRA funding per capita. The EIM measure is never significant and the point estimates are always negative. A political variable that is always estimated to have both statistically and economically significant effects is the extraction capital measure (RFEDAID). The coefficient point estimates are 1.604 and 0.929 for the FUNDAN and FUNDAV regressions, respectively. Both estimates are significant at the 1% level. The standard deviation of RFEDAID is 0.136 and is associated with a just under 60% of a standard deviation increase in logged funds announced per capita and just over 70% of a standard deviation increase in logged funds made available. US states previous skill at capturing federal funds appears to be an empirically important determinant of ARRA funding. Appropriations committee membership has no statistically significant effect on ARRA funding in the baseline regressions. Columns 3 and 4 report analogous regressions where, instead of total appropriations committee membership (HAC and SAC), we consider party affiliation. Democratic committee membership (HACD and SACD) and Republican committee membership (HACR and SACR) for each state are included separately in both regressions. Republican committee membership is not statistically significant in either the FUNDAN or FUNDAV regression. (To conserve space, HACR and SACR are not shown in Table 7 though they are part of the estimation.) As to whether majority party membership is an important determinant of ARRA funding in its own right, the evidence is mixed. SACD is never significant. HACD is statistically significant at the 1% level, but only for the FUNDAV regression. In that case the point estimate is large (0.077). Having an additional Democratic member on the House appropriations committee is associated with an almost 8% increase in funds made available per capita. All other results remain basically unchanged from the baseline regressions. 8 Finally, columns 5 and 6 perform a robustness check by replacing UNEMP and ln(gdp) with UNEMP and GDP. The latter did not register significant correlations with ARRA funds in Table 6. However, policymakers consideration of changes in economic conditions remains intuitively appealing. The column 5 and 6 results demonstrate that this intuition still does not pan out: neither UNEMP nor GDP is statistically significant. Most other results remain substantively the same. (An exception is that HACD is no longer significant.) 4.3 Department panel results; authorization and subcommittees In evaluating the congressional dominance model, our estimation of committee representation effects was limited above to appropriations committee membership. However, it is possible that important effects are associated instead with appropriation at the subcommittee level or authorization committees. To explore this possibility we examine data on 8 While not reported here, we also explored state-government partisanship effects by including the political party affiliation of every state s governor. This binary variable was never statistically significant in regressions.

Summary statistics for state-level per capita Recovery and Reinvestment Department funds mea- Table 8 sures (Bil $) FUNDAN FUNDAV HHS Mean $144 $171 Standard deviation $90 $56 Standard deviation (natural log) 0.397 0.306 ED Mean $269 $218 Standard deviation $21 $27 Standard deviation (natural log) 0.077 0.149 TR Mean $149 $114 Standard deviation $75 $59 Standard deviation (natural log) 0.389 0.378 EN Mean $83 $43 Standard deviation $91 $18 Standard deviation (natural log) 0.728 0.359 Notes. Source: www.recovery.gov. Data was collected when website stated its last update was April 7, 2010. Coefficients of Variation are computed as standard deviations divided by means ARRA funds associated with four different federal Departments: Health and Human Services (HHS), Education (ED), Transportation (TR) and Energy (EN). Table 8 reports summary statistics on ARRA funds measures for each of these departments at the state level. We collect data on each state s membership not only on House and Senate (i) appropriations committees (HAC and SAC), but also the (ii) appropriations subcommittees (HAPSUB and SAPSUB), (iii) authorization committees (HAUTH and SAUTH) and (iv) authorization subcommittees (HAUTHS) associated with each federal department. (Note that there US Senate rules do not provide for the establishment of authorization subcommittees.) Table 9 reports these various committees and subcommittees by department. Table 10 reports on panel regressions that include all variables from the Table 7 regressions plus Democratic and Republican HAPSUB, SAPSUB, HAUTH, SAUTH, and HAUTHS memberships. There is no time dimension, as one would typically associate with a panel of data, but rather the dimensions are departments by states: 4 50 = 200 observations. Estimation is by OLS with fixed effects associated with each department. Since heteroscedasticity is again a likely concern, the reported standard errors are White heteroscedasticity-consistent (across US state observations). RFEDAID is once again statistically significant at the one percent level for both FUN- DAN and FUNDAV. The estimated effects are both economically large. (Point estimates are, respectively, 1.136 and 1.012. 9 ) A state s previously-effectiveness in extracting federal funds is an important determinant of the cross-state distribution of ARRA funding. The political variables included from our previous regressions (see Table 7) are rarely statistically significant and in almost all of those cases economically small. The exception is OBAMA08 in the FUNDAV regression. In this case, a state s having voted for Obama in 2008 is associated with close to 13% reduction in ARRA funds per capita from a given 9 Neither can be distinguished from unity at the 10% significance level.

Table 9 Congressional committees and subcommittees Agency Appropriations Subcommittees Authorization Subcommittees HHS Senate appropriations House appropriations Labor, health and human services, education, and related agencies Labor, health and human services, education, and related agencies Senate Committee on Health, Education, Labor, and Pensions Senate Special Committee on Aging House Committee on Energy and Commerce Health oversight and investigations House Committee on Oversight and Governmental Reform House Committee on Small Business House Committee on Ways and Means ED Senate appropriations House appropriations Labor, health and human services, education, and related agencies Labor, health and human services, education, and related agencies Senate Committee on Health, Education, Labor, and Pensions House Education and Labor Committee TR Senate appropriations House appropriations Transportation, housing and urban development, and related agencies Transportation, housing and urban development, and related agencies Senate Committee on Commerce, Science, and Transportation House Committee on Transportation and Infrastructure EN Senate appropriations House appropriations Energy and water development Senate Committee on Energy and Natural Resources Energy and water development, and related agencies House Science and Technology Committee Energy and environment

Table 10 Panel regressions of HHR, ED, TR, and EN per capita funds measures on political controls, demographic controls, and selected Keynesian controls Variable ln(fundan) ln(fundav) RFEDAID 1.136 1.012 (0.357) (0.262) HACD 0.009 0.083 (0.095) (0.075) SACD 0.113 0.045 (0.060) (0.071) AVHOUSET 0.009 0.001 (0.004) (0.000) AVSENATET 0.009 0.003 (0.003) (0.002) OBAMA08 0.020 0.127 (0.089) (0.042) EIM 0.022 0.019 (0.014) (0.009) Democrats Republicans Democrats Republicans HAPSUB 0.041 0.116 0.036 0.057 (0.018) (0.122) (0.015) (0.043) SAPSUB 0.043 0.056 0.048 0.043 (0.044) (0.090) (0.063) (0.039) HAUTH 0.065 0.047 0.034 0.036 (0.022) (0.019) (0.018) (0.031) SAUTH 0.056 0.007 0.078 0.033 (0.026) (0.083) (0.052) (0.054) HAUTHS 0.090 0.059 0.005 0.047 (0.114) (0.060) (0.022) (0.071) UNEMP 0.012 0.012 (0.023) (0.009) ln(gdp) 0.349 0.643 (0.211) (0.236) R 2 0.683 0.871 Notes. White heteroscedasticity-consistent (across US state observations) standard errors in parentheses. Fixed department effects are included. Statistical significance at the 10 percent, 5 percent, and 1 percent levels is denoted, respectively, by,, and. Regressions include constants and the same demographic controls as table 8 regressions (not reported). Regressions also include HACR and SACR (not reported); neither is ever statistically significant department. This result is not supportive of the electoral vote maximization model. Only AVHOUSET is significant in both regressions but its coefficient differs in sign from the FUNDAN (positive) to the FUNDAV (negative) regression. On the other hand, subcommittee and authorization variables appear to matter. Democratic membership on either a House subcommittee on appropriations or an authorization

committee is significantly associated with greater ARRA funding. For either FUNDAN or FUNDAV an additional Democratic appropriations subcommittee member is associated with 4% more ARRA funds per capita from a given department. An additional Democratic authorization committee member is associated with about 6.5% (FUNDAV) or 3.4% (FUN- DAN) more funds per capita. (In the FUNDAN regression, having an additional Republican House authorization committee member is associated with a significantly smaller share of ARRA funds.) Evidence for the importance of Senate subcommittee and authorization committee membership is more scant. Only Democratic SAUTH is significant and only for the FUNDAN regression. In that case the point estimate is positive and associates an additional committee member with about 5.6% more per capita ARRA funds from a given department. Keynesian variables are now never both statistically significant and of the correct sign. Only the level of GDP per capita is significant and only in the FUNDAV regression. Its coefficient, in that case, is positive. We also again tried substituting UNEMP and GDP for the levels. Though not reported in Table 10, the Keynesian variables in change/growth rate form did not enter statistically significantly in either regression. 5 Concluding discussion We have examined the US state-level pattern of American Recovery and Reinvestment Act (ARRA) spending. We related spending to (1) Keynesian determinants of countercyclical policy, (2) congressional power and dominance, and (3) presidential electoral vote importance. The ARRA appears to be, in practice, poorly designed countercyclical stimulus. There is some evidence that high unemployment states have received more funds per capita. However, that result is quite fragile across specifications. The more robust finding is that higher per capita income states received have received more funding. Also, even if we take the unemployment effect at face value, it is at least offset by the (perverse from the perspective of countercyclical policy) GDP per capita level effect. Alternatively, we have documented statistically significant effects associated with majority party House of Representative appropriations subcommittee and authorization committee membership. The evidence suggests that Democratic members of these committees were able to channel more funds towards their respective states. The most striking result is that the elasticity of ARRA spending with respect to the pre- ARRA ratio of federal grants and payments to federal taxes paid is estimated to be greater than unity in most specifications. States previously capturing large amounts of federal funds have continued to do so under the ARRA stimulus. We refer to this previous, demonstrated ability to capture federal funds as extraction capital. By this we mean the accumulated political apparatus and skills aimed at capturing federal funds. The importance of extraction capital is very robust across specifications. In terms of federal funds, the rich do appear to get richer. Acknowledgements We thank Brandon Brice for excellent research assistance. References Anderson, G. M., & Tollison, R. D. (1991). Congressional influence and patterns of New Deal spending. The Journal of Law & Economics, 35, 161 175.