THE RELATIVE IMPACT OF ECONOMIC AND POLITICAL FACTORS IN U.S. GUBERNATORIAL ELECTIONS: A FIXED EFFECTS, INCOME-CONTINGENT APPROACH Joel Ryan B.F.A., Webster University, 2004 THESIS Submitted in partial satisfaction of the requirements for the degree of MASTER OF ARTS in ECONOMICS at CALIFORNIA STATE UNIVERSITY, SACRAMENTO SUMMER 2010 THE RELATIVE IMPACT OF ECONOMIC AND POLITICAL FACTORS IN U.S. GUBERNATORIAL ELECTIONS: A FIXED EFFECTS, INCOME-CONTINGENT APPROACH A Thesis by Joel Ryan Approved by: ____________________________________, Committee Chair Mark Siegler, Ph.D. ____________________________________, Second Reader Ta-Chen Wang, Ph.D. ______________________________ Date ii Student: Joel Ryan I certify that this student has met the requirements of format contained in the University format manual, and that this thesis is suitable for shelving in the Library and credit is to be awarded for the thesis. ____________________________, Graduate Coordinator Jonathan D. Kaplan, Ph.D. Department of Economics iii ______________________ Date Abstract of THE RELATIVE IMPACT OF ECONOMIC AND POLITICAL FACTORS IN U.S. GUBERNATORIAL ELECTIONS: A FIXED EFFECTS, INCOME-CONTINGENT APPROACH by Joel Ryan Abstract: Previous research has found that state-level election outcomes in U.S. presidential elections display an “income-contingent” pattern: voters in high-income states are influenced by economic factors like inflation, while voters in low-income states are persuaded more by income growth. This thesis tests whether there are also income-contingent voting patterns in U.S. gubernatorial elections. Additionally, it extends previous work on gubernatorial elections in many directions by including data from far more elections, introducing new explanatory variables, and improved econometric methods. The evidence presented in this thesis shows that there are also income-contingent voting patterns in U.S. gubernatorial elections, although the factors that depend on income differ from what has been found in presidential elections. In addition, political factors, such as challenger candidate spending, whether an incumbent is running, presidential approval, and whether a candidate is of the same political party as the U.S. President also influence the outcomes of gubernatorial elections.. ____________________________________, Committee Chair Mark Siegler, Ph.D. ______________________________ Date iv ACKNOWLEDGEMENTS I would like to thank Professor Mark Siegler and Professor Ta-Chen Wang for their invaluable guidance, patience, and encouragement as I worked to complete my thesis. v TABLE OF CONTENTS Page Acknowledgments............................................................................................................... v List of Tables ................................................................................................................... vii List of Figures ................................................................................................................. viii Chapter 1. INTRODUCTION ..........................................................................................................1 2. LITERATURE REVIEW ...............................................................................................5 Main Concepts – The “Rational Voter” Hypothesis and Retrospective Voting ..........6 Individual Survey Data versus Aggregate Research ....................................................8 Literature Influencing My Approach..........................................................................11 3. DATA ...........................................................................................................................14 4. EMPIRICAL MODEL ..................................................................................................27 5. RESULTS .....................................................................................................................31 Results for all States (Regardless of Income).............................................................33 Economic Impacts on the Vote ..............................................................................33 Political Impacts on the Vote .................................................................................35 Results for Low-Income vs. High-Income States.......................................................42 6. CONCLUSION .............................................................................................................48 References ..........................................................................................................................51 vi LIST OF TABLES Page Table 1: Top Ten States, 1978 and 2008 (ranked by per capita personal income) ............15 Table 2: Variable Descriptions and Summary Statistics, 1976-2008 ................................17 Table 3: Variable Sources and Web Links ........................................................................18 Table 4: Variables and Their Statistical Significance in Prior Gubernatorial Election Research ...............................................................................................................20 Table 5: Regression Results ...............................................................................................32 Table 6: Regression Results (Low-Income Versus High-Income States) .........................43 vii LIST OF FIGURES Page Figure 1: Growth in Education Spending and the Vote .....................................................22 Figure 2: One-Party Control and its Impact on the Vote ...................................................23 Figure 3: One-Party Control and its Impact on the Vote (less 16 outliers) .......................24 Figure 4: Inflation and the Vote, All States .......................................................................25 Figure 5: Inflation and the Vote, Y1 Group (low-income states) ......................................26 Figure 6: Inflation and the Vote, Y2 Group (high-income states) .....................................26 viii 1 Chapter 1 INTRODUCTION Why study gubernatorial elections? Three reasons: governors have growing power in their states, influence in national policy debates, and can affect their state’s economy (Leal, 2006). Many twentieth century state-level election reforms increased the power of governors over their states. For example, since the 1940s and 1950s, new state laws have passed to: isolate gubernatorial elections from the presidential ‘coattail’ effect by moving these elections to non-presidential election years; increase gubernatorial term limits from two to four years; and to allow for self-succession (Chubb, 1988). Besides the states’ legal expansion of gubernatorial power, the governor controls many important levers in the state government. The veto and appointment powers that governors exert give them some muscle in state governments. Also, gubernatorial staffing levels and budget authority increased over the last century (Adams and Kenny, 1989). These growing powers and authority help governors shape national-level policy debates. Through their action at the state level, governors and states have influenced national debates over environmental, health care, welfare, and education policy1. With the power to influence many national policies, U.S. governors have also elevated themselves to the national stage. Starting with Theodore Roosevelt, more than a third of U.S. presidents have previously been governors. In fact, three of the five current living presidents were once governors.2 1 See Sribnick (2008) for examples of how governors and states have influenced these national policies. George W. Bush (governor of Texas from 1995-2000), Bill Clinton (governor of Arkansas from 1979-81 and 1983-92), and Jimmy Carter (governor of Georgia from 1971-5). Only Barack Obama and George H.W. Bush have not been governors. 2 2 It is also important to study how governors get elected because governors may influence their state’s economy. State-level taxes and spending affect state economic prosperity partly due to where businesses decide to locate (Adams and Kenny, 1989). These tax and spending policies can also lure private investment and labor (Stein, 1990). Governors can even serve as a public face to promote their state and its economic agenda. In several major U.S. cities, billboards showing California Governor Arnold Schwarzenegger’s visage, and proclaiming “Arnold Says: California Wants Your Business” is an example of one governor’s attempt to influence his state’s economy, whether effective or not (Sribnick, 2008, p. 121). In fact, voters expect a certain level of gubernatorial accountability, so much so that they often partly blame governors for many state problems, including the economy (Dometrius, 1999). My thesis discusses whether voters hold their governors accountable for these economic changes, and reveals what political factors help these influential men and women win their seats. In order to uncover what factors affect voters’ choice for U.S. President, Haynes and Stone (2008) found that low-income states vote differently than high-income states. Their results indicate “income-contingent” voting patterns: voters in low-income states respond more to income growth, while voters in high-income states do not even consider it when voting for U.S. President. I will extend the income-contingent voting hypothesis to gubernatorial elections to find whether voters in low-income states vote differently than their high-income counterparts. My empirical model improves upon Haynes and 3 Stone’s (2008) regression, and their disaggregate approach to the data.3 In addition to testing their hypothesis on gubernatorial elections, I also utilize a poor-state dummy variable to improve Haynes and Stone’s (2008) econometrics. Recent gubernatorial election work by Leal (2006) and Cummins (2009) contribute to our understanding of how our powerful governors are elected. Leal (2006) and Cummins (2009) are unique in finding that state-level crime rates affect gubernatorial election outcomes. They also contribute to the most recent literature by agreeing that campaign expenditures impact the vote for governor. Therefore as in Leal (2006) and Cummins (2009), I use crime rates and campaign expenditures as regressors. However, unlike Leal (2006) and Cummins (2009), I verify these findings with fixed effects. I also expand their data sets. Leal (2006) uses campaign expenditure data from 1980 to 1996 (185 data points), while my thesis more than doubles these observations to 392, and encompasses a much broader time period from 1976 to 2008. Similarly, Cummins (2009) works with a smaller period of data, including only the years from 1986 to 2004. Leal (2006) and Cummins (2009) do not use state and year fixed effects in their regressions, so I will improve the statistical strength of these and prior models. My model also re-tests the relatively new population variable, and adds brandnew variables such as real per-pupil education spending growth, duration of one-party control, and state-level inflation from Berry et al. (2000). Several other variables found to be significant in prior gubernatorial election research are also included. “Disaggregate approach” refers to Haynes and Stone’s (2008) separation of the 50 U.S. states into five groups based on per capita personal income. Then based on those five disaggregate groups, they run five separate regressions to achieve results for each group. 3 4 My results indicate that some level of income-contingent voting patterns exist as shown in Haynes and Stone (2008), but that new variables for crime rate growth, education spending growth, and duration of one-party control do not explain how voters choose their governors. In fact, only four factors (challenger party campaign spending, presidential approval in the month prior to the election, a dummy variable for gubernatorial incumbent candidates, and another for whether the incumbent party candidate shares the same political party as the President) consistently influence elections. These results support prior findings that national (specifically presidential) factors, campaign spending, and incumbency influence gubernatorial election outcomes. However, I do not find support for research that suggests state-level economic conditions or crime rates have an impact on these elections. This thesis is structured as follows: Chapter 2 reviews the relevant election outcome literature; Chapter 3 provides information about data sources, web links, descriptions, summary statistics, and other important notes about the data; Chapter 4 discusses the empirical model; Chapter 5 reveals the results of the regression; and Chapter 6 brings this information together for a final conclusion. 5 Chapter 2 LITERATURE REVIEW When choosing U.S. governors, what does the literature say about how voters in gubernatorial elections respond to state-level economic conditions (inflation, income growth, and unemployment) and political factors (crime rates and incumbency)? In addition, do poorer states vote on these matters differently than richer states? As far as economic matters are concerned, most gubernatorial election research finds that voters do not respond to state-level economic conditions when casting their ballots for governor. A few studies, however, find the opposite to be true.4 Some variables like per-pupil education spending, state-level inflation, and duration of one-party control have never been tested in gubernatorial research, but some literature influences the choice to include them in my research. Before reviewing the literature regarding on these variables and more, it is important to give background on some main concepts. My literature review is divided as follows: the first section explains the basic theoretical concepts found in all election literature; the second section discusses where/how my work fits into the entire body of gubernatorial research, which is divided 4 State-Level Unemployment: Levernier (1992, 1993), Kone and Winters (1993), Leyden and Borrelli (1995), Bardwell (2005) and Leal (2006) do not find unemployment to affect the gubernatorial vote. Only Cummins (2009), Cohen and King (2004), and Case (1994) find statistical significance. State-Level Income: Leal (2006), King (2001), Case (1994), and Kone and Winters (1993) find no added benefit/punishment for candidates for governor based on income growth. Chubb (1988) finds statistical, but not actual significance (the coefficient on income growth was very small). Niemi et al. (1995) found significance in disposable income growth over a two-year period. Adams and Kenny (1989) did not find either state income, or the difference between state income and national income, statistically significant. However, they find the “average deviation from predicted state personal income” over the governor’s term in office significantly affects the votes for governor. Inflation: Cohen and King (2004) found percentage changes in the national Consumer Price Index had an impact on the gubernatorial vote. 6 into two groups (individual survey and aggregate data research); and, the third section describes the specific literature that influenced my thesis. Main Concepts – The “Rational Voter” Hypothesis and Retrospective Voting The rational voter hypothesis explored by pioneers Pearson and Myers (1948) and Kramer (1971) form the foundation for nearly all election research. Kramer (1971) describes the rational voter hypothesis: “the notion that a vote represents a decision or rational choice between alternatives” (p. 131). Based upon the “rational voter” concept, Pearson and Myers (1948) hypothesize that those incumbent presidential parties in office during good economic times will stay in power, while the ones holding the bag in recessionary times will lose the office. The concept of “retrospective voting” evolves from the rational voter hypothesis. In retrospective voting, voters view changes to economic and noneconomic conditions occurring over a range of time, from the past to the present, to retrospectively evaluate how to vote, and to predict the potential future consequences of keeping or dumping the incumbent gubernatorial party. Empirical research has interpreted “retrospective voting” in a few different ways. Researchers have used independent variables that measure oneyear, two-year, and four-year (or one gubernatorial term) economic and noneconomic changes, to test retrospective voting. As an example of the wide-ranging types of retrospective voting that researchers have tested, state-level unemployment variable has been defined many different ways. Cohen and King (2004) use a one month lag measurement, Leal (2006) uses a one-year change in the rate, and Cummins (2009) uses a 7 percent change in the rate from the previous election to the current one. In this thesis, some of my regressors resulted in many different examples of retrospective evaluations: from small retrospectives (i.e. annual real per capita income growth), to two and fouryear changes in crime rates and education spending, to a duration variable which measures whether voters evaluate how many successive terms the incumbent party has been in the governor’s office. The retrospective voting concept first appears in Fiorina (1978). Fiorina (1978) states that retrospective voting “looks at the results rather than the policies and events which produce them” (p.430). My model (discussed in more detail in Chapter 4) tests retrospective voting because it does not, for example, look at the effects of a governor’s tax, education and/or economic policies themselves; rather, my model evaluates the results of those policies. The concepts of “rational voter” and “retrospective voting” form an umbrella over the modern election research. Whether voters’ decisions, based on retrospective evaluations, are rational is debatable and perhaps a good reason why the term rational doesn’t appear in the research as often as it did before 1980. Retrospective voting better describes this type of voter behavior. The next section explores the two types of gubernatorial election literature, which evaluates retrospective voting using either individual survey or aggregate data. My thesis falls squarely into the aggregate research approach. 8 Individual Survey Data versus Aggregate Research To determine the impacts of economic factors on the gubernatorial vote, the aggregate research relies mostly upon actual economic data (i.e. unemployment and growth rates); the individual-level research uses surveys (usually exit surveys where voters are asked about the state’s economy and the level of economic blame assigned to the governor). An important distinction between the two types of research: the survey data measures the effects of voters’ subjective opinions, while the aggregate data measures the effects of objective economic and noneconomic measurements, on the vote for governor. As Leal (2006) puts it “neither methodology holds a dramatic advantage other the other” (p. 46).5 Research has shown that voters will not just rely on their personal opinions on economic conditions, but will engage in “sociotropic” behavior: voters exhibiting this behavior will vote based on how the general health of the economy, instead of their own feelings about it.6 Therefore, I chose to work with aggregate data as it still holds the power to reveal the motivating factors behind voters’ choices for U.S. governors. Fiorina (1978) pioneered the use of individual-level, or survey data, to study presidential and congressional elections. Prior to Fiorina (1978), the typical approach used aggregate measurements. In gubernatorial election research, which originated in the 1980s, several papers use individual-level data, while several others focus on aggregate- 5 Kinder and Kiewiet (1979, 1981) affirm the value of aggregate studies over individual-level studies by demonstrating that voters sometimes discount their personal economic experiences, and vote according to actual economic conditions. 6 Schlozman and Verba (1979) found that unemployed voters did not exact a political price on candidates based on their personal unemployment experience. 9 level data (e.g. inflation, unemployment rates, etc.).7 On one hand, aggregate-level research has found that actual state-level economic measurements either do not have any effect, or they do not have as much of an impact as national economic factors on the vote. On the other hand, working with individual-level surveys, researchers have found that voters’ evaluation of state-level economic conditions do impact their votes. Since my research uses aggregate data, it is important to remember this difference. While most, but not all, of the aggregate studies find that state-level income has no voting repercussions, individual studies point to the opposite conclusion.8 One similarity between the two types of literature is the choice of presidential approval as a variable. Several individual survey studies (Atkeson and Partin (1995), Carsey and Wright (1998) and Svodoba (1995)) and many aggregate studies (see Table 4 in Chapter 3) find that presidential approval affects the gubernatorial vote. This variable is included in my model as well. Two studies illustrate aggregate findings. Kenney (1983) tests 14 states and finds state-level measures of unemployment, income, and inflation only impact three out of the 14 states’ votes for governor. Kenney (1983) conducted his regressions separately for each state, so the sample size is less than 16 in all cases. Therefore, Kenney’s (1983) results should be viewed with caution. In another aggregate study, Chubb (1988) simultaneously explores the effects of national economic conditions and state economic 7 Individual Survey Research: Atkeson and Partin (1995), Carsey and Wright (1998), Stein (1990), Svodoba (1995), Howell and Vanderleeuw (1990), Niemi et al. (1995), and Partin (1995) Aggregate-level Research: Peltzman (1987), Chubb (1988), Leyden and Borrelli (1995), Holbrook-Provow (1987), Levernier (1992, 1993), Adams and Kenny (1989), Kenney (1983), and Kone and Winters (1993) 8 Howell and Vanderleeuw (1990) concluded that voters’ opinions on how the governor handled the state’s economy had a direct impact on his/her approval ratings. Stein (1990) found that when voters assigned responsibility to the governor and his/her party for good (or bad) economic conditions, they held him/her accountable at the ballot box. 10 conditions on gubernatorial election outcomes between 1940 and 1982. Chubb (1988) found that national income growth affected gubernatorial election candidates of the president’s party four times as much as state-level income growth affected the gubernatorial incumbent party’s candidate.9 In addition, the coefficient on state-level income was very small. Many observations used in Chubb (1988) include southern states, where it was common prior to the 1970s, for the Democratic candidate to win between 90 and 100 percent of the vote. Since election observations used in my thesis exhibit much more competition between the two major political parties, the results should be more accurate. I will test state-level economic conditions, and national measurements (like presidential approval and a dummy variable for candidates sharing the same party as the President). Also, I improve Kenney’s (1983) work by testing unemployment, income, and inflation with a much larger data set. Each of my regressions will include at least 168 observations, more than ten times the observations used in Kenney’s (1983) regressions. The body of gubernatorial election literature, divided in two camps, show mixed results on the effects of state-level income, but similar results on presidential approval ratings impact. Additionally, some of the prior research has not used ideal research methods, or data, where my thesis should improve the results. The next section will highlight the specific literature that influenced my approach. 9 Holbrook-Provow (1987) also finds that national economic factors contribute to the success or failure of the incumbent presidential party in its pursuit of gubernatorial seats. Peltzman (1987) agrees and finds that national-level, not state-level, economic variables influence the election outcomes for governors: “The ... voter apparently understands that governors have little influence on the growth of state income, and he or she accordingly rewards (penalizes) the party of the incumbent president for good (bad) [national per capita income growth]”. 11 Literature Influencing My Approach In an effort to uncover the economic parameters that influence the complex voting decision, one recent paper on presidential elections uses an interesting and unique approach. Haynes and Stone (2008) divide the states into different income-level groups to see if low-income and high-income states vote differently for President. Using five groups of U.S. states, classified by the ten highest income states to the ten lowest income states, Haynes and Stone (2008) determine that the poorest states make presidential voting determinations based only on income growth, while the richest states consider many more factors including stock price changes, inflation, military spending, and the number of terms of one-party control.10 Haynes and Stone (2008) refer to these phenomena as “income-contingent” voting patterns. They conclude that poorer voters prioritize employment status over asset holdings, and wealthier voters do the opposite. Jayadev (2006) states that lower-income adults care less about inflation than their higherincome counterparts, and more about unemployment (p. 71). Leal (2006) stresses the importance of including campaign expenditures, or if not, risk model misspecification (p. 42). Besides the campaign expenditure variables, Leal (2006) notes that, according to Tidmarch et al. (1984), factors like crime and education typically receive more media attention than other matters in their gubernatorial election news coverage. And, Dometrius (1999) states, “those likely [issues] to show up Besides a voter’s income, other determining factors may be at play here. Voters in wealthier states may consider a wider variety of factors (such as military spending) in their voting decisions because their states’ higher incomes may indicate higher voter educational attainment or higher voter sophistication. For example, Cummins (2009) found that voters in states with moderate to higher percentages of college graduates evaluated crime rates in determining their votes for governor, with the highest educated states considering the crime rate more than the moderately educated states. States with the lowest educational attainment did not consider crime rates at all. 10 12 consistently on the agenda of governors – have been crime, education, [and] highways” (p. 60).11 So, in addition to economic variables, the non-economic variables of crime and education should reveal important information about voters’ decisions. As a result, Leal (2006) attempts an education variable for SAT scores, but he did not find statistical significance. However, Leal (2006) finds that changes in the state’s crime rate affected gubernatorial races. Cummins (2009) explores the crime variable in more depth than Leal (2006). Cummins (2009) investigates whether the incumbent political party, not just an incumbent, is punished (rewarded) for higher (lower) crime rates. Cummins (2009) concludes that while voters hold incumbents more accountable, the incumbent party is still held responsible for changes in crime rates that precede the gubernatorial election. Also, he finds that violent crimes affect the vote, while property crimes do not. Since crime rates are relatively brand-new to the research, I will test Cummins’s (2009) conclusions that incumbent gubernatorial parties are answerable to changes in a state’s violent crime rate. I also add fourteen years to the data set: my range of observation is from 1976 to 2008, while Cummins data set is from 1986 to 2004. Since previous economic literature has not uncovered Haynes and Stone’s (2008) “income-contingent” voting patterns, my research tests these voting patterns to verify their reliability and applicability on the gubernatorial vote. This thesis will differ from Haynes and Stone (2008) because I divide the 50 U.S. states into two income groups, rather than five. This increases the sample size of the subgroups, which leads to more 11 Highways are excluded due to the difficulty of quantifying that data. 13 precise estimates of the political and economic variables being tested. Additionally, I use a poor-state dummy variable to improve econometric comparisons between the low- and high-income states. I also borrow the duration of one-party control variable from Haynes and Stone (2008) and other presidential election research, since it has not been tested by gubernatorial election researchers. Additionally, Haynes and Stone’s (2008) use of a national inflation measure influenced my search for a state-level measure of inflation. I found a brand-new inflation measurement to add to the gubernatorial election literature. The inflation data are from Berry et al. (2000). Leal (2006) and Cummins (2009) also influenced the addition of other characteristics to my empirical model. I include both incumbent and challenger party campaign spending in my empirical model. In response to Leal (2006), Tidmarch (1984), and Dometrius (1999), I added regressors for education and crime. I choose changes in each state’s real per-pupil education spending, rather than SAT scores, since the media sometimes report per-pupil education spending, and their state’s ranking, when reporting on education. This enables us to test Leal’s (2006) hypothesis that education impacts voters’ choice for governor. My thesis also increases the sample size and date range used by Leal (2006). In the next chapter, I will discuss the data used to test: whether voters make retrospective evaluations when voting for governor, the new concept of “incomecontingent” voting introduced by Haynes and Stone (2008), the crime variable explored by Cummins (2009), and a new education variable inspired by Leal’s (2006) attempt to discover relationships between education and the gubernatorial vote. 14 Chapter 3 DATA This chapter introduces the data descriptions, summary statistics, sources, web links, and relationships used in my thesis. Economic and political factors, from the years 1976 to 2008, are regressed to determine their effects on the gubernatorial vote. The elections data exclude the samples with missing campaign expenditure data, and when third parties won more than 15 percent of the vote or when third parties represent the incumbent gubernatorial party.12 After the data were filtered, some Louisiana elections were excluded because of that state’s unique open primary-style general elections. Louisiana elections from 1991, 1995, and 2003 were kept because they resulted in a runoff, mimicking a normal election. To test whether voters in low-income states vote differently than voters in highincome states, it is helpful to group the states. Unlike Haynes and Stone (2008), I divide states into two income groups, rather than five. The per capita personal income data are from the Bureau of Economic Analysis: Regional Economic Accounts SA1-3 Personal Income Summary.13 The low-income group is comprised of states below the per capita personal income average in 1993, the mid-sample year; the high-income states are above that average. In addition to grouping the states based on their income levels, and then 12 Other researchers have filtered elections resulting in high third party vote percentages, such as in Leal (2006). Leal (2006) excludes elections when third parties won more than ten percent of the vote 13 A web link to BEA’s SA1-3 Personal Income Summary: http://bea.gov/regional/spi/default.cfm?selTable=summary 15 separating the regression, I use a poor-state dummy variable for one regression and then interact that term with all of the economic and political variables.14 Table 1 Top Ten States, 1978 and 2008 (ranked by per capita personal income) 1 2 3 4 5 6 7 8 9 10 1978 Alaska Nevada* Connecticut California Wyoming New Jersey Illinois* Hawaii* Maryland New York 1 2 3 4 5 6 7 8 9 10 2008 Connecticut New Jersey Massachusetts* New York Wyoming Maryland Virginia* Alaska California New Hampshire* *states that migrated to/from the top ten. Table 1 shows the top ten states with the highest per capita personal income in the election years of 1978 and 2008. Although 30 years separate the two observations, the 2008 ‘top ten’ group has remained mostly stable, adding only three states: Virginia and New Hampshire (formerly in the top thirty states), and Massachusetts (formerly in the top twenty states). Nevada, Illinois, and Hawaii dropped out of the top ten, but not far: only to seventeenth, fourteenth, and fifteenth places, respectively. My state classifications are not much different than Haynes and Stone’s (2008) classifications, so comparisons between our results can be made. More than 75 percent of Haynes and Stone’s (2008) top 20 per capita personal income states, from their sample year 1960, were found in my two income groups from 1993 (their top 20 states match my 14 The poor-state dummy variable is 1 for states below the median per capita personal income level. The variable was allowed to change over the years 1976 to 2008 if a state rose above or fell below the median level as time progressed. 16 top 21 states except in five cases, Hawaii, Virginia, New Hampshire, Minnesota and Florida). On the following page, Table 2 presents data descriptions and summary statistics. Table 3 reports the data sources and web links. 17 Table 2 Variable Descriptions and Summary Statistics, 1976-2008 Variable Description Mean SD Min Max Incumbent Party Candidate Vote* The incumbent gubernatorial party candidate’s share of the gubernatorial vote 53.80 9.62 25.85 82.37 Normal Vote* Average incumbent party vote in the last two gubernatorial elections 1 if an incumbent is running, zero otherwise 53.89 7.96 27.46 78.84 0.58 0.49 0.00 1.00 Number of successive terms the incumbent party has been in the governor’s office. Twoyear terms = 0.5; four-year terms = 1.0 The percentage change in statewide violent crime rates per 100K people, from the previous election year to the present one. 3.34 5.38 0.50 32.5 4.55 20.53 -48.10 110.76 Per-Pupil Education Spending Growth Rate* The percentage change in real state-level perpupil education spending from the previous election year to the present one. 10.16 11.13 -16.55 58.24 State-Level Income Growth* The annual growth rate of per capita real GDP by state in the year of the election 1.00 3.53 -30.28 10.18 Relative Unemployment One month prior’s state unemp. rate minus one month prior’s national unemp. rate -0.39 1.48 -4.90 6.10 State-Level Inflation Rate* 4.29 2.53 0.68 12.48 14.94 1.02 12.97 17.40 Challenger Party Campaign Spending The change in the inflation rate, from the previous election to the present one. Using a state-level index from Berry et al. (2000), updated through 2007. Extended into 2008 using CPI. Dummy variable. 1 for states above the population median of 3,466,568, zero otherwise Natural log of real challenger party candidate spending in primary and general elections divided by the state’s population 1.23 1.12 0.00 8.99 Incumbent Party Campaign Spending Natural log of real incumbent party candidate spending in primary and general elections divided by the state’s population 1.52 1.24 0.03 16.32 Presidential Approval Rating* Average presidential approval rating by major pollsters in the month prior to the election. If the incumbent party is the same party as the President, Approval Rating = Approval Rating. Otherwise, Approval Rating = 100 – Approval Rating. 1 if the incumbent party is the same as the President’s Party, zero otherwise. 49.05 11.19 26.50 88.00 0.54 0.50 0.00 1.00 Incumbent Running Duration of OneParty Control State-Level Violent Crime Growth Rate* High Population Same Party as President *all rates/percentages were multiplied by 100 18 Table 3 Variable Sources and Web Links Variable(s) Political Variables (Incumbent Party Candidate Vote; Normal Vote; Incumbent Running; Duration of OneParty Control; and Same Party as President) Source(s) Web Link(s) All political variables are from the following three sources: 1. Dave Leip's Atlas of U.S. Presidential Elections, for years 1990-2008 1.http://uselectionatlas.org/RESULTS/ 2. Glashan (1979) for years 1976-1977 and Mullaney (1988) for years 1978-1987 2. none 3. Wikipedia, The Free Encyclopedia 3. e.g. http://en.wikipedia.org/wiki/ List_of_Governors_of_Wyoming Presidential Approval Rating Roper Center, Public Opinion Archives, University of Connecticut http://webapps.ropercenter.uconn.edu/ CFIDE/roper/presidential/webroot/ presidential_rating.cfm State-Level Violent Crime Growth Rate U.S. Census Bureau, Statistical Abstract of the United States http://www.census.gov/compendia/ statab/past_years.html Per-Pupil Education Spending Growth Rate U.S. Census Bureau, Statistical Abstract of the United States and for missing data, the National Education Association, NEA Research Rankings & Estimates database http://www.census.gov/compendia/ statab/past_years.html State-Level Income Growth Bureau of Economic Analysis: Regional Economic Accounts http://www.bea.gov/regional/gsp/ State-Level Inflation Rate 1. Berry et al. (2000) 1976-2007 2. Bureau of Labor Statistics Relative Unemployment Rate 1. U.S. Department of Labor, Bureau of Labor Statistics as reported in Economic Research Federal Reserve Bank of St. Louis 2. U.S. Department of Labor, Bureau of Labor Statistics Local Area Unemployment Statistics 1.http://mailer.fsu.edu/~wberry/garnetwberry/statecpi2007.zip 2. http://www.bls.gov/CPI/ 1. U.S. Civilian Unemployment Rate from http://research.stlouisfed.org/ fred2/series/UNRATE 2. Unemployment rates for all 50 U.S. states from http://www.bls.gov/lau/ Challenger & Incumbent Party Campaign Spending The Gubernatorial Campaign Finance Database http://www.unc.edu/~beyle/guber.html High Population Bureau of Economic Analysis: Regional Economic Accounts http://bea.gov/regional/spi/ default.cfm?selTable=summary 19 Table 4 reports the independent variables used by previous researchers. Each row identifies (for any one variable) whether prior studies have found their impact on the gubernatorial vote to be statistically significant, or not. In a few cases, the research reported mixed results. S S S S S S N N S Leal (2006) S N Bardwell (2005) N S S Lowry et al. (1998) M S Leyden & Borrelli (1995) Incumbent N S N Spending* Presidential Approval* S Same Party as S M S President* Presidential N N Election Year Ideology S Partisanship S S Unified Gov’t N M N Nat’l Econ. M S Measures Candidate’s Age S= statistically significant; N=not statistically significant; M=mixed results *used as regressors in my research Challener Spending* NormalVote* Incumbent Running* Crime Growth* Growth* Unemployment* Inflation* Population* Variable Cummins (2009) N N S N N S Levernier (1992, 1993) S N N N S Kone & Winters (1993) S S N M S Chubb (1988) S S S S N S S S M S Cohen & King (2004) S S S M King (2001) Table 4 Variables and Their Statistical Significance in Prior Gubernatorial Election Research S N S N Case (1994) S S S S S Niemi et al. (1995) M 20 21 Table 4 shows some variables not used in my thesis. Rather than using state ideology or state partisanship variables, I used the average of the previous two elections’ results for the incumbent party candidate as a proxy. Prior research has identified this average as a statistically significant variable. Partisanship and ideology variables were available, but the restricted nature of this data would have reduced my sample size. The presidential election year and unified government dummy variables have not shown statistical significance in most prior work, and, therefore, were not selected for testing.15 National economic measures were not selected since my thesis focuses mainly on testing state-level economic measurements. Finally, although candidate’s age was not chosen for testing, it should be tested in future work because it established its worth in Case (1994). The unified government dummy variable is 1 when the governor’s party controls the majorities of both state houses. And, zero otherwise. 15 22 New to the gubernatorial election research, real per-pupil education spending tests the hypothesis that higher education spending growth increases the incumbent party’s share of the vote. As seen in Figure 1, this may not be the case. As the incumbent party increases per-pupil education spending, their vote share declines slightly, but this finding is not statistically different from zero. The regression results in Chapter 5 will reveal whether ceteris paribus, the lack of any relationship between education spending and the vote remains true. Figure 1. Growth in Education Spending and the Vote 90 IncVote = 54.07 - 0.03*EGrowth (0.04) 80 Incumbent Vote 70 60 50 40 30 20 -20 -10 0 10 20 30 40 Education Spending Growth 50 60 23 Another new variable, measures the effects of the consecutive terms that one party has controlled the governor’s office. As seen in Figure 2, it appears that this has no impact on the vote that the incumbent party receives on Election Day. However, since many Southern states had one-party rule by the Democratic Party for nearly 100 years, there is a set of outlier data. Figure 2 shows the outliers on the right: all Southern states with more than 100 years of Democratic gubernatorial control.16 Figure 2. One-Party Control and its Impact on the Vote 90 IncVote = 53.60 + 0.06*Duration (0.09) Incumbent Party Vote 80 70 60 50 40 30 20 0 5 10 15 20 25 30 35 Number of Four-Year Terms of One-Party Control 16 Georgia (seven occurrences), Mississippi (four occurrences), Alabama (three occurrences), Texas (one occurrence), and Louisiana (one occurrence). 24 In Figure 3, these outliers are excluded, but the duration of one-party control variable still appears to have no impact. Once again, despite this apparent lack of relationships, we need to see if they hold up in a regression that holds all other variables constant. Chapter 5 reveals the results. Figure 3. One-Party Control and its Impact on the Vote (less 16 outliers) 90 IncVote = 54.44 - 0.33*Duration (0.30) Incumbent Party Vote 80 70 60 50 40 30 20 0 2 4 6 8 10 12 Number of Four-Year Terms of One-Party Control The last new variable that I add to the gubernatorial research, state-level inflation, comes from a state-level inflation index found in Berry et al. (2000). Haynes and Stone (2008) found “income-contingent” voting patterns partially based on their inflation coefficients: specifically, they found that high-income state voters vote negatively in relationship to higher inflation, and low-income voters do not respond at all. Figures 4 through 6 show state-level inflation’s impact in the gubernatorial vote in all states (Figure 25 4) and the two income groups, Y1 – the low-income states (Figure 5) and Y2 – the highincome states (Figure 6). All three groups show no relationship between inflation and the gubernatorial vote. This preliminary look does not indicate that the income groups vote differently based on inflation. None of the inflation coefficients are statistically significant. The next chapter will introduce the model used to determine whether these new variables impact the vote, and whether the vote is “income-contingent”. Figure 4. Inflation and the Vote, All States 90 IncVote = 54.09 - 0.02*Inflation (0.04) Incumbent Party Vote 80 70 60 50 40 30 20 0 10 20 30 Inflation 40 50 60 26 Figure 5. Inflation and the Vote, Y1 Group (low income states) 90 IncVote = 53.77+0.01*Inflation (0.05) Incumbent Party Vote 80 70 60 50 40 30 0 10 20 30 40 50 Inflation Figure 6. Inflation and the Vote, Y2 Group (high income states) 90 IncVote = 54.50-0.04*Inflation (0.06) Incumbent Party Vote 80 70 60 50 40 30 20 0 10 20 30 Inflation 40 50 60 27 Chapter 4 EMPIRICAL MODEL Fair’s (1978) presidential election research provides the basis for the empirical model used in my thesis. His model equates voter utility to economic and noneconomic factors, so that: E(Utility) = U [Economic Factors, Non-Economic Factors] In the model, which also allows for an analysis of the retrospective voting hypothesis described in Chapter 2, voters choose the candidate that best maximizes their expected future utility. As the relationship E(Utility) = U [Economic Factors, NonEconomic Factors] illustrates, this voter utility derives from a combination of economic and noneconomic factors. In order for voters to predict their future utility with the Democratic or Republican gubernatorial candidate, they may rely on their positive or negative judgments of how economic and noneconomic conditions changed during the gubernatorial incumbent party’s term in office. Additionally, non-economic factors such as political circumstances may play a role (e.g. whether an incumbent is running, or is in the same political party as the President) in how voters evaluate their expected future utility. I also add campaign expenditures, as Leal (2006) notes, because excluding this independent variable can lead to model misspecification. 28 This model tests what economic and political (non-economic) factors influence voters’ decisions on how to vote in order to maximize their expected future utility.17 The following equation shows that as the voter’s predicted utility from the challenger party candidate falls, the dependent variable, the vote for the incumbent party rises: Incumbent Party Vote = prob [U(Challenger Party Candidate) < U(Incumbent Party Candidate)] The dependent variable is the vote share of the incumbent gubernatorial party’s candidate for governor. Multiple researchers have also used this dependent variable.18 It is simpler and easier to interpret than an alternative dependent variable, of Democratic share of the gubernatorial vote, which involves interacting each of the independent variables with a 1 for Democrat incumbents and a -1 for Republican incumbents and leads to multicollinearity. Below is my basic regression: Incumbent Party Vote = f [C, Normal Vote (+), Incumbent Running (+), Duration of One-Party Control (–), State-Level Violent Crime Growth Rate (–), Per-Pupil Education Spending Growth Rate (–), State-Level Income Growth (+), Relative Unemployment Rate (–), State-Level Inflation Rate (–), High Population (+), Challenger Party Campaign Spending (–), Incumbent Party Campaign Spending (+), Presidential Approval Rating (+), Same Party as President (–)] 17 Economic Factors: State-Level Income Growth, Relative Unemployment Rate, State-Level Inflation Political Factors: Normal Vote, Incumbent Running, Duration of One-Party Control, State-Level Violent Crime Growth Rate, Per-Pupil Education Spending Growth Rate, High Population, Challenger Party Campaign Spending, Incumbent Party Campaign Spending, Presidential Approval Rating, and Same Party as President. 18 Cummins (2009), Leal (2006), Bardwell (2005), and Leyden and Borrelli (1995) 29 Expected signs agree with previous election research. The expected signs on regressors measuring the duration of one-party control and the per-pupil education spending growth (both new variables to this research) are based on preliminary tests done in Chapter 3. My model will improve upon Haynes and Stone’s (2008) econometrics. Haynes and Stone (2008) divide the regression into low and high-income groups, rather than running one large regression with all states. Therefore, comparisons of coefficients across the separate regressions are difficult to interpret, and unreliable. So, as an alternative to using high- and low-income state groups, I use a preferred method of a low-income state dummy variable in the larger combined regression, so that the coefficients are more easily compared between low and high-income states. I will test this gubernatorial income-contingent voting hypothesis: On factors related the vote, voters in lower-income states vote differently than voters in higher-income states. In line with Haynes and Stone (2008), I expect that low-income state voters will vote differently on some factors compared to high-income voters. I also predict that, similar to Cummins (2009), due to the likelihood of lower education levels in low-income states, my expectations are that voters from poorer states will be less sophisticated in their choices.19 For example, voters in richer states, with their assumed higher-levels of voting sophistication, will conversely be more influenced by issues such as crime and 19 Cummins (2009) describes the literature supporting the correlation between low-education states and lower voter sophistication. I found a link between low-education to low-income states; seven of the top ten low-income states show up in the ten states with the lowest educational attainment (as measured by the percentage of the population with a Bachelor’s Degree or more, in the year 1990). These states are Alabama, Arkansas, Kentucky, Louisiana, Mississippi, South Carolina, and West Virginia. 30 education, represented by variables for the state-level violent crime and per-pupil education spending growth rates. I will also test the effects of relatively new independent variables (population, campaign spending, and the state-level crime growth rates), and the brand-new regressors (duration of one-party rule, state-level inflation, and the growth in education spending) on the vote for governor. If the results are not in line with my expectations, the model still re-tests many historically significant variables, using fixed effects.20 A new story will likely emerge, which illustrates how voters choose their governors. The next chapter will present these results. 20 This thesis also employs random effects, White cross-section robust standard errors, and a time trend to ensure the strength of the results. 31 Chapter 5 RESULTS The empirical model in Chapter 4 suggests that economic and political factors can influence election outcomes. In this chapter, I use the empirical model to test for these influences, and whether low-income and high-income state voters weigh these economic and political factors differently in making their choice for governor. I estimate fixedeffects regressions and explain the results. The first section discusses the results for all states, regardless of income; the second section analyzes the results for two groups of states: low-income and high-income. In Table 5, the “All Races” column presents OLS estimates for all 392 gubernatorial election observations. Regardless of a state’s per capita income level, among the 13 variables chosen for testing, voters’ choice for governor is influenced by four factors: challenger party campaign spending, presidential approval, whether the incumbent party candidate is in the same political party as the President, and whether the incumbent governor is in the race. The remaining columns in Table 5 present the estimates for U.S. states divided into two time periods: the earliest half of the sample group (from 1976 to 1992) and the second half of the sample period (from 1993 to 2008). All regressions in Table 5 employ fixed effects, and White cross-section robust standard errors. 32 Table 5 Regression Results: Incumbent Party Candidate’s Share of Gubernatorial Vote, 1976-200821 Variable Intercept All Races 44.98*** (5.22) All Races (1976-1992) 54.38*** (6.25) All Races (1993-2008) 53.05*** (6.73) 0.07 (0.07) 0.02 (0.09) -0.11 (0.09) 6.86*** (0.87) 5.72*** (1.09) 6.88*** (1.46) Duration of One-Party Control State Violent Crime Growth -0.10 (0.14) 0.09 (0.17) -0.37** (0.17) -0.01 (0.03) -0.06** (0.03) -0.01 (0.04) Education Spending Growth Rate State Income Growth 0.05 (0.04) 0.04 (0.03) -0.05 (0.07) 0.15 (0.13) 0.02 (0.18) -0.02 (0.30) Relative Unemployment -0.20 (0.44) -0.52 (0.37) -1.65* (0.91) State Inflation Rate 0.03 (0.11) -0.04 (0.07) 0.03 (0.18) High Population 0.06 (2.66) 6.56 (4.28) 1.70 (1.96) Challenger Spending -3.68*** (1.09) -6.23*** (0.67) -2.53*** (0.91) Incumbent Spending -0.01 (0.44) -0.01 (0.31) 0.34 (0.66) Presidential Approval 0.13*** (0.05) 0.00 (0.05) 0.14*** (0.05) Same Party as President -3.18*** (1.25) -1.18 (1.59) -3.27* (1.71) Adjusted R2 S.E. of Reg. n Fixed Effects 0.39 7.51 392 Both 0.47 6.98 202 State only 0.50 6.86 190 State only Normal Vote Incumbent Running 21 Estimated with fixed effects and White cross-section robust standard errors. All regressions have normally distributed residuals and passed the Jarque-Bera normality test. The residuals do not exhibit autocorrelation or partial autocorrelation. The squared residuals do not show signs of ARCH. 33 The results in Table 5 were tested using a variety of econometric tools including: the Jarque-Bera Normality Test (for normally distributed residuals); Correlograms and QStatistics (to test residuals for autocorrelation and partial autocorrelation; and to test squared residuals for autoregressive conditional heteroskedasticity); Variance Inflation Factors (to check the regressors for multicollinearity); and F-Tests (to test whether the group of economic variables is jointly significant, since this group of variables may be highly collinear). The regression was also run in a number of different combinations (removing and adding variables, adding a time trend, and using random instead of fixed effects) to check the robustness of the significant coefficients. In nearly all cases, test results were normal and coefficients maintained their significance and sign.22 Results for all States (Regardless of Income) This section reviews the aggregate regression results for all 50 U.S. states, as shown in the first column of Table 5. Economic Impacts on the Vote As shown in Table 5, variables measuring the effects of economic changes (statelevel income growth, unemployment, and inflation) did not have a statistically significant impact on the gubernatorial vote. Even when using different measurements for unemployment, such as level, one-year, and four-year changes did not result in 22 Variance Inflation Factors testing for multicollinearity for two variables (inflation and population) were problematic. Inflation (originally measured as a one-year change) resulted in an unacceptably high factor of 20. Using four-year change in inflation reduced this VIF to 12.5. While not low enough (VIFs should be lower than 5 or 10), when inflation is removed from the regression, the other coefficients and their significance do not change. Population, and log of population, resulted in a VIF near 100! Using a dummy for High Population (1 for states above the population median, zero otherwise) reduced the VIF to 12.5. Although too high, removing this variable also yields the same results. 34 statistically significant coefficients.23 Additionally, after removing all of the noneconomic variables from the regression, and re-estimating it with only the economic variables, the regression does not yield significant results. The F-test indicates that we cannot reject the null hypothesis that the economic factors are jointly equal to zero.24 This is not surprising considering the disagreement in prior research over whether state-level economic variables (especially changes in unemployment and income) have any impact at all on gubernatorial elections. The evidence weighs in the direction of no impact for these two variables.25 However, regarding inflation, evidence is lacking. Prior research has not attempted to use a state-level measure of inflation. I use a formerly untried statelevel index from Berry et al. (2000) for state-level inflation measurements, but it did not show any statistical significance. Since the Berry et al. (2000) inflation data is new to the research, more exploration of the variable should be done.26 Results for inflation, unemployment, and income do not square with the retrospective voting concept. The results show that voters do not look “retrospectively” at unemployment, inflation, and income changes to decide how to vote. Overall, the lack of 23 When fixed effects were removed, unemployment was significant and the expected negative sign, such as in the latter half of the sample period in Table 5, but it was not consistent in a variety of different specifications. 24 The F-Statistic is 1.00 and the p-value is 0.48 25 State-Level Unemployment: Levernier (1992, 1993), Kone and Winters (1993), Leyden and Borrelli (1995), Bardwell (2005) and Leal (2006) do not find unemployment to affect the gubernatorial vote. Only Cummins (2009), Cohen and King (2004), and Case (1994) find statistical significance. State-Level Income: Leal (2006), King (2001), Case (1994), and Kone and Winters (1993) find no added benefit/punishment for candidates for governor based on income growth. Chubb (1988) finds statistical, but not actual significance (the coefficient on income growth was very small). Niemi et al. (1995) found significance in disposable income growth over a two-year period. Inflation: Cohen and King (2004) found percentage changes in national Consumer Price Index had an impact on the gubernatorial vote. 26 Berry et al. (2000) extensively tested their state cost-of-living indexes for reliability and validity. Other research has used their state-level cost-of-living indexes, but not election research. 35 significance in state-level economic variables influencing the gubernatorial vote throws cold water on the hypothesis that state-level economic factors can influence elections. The significance in measurements of satisfaction with the President’s party, which undoubtedly contain some measure of voters’ approval of the President’s handling of the national economy, may support the theory that national (and not state) economic measurements influence the vote. Political Impacts on the Vote Similar to economic variables, some political factors did not show significance. These statistically insignificant variables include the state-level violent crime growth rate, per-pupil education spending growth rate, duration of one-party control, the average of the past two election results for the incumbent party, a dummy variable for highpopulation states, and incumbent party campaign spending. However, some political factors did show significance including: challenger party campaign spending, the presidential approval rating, whether the incumbent party is the same party as the President, and a dummy variable for incumbent candidates. First, discussion centers on the insignificant political factors, followed by the significant political factors. My results indicate no relationship between violent crime growth rates over a four-year period (or two-year period in states where gubernatorial elections are held every other year, such as in New Hampshire) and the gubernatorial vote outcome. However, when running my regression for the years 1976 to 1993, prior to the 1994 federal crime law, state-level violent crime growth rates are statistically significant at the 36 five-percent level, and the expected negative sign. Results indicate that for every fifteen percent increase (decrease) in the violent crime growth rate over the gubernatorial term, the incumbent party vote decreases (increases) by one percent. Cummins (2009) notes a similar finding that the pre-1994 effects of crime rates on gubernatorial elections were stronger. Federal involvement in may have “muddied the waters” with respect to who is held accountable for changes in the crime rate: the federal or the state government (Cummins, 2009, p. 645). However, Cummins (2009) still finds that crime is significant post-1994. Consequently, contrary to Cummins (2009) and Leal (2006), my regression results do not support that crime rates continue to impact gubernatorial races beyond 1994. This thesis does not agree with the hypothesis that crime rates impact gubernatorial races.27 Another variable not showing significance, and one that is brand-new to the gubernatorial election literature, is real per-pupil education spending growth over the incumbent party’s term in office. Additionally, when using level education spending (instead of growth), the coefficient remains insignificant. Therefore, education spending does not have a positive or negative effect on the vote. Leal (2006) also found that an education measurement (changes in SAT scores) did not impact candidates for governor. Per-pupil education spending could theoretically impact the vote for governor because education is one of the media’s favorite topics when discussing the gubernatorial election (Tidmarch et al., 1984). Also, the there is a reasonable expectation that some voters will 27 Also, unlike Cummins (2009), even when disaggregating the states into two groups based on educational attainment, neither the highly-educated states’ nor the lower-educated states’ voters appear to consider crime rates when choosing their governors. 37 hold the governor accountable for issues under state control like crime and education (Atkeson and Partin, 1995). However, as per-pupil education spending and SAT score growth do not appear to impact the vote for governor, perhaps another measurement such as graduation rates could show significance. One political variable measures the duration one political party has occupied the Governor’s Mansion without interruption. This variable did not produce statistically significant coefficients in the “All Races” column. The duration of one-party control variable is new to the literature, as it appears untested in prior work. It shows up in presidential election research, so it seemed like a reasonable addition to another executive-type election, such as a gubernatorial election.28 Governors have high visibility and executive power in the state, so it is reasonable to suspect voters may want a change in party from time to time. Glaring exceptions to this desire for a “change of pace” are 16 election data points where duration of one-party control exceeds 25, or 100 years, of single-party rule. These observations all occur in Southern “Dixiecrat” states like Georgia. In Georgia’s case, until 2003 it had not had a Republican governor since Reconstruction! When these 16 outliers are removed, the results in the “All Races” column change slightly; the regressors maintain their significance and signs, except for the duration of one-party control variable, which is significant at the ten percent level and the expected negative sign.29 It represents a fall of about 0.6 percentage points for each full four-year term of one-party control. This may seem small, but according to this See Haynes and Stone (2008) for an example of the duration variable’s use in the presidential election research. 29 Since these 16 observations all occur in the low-income states group, I also re-estimated the low-income states regression in Table 6, but we get the same results. 28 38 model, in states like Hawaii in 2002, where Duration of One-party Control was as high as 10, the challenger party candidate lost six percentage points. This finding and variable are new to the literature. A variable representing the average of the past two incumbent party gubernatorial vote results does not show any statistically significant effect on the gubernatorial vote. As shown in Table 4, prior literature has found that in many cases this variable is significant. My thesis’ use of fixed effects may explain the difference between my results and past results. Previous research has not selected population very often for testing, but my results show that a variable for higher population does not affect the incumbent party’s candidate.30 Partin (2002) argues that higher population positively affects the gubernatorial incumbent party’s vote because the incumbent party has control over the governor’s office, which makes it easier for that party, in highly populated states to reach out to a multitude of large media markets, and therefore, reach the voters to influence their vote. In contrast, it is more difficult for most challengers to use larger populations to their advantage. My results do not confirm Partin’s (2002) hypothesis. Challenger party campaign spending strongly affects the vote for governor. No impact was found for incumbent party spending. So, if California’s 2010 Republican candidate Meg Whitman truly spends up to $150 million of her own money on the election, she should not expect a benefit on Election Day because she is the incumbent party candidate. Fortunately for California’s 2010 Democratic candidate Jerry Brown, 30 Cummins (2009) found population to be statistically significant, with a positive sign. Cohen and King (2004) did not find population to be statistically significant. 39 challenger party spending has an impact equal to four percentage points for each additional dollar spent per voter. In California’s 2010 election, that means that $37 million additional challenger party candidate spending nets the candidate (Jerry Brown) four percentage points. The lack of any impact from incumbent party campaign spending is not new to the literature. For example, Leal (2006) and Bardwell (2005) use TSLS and instrumental variables for incumbent party spending, but they disagree on whether incumbent spending affects the vote. However, some specification concerns exist regarding the instruments chosen for their TSLS regressions.31 Similar to Cummins (2009) who does not use instrumental variables, or TSLS, I find no effect from incumbent party spending. Other literature has found campaign spending to be significant, and agrees that challenger spending has a much more substantial impact than incumbent spending.32 One reason for challenger spending dominance could be that a dummy variable for incumbency already measures the benefit of being an incumbent in the regression. The President’s approval rating is highly significant and the expected positive sign. I transformed actual presidential approval so that increases benefit the dependent variable, the vote for the incumbent party candidate. When the incumbent gubernatorial party is Democratic, approval equals the actual approval of a Democratic president, or in 31 I did not subject incumbent party spending to TSLS regression as Bardwell (2005) and Leal (2006). The instrumental variables they chose did not appear to be unrelated to the dependent variable as necessary. Some of the instrumental variables such as a candidate’s age, population, normal party vote, and the state’s unemployment rate have all been shown to be significant variables in the predicting the vote for governor. Since this subject goes beyond the scope of my thesis, I will not explore it more depth. Future research should seek out better instruments to perform a proper TSLS regression for incumbent party spending. 32 Cummins (2009) and Bardwell (2005) found challenger spending significant, but not incumbent spending. Leal (2006) found both types of spending were statistically significant. Partin (2002) found that challenger spending better measures the competitiveness of a gubernatorial election. 40 the case of a Republican president, approval equals 100 minus the actual approval rating.33 For example, when the incumbent gubernatorial party is Democratic, George Bush’s 2002 approval rating is transformed to 35.2 (64.8 when the incumbent gubernatorial party is Republican), and his 2008 approval equals 73.5 (26.5 for a Republican gubernatorial incumbent party). This increase of approximately 40 points translates into more than a five-percent benefit to Democratic gubernatorial candidates running in 2008 compared to those in 2002. In fact, Democrats did very well in 2008, winning seven of ten contests included in the sample. Conversely, in 2002, Republicans won 21 of 36 elections. Table 4 shows that prior gubernatorial election research agrees on the significance of presidential approval on gubernatorial elections.34 One way to show dissatisfaction with the President is to vote against his party in gubernatorial elections. In fact, since many voters do not absorb a lot of information prior to voting, and may not differentiate much between national and state politicians, these voters may associate all politicians of one party with the President’s party, and therefore that party receives a backlash, or reward in gubernatorial elections (Carsey and Wright, 1998). In fact, not only do presidential approval ratings matter, but holding these ratings constant, so does whether the incumbent gubernatorial party is in the same party as the President. Results indicate that this condition is statistically significant, and accounts for a loss of more than three percentage points for the incumbent party candidate. As shown in Table 4, prior research concurs with this finding. In an often-cited example of how 33 The opposite transformation occurs when the incumbent gubernatorial party is Republican. Cummins (2009), King (2001), Cohen and King (2004) and Leyden and Borrelli (1995) found presidential approval to be both positive and significant. 34 41 sharing the President’s Party results in a loss of votes, in New Jersey and Virginia (the only two states that have gubernatorial elections in the odd-numbered year after the presidential election) the last six consecutive elections have brought a candidate to the governor’s office of a different party than the President. The theory that voters desire divided government, that is the President of one party and other sub-national elected officers of the opposing party, is supported in these results. Additionally, as can be seen in the third column of Table 5 (the latter half of the sample period, 1993 to 2008), this factor and presidential approval ratings may have a stronger effect on contemporary gubernatorial elections, than those elections prior to 1993. The benefit of being an incumbent gubernatorial candidate is large and significant. An incumbent candidate can expect a five-percent boost to his/her vote outcome on Election Day. This agrees with the same well-established pattern shown in past research.35 Many reasons exist for an incumbent boost. Incumbents usually receive a higher percent of the vote on Election Day than non-incumbents (Cummins, 2009). In most cases, voters have more familiarity with incumbents than non-incumbents. The next section discusses four variables, which showed considerable statistical strength across all of the regressions that were tested.36 Further, this section discusses tests whether a hypothesis put forward by Haynes and Stone (2008) called “incomecontingent voting patterns” reoccurs in my research or not. 35 Cummins (2009), Lowry et al. (1998), Leyden and Borrelli (1995), Niemi et al. (1995), Kone and Winters (1993), Levernier (1992,1993), and Chubb (1988) 36 These four variables (a dummy variable for incumbency, challenger party campaign spending, a dummy variable for incumbent party candidates in the same party as the President, and presidential approval) kept their significance with no fixed effects, with the addition of a time trend, random effects, White crosssection robust standard errors, and both state and year fixed effects. 42 Results for Low-Income Vs. High-Income States Table 6 presents the estimates for U.S. states divided into groups based upon their per capita personal income. The low-income group is for states below the per capita personal income average in 1993, the mid-sample year; the high-income states are above that average.37 The remaining columns in Table 6 show results for when a poor-state dummy variable is added to the regression, and then interacted with all of the economic and political variables. The last column reports results without using state or year fixed effects. 1993’s average per capita personal income, averaged between all 50 states, is $20,505. Low-income states are Alabama, North Dakota, South Carolina, Louisiana, Kentucky, New Mexico, Utah, Arkansas, West Virginia, Mississippi, Vermont, Texas, Tennessee, Iowa, Maine, Arizona, South Dakota, Idaho, Montana, Oklahoma, Wisconsin, Kansas, Oregon, Wyoming, Missouri, Indiana, Georgia, Nebraska, and North Carolina; High-income states are Ohio, California, Virginia, New Hampshire, Colorado, Washington, Pennsylvania, Minnesota, Rhode Island, Florida, Michigan, Connecticut, New Jersey, Massachusetts, New York, Maryland, Hawaii, Alaska, Illinois, Delaware and Nevada. 37 43 Table 6 Regression Results (Low-Income Versus High-Income States): Incumbent Party Candidate’s Share of Gubernatorial Vote, 1976-200838 LowIncome 40.52*** (7.50) HighIncome 54.01*** (8.98) All Races (with f.e.) 55.97*** (7.32) Interacted Terms All Races (no f.e.) 52.33*** (4.05) Interacted Terms Normal Vote 0.21** (0.10) -0.11 (0.10) -0.03 (0.12) 0.21 (0.20) 0.01 (0.08) 0.19 (0.14) Incumbent Running 6.86*** (1.13) 4.36** (1.87) 5.98*** (1.77) 1.01 (2.38) 5.77*** (1.22) 1.57 (1.96) Duration of One-Party Control State Violent Crime Growth -0.10 (0.18) -0.93 (0.58) -0.82 (0.59) 0.68 (0.61) -0.07 (0.24) 0.16 (0.28) -0.01 (0.03) -0.06 (0.04) 0.00 (0.04) -0.02 (0.05) -0.01 (0.04) 0.00 (0.04) Education Spending Growth Rate State Income Growth 0.03 (0.05) 0.08 (0.06) 0.09 (0.06) -0.05 (0.07) 0.07 (0.05) -0.08 (0.05) 0.26 (0.26) 0.06 (0.12) 0.08 (0.15) 0.09 (0.23) 0.06 (0.13) -0.10 (0.20) Relative Unemployment 0.19 (0.73) 0.09 (0.60) -0.55 (0.45) 0.77** (0.39) -0.55* (0.32) 0.15 (0.36) State Inflation Rate -0.01 (0.20) 0.14 (0.17) -0.06 (0.12) 0.19*** (0.07) -0.09** (0.04) 0.11* (0.06) High Population -0.86 (3.70) 3.07 (3.57) 1.25 (2.44) -0.12 (4.01) -3.11*** (1.21) 1.92 (1.73) Challenger Spending -3.67*** (1.12) -3.63*** (1.17) -3.95*** (1.38) 0.64 (0.99) -3.77*** (0.76) 0.59 (0.69) Incumbent Spending 0.25 (0.36) -1.09 (0.98) -0.93 (0.92) 1.01 (0.87) 0.80 (0.66) -0.40 (0.66) Presidential Approval 0.08* (0.04) 0.17*** (0.06) 0.12** (0.06) 0.00 (0.06) 0.12*** (0.04) -0.04 (0.06) Same Party as President -2.61** (1.15) -5.29*** (1.98) -4.75*** (1.54) 2.71** (1.34) -4.41*** (1.09) 1.93 (1.22) Adjusted R2 SE of Reg. n Fixed Effects 0.36 7.57 224 Both 0.48 7.11 168 Both 0.39 7.49 392 Both Variable Intercept 38 0.40 7.47 392 None Estimated with White cross-section robust standard errors. All regressions have normally distributed residuals and passed the Jarque-Bera normality test. The residuals do not exhibit autocorrelation or partial autocorrelation. The squared residuals do not show signs of ARCH. 44 Similar to Haynes and Stone (2008), income patterns appear in three of the four statistically significant coefficients, and in another coefficient not significant, from the “All Races” column in Table 5. Income-contingent voting patterns imply that voters in lower-income states are influenced differently by the typical factors that impact vote choice than voters in higher-income states. In Table 6, it appears that income patterns may exist in the variables measuring presidential approval ratings, the normal incumbent party vote, inflation, and two dummy variables: in one that tests whether the incumbent gubernatorial party’s candidate is the same party as the President’s, and in another that measures the effects of incumbent candidates in the race. The advantage of being an incumbent candidate in lower-income states is 2.5 additional percentage points beyond incumbent candidates in higher-income states. Every prior study has shown a strongly positive and statistically significant boost for incumbent candidates, but none of these studies has disaggregated states into separate income groups to test for potential differences among poor and rich state voters. One potential explanation for the 2.5 point difference is voter sophistication: the concept that voting on issues increases with more educated voters and more educated states, as shown in Cummins (2009). If richer-state voters base their votes more on issues than poorer-state voters, then unless the governor is an ‘issue’ himself (i.e. scandal plagued, etc.), it makes sense that incumbents do not receive as much of a boost in the higher-income states. This effect also appears when rather than classifying the U.S. states by state-level per capita personal income, I divide the states into two groups based on educational attainment (percentage of state population with a Bachelor’s degree or higher). There is a similar 45 added benefit of two percent for the incumbent candidate in lower-educated states when compared to higher-educated states. The lower-income/lower-education voters do not engage in the issues at the same level/sophistication as higher-income/higher-education voters, and the well-known incumbent candidate benefits. Examining the incumbency variable further, by interacting it with a poor-state dummy variable, we do not find the same result. No difference in the power of incumbency appears between states based on their income. Therefore, this finding should be viewed with some caution. The Normal Vote variable measures the average of the last two election results for the incumbent’s party. Lower-income states do appear to exhibit a consistent pattern in their vote for one party over the other. In low-income states, when the average of the last two election results for the incumbent’s party is ten points higher, the incumbent party will receive a two-percent jump on Election Day. Higher-income states do not show this pattern: Normal Vote is not statistically significant and the sign is negative. The impact of presidential approval in higher-income states, a 1.7 percent increase for the presidential party’s gubernatorial candidate for each ten-percent increase in the President’s approval, is more than double that of lower-income states’ 0.8-percent increase.39 Likewise, higher-income states tend to more strongly punish (reward) gubernatorial candidates who are in the same (opposite) political party as the President. In high-income states, candidates in the same party as the President lose almost three additional percentage points more than candidates with the same handicap in the lowest- 39 This finding does not remain significant when the presidential approval variable is interacted with a poorstate dummy variable. 46 income tier of states.40 In gubernatorial elections, national political considerations appear more salient in higher-income states, and prior research has not uncovered this pattern. Previous work has shown that the President’s Party loses sub-national elections during midterm years.41 Now, it seems that this hypothesis is truer in higher-income states than it is in lower-income states. There does not appear to be a theoretical reason for this difference. Looking at the states separated by educational attainment, this same pattern is not revealed. Since better-educated voters do not vote differently on presidential approval than less-educated voters, a link to voter sophistication cannot be made in this case.42 My results differ from Haynes and Stone (2008), but also share some similarities. First, Haynes and Stone (2008) uncover a remarkably consistent income-contingent voting pattern in all variables without a single exception, such as the challenger spending exception that I found. My results indicate that voters in the low- and high-income groups are equally influenced by per capita challenger campaign spending. One similarity between their results and mine is that voters in low-income states punish incumbents less for higher inflation than voters in high-income states. When inflation was interacted with a poor-state dummy variable, it is statistically significant and positive.43 To conclude, Table 5 results confirm much of the past research on the economic and political factors that affect gubernatorial elections, but also refute some previous findings. Table 6 lends some credibility to the income-contingent voting patterns 40 This finding remains true even when the variable is interacted with a poor-state dummy variable. Campbell (1966) and Tufte (1975) 42 The coefficient on presidential approval is 0.14 for both educationally ranked state groups. Also, between the two groups, the Same Party as President coefficient only differs by one percent, not the nearly 2.5 percent difference found in the income ranked groups. 43 In Table 6, unemployment appears to affect high- and low-income state voters differently, but in a number of different specifications, the results do not hold up. 41 47 hypothesis put forward by Haynes and Stone (2008). Three of my four most significant variables, and the inflation variable, illustrate their case. So, these patterns may exist in gubernatorial elections in the same manner as presidential elections. In addition to finding some evidence for these patterns, my work tests three previously untried variables on the gubernatorial vote: duration of one-party control, a state-level inflation measurement, and growth in real per-pupil education spending. These three measures do not impact voters’ choices on Election Day, except in one case for the duration of oneparty control variable when 16 outliers are excluded.44 Finally, the regression results in Table 5 conflict with Leal’s (2006) and Cummins’s (2009) findings that crime rate growth impacts the gubernatorial vote, but agree with Cummins (2009) that crime rates affected elections prior to 1994. The next chapter will make conclusions about the results, and offer ideas for future research. 44 These results did not hold up under other regressions. In fact the p-value was very close to ten percent at 0.1014. So, due to the lack of more robust results, the duration of one-party control variable cannot be confirmed significant indicator of the gubernatorial vote. 48 Chapter 6 CONCLUSION To answer how people win the increasingly powerful position of Governor, this thesis uses a fixed effects model, and economic and political data from 1976 to 2008. Improvements and changes from prior research include the use of state fixed effects, robust econometric testing, and a much larger data set. I use the model in Chapter 4 to test the brand-new variables (inflation, education, and one-party control), some relatively new variables (campaign spending, crime, and population), retrospective voting, and Haynes and Stone’s (2008) income-contingent voting patterns. Three new variables this thesis adds to the research, inflation, per-pupil education spending growth, and duration of one-party control, do not exhibit robust, significant effects on voters’ choices. Among the relatively new variables of violent crime rate growth, population, and campaign spending, I confirm Cummins’s (2009) and Leal’s (2006) findings regarding the statistical and actual significance of challenger campaign spending on the vote, but dispute their results that crime rates affect the vote. Results indicate that state-level economic variables for income, unemployment, and inflation do not impact the gubernatorial vote, which casts doubt on whether voters look backwards – the concept known as “retrospective voting”. However, the results in this thesis support Haynes and Stone’s (2008) income-contingent voting pattern hypothesis in two variables – high-income state voters more heavily weigh their voting decision based on whether the incumbent party shares the President’s Party, and low-income voters punish the incumbent party less for increases in inflation. 49 I have many ideas for future research as shown below. Since the results indicate some success using disaggregated U.S. states, future research can disaggregate the states in different ways to uncover the impact of their group characteristics. Other possible ways to group the states include educational attainment and percentages of different racial groups. The age of candidates has only been tested and found significant in Case (1994). The addition of this variable could improve future gubernatorial election models and results. Since economic variables did not show significance in my results, future work could research whether there are asymmetric economic effects on the vote (i.e. voters’ choices are affected by economic downturns, but not by economic improvements, or vice versa). Bloom and Price (1975) show asymmetric economic impacts on presidential elections. Future research should carefully choose instrumental variables for incumbent party campaign spending, since the research described in Chapter 5 used instrumental variables that were endogenous. Since Leal (2006) and I cannot find an education variable that affects the vote, perhaps graduation rates could have some impact on the incumbent gubernatorial party. These have not yet been tested. Elements from this thesis could be combined with elements of two other unorthodox, brand-new papers: Olivola and Todorov (2010) find that a 50 candidate’s appearance, specifically their face, can predict their success on Election Day. Healy and Malhotra (2009) find that local football team wins can improve voter moods and increase the vote for the incumbent senatorial, gubernatorial, and presidential candidates! Finally, while it appears that some of the factors affecting gubernatorial vote are settled (presidential approval ratings, sharing the same party as the President, incumbents running, and challenger party campaign spending), many other factors could be uncovered, and re-tested as noted in the suggestions for future research. Understanding the power and influence that governors have, and once researchers have a firm understanding of how our governors get elected, then perhaps people can use this information to their preferred candidate’s benefit. They might even go out to support their local football team to help the incumbent candidate win. Or, maybe they will make campaign contributions to the challenger candidate. It depends what side they’re on. For researchers, the intersection of elections and economics provides an endless source of study about how economics does or does not affect voters at the polling booth. 51 REFERENCES Adams, J., & Kenny, L. (1989). The Retention of State Governors. Public Choice, 62(1), 1-13. Retrieved from EconLit database. Atkeson, L., & Partin, R. (1995). Economic and Referendum Voting: A Comparison of Gubernatorial and Senatorial Elections. American Political Science Review, 89(1), 99107. Retrieved from EconLit database. Bardwell, K. (2005). Reevaluating Spending in Gubernatorial Races: Job Approval as a Baseline for Spending Effects. Political Research Quarterly, 58(1), 97-105. Retrieved from America: History & Life database. Berry, W., Fording, R., & Hanson, R. (2000). An Annual Cost of Living Index for the American States, 1960-1995. Journal of Politics, 62(2), 550. Retrieved from Academic Search Premier database. Beyle, Thad, ed. 1992. Governors and Hard Times. Washington, DC:Congressional Quarterly Press. Bloom, H., & Price, H. (1975). Voter Response to Short-Run Economic Conditions: The Asymmetric Effect of Prosperity and Recession. American Political Science Review, 69(4), 1240-1254. Retrieved from EconLit database. Campbell, Angus. 1966. Surge and Decline: A Study of Electoral Change. In Elections and the Political Order, edited by Angus Campbell, Phillip Converse, Warren Miller, and Donald Stokes. New York: Wiley. Carsey, T., & Wright, G. (1998). State and national factors in gubernatorial and senatorial elections. American Journal of Political Science, 42(3), 994. Retrieved from Academic Search Premier database. Case, A. (1994). Taxes and the electoral cycle: How sensitive are governors to coming elections?. Business Review (Federal Reserve Bank of Philadelphia), 17. Retrieved from Academic Search Premier database. Chubb, J. (1988). Institutions, the Economy, and the Dynamics of State Elections. American Political Science Review, 82(1), 133-154. Retrieved from America: History & Life database. Cohen, J., & King, J. (2004). Relative Unemployment and Gubernatorial Popularity. Journal of Politics, 66(4), 1267-1282. 52 Cummins, J. (2009). Issue Voting and Crime in Gubernatorial Elections. Social Science Quarterly, 90(3), 632-651. Retrieved from EconLit database. Dometrius, Nelson. 1999. “Governors: Their Heritage and Future.” In American State and Local Politics: Directions for the 21st Century, edited by Ronald Weber and Paul Brace. New York: Chatham House Publishers, Seven Bridges Press. Fair, R. (1978). The Effect of Economic Events on Votes for President. Review of Economics & Statistics, 60(2), 159. Retrieved from Business Source Premier database. Fiorina, M. (1978). Economic Retrospective Voting in American National Elections: A Micro-Analysis. American Journal of Political Science, 22(2), 426. Retrieved from Academic Search Premier database. Glashan, Roy R. 1979. American Governors and Gubernatorial Elections, 1775-1978. Westport, Connecticut : Meckler Books Haynes, S., & Stone, J. (2008). A Disaggregate Approach to Economic Models of Voting in U.S. Presidential Elections: Forecasts of the 2008 Election. Economics Bulletin, 4(28), 1-11. Retrieved from EconLit database. Healy, A., & Malhotra, N. (2009). Euphoria and Retrospective Voting: The Impact of College Football Outcomes on Incumbent Reelection. Conference Papers -- Midwestern Political Science Association, 1. Retrieved from Academic Search Premier database. Holbrook-Provow, T. (1987). National Factors in Gubernatorial Elections. American Politics Quarterly, 15(4), 471. Retrieved from Academic Search Premier database. Howell, S., & Vanderleeuw, J. (1990). Economic Effects on State Governors. American Politics Quarterly, 18(2), 158. Retrieved from Academic Search Premier database. Jayadev, A. (2006). Differing Preferences between Anti-Inflation and AntiUnemployment Policy Among the Rich and the Poor. Economics Letters, 91(1), 67-71. Kenney, P. (1983). The Effect of State Economic Conditions on the Vote for Governor. Social Science Quarterly, 64(1), 154-162. Retrieved from EconLit database. Kinder, D., & Kiewiet, D. (1979). Economic Discontent and Political Behavior: The Role of Personal Grievances and Collective Economic Judgments in Congressional Voting. American Journal of Political Science, 23(3), 495. Retrieved from Academic Search Premier database. 53 Kinder, D., & Kiewiet, D. (1981). Sociotropic Politics: The American Case. British Journal of Political Science, 11(2), 129-161. Retrieved from America: History & Life database. King, J. (2001). Incumbent Popularity and Vote Choice in Gubernatorial Elections. Journal of Politics, 63(2), 585. Retrieved from Academic Search Premier database. Kone, S., & Winters, R. (1993). Taxes and voting: Electoral retribution in the American states. Journal of Politics, 55(1), 22. Retrieved from Academic Search Premier database. Kramer, G. (1971). Short-Term Fluctuations in U. S. Voting Behavior, 1896-1964. American Political Science Review, 65(1), 131-143. Retrieved from EconLit database. Leal, David L. 2006. Electing America’s Governors: The Politics of Executive Elections. New York: Palgrave MacMillan. Levernier, W. (1992). The Effect of Relative Economic Performance on the Outcome of Gubernational Elections. Public Choice, 74(2), 181-190. Retrieved from EconLit database. Levernier, W. (1993). Election Outcomes and Economic Conditions: An Application of a Logit Model. Journal of Economics & Finance, 17(1), 115. Retrieved from Business Source Premier database. Leyden, K., & Borrelli, S. (1995). The Effect of State Economic Conditions on Gubernatorial Elections: Does Unified Government Make a Difference?. Political Research Quarterly, 48(2), 275. Retrieved from Academic Search Premier database. Lowry, R., & Alt, J. (1998). Fiscal Policy Outcomes and Electoral Accountability in American States. American Political Science Review, 92(4), 759. Retrieved from Academic Search Premier database. Mullaney, Marie Marmo. 1988. American Governors and Gubernatorial Elections, 19791987. Westport, Connecticut: Meckler. Niemi, R., & Stanley, H. (1995). State economies and state taxes: Do voters hold governors accountable?. American Journal of Political Science, 39(4), 936. Retrieved from Academic Search Premier database. Olivola, C. & Todorov, A. (2010). Elected in 100 milliseconds: Appearance-Based Trait Inferences and Voting. Journal of Nonverbal Behavior. 34(2), 83. 54 Partin, R. (2002). Assessing the Impact of Campaign Spending in Governors' Races. Political Research Quarterly, 55(1), 213. Retrieved from America: History & Life database. Pearson, F. A. & Myers, W. I. (1948). Prices and Presidents. Farm Economics. Ithaca: New York State College of Agriculture, 163 (September), 4210-4218. Peltzman, S. (1987). Economic Conditions and Gubernatorial Elections. American Economic Review, 77(2), 293-297. Retrieved from EconLit database. Sribnick, Craig ed. 2008. A Legacy of Innovation: Governors and Public Policy. Philadelphia: University of Pennsylvania Press. Schlozman, Kay Lehman, and Sidney Verba. 1979. Injury to Insult: Unemployment, Class, and Political Response. Cambridge, Massachusetts. Harvard University Press Stein, R. (1990). Economic Voting for Governor and U.S. Senator: The Electoral Consequences of Federalism. Journal of Politics, 52(1), 29. Retrieved from Academic Search Premier database. Svoboda, C. (1995). Retrospective voting in gubernatorial elections: 1982 and 1986. Political Research Quarterly, 48(1), 135. Retrieved from Academic Search Premier database. Tidmarch, C., Hyman, L., & Sorkin, J. (1984). Press Issue Agendas in the 1982 Congressional and Gubernatorial Election Campaigns. Journal of Politics, 46(4), 1226. Retrieved from Academic Search Premier database. Tufte, E. (1975). Determinants of the Outcomes of Midterm Congressional Elections. American Political Science Review, 69(3), 812-826. Retrieved from America: History & Life database.