Tax policy endogeneity: evidence from R&D tax credits

ABSTRACT

This paper estimates the causal effect of research and development (R&D) tax incentives on R&D expenditures using new data on U.S. states. Identifying tax variation comes from changes in federal corporate tax laws that heterogeneously and, due to the simultaneity of state and federal corporate taxes, automatically affect state-level tax laws. Instrumental variables regressions indicate that a 1% increase in R&D tax incentives causes a statistically significant 2.8–3.8% increase in R&D. Alternatively, ordinary least squares (OLS) regressions of R&D expenditures on R&D tax incentives, which do not correct for the policy endogeneity of R&D tax incentives, indicate that a 1% increase in R&D tax incentives causes a statistically insignificant 0.4–0.7% increase in R&D. One possible explanation for these results is that tax policies are implemented before an economic downturn.

KEYWORDS:

• Corporate tax
• fiscal policy
• R&D price elasticity
• research and development
• tax credits
• policy endogeneity

JEL CLASSIFICATION:

• H20
• H25
• H32
• H71
• K34
• O38

Acknowledgments

The views expressed in this paper are mine and are not necessarily those of the university or the Board of Governors of the Federal Reserve System. I thank anonymous referees, Ryan Baranowski, Marianne P. Bitler, David Brownstone, Christopher S. Carpenter, Linda R. Cohen, Kenneth A. Couch, Theodore F. Figinski, Christopher Karlsten, Bree J. Lang, Sarah B. Lawsky, David Licata, David Neumark, George C. Saioc, Manisha Shah, and seminar participants at the All California Labor Conference, APPAM, Bates White, Cal Poly-SLO, CFPB, CLEA, DOJ, FDIC, FRB, Georgia Tech, HKU, NTAs, NTU, Oregon State, RAND, RPI, UB, UBC, UC-Irvine, and WEAI for valuable comments. I also thank Amanda G. Bauer and Anthony Marcozzi for valuable research assistance, Ellen Augustiniak and Jenny Chang for aid in interpreting tax laws, Jonah B. Gelbach and Jennifer Graves for assistance with bootstrapping, Daniel J. Wilson for information on tax data, and Nirmala Kannankutty and Raymond Wolfe for support with National Science Foundation data. I am responsible for any errors.

Disclosure statement

No potential conflict of interest was reported by the authors.

ORCID

Andrew C. Chang http://orcid.org/0000-0002-9769-789X

Notes

1 For non-U.S. data, work by Bloom, Griffith, and Van Reenen (2002), Falk (2006), Lokshin and Mohnen (2012), Crespi et al. (2016), and Thomson (Forthcoming), among others, also find relatively inelastic estimates, while several studies on U.K. data by Fowkes, Sousa, and Duncan (2015), Guceri and Liu (2015) and Dechezleprêtre et al. (2016) find larger effects. See also Cerulli (2010) and Zúñiga-Vicente et al. (2014) for a more recent reviews.

2 An example of this type of preemptive tax incentive behavior is from December 2016, when then U.S. President-elect Trump and then U.S. Vice President-elect Pence offered Carrier tax incentives to prevent Carrier from moving some of its operations from the U.S. to Mexico.

3 To maximize the public benefit of these data, I will make my data and code freely available upon publication acceptance of this article, whether or not the journal that accepts this paper requires authors to share their data and code, following the recommendations of Chang and Li (2015, 2017). My data and code will also allow other researchers to verify my results, which will boost this paper's scientific integrity and contribution.

4 This model is analogous to the panel data models of Bloom, Griffith, and Van Reenen (2002) and Wilson (2009).

5 R&D data are available since 1963, but I focus on the period since the introduction of the federal R&D tax credit, following previous studies of state R&D tax incentives (Paff 2005; Wu 2005; Wilson 2009). The introduction of the federal R&D tax credit in 1981 created strong incentives for firms to relabel expenditures as R&D and creates a potential measurement break between the pre-credit era and the post-credit era (Eisner, Albert, and Sullivan 1986; Hall and Van Reenen 2000). While subsequent revisions to the tax code that increased the generosity of the federal R&D tax credit could have strengthened the relabeling incentive, starting in 1981 firms already had the incentive to relabel their expenditures as R&D.

6 For the period from 2000 to 2006 the NSF provides imputed observations of R&D for states that are not in the data for the 1980s and 1990s. For this paper, I cannot use the states that appear in the sample after 2000 because the variation I use for identification is in the 1980s and 1990s. The 21 states in my sample are: Alabama, Arizona, California, Colorado, Connecticut, Florida, Illinois, Indiana, Maryland, Massachusetts, Michigan, Minnesota, New Jersey, New York, North Carolina, Ohio, Oregon, Pennsylvania, Texas, Virginia, and Wisconsin. This sample of high-R&D states makes up 80–90% of R&D after 2000.

7 See Barlevy (2007), Ouyang (2011), or Chang (2013) for research into macroeconomic determinants of R&D.

8 There is a large literature debating whether public funds complement or substitute for private funds. See David, Hall, and Toole (2000) for a review.

9 See the review in Brown, Plewes, and Gerstein (2005) for details on the differences between these two sources of data. The results report estimates using obligation data to maximize the sample size. The results are insensitive to both measurements of federal R&D expenditures.

10 The raw data for most of the variables are non-stationary. However, the time dummies and state fixed effects detrend all of the variables (Cameron and Trivedi 2005). Panel unit root tests (Said and Dickey 1984; Levin, Lin, and Chu 2002) on the detrended variables support stationarity for all variables.

11 As a robustness check, I also attempt to correct for potential Nickell bias with both the one-step and two-step Blundell and Bond (1998) generalized method of moments estimators, transforming the instrumenting equation using the orthogonal deviations transformation (Arellano and Bover 1995) the three bias-corrections of the bias-corrected least squares (LSDVC) estimators of Bruno (2005a,b). Unfortunately, both the Blundell and Bond (1998) and Bruno (2005a,b) LSDVC estimators generate imprecise estimates.

12 Firms above their minimum taxable income amount can reduce their tax liability by increasing R&D because R&D is fully deductible.

13 I model the representative firm because the NSF's R&D data are at the state level.

14 Taking into account the discounted sum of future changes in taxes is necessary because R&D tax credits are occasionally calculated as a credit amount over an M-year moving-average base of previous R&D expenditures. This calculation implies that taking R&D tax credits in period t can affect the ability of a firm to take a credit in future periods. The model takes into account future changes in taxes only when they would be affected by a moving-average base, which is at most four years into the future.

15 The theoretical rationale behind discounting future periods with the S&P 500 is the opportunity cost of a firm’s funds. A firm deciding to undertake R&D could instead fund some outside investment, with the S&P being a representative indicator of the available market rate of return.

16 Equation (2) discounts changes in the tax liability of future periods using the actual realized interest rate. The assumption behind this formulation is firms correctly anticipate the interest rate with certainty and follows Wilson (2009). As a robustness check, I also discount future periods by assuming that firms in period t use the interest rate from period t – 1 to form future expectations of the interest rate. This alternative formulation gives similar results.

17 The tax rates described by tax laws are called statutory rates.

18 At the end of my sample in 2006, the average effective state R&D tax incentive is worth about one-half of the federal R&D tax incentive. Therefore, firms have a strong incentive to take into account state-level R&D tax incentives.

19 Examples of studies that research how state characteristics affect taxes include: Berry and Berry (1992, 1994) for electoral cycles, Stratmann (1992, 1995) for strategic coalitions among politicians, Poterba (1994) for balanced budget rules, Crain and Muris (1995) and Gilligan and Matsusaka (2001) for legislative structure, Swank and Steinmo (2002) for unemployment and capital mobility, and Aidt and Jensen (2012) for fiscal spending pressure and tax collection costs.

20 By extension, trying to infer policy endogeneity based on correlations between observable variables and the policy is not a meaningful exercise.

21 This practice is also called logrolling.

22 Gruber and Saez (2002) isolate exogenous changes in personal income tax rates arising from variation in tax laws at time t by conditioning on the previous period's income. Their exogenous changes in personal income tax rates reflect policy decisions at a higher level (federal government) than the unit of observation (individual). I take the analogous approach and create exogenous R&D tax incentives from variation in federal tax laws at time t by conditioning on the previous period's state tax laws. My exogenous changes also reflect law changes at a higher level (country) than the unit of observation (state). The general identification strategy of using federal laws for policy variation across states has been used in other ways, such as analyzing minimum wages (Card 1992).

23 Section 5.2 also presents Difference-in-Sargan overidentification validity tests following a format similar to that in Weber (2014). With these overidentification tests, I am unable to reject the validity of my instrument.

24 Examples of these preambles are available on my website.

25 With state fixed effects and time dummies, identifying variation comes from mean deviations in R&D tax incentive rates, not from large shifts that affect all states equally. The robustness checks appendix confirms that the main results are not sensitive to the large increase in rates from the introduction of the federal R&D tax credit in 1981.

26 Attenuated estimates with improved precision when including the lagged dependent variable are consistent with Bloom, Griffith, and Van Reenen (2002)'s cross-country study of R&D tax credits.

27 This group consists of: Arizona, California, Connecticut, Indiana, Massachusetts, Minnesota, New Jersey, New York, Pennsylvania, and Wisconsin.

28 This group consists of: California, Colorado, Connecticut, Indiana, Minnesota, New York, North Carolina, Pennsylvania, Oregon, and Wisconsin.

29 Researchers may be concerned that firms anticipated PL 101-239. However, anticipation of PL 101-239 would bias the elasticity estimates toward zero. In 1989, the federal R&D tax credit was a credit amount for R&D over a three-year moving-average base of R&D. The moving-average base created a disincentive for firms to claim the R&D tax credit as taking a credit in a given year would reduce the allowable credit for the next three years. PL 101-239 removed the moving-average base amount and the opportunity cost of claiming the R&D tax credit. If firms anticipated this policy change in 1989, then more firms would have claimed the R&D tax credit in 1989, perhaps at the expense of R&D they would have claimed in 1990, which would bias the estimate of the effect of PL 101-239 in 1990 toward zero.

30 These results are also consistent with the growth theoretical results of Yang (2005) and the empirical results on US GDP by Romer and Romer (2010). Yang (2005) is a theory paper that simulates growth models. The paper shows that calibrated models that omit preemptive tax policies are misspecified. Romer and Romer (2010) use narrative information on federal taxes to separate endogenously determined taxes from exogenously determined taxes to identify the effect of tax changes on gross domestic product (GDP). With vector autoregressions, Romer and Romer (2010) find the endogenous tax variation leads to underestimates of the effect of taxes on GDP.

31 The federal government has allowed these deductions since prior to the beginning of the R&D data from the National Science Foundation.

32 Hall (1993) notes that the majority of R&D firms have R&D levels above their base amounts. Mamuneas and Nadiri (1996) and Wilson (2009) also employ the assumption of R&D levels over the base amounts.

33 In the R&D sample, Connecticut and Maryland are exceptions. Connecticut has had two R&D credits since 1993: a 20% credit for QREs over a one-year moving average (Connecticut General Statutes § 12-217j) and a level credit for QREs below the moving average (Connecticut General Statutes § 12-217n). The level credit is tiered at 1%, 2%, 4%, and 6% based on the firm's level of QREs. In addition, the firm may take only one-third of the level credit in the tax year that it incurs the R&D expenditures. The remainder must be deferred until the next tax period. Transitional provisions were in place from 1993–1994. Like Connecticut, Maryland has two R&D credits that work in tandem and have been in place since 2000 (Maryland Tax-General Code § 10-721). The first component is a 10% credit for QREs above a four-year moving average of QREs. The second component is a 3% credit for QREs that do not qualify for the 10% credit component. I model both of these alternative mechanisms.

34 Some states impose a maximum credit amount a firm can claim that is not dependent on the firm's taxable income, a statewide limit on the amount of R&D tax credit that can be claimed by all firms in the state each year, or both a firm-specific maximum and a statewide maximum. The firm-specific limit on R&D tax credits is equivalent to a marginal rate of zero for the top tier. I assume that the statewide limit provision is not binding, following Wu (2005); Wilson (2009).

35 Equation (2(2) ) assumes firms have sufficient taxable income to claim R&D tax incentives, consistent with previous studies of R&D tax incentives. A dummy variable for whether a state has a refundable R&D tax credit (tax credits that can be claimed for any level of taxable income) or allows firms to sell tax credits to other firms has no effect on the results.

36 The results are also robust to adding state-specific linear time trends, the rate of growth of GSP, and the first lag of the rate of growth of GSP as controls.

37 The endogenous R&D tax incentive rate driven by both state and federal laws gives inelastic to approximately unit elastic point estimates for all robustness checks.

38 A 5% sample trim (2.5% from each tail) yields similar estimates.

39 Weighting states by average GSP from 1981–2006 also gives similar results.

40 One control I do not consider is some type of geographically-weighted or proximity measure of an out-of-state subsidy rate. For example, adding the weighted R&D tax credit subsidies of Arizona, Nevada, and Oregon as a control observation for California. Geographically-weighted measures most likely ignore or mis-attribute R&D reallocation, particularly within firms situated in multiple states, which would lead to measurement error in a right-hand side variable. For example, the aircraft producer Boeing has manufacturing plants in Washington state and South Carolina. These states are on opposite sides of the U.S., but for purposes of R&D allocation Boeing may want to conduct R&D between these states due to existing infrastructure and human capital while also taking into account R&D tax incentives. However, geographic proximity measures will miss this link. My future work will consider modeling R&D mobility across states. This paper focuses on the within-state response.

41 Estimating separate policy variables and separate controls for each census region gives imprecise estimates.

42 The clustered standard errors imply rejection at the 5% level or lower for the key coefficient in the preferred models. I also check the rejection rates, following the recommendation of Cameron, Gelbach, and Miller (2008), by bootstrapping the t-statistic using the wild cluster bootstrap-t procedure (Brownstone and Valletta 2001). I use Rademacher weights with 1000 replications for each test and impose the null hypothesis that the tax policy variable is zero, as advocated by Davidson and MacKinnon (1999) and Cameron, Gelbach, and Miller (2008). The bootstrap blocks are states. The hypothesis test of vs. yields p-values between 0.09 and 0.10 for the preferred model's key regressors.HA : ɣ > 0

43 Specifically, I compute OutputTaxIncentiveRate with the model in Appendix 1 without the terms for R&D-specific tax incentives.

44 Calculating OutputTaxIncentiveRate by isolating only state-level tax variation from federal tax laws in the cost of output with the analogous definition from Equation (3(3) ) also gives similar results.

45 These two classes are themselves interdependent, but I separate them for the sake of exposition. See the model in Appendix 1.

46 The presence of R&D tax credits and state deductions for state corporate income taxes complicates the intuition, but the main point is the same.

47 See Guenther (2006) for a review of the federal R&D tax credit.

48 Treating tax credits as taxable income is called credit recapture.

49 In the interest of brevity I simplified this discussion slightly. Some states have specific provisions that override what the base would predict. See Appendix 1 for details. A detailed example of how a federal tax law passes through to state tax law is available on my website.

Additional information

Funding

This research was supported by a grant from the Department of Economics at the University of California - Irvine.