Economic development program spending in the US: is there club convergence?

ABSTRACT

This paper provides a novel investigation of whether budgetary spending on economic development programs converges across US states. States use a wide array of tax and subsidy programs to try to attract firms in a highly competitive environment. If states engage in strategic tax and incentive competition as previous literature suggests, we should expect economic development spending to converge over time. Using a national database of state ‘out of pocket’ economic development expenditures, we apply the panel convergence method developed by Phillips and Sul (2007, 2009) which endogenously identifies the number and members of convergence clubs. We find that states flock together in three spending clubs which reflect socioeconomic characteristics. The existence of multiple clubs with heterogeneous spending patterns reveals the complexity of state-level budgetary efforts put towards economic development programs.

KEYWORDS:

• Economic development spending
• club convergence
• fiscal competition
• state government finance

JEL:

• H25
• H71
• R1

Disclosure statement

The authors have no financial arrangements that might give rise to conflicts of interest with respect to the research reported in this paper.

Notes

¹ “Incentive Programs Currently Offered in Each State, by Number of Programs” downloaded from www.stateincentives.org on August 9, 2021.

² For articles that review the empirical literature on incentive efficacy, see Peters and Fisher (2004), Busso, Gregory, and Kline (2013), Bartik (2019), Giroud and Rauh (2019), Mast (2020), Slattery (2020), and Donegan, Lester, and Lowe (2021).

³ There is a large literature on why states offer incentives and the role of yardstick competition and political economy motives. See for instance, Calcagno and Hefner (2018). Research typically focuses on program implementation. For instance, Man (1999) and Byrne (2005) find evidence of strategic interaction in municipal tax increment financing adoption decisions in Indiana cities and the Chicago metro area, respectively.

⁴ Wang (2018) uses this data for robustness tests in her analysis of strategic interactions of state economic development incentives, which our analysis builds upon.

⁵ Slattery (2020) shows that on average, budgetary spending accounts for roughly 62% of total incentive spending for the 2007–2014 period, which overlaps our sample period (2008–2018).

⁶ β-convergence is a necessary but not sufficient condition for $\sigma$-convergence (Sala-i-Martin 1996).

⁷ PS recommends setting r = 0.3 for small and moderate samples (T ≤ 50).

⁸ Phillips and Sul (2009) recommend setting c = 0 for small samples (T ≤ 50) to ensure that it is highly conservative.

⁹ For example, only tax expenditures of $5 million or more are reported in California (Boddupalli, Sammartino, and Toder 2020). Among states that do not report yearly are Iowa, Washington, Kentucky, and Oregon, which report every 5 years, 4 years, 2 years, and 2 years, respectively (Boddupalli, Sammartino, and Toder 2020).

¹⁰ C2ER has some data for 2007 but it is incomplete. It is currently undergoing backdating of 2007 records to ensure consistency and accuracy. Thus, our analysis begins in 2008, when their most comprehensive data coverage starts.

¹¹ The database is available at https://www.goodjobsfirst.org/subsidy-tracker.

¹² Please refer to Wang (2016, 2018) for more details.

¹³ Excluded states are Alaska, Arkansas, Delaware, Hawaii, Idaho, Kansas, Maine, Mississippi, Montana, New Hampshire, North Dakota, Oklahoma, Rhode Island, South Dakota, Utah, Vermont, West Virginia, and Wyoming.

¹⁴ She collected the data from 2000–2018 and at the time of writing this article, has compiled comprehensive coverage from 2007 through 2014. See Slattery (2020) for a detailed discussion of her collection method.

¹⁵ We also investigated estimates omitting the three support categories (program support, administration, and other program areas) because these do not function like direct incentive expenditures. We found similar evidence of club convergence.

¹⁶ The range of relative transition paths in the high-spending club got a bit smaller over time but did not narrow conspicuously as in the other clubs. In the robustness check of using an alternative economic development spending measure, relative transition paths in the high-spending club displayed a more apparent decrease in dispersion in Figure 5.

¹⁷ A critical assumption of the multinomial logit model is the independence of irrelative alternatives (IIA), which requires that alternatives facing an individual must be sufficiently different from each other. In particularly, we exclude the three clubs one at a time and compare the coefficients from the subsamples with those from the full sample. In each case, we fail to reject the null hypothesis that the coefficients do not change, which suggests that IIA is not violated, and the use of multinomial logit is appropriate. We also considered ordered logit model and multinomial probit model, and our results stay qualitatively similar.

¹⁸ Demographic characteristics such as the fractions of elder and young population are excluded due to their statistically insignificant impacts. When demographic variables are included, coefficient estimates for other variables remain qualitatively similar.

¹⁹ We also considered a few other political variables including governor party affiliation, party control of state legislature, and state control (whether the governor and state legislature belong to the same party or not). However, none of these variables are found to have a statistically significant impact on club classification. We drop them in our main results presentation. Note that our results do not change qualitatively when these variables are included.

²⁰ https://fred.stlouisfed.org/release/tables?rid=330&eid=391444&od=2008-01-01#.

²¹ https://www.taxpolicycenter.org/statistics/state-and-local-direct-general-expenditures-capita; https://www.ncsl.org/research/labor-and-employment/right-to-work-laws-and-bills.aspx.

²² Alternatively, β-convergence can be tested by relating the growth rate of a variable (over the entire sample period) to its initial value. In our case, results are qualitatively similar regardless of which method we use.

²³ The coefficient estimate is marginally significant for club 1, at 12% level.

²⁴ Coefficient of variation (CV) has also been used in testing $\sigma$-convergence. Both measures produce similar results for our sample.