Public expenditures and the unemployment rate in the American states: panel evidence

Abstract

We reexamine the Unemployment Rate (UR) – government expenditure nexus in a panel of 50 State and Local Governments (SLGs) over the period 1977–2006 to provide new pre-recession empirical evidence that helps put the expectations on the effects of the federal relief to SLGs in a broader context. We found that: (1) per capita real public spending (total and capital, assistance and subsidies, wages and salaries, and social insurance categories) was part of a cointegrating relationship with UR and real per capita state personal income. (2) With the exception of social insurance, other spending variables, when statistically significant, actually had a depressing effect on UR. The magnitude of this effect, however, was generally small. UR was most sensitive to increases in wages and salaries. (3) Long-term causality analysis based on panel error-correction coefficients provided consistent evidence of a causal effect from spending to UR, but less consistent evidence of such effect in the opposite direction. Social insurance, however, drove UR. (4) The size of the error-correction coefficients suggested a slow response of UR to deviations from the cointegrating relationship. (5) The marginal effect of spending on UR increased with the amount of the federal grants received. Our results suggest that public spending may not serve as a quick fix in relation to UR. They also seem to favour allocation of the federal funds to wage and salaries and assistance and subsidies, but not to capital and social insurance expenditures to lower UR.

Keywords:

• state and local governments
• state unemployment rate
• government expenditures
• panel cointegration and causality

JEL Classification::

• H7
• C22

Notes

¹ See, for example, Stone (2009). In this connection, Krugman (2009) contends that ‘Without the recovery act, the free fall would probably have continued, as unemployed workers slashed their spending, cash-strapped state and local governments engaged in mass layoffs, and more. The stimulus didn’t completely eliminate these effects, but it was enough to break the vicious circle of economic decline. Aid to the unemployed and help for state and local governments were probably the most important factors. If you want to see the recovery act in action, visit a classroom: your local school probably would have had to fire a lot of teachers if the stimulus hadn’t been enacted.’ (Emphasis added).

² See, for example, Riedel (2008) and Ferrara (2009).

³ For detailed discussions of the classes of theoretical models reviewed here, empirical evidence of the effects of fiscal policy and a comprehensive list of original references see Hebous (2011).

⁴ Barro (1990) contends that the positive effects of productive government expenditures on growth (for example, resulting from positive externalities, development of legal, administrative and economic infrastructure) may begin to be dominated by their negative effects (through, for example, excessive and distortionary taxes and government inefficiencies) as the size of the government grows too large.

⁵ The underlying argument is that the positive wealth effect of public spending offsets the negative wealth effect of higher wage taxation leaving the substitution effect of the latter to reduce the labour hours.

⁶ Several post-stimulus period studies have yielded generally favourable but, in some cases, conflicting findings. Chodorow-Reich et al. (2010) estimate that an additional $100 000 in Medicare funding injected into a state was associated with an additional 3.5 job-year less than half of which was in government health and education sectors. Clemens and Miran's (2010) baseline estimates indicate a spending multiplier equal to 1.7 and that states saved one job by avoiding mid-year cuts in spending of $25 000, Aizenman and Pasricha's analysis (2010), on the other hand, suggests that the expansionary effect of the stimulus expenditure merely offset the contractionary effect of the collapse in state expenditures, thus, yielding a nearly zero net effect in 2009. Feyrer and Sacerdote (2010) report the results favouring support programs for low-income households and infrastructure spending over grants to states for education and justice in terms of their expansionary employment effects. Wilson (2011) finds that the employment impact of increased infrastructure spending attributable to ARRA was large particularly in relation to SLG employment. Furthermore, general fiscal aid showed to have a more favourable employment effect than restricted aid such as Medicaid.

⁷ In a seminal work, Granger and Newbold (1974) show that if variables are nonstationary in level (or contain a ‘unit root’) their moments are time dependent and standard statistical methods involving such level variables lead to spurious regression results and inconsistent parameter estimates. Engle and Granger (1987) also demonstrate that if nonstationary variables of the same order of integration (first-difference stationary, for example) share a stable long-term (cointegrating) relationship, then level relationships can be used to draw correct statistical inferences. In this case the estimated slope coefficients measure long-term equilibrium responses as opposed to responses to cyclical changes. The authors propose an Error-Correction Model (ECM) which allows to exploit the information provided by the cointegrating equation in explaining the short-term dynamics of the variables included in it (see Section III).

⁸ These variables range from general state economic conditions to specific ones such as educational attainment, earnings per worker and percentages of population residing in metropolitan areas, living in poverty, and living in owner occupied residences (a mobility factor).

⁹ This follows from the fact that (TEXPC) = (TEX/PI)(PI/POP) where TEX, PI and POP are total government expenditures, personal income and population, respectively. Taking logs, we have ln(TEXPC) = ln(TEX/PI) + ln(PI/POP).

¹⁰ ‘Economic’ expenditure categories are defined based on economic characteristic or type of spending rather than its purpose as in ‘functional’ categories.

¹¹ Note that wages and salaries category is an ‘exhibit’ item under the economic categories of spending. It represents the sum of the labour compensation parts of all other categories. Given the importance of the public sector employment noted above, however, we believe its inclusion in our analysis as a ‘separate’ category may be informative. Also, as pointed out by a referee, this category does not represent contracted work and, therefore, may not fully capture the effect of the public sector on the labour market.

¹² The Bureau of Labor Statistics data indicate that it took nearly 30 months before state government employment fell below its level at the start of the recent recession. Boyd and Dadayan (2010) explain some of the reasons for this rather long lag as follows: ‘That reflects, among other things, the stable or sometimes rising demand for services provided by government, the lengthy and contentious political and budgetary decision-making processes in government, and lags in how the finances of different levels of government are affected by recessions’.

¹³ The data for nominal total state-local expenditure and its sub-categories were taken from US Census Bureau: State and Local Government Finances. We used the deflator for overall state-local government consumption and investment expenditures (2000=100) and state population size to construct expenditure variables on a real per capita basis. Personal state income, however, was adjusted for inflation using US GDP deflator (2000=100). The source for the deflators and state population size was Bureau of Economic Analysis (US Department of Commerce): Regional Economic Accounts. The annual state unemployment rate was available on a seasonally unadjusted basis only. The source for this variable was Bureau of Labor Statistics (US Department of Labor).

¹⁴ Monte Carlo experiments conducted by Pesaran (2007) suggested that the CADF tests have satisfactory size and power even for relatively small values of N and T.

¹⁵ The exception is when INSPC is the dependent variable. Note that the sign of the error-correction coefficient is positive and insignificant in this case implying no long-term causal effect from UR to INSPC.

¹⁶ Note that, with a few exceptions, we are unable to reject the hypothesis of no long-term causal relationship with PIPC as the dependent variable (result are not shown).

¹⁷ The interaction term estimated using DOLS are all negative but insignificant. These results (available upon request) should be treated with caution due to a high degree of multicollinearity.