Does Corruption Lead to Lower Subnational Credit Ratings? Fiscal Dependence, Market Reputation, and the Cost of Debt

Abstract While studies show a consistent negative relationship between the level of corruption and range indicators of national-level economic performance, including sovereign credit ratings, we know less about the relationship between corruption and subnational credit ratings. This study suggests that federal transfers allow states with higher levels of corruption to retain good credit ratings, despite the negative economic implications of corruption more broadly, which also allows them to continue to borrow at low costs. Using data on corruption conviction in US states and credit ratings between 2001 and 2015, we show that corruption does not directly reduce credit ratings on average. We find, however, heterogeneous effects, in that there is a negative effect of corruption on credit ratings only in states that have a comparatively low level of fiscal dependence on federal transfers. This suggest that while less dependent states are punished by international assessors when seen as more corrupt, corruption does not affect the ratings of states with higher levels of fiscal dependence on federal revenue.


Introduction
During the last few decades, international organizations, policy makers, and experts have expressed an increased concern about the negative economic effect of corruption. While early studies seem to indicate that corruption may help business navigate red tape and thereby improve economic performance, 1 the balance of evidence to date suggests that the negative economic consequences outweigh any potential benefits. 2 Corruption reduces government revenue from taxable sources, 3 increases government expenditures by reducing the productivity of government spending, 4 decreases the rate of growth, 5 and increases public deficits 6 and public debt. 7 International credit rating agencies, such as Standard & Poor's and Moody's, have also been sensitive to the fact that "institutions matter" for economic performance, 8 and some agencies even incorporate widely used cross-country comparative measures of corruption in their sovereign credit rating indices. 9 Studies suggest that corrupt countries receive lower credit ratings, 10 which increases the cost of borrowing. With few notable exceptions, 11 however, studies on the link between corruption and credit ratings have been at the national level. Thus, despite the growing levels and importance of subnational borrowing, 12 in addition to well documented subnational variation in corruption and quality of institutions, 13 we know comparatively little about the effect of corruption on subnational credit ratings. This paper suggests that corruption does not reduce credit ratings to an equal extent across all US states, and, therefore, that the link between corruption and credit ratings is far from as straightforward as previous studies suggest. Specifically, we suggest that the credit ratings of states that receive higher levels of federal fiscal transfers, are largely unaffected by higher levels of corruption. Fiscal dependence on the federal government thereby seems to provide the kind of debt repayment guarantee that states need in order to maintain a good credit reputation, despite the fact that political corruption fuels the misallocation of government funds and tends to lead to lower ratings at the national level. If credit rating agencies perceive fiscal dependence as a guarantee of debt repayment, 14 and therefore refrain to punish the fiscal excesses of corrupt states, corrupt states are allowed to continue to borrow on favorable terms.
Using data on corruption convictions in US states and bond ratings between 2001 and 2015, we show that corruption does not directly reduce subnational credit ratings on average. The effects of corruption on subnational credit ratings are, however, heterogeneous, and vary depending on a state's level of fiscal dependence on the central government. Empirically, we find a negative effect of corruption on credit ratings only in states that receive a comparatively low level of federal transfers.
We thereby seek to make several contributions to the literature on corruption and credit ratings. First, our focus on the subnational level offers several advantages. As noted by studies on subnational debt, the subnational level is important because of its growing level of borrowing, 15 but also because this level of government is often closely involved in public service delivery. Cutting expenditures often has important consequences in the form of abandoned infrastructure projects, or layoffs of social workers, teachers, and police officers. Despite this, we know comparatively little on the link between corruption and credit ratings at the subnational level. Our sample also provides ample variation on our key variables both across US states, as well as within them over time. More importantly, the subnational analysis might provide a more valid comparison of the dynamics at play across units, as many unobserved cultural and institutional factors are "naturally controlled for" when comparing units within countries. 16 Moreover, our focus also allows us to use objective data on corruption levels rather than measures based on expert perceptions.
Second, to the best of our knowledge, this is the first study that investigates how the overall level of fiscal dependence affects the creditworthiness of corrupt subnational units. We thereby seek to add to the rich literature on fiscal dependence and decentralization. While the debate on whether external aid may or may not produce better economic performance among corrupt governments is ongoing, 17 the evidence on the beneficial effects of fiscal dependence and decentralization is decidedly mixed. 18 An important body of work suggests that when central governments commit to substantial co-financing, they are often both constitutionally and politically constrained from ignoring the fiscal problems of subnational units. 19 In other words, fiscal dependence may lead to bailout expectations and reduced risk of standalone default, at least under some circumstances. 20 While fiscal guarantees may not be expected in all situations, we suggest that maintaining a high level of fiscal transfers despite higher levels of corruption or economic mismanagement can send a strong signal to rating agencies about 12 Ahrend et al., 2013. 13 Charron et al., 2014. 14 Rodden, 2006;Von Hagen and Eichengreen, 1996. 15 Ahrend et al., 2013. 16 Snyder, 2001 Svensson, 1999;Asongu and Jellal, 2013. 18 e.g., Rodden, 2002;Rodríguez-Pose and Krøijer, 2009;Martinez-Vazquez et al., 2017;Fisman and Gatti, 2002. 19 Rodden, 2002. 20 Hanniman (2018) suggests that only stable and predictable payments support local creditworthiness. Some of these related studies do not use credit ratings as their dependent variable but interest rates. the central government's commitment to act as a guarantor of timely debt repayment. This allows corrupt states to maintain favorable ratings and continue to borrow at low costs.
Corruption, credit ratings, and fiscal transfers Credit ratings have an important impact on government's ability to finance its budget. Good credit ratings allow governments to borrow money on more favorable terms. 21 Credit ratings provide an external expectation on a state's perceived ability and willingness to repay its debts, which best represents our concept of creditworthiness. In particular, at the subnational level where information may be less available, bond credit ratings provide domestic and international investors key additional data used to assess investment strategies. Although they have received ample international critique from a host of critics and scholars, 22 these ratings remain highly salient to a state's reputation, and changes in ratings often warrant "front page news." 23 Rating agencies exercise a unique form of market based authority by assigning ratings that determine the creditworthiness of debt issuers, including sovereign governments, and the default risk associated with their bonds. 24 They have been described as a market actor that combines "the normative market authority and the moral authority of the non-state, non-self interested referee," exercising significant authority in the world economy. 25 A great amount of scholarly attention has therefore been devoted to analyzing the determinants of credit ratings, and not least political determinants of credit ratings. 26 The literature on the effect of the quality of institutions, e.g., factors such as corruption or rule of law, have provided evidence that better functioning institutions are strongly associated with higher credit ratings from international assessors. 27 Scholars have noted that this relationship is due to the decrease in perceived risk of default among countries with better functioning institutions-leaders operating a strong rule of law system with lower corruption are more constrained from malfeasance, which in turn makes their commitments more credible. 28 Moreover, the cost of publicly funded projects increases if civic servants receive graft, thus increasing debt and reducing a state's capacity to repay. Corrupt governments tend to direct public resources toward sectors that provide the greatest opportunities for rent seeking, such as large capital investments or infrastructure, 29 rather than the ones that provide the greatest returns for the public good. Corruption also leads to tax evasion, 30 which further undermines fiscal capacity. Several studies have found that corruption increases investors' uncertainty, and thus leads to higher borrowing costs. 31 At the country level, the negative relationship between corruption and credit ratings has been consistent across multiple samples and model specifications. 32 Thus, our data and sample of US states presents an empirical puzzle, which is elucidated in figure 1. On the left side, we plot a measure of (control of) corruption risk on the x-axis (from International Country Risk Guide [ICRG]), and the Standard & Poor's sovereign credit rating on the y-axis. As per the previous literature, we observe a strong and positive relationship between these two variables, with the corruption measure explaining 75 percent of the variation in cross-national credit ratings. On the other hand, using data from the same year (2014), we find that our measure of corruption is essentially orthogonal to state-level credit ratings; 21 Kerwer, 2005;Sinclair, 2008. 22 e.g., Mennillo and Sinclair, 2019. 23 Abdelal and Blyth, 2015. 24 Sinclair, 2008. 25 Cutler, Haufler and Porter, 1999;Hall and Biersteker, 2002. 26 Cantor and Packer, 1996;Sinclair, 2008;Bruner and Abdelal, 2005;Archer et al., 2007;Biglaiser and Staats, 2012;Barta and Johnston, 2018;Barta and Makszin, 2020. 27 Mellios and Paget-Blanc, 2006. 28 Butler et al., 2009. 29 Tanzi and Davoodi, 2002;Kaufmann, 2001;Liu and Mikesel, 2014. 30 e.g., Richey, 2010;Matsaganis et al., 2012;Litina and Palivos, 2016;Baum et al., 2017. only explaining 5 percent of the variation (p = 0.00). Our central argument in addressing this puzzle is that the effect of corruption on credit ratings is heterogeneous-conditional on a state's fiscal dependence to the central government.
Thus, the relationship between corruption and credit ratings may not be as straightforward and direct as previous studies suggest. Although fewer studies investigate the effect of corruption on subnational credit ratings directly, a number of studies point to the fact that corruption may undermine credit ratings also at the subnational or municipal level. 33 In the United States, both Depken and Lafountain (2006) and Butler et al. (2009) show that corruption is associated with lower subnational credit ratings.  Bastida et al., 2015 suggests that only stable and predictable payments support local creditworthiness. Some of these related studies do not use credit ratings as their dependent variable but interest rates." However, despite the fact that claims related to the negative effect of corruption on a range of economic variables, including credit ratings, enjoy considerable empirical support and theoretical clout, these results do not necessarily hold across time and space at the subnational level. Building on the vast literature on fiscal dependence and decentralization, we suggest that one factor that may contribute to maintaining good credit ratings among corrupt subnational units is the relative level of fiscal transfers and fiscal dependence. A broad body of studies suggest that fiscal dependency fuels bailout expectations, and that government that co-finances the budget of subnational units will signal both ability and willingness to repay debts. 34 Rodden (2002, 670) highlights this key dilemma of fiscal federalism and contends that when "constitutionally or politically constrained central governments take on heavy co-financing obligations, they often cannot credibly commit to ignore the fiscal problems of lower level government." Moreover, Rodden (2006) notes explicitly that international ratings agencies grade lesser developed regions more favorably when the federal government has an established equalization system of transfers across subnational units, in that "they provide a safety net of varying importance during difficult times." 35 Other studies, similarly argue that fiscal dependence may create bailout expectations and implicitly guarantee debt repayment. This compensation effect is driven by federal governments' wish to maintain their own ratings as well as avoid being punished by the electorate for failures in local public service delivery. 36 Subnational creditworthiness becomes less about the solvency of the subnational borrower and more about the creditworthiness of the implicit guarantor-the federal government in this case. 37 Thus, we suggest that a higher degree of federal dependence serves as a buffer mechanism against the negative effect of corruption on subnational credit ratings. In turn, the effect of corruption is anticipated to be heterogeneous-dependent on the level of federal dependence-not direct, as suggested by previous studies.
A simple two-state comparison with similar, high-profile corruption cases provides an illustrative example of our main point. On the one hand, the state of Illinois is, and has been, one of least dependent states on federal transfers, ranking usually in the bottom third of all states in terms of budget dependence annually. In 2008, former governor Rod Blagojevich was implicated in a corruption scandal where he attempted the sale of then Senator Obama's senate seat, along with extortion and lying to federal agents, which resulted in a conviction and prison sentence for Blagojevich. The corruption conviction in this case led to a near immediate downgrade from Standard & Poor's in the following year of 2009. On the other hand, Alabama is a state that is highly dependent on annual federal transfers, usually in ranking among the top 10 percent of states annually. However, a 2006 conviction of former governor Don Siegelman on several counts of bribery and obstruction led to no subsequent change in Alabama's credit rating.
This leads to the following two hypotheses: H1. Corruption does not have a direct effect on subnational credit ratings in the United States.
H2. Corruption has a negative effect on credit ratings only in states with low levels of fiscal dependence.

Sample and data
Our empirical hypotheses are tested via a subnational analysis of US states from 2001-15. This sample offers several key advantages to test the hypotheses put forth in this study. First, the subnational analysis allows us to "naturally control for" unobserved cultural and institutional factors, which can lead to a more valid comparison. 38 Second, as we show in this section, there is ample variation on our key variables both across states, as well as within with them. Third, with some notable exceptions 39 34 Rodden 2002Rodden , 2006Rodden, Eskeland, and Litvack, 2003;Ahrend, 2012;Wildasin, 2004;Bordignon and Turati, 2009;Vigneault, 2010;Escolano et al., 2012. 35  most of the literature on the relationship between institutional quality and credit ratings is at the country-level, our study provides a clear empirical compliment.

The dependent variable
The outcome variable of this study is bond credit rating of US states between 2001 and 2015. Rather than using actual debt or default or interest rates, the credit rating represents the expert, external expectation of a state's ability and willingness to repay its debt, which best represents our concept of creditworthiness. In particular, at the subnational level, where information may be less available, bond credit ratings provide domestic and international investors key additional data used to assess investment strategies. Although ratings have not been without controversy 40 these ratings provide valuable information to the credit market 41 and remain highly salient to a state's reputation. 42 While all of the "big three" agencies (Standard & Poor, Moody's, and Fitch) rate some states in some years, only Standard & Poor provide data for all states over a significant time period annually, and thus we elect to use their measure to maximize our sample coverage. 43 The corresponding ratings among the three agencies for available states is remarkably high, however. 44 The Standard & Poor ratings are on a possible 23-point scale, with the coveted "AAA" rating being the max value (representing the highest assessment of creditworthiness) and "SD" being the lowest (the lowest assessment of creditworthiness), with higher scores resulting in a lower interest rate on debt. Similar to previous studies, we transform the ratings into an ordinal scale. As most state-years are at least a grade of "A" or above (save Illinois and California for two and three years, respectively), the "effective" range of our sample is a 6-point scale, with higher values equating to high levels of creditworthiness. Thus, while US municipalities may default on their debt, it is important to note that US states have not historically defaulted.
Although the grand mean of our outcome variable is relatively stable over time in our sample, we do observe temporal changes in ratings in 36 of the 50 states (72 percent) and the cross-sectional variation is noteworthy, with some states achieving and maintaining top ratings, while others are rated significantly lower. Figure 2 shows the average rating score by US state for our time period, with lighter shades equating to better average ratings:

Corruption
Corruption is defined most commonly as "the abuse of public office for private gain." 45 However, due to the clandestine nature of the act, it is all but impossible to measure directly, thus contemporary measures (whether "objective" or "subjective" ones) are indirect measures. The literature on corruption's effect on credit ratings varies depending largely on the level of analysis, mainly due to lack of data available across multiple levels of governance. For example, cross-country level studies mainly employ expert assessment measures, such as Transparency International's Corruption Perceptions Index or the World Governance Indices. 46 Yet for subnational level analysis, these expert perceptions measures are not available to the same degree as the country level. Moreover, such expert-based perceptions measures are highly problematic in explaining credit and bond ratings, as Panizza (2017, 27) (2019). While these two agencies use different methodologies, correlation between ratings from these two agencies is 0.98  sovereign raters such as Moody's and Fitch actually incorporate the World Governance Indicator of corruption in their ratings scores, thus such measures introduce endogeneity, per definition. And, for US states, such data is only available for a single year. 47 We therefore rely on an objective measure of corruption convictions commonly employed in subnational studies of US states. 48 Our measure is federal convictions for all federal, state, and local public officials for each state-year from 2001-15 as reported by Public Integrity Section (PIS), who define corruption similar to the literature as, "crimes involving abuses of the public trust by government officials." These convictions include various forms of corruption including accepting bribes, awarding government contracts to vendors without competitive biddings, fraud or campaign-finance violations, and obstruction of justice. For purposes of comparability, we take the per capita (100,000 inhabitants) number of PIS reported convictions. We find considerable variation across states in this measure, with South Dakota and Louisiana reporting roughly 1 corruption conviction per 100,000 inhabitants, while Utah and Oregon have just 0.10, and New Hampshire just 0.05 convictions per 100,000 residents. We find, however, that observations on the high end of this measure constitute significant leverage outliers (> 3 standard deviations above the variable mean). Thus, to deal with the skewed nature of the distribution of the variable and to avoid misleading findings from such outlying observations, we transform this measure to the cube root, as previous studies have done to adjust for this issue. 49 One might express concern over the validity of an objective proxy of corruption, such as convictions, which could simply be measuring rule of law enforcement or differences in media oversight. However, Liu et al. (2017) show that such data is not statistically associated with state-level capacity factors, such as caseloads, number of judges, or state judiciary expenditures, thus alleviating some of these concerns. 50 Some scholars argue that using federal convictions for subnational data may be problematic because nonfederal officials commit different crimes than federal ones. 51 However, using federal convictions ensures that the laws under which convictions occur are uniform across different states, 52 and the measure is a consistent metric over time, as opposed to perceptions measures, which are far less certain in terms of temporal comparisons. Finally, as the literature on US state-level corruption uses this measure consistently, our use of convictions data allows for a more valid comparison of our results with previous studies.

Fiscal dependence
Our measure of fiscal dependence is the total amount of federal revenue to the state, most of which is in the form of grants, divided by the total amount of state revenue annually. Data are taken from the US Census Bureau's Annual Survey of State Government Finances and Annual Survey of State and Local Government Finances and organized by PEW research. 53 The measure aggregates all types of federal transfers that complement a state's budget in paying for health services, such as Medicaid, education and training programs, transportation, and infrastructure. The transfer system is largely progressive, with poorer states receiving a greater proportion of their budgets from federal revenues on average, yet the correlation between federal dependence and GDP per capita is -0.53, and thus economic development does not exclusively explain a state's level of dependence. The measure also captures the stable, annual equalizing payments that are argued to be more important for international creditworthiness of subnational actors. 54 The data show remarkable variation across states. 55 For example, in our latest year (2015), Virginia and Hawaii receive 22 and 23 percent of their budget respectively from federal revenues, while Louisiana and Mississippi rely on roughly double that percentage, over 42 percent, respectively. We also see significant changes over time within states. For example, the dependence on federal revenues in Arizona went from just 30 percent in 2008 to 46 percent in 2010, a much larger increase from the national average of 28 percent to 35 percent during the same period. The share of Kansas's federal revenues has dropped to under 25 percent since the 2009 recession, while in Kentucky; dependence on federal transfers increased to over 40 percent and have remained roughly so since.

Additional controls
As our study employs observational data, we add additional control variables to our model to reduce the possibility of drawing invalid inferences from a spurious relationship between our main variables of interest. We are particularly interested in factors that can confound our relationship between corruption and credit ratings, along with the conditional effects based on transfers. In the literature, such standard controls include gross public debt as percentage of GDP, GDP per capita, unemployment, total spending as % GDP, the population of a state. 56 Based on previous findings, we anticipate that all control variables, save GDP growth, GDP per capita, and the size of population, have a negative effect on the ability to debt repayment, thus we anticipate negative coefficients for credit ratings. Moreover, our models control for the amount of economic freedom in a state (from the Cato Institute 57 ), 50 However, Alt and Lassen (2012) show that greater prosecutor resources result in more convictions for corruption, though caseload itself may not influence the outcome of particular case. 51 Cordis and Milyo, 2016. 52 Depken and Lafountain, 2006. 53 https://www.pewtrusts.org/en/research-and-analysis/data-visualizations/2014/fiscal-50#ind1. 54 Hanniman, 2018. 55 In appendix G we provide more detailed discussion about federal transfers. We discuss the consistency of federal transfers over time. In addition, we provide bivariate relationship between federal transfers and credit ratings. 56 Cantor and Packer, 1996;Archer et al., 2007;Depken and Lafountain, 2006. 57 The variable is a composite index of a host of factors representing fiscal and regulatory freedom of a state, annually (https:// www.freedominthe50states.org/). We thank an anonymous reviewer at Business and Politics for this suggestion. which proxies for the overall business climate in a state. Summary statistics for all variables are found in the appendix A.

Diagnostics tests and estimation strategy
In the literature, credit ratings are modeled in several ways. First, some scholars elect to use either tobit estimation, 58 or ordered probit/logit 59 due to the limited, ordered nature of the measure. Others choose to model variation in credit ratings continuously, via linear regression. 60 While exact model predictions would be expected to vary based on the estimation choice, these approaches produce similar substantive effects irrespective of this decision, thus we present linear models for simplicity, with alterative estimations in appendix F.
As our data is time series cross-sectional (TSCS), we perform the standard diagnostic tests of problems commonly associated with such data. A Dickey-Fuller test for stationarity shows no unit root in the dependent variable, implying that the sample mean and variance do not significantly vary over time. However, we find the presence of positive serial correlation and heteroscedasticity in the data, and thus similar to most previous panel data studies, 61 we employ dynamic models with a lagged dependent variable and robust standard errors, clustered by states. To arrive at the proper lag structure for the dependent variable, we follow Hendry (1995) and "test down" (via cs) to assess how many lags to include, which in our case is one lag.
We account for the unit heterogeneity of states over time and omitted, time-invariant variables from the model in several ways to test the consistency of our findings. First, a Hausman test favors fixed effects models, which account for the within-state variation of the dependent and explanatory variables. Our fixed effects estimation to test H1 is thus the following: Where credit ratings in each state-year (CR it ) are explained by corruption (C) and a 'k' number of covariates, δ it , which includes a lagged dependent variable (CR it−1 ). 62 u i are the state-specific, timeinvariant effects, and ϵ it is our error term. The fixed effects transformation subtracts the within-unit average from each observation for each time point, and thus washes out unit specific effects, and also drops time invariant covariates.
However, as our effective time series is fifteen years, a somewhat short period to analyze relatively slow moving variables such as credit ratings and corruption, relying on solely within-state variation provides a "tougher test' and oftentimes less efficient estimates. 63 As the between-state variation in the dependent variable is more than twice that of the average within-state variation, we present partialpooled and pooled models with alternative estimators to incorporate more directly this between-state variation. Thus, second, we report random intercept models. However, it is well understood that the inclusion of a lagged dependent variable in a random effects model violates a key assumption of the model, in that the random intercept (μ i ), which represents the unobserved state-level, static variables, is assumed to be orthogonal with the independent variables on the right side of the model. Yet the inclusion of a lag of our outcome variable, credit ratings, on the right side (CR it−1 ) is necessarily correlated with μ i and will bias CR it−1 upward and estimates of other explanatory variables downward. 64 58 Depken and LaFountain, 2006. 59 Grizzle, 2010; Pérez-Balsalobre and Llano-Verduras, 2020. 60 Archer et al., 2007;Cantor and Packer, 1996. 61 See Biglaiser et al., 2008;Liu et al., 2017. 62 The inclusion of lagged dependent variable is a standard procedure in studies on credit ratings (e.g., Biglaiser and Staats, 2012;Shea and Solis, 2018). 63 Troeger, 2019. 64 Allison et al., 2017. To address this issue, we provide our random effects estimates based on a technique based on Anderson-Hsiao (1981), which is developed for dynamic panel models via the use of an instrumental (IV) estimator. In this case, we employ the second lag of the dependent variable (CR it−2 ), which is exogenous to μ i 65 when controlling for the first lag. The generalized, two-stage estimator (G2SLS) is particularly appropriate when the number of units is relatively large and the number of time units are relatively small, and no second-order autocorrelation, as in our case.
For example, our model specification to test H2 with this estimation is as follows: Where the following first stage estimates are: Where we run a first stage model in which the lag of our dependent variable (CR it−1 ) is instrumented with the second lag (CR it−2 ), along with the battery of "k" number of explanatory variables (θk it ) in the second stage model. In the second stage model, we then regress credit ratings CR it on its first year lag (modeled as endogenous), along with the main variables of interest (corruption and transfers), and to test H2, we include β 4 to test the interaction. δk it represents our battery of control variables, while μ i is the random intercept for each state and ϵ ij is our model error term. Since credit rating agencies can react with a rating changes at any time (they do not have to wait to the end of a fiscal year, for example) we model our explanatory variables as in the same time as the outcome variable, with the exception of growth, which we expect to have a time lagged effect.
Finally, we use Prais-Winsten models, adjusting for first order autocorrelation and panel corrected standard errors, 66 to provide further robustness checks of our FE and RE estimation.

Results
We begin with a set of models to test the first hypotheses in table 1-that of the direct effect of corruption on credit ratings. The first two models show the within estimates from the fixed effects models, while models 3 and 4 show the random effect generalized IV estimates. Model 5 presents the pooled Prais-Winsten (PW) estimates without a lagged dependent variable, adjusting for first order autocorrelation and includes state-clustered standard errors. We find, in fact, a consistent small, positive effect of state-level corruption on our outcome variable, yet in all models, the effect is statistically negligible. Overall, however, we find no evidence of a direct effect of corruption on state-level credit ratings, which supports our H1. However, this finding runs contrary to the earlier findings of Depken and Lafountain (2006). This discrepancy could be due to their analysis relying on a different set of control variables, an earlier time period (1995)(1996)(1997)(1998)(1999)(2000), as well as using a pooled linear estimation model, which does not account for structural heterogeneity across states, which we model directly. Another reason could be that rating agencies do not have fixed criteria of assessment over time 67 and the global financial crisis could play a significant role in shaping these new criteria.
In H2, we posit that corruption has heterogonous effects on our dependent variable, depending on levels of fiscal dependence. In table 2, we provide a similar set of estimation strategies in models 1-5 as per table 1.
The interaction term is the coefficient of interest, as this provides the test of whether the effect of corruption on credit ratings changes significantly as a function of fiscal dependence. In this case, we find quite consistent results across the models that adjust for different estimators and specification of control variables. The models reveal that the interaction is positive in all cases, implying that corruption's effect on our dependent variable in fact increases positively as a function of higher levels of fiscal 65 Baltagi, 2013. 66 Beck and Katz, 1995. dependence. The effect is robust to specification and control variables. In all models 1-4, the effect is statistically significant irrespective of specification. In the pooled Prais -Winsten estimation in model 5, the effect falls short of the 90 percent level of confidence (p = 0.11). Yet the Durbin-Watson statistic indicates there is even some positive autocorrelation present in the model post-adjustment, which Note: Dependent variable is subnational credit ratings with clustered standard errors by state are in parenthesis. All models include time (year) fixed effects. Models 1 and 2 are fixed effects (within) estimate, while models 3 and 4 use a generalized, 2-stage random effects estimation (REG2SLS), in which the lagged dependent variable is instrumented with its second order lag (first stage estimates shown only). Model 5 is a pooled Prais-Winsten model, adjusting for first order autocorrelation and includes state-clustered standard errors. Durbin-Watson statistics range from 0-4, with "2" indicating no autocorrelation, and show the original (orig) and the model transformed (adj) statistics. All models account for year effects in credit ratings. ***p < 0.01, **p < 0.05, *p < 0.1. Note: Dependent variable is subnational credit ratings with clustered standard errors by state are in parenthesis. All models include time (year) fixed effects. Models 1 and 2 are fixed effects (within) estimate, while models 3 and 4 use a generalized, 2-stage random effects estimation (REG2SLS), in which the lagged dependent variable is instrumented with its second order lag (first-stage estimates shown only). Model 5 is a pooled Prais-Winsten model, adjusting for first order autocorrelation and includes state-clustered standard errors. Durbin-Watson statistics range from 0-4, with "2" indicating no autocorrelation, and show the original (orig) and the model transformed (adj) statistics. All models account for year effects in credit ratings. ***p < 0.01, **p < 0.05, *p < 0.1.
affects the efficiency of the standard errors, and could explain the higher p-value. Models 2 and 4 show that the effects are affected by the control variables as magnitude of the integration effects decrease by a magnitude of 0.62 and 0.24, respectively, yet the effects remain significant and in the expected direction. Models 1 and 2 that provide the within-state estimates from the fixed effects models show the greatest magnitude in terms of this effect. The IV random effects and pooled PW models however show similar substantive effects. Yet as Angrist and Pischke (2008) point out, this discrepancy is common, and in these cases the "true" effect lies between the FE estimate (which is usually too high) and the pooled estimate (which is usually too low). Thus, overall, we interpret this set of results as providing empirical evidence for H2. With respect to our control variables, the coefficients are largely in the expected direction according to previous findings from the literature. We find that higher values of unemployment have a negative and significant effect on credit ratings, which is robust across all models. This is similar to previous studies at the US state level 68 and cross-country analyses. 69 As anticipated, we find that higher levels of economic freedom are associated with better ratings in most models. We find mixed support that debt is a significant predictor of subnational credit ratings, also consistent with the literature. 70 Yet the mixed findings across models could be due to various types of debt having heterogeneous effects on bond credit ratings in subnational governments. 71 Not surprisingly, due to the strong presence of first order autocorrelation, as credit ratings t−1 are a significant predictor of current ratings. Other factors, such as spending levels of sates, overall economic development, and population are statistically negligible.
Checks for nonlinearity As our interaction models in table 2 assume a constant, linear effect of corruption moderated by transfers, we check to see if that assumption holds, or whether the interaction is in fact non-linear. Our anticipation is that corruption has a negative effect on credit ratings at low levels of federal dependence, yet is negligible at higher levels.
We take a pragmatic approach to testing this via splitting our sample into four ranked quartiles according to a state-year's level of fiscal federal dependence throughout the sample. Quartile 1 represents those state-years with the lowest levels of dependence, while quartile 4 represents the highest dependence. We run both fixed and random (2-stage instrumental) models for each sub-sample, reported in table 3. As anticipated, we find that the effect of corruption on credit ratings is indeed nonlinear in that the significance levels only apply to low levels of transfers. We find a robust, negative effect of corruption on credit ratings for the bottom quartile (the least federally dependent) for both estimation approaches, while we observe positive, yet statistically negligible effects for the three subsequent quartiles. In addition to the split-sample approach we report here, we also employ a nonlinear check of our interaction via an approach developed by Hainmueller, Mummolo, and Xu (2019), found in appendix F. The estimates from their approach are essentially similar to those in table 2.
In substantive terms of the sample, what we observe is that among states with low levels of dependence (in the bottom quartile), we find that states such as Illinois, Massachusetts, and New Jersey have multiple years of higher than average corruption and lower than average credit ratings. While conversely, low dependence sates such as Colorado, Nevada, North Dakota, and Minnesota have low corruption and high credit ratings. Such patterns at this low level of federal dependence elucidate the negative and significant conditional relationship between corruption and credit ratings.

Further robustness checks
In addition to checking for nonlinearity of our findings for H2, we explore further checks of robustness (see appendix D). We check for several alternative specifications. First, we check whether our results 68 Depken and Lafountain, 2006;Liu et al., 2017. 69 Afonso et al., 2011. 70 Cantor and Packer, 1996; Depken and Lafountain, 2006;Grizzle, 2010. 71 Pérez-Balsalobre and Llano-Verduras, 2020. for table 2 hold if we include a measure of partisanship. 72 Namely, we include the partisan affiliation of the state's governor, the president's party affiliation, and whether the two are aligned. Similar to our theory regarding transfers, we would expect that partisan alignment would signal to external creditors that a state is more likely to receive financial backing if they are politically aligned with the federal executive. Table A4 (in appendix D) shows that the main results from table 2 hold in the main fixed effects models, while the effects of partisanship are negligible across the models. Next, we re-run the models using an alternative dependent variable, namely the two-year moving average of a state's credit rating. 73 We do this because ratings can be adjusted anytime during a given year, and therefore a yearly summary may lack precision in terms of effects from the explanatory variables. We re-run our tests of H2 in particular with this dependent variable across our three main specification, and find our original estimates to be robust (see appendix D, table A5).

Conclusion
This article investigates the link between corruption and subnational creditworthiness. Using data on US states, we show support for our first hypothesis that corruption is largely unrelated to credit ratings in the aggregate sample. However, as per our second hypothesis, we find a clear negative effect of corruption only in states with a comparatively high level of fiscal autonomy. Our results for both hypotheses are robust and consistent across several model specifications and different estimators. This suggests that fiscal dependence allows corrupt subnational units to continue to borrow on favorable terms, and that the federal government is seen to implicitly guarantee the repayment of debt in fiscally dependent states, which shields more dependent states from the negative effects of corruption on credit ratings.
Our findings thereby contribute to the emerging debate on subnational creditworthiness by suggesting that corruption does not always undermine ratings, and that such effects may both shift over time and between contexts. This finding thereby adds some nuance to the emerging consensus on the detrimental effect of corruption on credit ratings. While corruption has important and severe negative consequences for economic performance, some subnational units are still seen as creditworthy and can thereby continue to borrow on favorable terms.
The evidence provided in this study also has implications for the debate on fiscal federalism, decentralization, and the impact of external assistance and aid more broadly. Potentially, fiscal dependence may be a precondition for corruption control and protect taxpayers from assuming an extra cost on debt repayment, as well as uphold vital public services and thereby protect vulnerable groups from being adversely affected by the consequences of corruption and mismanagement. However, favorable credit ratings and low-cost borrowing may also potentially contribute toward insulating corrupt officials from demands for accountability and even empower corrupt governments. Fiscal dependence may, much in line with ideas of external aid "resource curse" literature, lead to a "fiscal illusion" 74 whereby voters see expenditures as disconnected from revenues, since revenues are generated by distant others. In other words, if state-level voters perceive that the cost of local consumption and public service expenditures can be secured without increasing local taxes, they may also not demand accountability for overspending. This allows government to continue to "expand their expenditures while externalizing the costs to others." 75 Several studies show that corrupt officials seen as competent to secure local economic benefits can induce loyalty among voters and local constituents. 76 If politicians are convicted for a form of corruption that does not necessarily directly or immediately affect voters, such corruption, sometimes go under the radar of voter punishment. 77 Corrupt subnational governments may thereby be allowed to continue to "overfish" the common pool resource. 78 Our findings also suggest several viable avenues for future research. First, it is worth investigating how shaping the composition of different types of federal transfers may have varying effect on keeping corrupt politicians in office. Our data does not allow us to distinguish between different forms of federal transfers, and some types of transfers may more effectively signal bailout expectations. 79 It would also be interesting to see how our results relate to other indicators of fiscal mismanagement, including, i.e., adherence to the rule of law or the ability of governments to pass the budgets on time. 80 Second, while we attempt to mitigate potential endogeneity by relying on a dynamic panel with an instrumental estimator and instrumental analyses, 81 we encourage future research to engage with other approaches to enhance possibilities to make causal inferences. Finally, it could be useful to determine whether the relationship between corruption and subnational creditworthiness in the United States is the same as in other federal countries, such as Canada or Germany.