Strabo 1 The Causes of State Level Corruption in the United States By: Mark M. Strabo mstrabo@princeton.edu Princeton University Princeton, New Jersey 12 January 2015
Strabo 2 Introduction The United States is far less immune to corruption than the average American likely believes. A lot of research has been dedicated to the causes of corruption at the cross-national level (Treisman, 2000; Ades and Di Tella, 1999; Lipset, 1960; Tanzi, 1994; Lipset and Lenz, 1999; Mauro, 1995; Jong-Sung and Khagram, 2005; Shabbir and Anwar, 2007; and Elbahnasawy and Revier, 2012), but much less analysis has provided insight into what causes the corruption at the state level, which directly or indirectly affects most Americans on a daily basis. Glaeser and Saks have most notably analyzed this, and I will reference them closely throughout this paper. Most recently, Governor Bob McDonnell of Virginia was charged with eleven counts of corruption, which sparked my interest in conducting research in this field. 1 Corruption is defined as a breach of public trust and/or abuse of position by federal, state, or local officials and their private sector accomplices. By broad definition a government official, whether elected, appointed or hired, may violate federal law when he/she asks, demands, solicits, accepts, or agrees to receive anything of value in return for being influenced in the performance of their official duties. 2 Corruption is historically known to be detrimental to the proper functioning of democratic government, and is associated with massive economic costs. Thus, identifying the causes of corruption is necessary to the success of our nation s most basic ideals. The primary question my research will examine is: What causes corruption at the state level? In my analysis I choose to use a concrete and normalized measure of corruption that can be compared across states: conviction data. This gives my measure the ability to more accurately quantify corruption than the opinion survey analysis that is most prevalent in the field. Opinion 1 Helderman and Zapotosky, Ex-Va. Governor Robert McDonnell Guilty of 11 Counts of Corruption. 2 Definition of Corruption - Cornell Law School.
Strabo 3 survey data analysis on corruption is susceptible to a variety of influences, including the seeming ambiguity of what qualifies as corruption across samples. Some argue that using conviction rates may not be fully accurate in this analysis, because states with truly high levels of corruption may be quantitatively low in our measure due to the fact that a state s justice system, which handles corruption charges, may be corrupt itself, and thus less likely to convict those guilty of corrupt behaviors. This potential issue with my data is accounted for by the use of a secondary corruption measure, which only analyzes convictions at the federal level. By its nature the federal justice system should be relatively immune to state corruption. If I were to use federal level corruption convictions only, however, this could introduce a separate potential problem to my analysis: poorer states tend to have fewer resources to handle corruption cases at the state or local level, and thus push these cases off to the federal system. 3 As a result, it is possible that in using only a federal level measure, poorer states would appear more corrupt, but merely as a function of their limited resources in handling the cases at the state and local levels, not because they are truly more corrupt. In other words, they could have very high corruption levels using the federal level index, but relatively normal or low levels of corruption using a combined index of local, state, and federal level cases. Having considered this, my primary measure of corruption, which I call corrup1, includes the local, state, and federal levels. By including one DV measure of the local, state, and federal level and a second DV measure of only the federal level, I am able to account for the respective measurement errors caused by the inherent characteristics of each measure, and analyzing the two measures will allow me to observe the causes of corruption, while accounting for these issues. As far as I am aware, I am the first to have included both of these measures in an analysis, making the results robust across measures and eliminating the 3 Census Bureau - Income Inequality Data.
Strabo 4 potential effects of inherent measurement error. In doing so, I am able to contribute to prior research in providing results that use an objective corruption measure of convictions, as is also done by Glaeser and Saks. In using two distinct measures, however, I am able to eliminate a critic s ability to argue that there is a misrepresented effect due to this measurement error. As a result of my research, I found strong statistical support for my first hypothesis (H1): states with higher levels of education are less corrupt. This finding is robust across many variables and the two separate measures of the dependent variable: corrup1 and corrup2. In my second hypothesis, H2, I argue that higher levels of income inequality are associated with higher levels of corruption; H2 was found to be statistically significant by the results of the corrup2 measure, but not corrup1. Thus, the lack of robustness across measures of the dependent variable makes one question the significance of the results. My third hypothesis (H3) argues that an increase in government share of the state GDP is correlated with an increase in corruption. H3 is supported by the data in terms of the direction of the effect on corruption, but is not profound enough to rule out the null hypothesis, which indicates that the same effect could reasonably be observed due to chance. Thus, the effect cannot be differentiated from noise. The results indicating the effect of education level on corruption helps to indicate that a policy focus should be placed on increased education, in addition to its widely accepted benefits to society, it seems to reduce corruption. Additionally, the robustness of education on corruption is indicated by its steady statistical results across multiple regression specifications and the two measures of corruption. My results in some ways align closely with the findings of Glaeser and Saks (2005), especially in regards to education having the largest and most significant effect on corruption. However, as I will discuss in more detail later on, the findings derived from my novel method of data analysis contribute to the field in a notable fashion.
Strabo 5 Theory In this paper, I will focus on the effect of three target variables on corruption. My first hypothesis (H1), grounded in an early study conducted by Glaeser and Saks, states that a more highly educated state populace is associated with lower levels of corruption within that state. More highly educated citizens are more likely to be able to discern the effects of corrupt activities, more likely to monitor those in positions of power capable of acting corruptly, and more capable of acting against perpetrators if corruption is detected. One can think of a government official s behavioral choice in acting corruptly as a cost-benefit game (Game Theory). Therefore, the official must weigh their individual costs and benefits of acting in corrupt ways. A highly educated populace increases the potential costs if caught (which in this case can include jail time, monetary fines, and/or loss of office), and also increases the probability of being caught. For these reasons I suspect that a more highly educated populace would be observed in states with lower levels of corruption. My second hypothesis (H2), extrapolated from a cross-national study by Jong-Sung and Khagram (2005), indicates that higher income inequality is associated with a higher level of corruption within the respective state 4. One can reasonably suspect that when income inequality is greater, there is a heightened individual benefit in committing corrupt actions that benefit the rich and a decreased cost of corruption due to the limited resources available to the lower end of the income spectrum. One can imagine that the probability of being caught is also lower in a state with high income inequality. Those more likely to discern corruption, and more capable of acting against, benefit from the corrupt action, while the lower end of the income spectrum is both largely unaware and incapable of acting in an effective way, even if made aware. 4 Jong-sung and Khagram, A Comparative Study of Inequality and Corruption.
Strabo 6 My third and final hypothesis (H3), which applies a cross-national study conducted by Tanzi (1994) to the state level, specifies that states whose governments consist of a larger share of the state GDP are associated with higher levels of corruption. 5 In such a scenario, officials in a position to act corruptly have a heightened economic incentive to do so when there is a larger financial incentive at play. A corrupt government official has easy access to individual monetary benefit, increasing the potential benefits of corruption, while the costs remain unchanged. Data In this paper I seek to identify the causes of corruption at the state level. In my three hypotheses I have identified the variables that I wish to focus on, and which I believe are affecting the corruption level: education, income inequality (gini), and the size of the government in the economy (percentgov). Education is operationalized using data from the 2010 U.S. Census, and indicates the percentage of the state population that holds a bachelor s degree or higher (Dataset 233). In order to analyze income inequality within my model, I operationalized this variable by using gini coefficient data published by the U.S. Census Bureau in the 2010 Census. 6 The gini coefficient is a summary measure of income inequality. The gini index varies from 0 to 1, 0 indicating perfect equality where there is a proportional distribution of income. A 1 indicates perfect inequality where one household has all the income and others have none. 7 My final target variable, the size of the government s role in the state s GDP (percentgov), is operationalized by dividing the log of the government GDP (in billions) by the log of the overall state GDP (in billions). I chose to operationalize this variable in such a way for a few specific reasons. First, due to the large variance in the size of each state s GDP and 5 Tanzi, Corruption, Governmental Activities, and Markets. 6 Census Bureau - Income Inequality Data. 7 Ibid.
Strabo 7 government GDP, I log each value, minimizing the extent to which outliers will skew the data. Then, taking the ratio of each logged value allows me to use a percent as the variable, which is easily regressed in my model. In doing this, I have limited the effect of outliers (such as California and New York), and I have normalized the variable for easy comparison across the states. In order to focus legitimate attention on my three target causes, I must also identify and account for other variables that can affect the outcome of this cost-benefit game that can be understood to influence government officials decisions in acting corruptly. I must also control for confounders, or those variables that may affect both my target independent variables and the dependent variable. In addition to the two measures of corruption, and the three target variables (education, gini, and percentgov), I include a series of independent variables that must be included in my data in order to draw sound conclusions from the results of my statistical analysis. The variable party is included in my data and is operationalized as the major political party (Republican or Democrat) that controlled the respective state in the 1996 presidential election. It is reasonable to think that the partisan roots of the state affect the education level, income inequality, and government role in the economy, while also independently being associated with levels of corruption. I chose the year 1996 because this election occurred directly before the time period in which my primary dependent variable (corrup1) was measured and falls within the time period during which my secondary dependent variable (corrup2) was measured. This variable will be more accurate for corrup1, but should still adequately control for party in the case of corrup2, because party, like most of the variables included in my model, are very sticky. Most of the variables controlled for are not subject to significant fluctuation over time. Including party in my model allows me to isolate and account for its potential effect.
Strabo 8 The dependent variable, corruption, is analyzed using two separate measures in our research. Doing so enables us to reach conclusions that are robust across different measures. The primary measure, corrup1, is data representing the guilty corruption convictions per million residents per year from 1998 to 2007, at the local, state, and federal level. I gathered this data from the Washington Post s publishing of the Department of Justice Census Bureau data. 8 Corruption crimes investigated by the Department of Justice include a wide array of topics such as conflict of interest, fraud, campaign-finance violations, and obstruction of justice. 9 My secondary measure of corruption, corrup2, is data originally published in the Glaeser and Saks paper, and indicates the rate of federal corruption charges per state, measured as average federal convictions per year from 1976 to 2002, divided by the average population of the respective states during this time period, in order to normalize for population. 10 This data is also derived from the Department of Justice. The measures include convictions of corruption ranging from police officers charged with extortion and soliciting bribes, to governors charged with campaign-finance violations. 11 Use of these two separate measures of corruption not only allows us to determine the robustness of effects across specifications, but also increases the credibility of the claims. Corrup1 alone is imperfect because its only includes state and local level convictions, and if a state is corrupt, it is possible that its justice system is also corrupt, skewing the corruption convictions to make the state look seemingly less corrupt than it is in actuality. A common solution to this is the use of only federal level convictions, which is why I use corrup2. However, utilizing solely federal level convictions can make poorer states appear more corrupt 8 Wilson, The Most Corrupt State(s) in America. 9 Glaeser and Saks, Corruption in America, 1057. 10 Glaeser and Saks, Corruption in America. 11 Ibid., 1057.
Strabo 9 than they are in reality (at least relative to other states), because often states that lack the resources to process corruption charges at the state or local level push these cases to the federal level for trial. 12 Doing so confounds the effect of poor states on corruption, simply because of skewed data analysis, rather than observed effect. The use of both measures allows us to analyze the effects transcending these potentially problematic aspects of the given measures. Map #1: Corruption Grades by State (The Washington Post) 13 12 Glaeser and Saks, Corruption in America. 13 Wilson, The Most Corrupt State(s) in America.
Strabo 10 Map #2: Varying Levels of Corruption in the United States 14 In my analysis, I have also included the variable region. The aim is to control for the geographical location of the state within the United States and is operationalized using the four Census Bureau region categories: Northeast (1), South (2), Midwest (3), and West (4). It is clear from the maps above that certain regions are prone to higher levels of corruption; it is unclear, however, if this is due to a regional effect, or another variable that is closely associated with region. Including region in my model allows me to control for the possibility of region driving the effect, therefore, I can better analyze the effect of the target variables. The variables population and govemploy are also included in my model. Population is operationalized as the log of the state s population (in thousands) and govemploy is operationalized as the log of the number of civilian government employees within the state (in 14 Liu and Mikesell, The Impact of Public Officials Corruption on the Size and Allocation of U.S. State Spending.
Strabo 11 thousands). Logging these variables, as previously mentioned, allows me to normalize the variable and use it for my cross state analysis by dampening the effect of outliers. For example, given that the population of Montana is significantly smaller than that of California, logging helps to mitigate the effect of this skewing my findings. Including the population variable allows us to account for the fact that the mere size of a state s populace likely impacts many of the independent variables and dependent variables separately, and thus could confound the results if not accounted for. Furthermore, govemploy, or the number of total civilian government employees, must be accounted for, although the dependent variables are normalized by population, this does not account entirely for the confounding effects. The entire population is not subject to the possibility of committing corruption as I define it; only civilian government employees are. Additionally, due to the nature of government, the number of government employees is not a fixed ratio to the size of the population. There are many government jobs that exist independent of population size; thus, accounting for the gross number of civilian employees allows me to better analyze the desired effects. It is not unreasonable to suspect, in fact, that states with smaller gross populations have larger percentages of their total populations employed by the government, due to the fixed nature of many government jobs. Including both population and govemploy, when regressed, allows me to normalize the corruption measures, and independent variables being examined, to be analyzed independent of the effects of these two variables. Studies of corruption, especially at the cross-national level, have widely accepted the role of religious and ethnic diversity in the level of corruption within a region (Treisman, 2000; Mauro, 1995; Jong-Sung and Khagram, 2005; Shabbir and Anwar, 2007; and Elbahnasawy and Revier, 2012). Religious Diversity (religious.diversity) is included in my model, and
Strabo 12 operationalized using data from the 2010 Census, as the percent of the state that is Christian (Dataset 77). Ethnic Diversity (ethnicdiversity) is operationalized as the percent of that state that is white, using data from the 2010 U.S. Census (Dataset 19). Accounting for these measures of diversity in my model allows me to control for their potentially confounding effects, as these variables can be associated with both the dependent and independent variables. Lastly, I have included the variable vote in my model, which is operationalized as the percentage of the state population that in the 2010 Census said to have voted in the most recent election (Dataset 400). Since voter turnout survey data is largely accepted to be subject to the social desirability bias, it is likely that these percentages are higher than the true value. It is reasonable to assume that this bias occurs across all the states. Therefore the data can still be legitimately analyzed, as I am concerned with the voter turnout percentages relative to other states. Thus, the effect of the bias should be insignificant for our purposes. While Glaeser and Saks did not include this variable in their analysis, I found controlling for it important. Generally, when voter turnout is low, those who vote have some heightened incentive to vote, encouraging them to vote when many do not. For example, in school board elections, especially if the election falls in an odd numbered year, it is widely understood that those who do vote are generally directly affected by the election. Teachers, families of teachers, and teacher unions all have very high turnouts, while the average voter turnout is low. Thus, the results of the election are skewed in favor of the teachers, who had a heightened incentive to vote, even if against the general public s desires. This same logic can be applied to corruption. Those who may benefit from a corrupt official being elected are more incentivized to show up at the polls; with low turnout numbers, their vote carries more significance. As a result, when voter turnout is low, it is more likely for a corrupt official to be elected. The voter turnout variable has another significant
Strabo 13 effect in our analysis, though an indirect one. Voter turnout rates are largely associated with the extent to which the populace is involved in monitoring and following the actions of public officials, thus in states with low voter turnout rates, it is not unreasonable to conclude that the actions of public officials (at least elected ones) are not as closely monitored and scrutinized by the public. In such a situation, the costs of acting in a corrupt manner, and the likeliness of being caught, both decrease, and thus will affect the behavior of public officials. It is clear that including voter turnout in my model allows me to control for the aforementioned effects associated with it, and thus better discern the effects of the target variables. Although confounding variables were attentively accounted for, and the potential effect of using different measures was carefully thought out before analysis, there are aspects of the data that may cause potential problems, and must be identified. One potential problem is that the majority of the independent variable data I use in my analysis is from the 2010 Census, which accounts for the period from 2000-2009, as the data is collected every ten years. My primary dependent variable (corrup1), however, begins in 1997, thus in a time period beginning shortly before the time period the IV data represents. This trend is especially true for the secondary dependent variable measure (corrup2), which measures a time period from 1976-2002. The effects of this issue should be minimal and insignificant however, because nearly all the IVs are very sticky, meaning shifts occur very slowly over time. Additionally, since the Census is only conducted every ten years, this is largely the best option, though not perfect. This problem becomes even less significant when one considers that what one is truly concerned with is the ranking of the states relative to each other, more so than an absolute numerical value, considering this, the variables are determined to be even more sticky. Within my model I have also not included a variable that accounts for the strictness of anti-corruption laws by state. This
Strabo 14 should be largely neutralized for corrup2, which deals with federal level convictions, but such a variable could be seen to have an effect in the corrup1 measure, which includes state and local convictions. My findings should not be largely affected by the absence of such a measure, however, due to the fact that both corrup1 and corrup2 are regressed in my analysis. If the effect of a variable is robust across these two measures, the absence of this variable should be deemed largely insignificant. At the very least, its absence may deter us from discerning the significant effect of other variables, but does not harm the results of the study that are found to be statistically significant. Another potentially problematic absence in the data set is the lack of a variable explicitly corresponding with media coverage within the state. The access to media and news coverage within a state is associated with the risk of being caught engaging in corrupt activity, and consequently the costs of corruption due to electoral and reputational repercussions. Although media coverage is absent from our analysis, the vote variable can be seen as performing a similar role in the analysis. In the case of voter turnout percentage, this variable provides an indirect (and at times direct) observation on the extent to which the population of a state holds their elected officials accountable and monitors their activity. The higher the turnout, the higher the political engagement by the populace, and thus the higher the likelihood of getting caught if engaging in corrupt activity, and the higher the costs if caught. Although not identical to media coverage, voter turnout can control for a similar effect.
Strabo 15 Analysis Figure 1: Education on both measures of corruption
Figure 2: Income Inequality on both measures of corruption Strabo 16
Strabo 17 Figure 3: Government GDP on both measures of corruption Before analyzing the statistical significance and quantitative size of the effect of our hypothesized variables on corruption, one can visualize the effects by referring to the correlation plots above, titled Figure 1, Figure 2, and Figure 3. Each figure represents the correlation between the indicated target variable, and the two measures of corruption: corrup1 on the left, and corrup2 on the right. Figure 1 depicts the correlation between education and corruption, indicating that the data confirms that higher levels of education within a state are associated with lower levels of corruption. It is clear, at the very least, that the direction of the effect of education on corruption in the data supports my hypothesis. Figure 2 depicts the correlation between the
Strabo 18 gini coefficient of a state (income inequality) and corruption, suggesting from the data that a higher gini coefficient is associated with higher levels of corruption. Again, the statistical significance and exact size of the effect cannot be discerned using this plot, but it is clear that at least the direction of the effect supports my hypothesis. Additionally, it is apparent that the direction, and relative size, of the effect is robust across both measures of the dependent variable. Lastly, Figure 3 depicts the correlation between the percent government share in the state GDP and corruption, indicating that a higher percentgov is loosely associated with a higher level of corruption. The correlation plots, with both measures of corruption, support the direction of my hypothesis. In order to discern the size and significance of the effect, I use a series of multivariate linear regressions. Regression #1: (Corrup1) education -0.204 * (0.098) gini 0.151 (0.123) percentgov 24.5 (14.9) Regression #2: (Corrup1) -0.200 * (0.091) 0.402 * (0.158) 39.9 * (17.9) Regression #3: (Corrup1) -0.368 * (0.146) 0.307 (0.212) -24.0 (121) 0.022 (0.106) -0.118 (0.154) -22.2 (16.3) religious.diversity N/A 0.052 (0.087) ethnicdiversity N/A -0.004 (0.137) population N/A -11.4 ** (3.63) party N/A N/A 1.14 (3.03) region (South) N/A N/A -8.05 (6.19) region (Midwest) N/A N/A -4.01 (5.57) region (West) N/A N/A -9.29 (5.67) vote N/A N/A 0.114 (0.125) govemploy N/A N/A 25.3 (36.5) Regression #4: (Corrup2) -0.349 * (0.142) 0.425 * (0.178) 38.5 ` (21.5) Regression #5: (Corrup2) -0.342 * (0.130) 0.613 * (0.227) 38.4 (25.7) Regression #6: (Corrup2) -0.604 ** (0.196) 0.596 * (0.286) -159 (163) 0.036 (0.143) -0.468 * (0.207) -46.5 * (21.9) N/A 0.118 (0.125) N/A -0.280 (0.197) N/A -14.5 ** (5.21) N/A N/A 5.58 ( 4.08) N/A N/A -15.2 (8.34) N/A N/A -3.45 (7.50) N/A N/A -14.6 (7.64) N/A N/A 0.115 (0.169) N/A N/A 71.0 (49.2) Table 1: Regression Output (Statistically Significant Key (P-value): `=0.1; * = 0.05; **=0.01
Strabo 19 Top number represents the regression coefficient for that specification; bottom number in parenthesis represents the corresponding standard error) The model just described was used to run a series of linear regressions on the effect of the indicated variables on corruption. Use of the linear regression model allowed me to discern the effect of each variable on corruption (the regression coefficient), the standard error of this estimate, and the statistical significance of the findings. The regression coefficient indicates the effect on the dependent variable that is observed by a one-unit increase in the respective independent variable, and the sign indicates the direction of the effect. The standard error indicates the degree to which this effect varies across trials, essentially indicating the accuracy of the estimated effect. Lastly, we are able to determine the statistical significance from the p-value. When P < 0.05, we say that the effect is statistically significant, meaning that we are able to rule out the null hypothesis of no effect because there is less than a 5 percent chance that we would observe the same results under the null hypothesis (i.e. due to chance, rather than under the effect of the IV). Table 1, pictured directly above, depicts the regression output of the series of regressions. I elected to run three regressions on each of the two corruption measures. Regression 1 includes merely my hypothesized independent variables (education, gini, percentgov) on corrup1. Regression 2 includes the components of regression 1, and adds religious diversity, ethnic diversity, and population as variables in the regression. Regression 3 regresses all the mentioned independent variables on corrup1, thus adding party, region, govemploy, and vote to regression 2. For regressions 4, 5, and 6 the same specifications are used as in regression 1, 2, and 3, respectively, but use the corrup2 measure of the dependent variable, rather than the corrup1 measure. The regression output displayed above in Table 1 offers us insight into the causes of corruption. In regression 1, only education was found to have a statistically significant effect,
Strabo 20 with a 0.204 decrease in corruption for a 1-unit increase in education. The standard error was relatively small at 0.098. The effects of gini and percentgov in regression 1 were largely insignificant and unhelpful given their large standard errors in relation to the coefficients. In regression 2, once the two diversity measures and population were controlled for, I observe nearly identical results for the effect of education, with a coefficient of -0.200 and a standard error of 0.091; in addition, gini, percentgov, and population have all become statistically significant. It is likely that controlling for population had a major effect on the observed effect of the gini and percentgov variables. The gini effect for regression 2 was found to have a 0.402 increase in corruption for a 1 unit increase in the gini value, which given the scale of the gini (0-1), is largely erroneous in actuality. A 1-unit increase in percentgov in regression 2 was found to have a 39.9 unit increase in corrup1, with a standard error of 17.9. This standard error is too large to discern reliable conclusions from the data. The insignificance of this observed effect is made evident in regression 3, when all variables are controlled for, the percentgov measure not only becomes statistically insignificant once again, but the direction of the effect shifts enormously from positive to negative. With all factors included, a 1-unit increase in percentgov is found to decrease corruption by 24.0 units, with a standard error of 121. Based on the corrup1 measure, we are unable to discern that percentgov has any effect on corruption. From regression 2 to regression 3, we observe that both population and gini loose their statistical significance with the addition of the final variables into the regression. I suspect that the addition of the govemploy variable causes population to lose its statistical significance, with corruption now controlled in relation to the number of people who hold a government position capable of committing corruption, not merely the number of citizens within the state, which is a less accurate control given the circumstances. We can reasonably suspect that the gini variable
Strabo 21 (income inequality), loses its statistical significance from regression 2 to regression 3 with the addition of the region and party variables. Both of these can be understood to be associated with the level of income inequality within a state, and thus we must control for them in order to accurately discern the effects of our hypothesized variables. H1, however, hypothesizing that higher levels of education are associated with lower levels of corruption, is largely supported by the output provided by the series of regressions on the corrup1 measure. Across all three specifications, the effect of education remains robust, with coefficients of: -0.204, -0.200, and - 0.368; and standard errors of: 0.098, 0.091, and 0.146, over regressions 1, 2, and 3, respectively. In all three cases the effect of education was found to be statistically significant at the p<0.05 level. In the cases of H2 and H3, we are unable to rule out the effects of either gini or percentgov from the null hypothesis. As an additional measure of robustness, the same three regression specifications were run using the alternative measure of corruption, corrup2. In regression 4, a 1-unit increase in education was found to be associated with a 0.349 decrease in corruption, with a standard error of 0.142. In regression 5, a 1-unit increase in education was found to be associated with a 0.342 decrease in corruption, with a standard error of 0.130. Lastly, in regression 6, a 1-unit increase in education was found to be associated with a 0.604 unit decrease in corruption, with a standard error of 0.196. In all of the regressions, education was found to be statistically significant at the 0.05 level. The effect of education on corruption remained largely constant across all the robustness checks: the three specifications and both measures of the dependent variable. The results of the data strongly support my hypothesis on the effects of education (H1). As is evident from Table 1, the corrup2 measure was no better in discerning an effect percentgov variable, however, it did indicate a statistically significant effect of income inequality
Strabo 22 on corruption. Income inequality was found to be statistically significant across all three specifications of the corrup2 measure, with 1-unit increase in gini being associated with effects of: 0.425, 0.613, and 0.596; and standard errors of: 0.178, 0.227, and 0.286, across regressions 4, 5, and 6, respectively. Although I am able to rule out the null hypothesis on the effect of income inequality on corruption for the corrup2 measure, I had differing results on the corrup1 measure, indicating that this effect is likely not robust across different measures of the dependent variable. Given that there is no clear rationale for why income inequality would have a statistically significant effect on the federal measurement of corruption, but not the state, local, and federal measures, I find it difficult to rule out this result from noise. In both measures of corruption, I was unable to rule out the null hypothesis of no effect for H3. Percentgov appears to have no discernable effect on corruption, with massive standard errors in both measures, and a sign change from the second to third specifications along both measures. The degree to which many of the variables regression output varies over the different specifications indicates to me that the results of this study can be largely sensitive to which variables are controlled for; this is likely due to the relatively small n value (fixed at fifty), and the large extent to which variables are associated with one another. The findings described by Saks and Glaeser indicated similar conclusions from their analysis. 15 Thus, the observed effect of education becomes even more significant given the extent to which its effect remained constant across specifications and robustness checks. Conclusion It is evident from the careful multivariate regression analysis conducted that the effect of education on corruption is statistically significant and robust across many specifications. This 15 Glaeser and Saks, Corruption in America, 1062.
Strabo 23 effect is observed across specifications with the inclusion of many different variables, and on both measures of the dependent variable, a further robustness check. Thus, such results overwhelmingly support my hypothesis on the association of higher levels of education with lower levels of corruption (H1). The second hypothesis (H2) was not supported as strongly as the effect of education. My prediction that an increase in the level of income inequality within a state would be associated with a higher level of corruption within that state was not supported across all specifications. H2 was supported when the regression analysis was run with the corrup2 measure of the dependent variable, in which case all three regression specifications indicated statistically significant results. However, for the corrup1 measure, none of the specifications indicated a statistically significant effect of income inequality on level of corruption. Such a discrepancy seems to indicate that the observed effect of income inequality is subject to differing measures of the dependent variable, and thus the significance of such results must be questioned. It is unclear why the corrup2 measure, including only federal convictions, would indicate income inequality as having a statistically significant effect on corruption, while the corrup1 measure, including local, state, and federal convictions, would not. Outside of measurement error, or discrepancy, there is no clear justification for why the federal level corruption measure would be associated with income inequality, while the local and state corruption measures are not. Lastly, H3 was not supported by the statistical analysis, since there was not found to be any significant correlation between government share in the state GDP and level of corruption. It is possible that the large error and discrepancy in size of the effect across specification is due to my omittance of a variable corresponding to the level of regulation within the state. With the level of regulation of state spending controlled for, the results may have turned out differently. One can imagine the importance of this for states like California and New York, which although they have massive
Strabo 24 GDPs, also have high regulation, seemingly making it more difficult for corruption to occur regarding state expenditures. In further research, inclusion of a proxy variable for regulation level within the state would seem valuable in analyzing the association between government share in GDP and corruption. Additionally, further studies may find substantial reasons to include a variable accounting for the media or news coverage within a state. In my analysis, the variable vote can seem to represent a proxy for the public s alertness to politics, which affects the cost of acting corruptly. Although the relationship between voter turnout and public attentiveness to politics is largely indisputable, inclusion of a variable more explicitly indicating the public s access to information may also serve a valuable purpose as a proxy, ultimately for our desired objective, which is measuring the likelihood that a corrupt politician will be held accountable for his actions. My research has critically examined the work of Glaeser and Saks, and others who have dedicated portions of their research to the causes of political corruption at the state level (Apergis et. al, 2011; Goel and Nelson, 1996), and has contributed to the field significantly by adding a new method to the field s analysis. Researchers have long accepted that measuring corruption is a difficult task. Survey research has been prevalent, especially at the international level, but critics, myself included, have suggested that due to the ambiguity of what constitutes corruption across nations, as well as survey error and social desirability bias, survey data may be inaccurate. State level researchers have identified such issues, and have turned to the objective measurement of corruption convictions (Glaeser and Saks, 2005; Apergis et. al, 2011; Goel and Nelson, 1996). Such a measurement is superficially very accurate, but a new measurement inaccuracy comes with it. As previously mentioned, it is not unreasonable to suspect that states that are highly corrupt, also may have corrupt justice systems. If such a situation were true, corrupt states would
Strabo 25 appear far less corrupt than they are in actuality. The corruption researchers response to this has been use of only federal level corruption conviction data, arguing that the federal justice system should be largely immune to varying state levels of corruption. Citing this argument, many of my predecessors, namely Glaeser and Saks, have used this federal level corruption measure as their dependent variable in their research. What has largely been ignored, however, is that there is inherent measurement error associated with the federal conviction measure that previous researchers have used. In using only a federal conviction measure of corruption, we are not accounting for the measurement error that occurs because poorer states do not have the funding to handle many corruption cases at the local or state level, and instead hand these cases off to the federal level. In such an instance, states with fewer resources dedicated to their justice system will inaccurately appear to be corrupt, while states with larger resources that handle these respective cases at the local or state level, will appear to be less corrupt. This discrepancy is not indicating that poorer states are more corrupt, but rather that poorer states may falsely appear more corrupt if a federal conviction measure is used. Acknowledging this potentially issue in the conviction data measure of corruption, in my research I have included a state, local, and federal measure of corruption (corrup1), in addition to a merely federal level measure of corruption (corrup2). In doing so, I am accounting for the aforementioned measurement errors associated with each measure of the dependent variable. If the effect of an independent variable is robust across both measures (and all specifications), we are able to, with a high degree of certainty, indicate the statistically significant correlation between the targeted independent variable and corruption, as the measurement error of the dependent variable would be mitigated. In addition, while Glaeser and Saks hypothesized that states with higher average levels of household income would be associated with higher levels of corruption, I hypothesized that a
Strabo 26 higher level of income inequality would be associated with a higher level of corruption. The former measure largely measures the average state wealth; I wish to analyze the effect of inequality of income on corruption. Clearly these are two distinctly different measures. One can reasonably see a scenario where, with high-income and low-income individuals (high inequality), there is a larger opportunity to benefit from corruption as a government official and there is a lower chance of being held accountable for such an act. My results indicate that high income inequality has a statistically significant association with higher corruption for the federal measure, but largely no statistical significance for the local, state, and federal measure (although the direction of the effect still supports H3). Lastly, I have included an additional variable in my analysis that was omitted by Glaeser and Saks. In their conclusion, Glaeser and Saks indicate that (Their) results can be interpreted as evidence that the costs of corruption (as influenced through the probability of being caught) matter more than the potential gains (as measure by the size of government or the number of government regulations that can be circumvented through bribery). 16 If such a statement were true, I found it quite important to include a variable that can serve as a proxy for the probability of being caught or held accountable for corruption: voter turnout. Voter turnout is not only directly associated with whether elected officials remain in office (and indirectly with whether those appointed remain in office), but also can be understood as a proxy for the public s attentiveness or awareness to politics and government actions. Glaeser and Saks, and multiple researchers before me (Apergis et. al, 2011; Goel and Nelson, 1996), have included neither this proxy, nor a measure of news coverage within that state in their models. Either of these, if not both of them, seem important if one is considering that a government official s decision in acting 16 Ibid., 1068.
Strabo 27 corruptly is motivated by the likelihood and costs of being caught. Inclusion of these techniques to my quantitative analysis, especially use of the secondary corruption measure, provides more accurate and more robust statistical results to the field on the causes of corruption in the United States. Through my analysis, I have thoroughly depicted the varying levels of corruption within the United States, and factors that are highly correlated with corruption. It is the negative consequences of corruption, however, that make this a crucially important area to research. Many researchers have studied the massive negative effects of corruption on the United States, and the everyday lives of its people including: large monetary costs, a decreased trust in government, and a feedback effect further decreasing education, among others (Johnson et. al, 2010; Dincer and Gunalp, 2011; Johnson et. al, 2013; Liu and Mikesell, 2014). A study conducted by Liu and Mikesell, and published in the Public Administration Review in the spring of 2014, found, That the nine most corrupt states could have spent $1,308 less annually per capita if they had succeeded in maintaining only an average level of corruption. 17 Such a statistic indicates the massive financial costs of corruption that burden everyday Americans, a cost that could be eliminated if the most corrupt states were merely average in their level of corruption. The aggregate fiscal cost of state level corruption is shocking to our Nation, which takes pride in the success of its democracy. Additionally, Glaeser and Saks point out the potential feedback effect of corruption, suggesting, If corrupt practices lead to less investment in human capital, the negative impact of low education levels could be self-reinforcing. 18 These are only a few 17 Liu and Mikesell, The Impact of Public Officials Corruption on the Size and Allocation of U.S. State Spending, 355. 18 Glaeser and Saks, Corruption in America, 1068.
Strabo 28 notable consequences of corruption, but are significant enough to motivate research into the causes of corruption, and ultimately how these causes can be mitigated. My analysis has indicated the effect of education on corruption levels, and the aforementioned research briefly illuminates a few major consequences of such corruption. We as a nation have a vested interested in reducing the levels of corruption within our states. Corruption levies a massive financial burden upon Americans, decreases the efficacy of our democratic system, and can even lead to a further reduction in the education level of the populace, thus creating a death spiral if not adequately addressed. It is clear therefore, that policy within the United States, as well as public attention, should be focused on increasing the education level of the populace. In addition to the long list of widely accepted societal and individual benefits education brings, it also is our most effective weapon against corruption. A study by Adserà et al argues that reducing corruption requires a populace that is capable of holding those in power accountable, due to the cost-benefit game that corrupt behavior can be understood as. 19 An increased level of education among the populace increases both the likelihood that a corrupt official would be caught, as well as the cost if caught. A successful focus on increasing education levels would have a great payout in the success and permanence of the United States. The nation would see economic prosperity it has never seen before, a massive reduction in corruption, a large increase in political participation, and a greater trust in our democratic system. Individuals would be greatly benefitted from a higher level of education, and one may suspect that we would see a decrease in crime, as well as an increase in per capita income. An increase in the level of education among the U.S. populace 19 Adserà, Boix, and Payne, Are You Being Served?.
Strabo 29 will decrease corruption, benefitting the citizens, the economy, and ultimately, the democratic government, of the United States of America. Appendix Visualization of Corrup1 Data
Strabo 30 The Washington Post 20 Operationalized Variables Key Variable Call Name Description 20 Wilson, The Most Corrupt State(s) in America.
Strabo 31 corrup1 Guilty corruption convictions, per million residents per year (1998-2007) at the local, state, and Federal level corrup2 Federal level corruption charges (Saks and Glaeser Data) Corruption Rate as average convictions per year from 1976-2002 divided by average population 1976-2002 Education % of population with Bachelor s degree or higher Party Party that controlled the state in the 1996 presidential election (1=Republican, 0=Democrat) Region U.S. Region the state is located within (Northeast=1, South=2, Midwest=3, West=4) Govemploy Civilian Government employees (in thousands) logged religious.diversity Percent of state that labels themselves as Christian Ethnicdiversity Percent of state that is white Gini Income inequality measure GDP State GDP (in billions of dollars) logged govgdp Government share of State GDP (in billions of dollars) logged Population Population (in thousands) logged Vote Percentgov state voting participation percentage Percent of the state GDP that is government spending (log(govgdp_/log(gdp)) Bibliography 2010 U.S. Census Datasets