Voter ID Laws and Voter Turnout

Transcription

1 Voter ID Laws and Voter Turnout Kyle A. Dropp 1 Do Voter Identification statutes reduce voter turnout? I demonstrate that the decadelong expansion of Voter ID statutes has demobilized Democratic-leaning individuals such as young adults, the poor, African Americans and renters by aggregating tens of millions of individual voting records over a series of four elections ( ). I use a difference-in-differences approach to compare turnout between 2004 and 2008 among subgroups in Voter ID states with broader statewide turnout and turnout in states with no election law policy change. Compared with previous survey-based and aggregate level studies, my approach is well suited for detecting small but politically meaningful effects across voter subgroups and significantly reduces measurement error. This research both clarifies mixed findings in the previous literature and introduces a new method to document subgroup effects for election law changes. 1 Ph.D. candidate, Department of Political Science, Stanford University, dropp@stanford.edu 1

2 Thirty states have implemented Election Day Voter Identification policies, more than twice as many as in Further, strict policies requiring a government-issued photo ID have become commonplace following the 2008 Supreme Court decision upholding an Indiana Voter ID statute, 3 and widespread Republicans electoral gains in 2010 in statehouses and governor s offices. In 2011 alone, lawmakers in 34 states introduced legislation to strengthen existing Voter ID laws or implement new statutes. 4 The Justice Department and federal courts are reviewing stringent statutes signed into law, and proponents and opponents alike are petitioning the Supreme Court for further guidance. 5 The question remains, has the expansion of Voter ID statutes reduced voter turnout? And, if so, which groups have been disparately impacted? Survey-based and aggregate level studies have yielded puzzling, often contradictory findings. 6 While some studies have found that Voter ID laws have no impact on turnout [Ansolabehere, 2009, Lott, 2006, Pastor et al., 2010, Mycoff et al., 2009, Milyo and Policy, 2007] others have concluded that Voter ID laws decrease turnout among voter subgroup such as young adults and minorities [Vercellotti and Anderson, 2006, Alvarez et al., 2007, Logan et al., 2007, Alvarez et al., 2011]. One explanation for My research examines the impact of Voter ID laws on voter turnout using both national voter file data and precinct level elections returns. I aggregate tens of millions of individual level voting records over a series of four elections ( ) using a national voter database. I isolate groups with low ID ownership rates such as low income individuals, African Americans, young adults, adults over 65, Hispanics and renters using demographic data contained in the voter files. Then, I use a difference-in-difference approach to compare the turnout of these groups before and after a Voter ID law change with broader statewide turnout and with turnout patterns among Americans in states where election law stayed the same. This paper has two principal findings. First, Voter ID statutes exert a small but politically meaningful demobilizing effect, especially among young adults, lower income Americans, African Americans and renters. The reported effects sizes are small but substantial enough to influence election outcomes in close races. My research is the Crawford v. Marion County 4 National Conference of State Legislatures elections/voter-id.aspx For a catalog of studies, see this link research_on_voter_id/ 2

3 first to demonstrate that Voter ID laws impact the participation of a broad swath of the electorate. Second, Voter ID laws are more likely to reduce turnout in midterm elections (compared with presidential contests) and multiple election cycles after implementation. This finding suggests that widespread mobilization efforts during presidential contests or enhanced grassroots efforts attracted to states after they enact a Voter ID law can offset the impact of election laws. This study proceeds as follows. First, I describe the widespread adoption of Voter ID statutes in the past decade and examine previous work on the impact of these election laws. I discuss hypotheses, research design, data sources and findings in that order. Finally, I briefly conclude. The widespread adoption of Voter ID statutes The successful, decade-long effort to strengthen Voter ID statutes has been fueled by a series of factors: the contested 2000 presidential campaign and the passage of the Help America Vote Act (HAVA), a concerted, coordinated effort among Republican state legislators and governors, unmistakable public support for ballot security measures and the widespread perception of pervasive voter fraud among Americans. In 2001, only one in four states required that voters provide an ID at the polls, according to the National Conference for State Legislatures. 7 Furthermore, none of these states turned away voters without an ID. Today, voters in 30 states are required to show a form of identification at the polls, and there has been a decisive trend toward strict policies requiring voters to present a government-issued photo identification [Alvarez et al., 2007, p. 8]. 8 Pending court challenges, that number could rise to 33 states for the November 2012 election. The Help America Vote Act (HAVA) of 2002 established minimum federal standards for first-time voters and caused many states to focus on ballot security measures. The act, which passed with overwhelming bi-partisan majorities in both the U.S. Senate and U.S. House, was signed into law by President George W. Bush in October The act required that first-time voters who register by mail document proof National Conference for State Legislatures elections/voter-id.aspx 3

4 of identify with a series of photo or non-photo IDs [Pear, 2002]. 9 Americans hold widespread concerns about ballot integrity and voter fraud, and Voter ID laws enjoy widespread public support. Seventy-eight percent of Americans said voters should be required to show an official photo identification on Election Day, according to a 2006 Pew poll, including 86 percent of Republicans and 71 percent of Democrats. 10 Fully 48 percent of Americans said that voter fraud people voting who are not eligible or voters casting multiple ballots is a major problem, according to a Washington Post poll Figure 1 on Page 5, taken from Alvarez et al. [2007], displays state Voter ID requirements from The darker the state, the stricter the ID requirement. 13 The plot demonstrates that states increasingly have implemented stricter Voter ID policies. In 2000, conservative states concentrated in the South such as Georgia, Louisiana and Texas required IDs. Voter ID statutes have diffused gradually to the Mountain States, then the Southwest and finally to parts of the Midwest. The recent adoption has been driven by Republican legislative gains in the 2010 midterm elections. In fact, the only regions largely unaffected by the nationwide drive to strengthen voter ID laws are the Northeast and the West Coast. 9 The minimum requirement for individuals who registered after December 31, 2002, was to provide their driver s license number or the last four digits of their Social Security number [Pear, 2002]. 10 Survey by Pew Research Center for the People and the Press, October 17-22, 2006 based on 2,006 telephone interviews. 11 The poll was conducted July 18-29, 2012, and based on 2,047 telephone interviews. 12 Thirty-three percent say it is a minor problem and 14 percent say it is not a problem. 13 White - name; light blue - signature; medium blue - matching signature; green - request id; blue - require non-photo id; dark blue - require ID plus signature; darkest blue - request photo ID; black - require photo ID. 4

5 Figure 1: Voter Identification Laws stringent Voter ID requirements. Darker shades correspond to more Figure 2 on Page 6, from the National Conference of State Legislatures, provides the current status of Voter Identification laws across the country. Thirty states now require all voters to show an ID prior to casting a ballot. The states shaded in green, so-called strict ID states, require voters to present a photo ID and do not provide another option such as stating the last four digits of your Social Security Number. Yellow states require a photo ID but have other options for casting a provisional ballot, blue states require a non-photo ID and gray states have the HAVA minimum requirement. 5

6 Figure 2: Voter Identification Laws Gray states have no requirement, blue states require a non-photo ID, yellow states require a photo ID, green have strict photo ID policies. Source: National Conference for State Legislatures. Overall, more states have adopted stricter Voter ID policies in recent years. Do these policies affect voter turnout? Literature Review A growing, methodologically diverse research literature on Voter Identification statutes and voter turnout has yielded mixed findings. These mixed findings have resulted from a combination of insufficiently granular data collection efforts, the diversity of Voter ID statute changes and relative novelty of such policy changes. Some studies have found that Voter ID laws have a minimal impact on turnout [Ansolabehere, 2009, Lott, 2006, Pastor et al., 2010, Mycoff et al., 2009, Milyo and Policy, 2007]. They argue that Voter ID statutes are inconsistently implemented, that most adults have suitable forms of identification, that individuals without IDs are not likely to vote, or that the available data, simply put, is not up to the task of answering this research question. Ansolabehere [2009] examines Current Population Survey (CPS) data and concludes 6

7 that [v]oter ID appears to present no real barrier to access [Ansolabehere, 2009]. Poll workers rarely ask for ID, he finds, and individuals almost never say they did not vote because they lacked in ID. Others have found, using largely the same methods and data, that Voter ID statutes decrease turnout and disproportionately impact subgroups that are less likely to have valid forms of identification [Vercellotti and Anderson, 2006, Alvarez et al., 2007, Logan et al., 2007, Alvarez et al., 2011]. These studies argue that Voter ID disproportionately affects certain subgroups such as young adults or minorities and that stricter requirements exert a more negative impact on voter turnout than more relaxed ones. Alvarez et al. [2007], for example, find evidence that the stricter voter identification requirements depress turnout to a greater extent for less educated and lower income populations [Alvarez et al., 2007, p. 1]. An overarching theme in the literature, however, is that the available data is not powerful enough to answer this question with confidence. Aggregate level studies using counties or states cannot document the impact of state-level interventions on subgroups, while individual level survey data often lacks power to make inferences about population segments. Some scholars, for example, call into question the use of cross-sectional data from the Current Population Survey and other data used to examine the impact of Voter ID statutes [Erikson and Minnite, 2009]. Erikson and Minnite [2009] analyze multiple years of CPS data and find no relationship between Voter ID laws and turnout, contrary to previous authors conclusions from analysis of the same data: [W]e see the existing science regarding voter suppression as incomplete and inconclusive. This is not because of any reason to doubt the suppression effect but rather because the data that have been analyzed do not allow a conclusive test [Erikson and Minnite, 2009, p. 98]. Second, the expansion of Voter ID statutes is a recent phenomenon, and scholars have not had many election cycles to examine their effects: [S]ince the changes in voter identification requirements have really only started since the passage of HAVA in 2002 and the law we are most interested in photo identification requirements was only implemented in 2006, we have only a small amount of information in the available data about how each of voter identification requirement might affect participation [Alvarez et al., 2011, p. 10]. Voter Identification Laws may exert their greatest impact years after implementation, when young adults and first-time voters must obtain an ID to cast their ballot. Further, the trend toward stricter laws in recent years means there have been even fewer data points to study the stringent 7

8 election policies most likely to adversely affect turnout. Why do Voter ID laws affect turnout? Heterogeneous Effects Scholars generally cite three explanations reasons why Voter ID statutes may reduce turnout. First, some Americans do not have up-to-date forms of photo or non-photo identification. Second, voters may not cast a ballot because they don t believe they have a suitable ID. Third, Voter Identification laws reduce voter impersonation at polling places. 14 This section highlights differential ID ownership rates and civic skills, two factors that Voter ID statutes may exert a heterogenous impact across the electorate. Cumulatively, I argue that Voter ID statutes will disproportionately affect turnout among young adults, college students, African Americans, Hispanics, renters and low income individuals. ID ownership and access Put simply, voters need an ID to cast a ballot in ID states. Acquiring an ID requires effort, 15 and there is ample evidence that certain subgroups are less likely to own IDs required to cast ballots in the Voter ID states. African Americans, Hispanics, college students, young adults, renters, and the poor are less likely to have valid forms of identification [Barreto et al., 2007, Sanchez et al., 2011, Pawasarat, 2005] Twenty-five percent of African Americans and 18 percent of adults 65 years and older lack the government-issued photo ID necessary to cast a ballot in stringent Voter ID states, compared with one in 10 adults overall, accord- 14 According to the National Conference for State Legislatures, Little evidence exists that fraud by impersonation at the polls is a common problem elect/canvass_apr_2012_no_29.pdf 15 There is evidence that the same groups with low ID ownership rates have a more difficult time acquiring valid IDs in policy states. A survey of voters found that hundreds of thousands of eligible voters face challenges in obtaining proper IDs to cast a ballot. 16 According to the Brennan Center, nearly 500,000 voters in 10 states with restrictive voter ID laws live in households without vehicles and reside at least 10 miles from an ID-issuing office open more than two days a week. 17 8

9 ing to the Brennan Center for Justice. 18 Pawasarat [2005], for example, finds that African Americans and Hispanics are less likely to have drivers licenses than whites in Wisconsin: Less than half (47 percent) of Milwaukee County African American adults and 43 percent of Hispanic adults have a valid drivers license compared to 85 percent of white adults [Pawasarat, 2005, p. 1]. On the other hand, white voters are more likely to have government issued IDs [Barreto et al., 2007]. Ninety-five percent of white registered voters have an up-todate government ID, 19 compared with 90 percent of African Americans, 89 percent of Latinos and 86 percent of Asian Americans, according to a 2008 study [Sanchez et al., 2011]. Further, residents who have moved are less likely to own suitable forms of identification: The population that changes residence frequently is most likely to have a drivers license address that differs from their current residence. This would include lower-income residents who rent and students and young adults living away from home [Pawasarat, 2005, p. 2]. Time, resources and civic skills Factors such as socioeconomic status, time and civic skills contribute to political activity. When there are changes in election administration policies, those with more skills, resources or flexibility in work schedule may adapt more readily than those without such skills: [The] presence or absence of resources contributes substantially to individual differences in participation. Resources are, in turn, not equally distributed; some socioeconomic groups are better endowed than others [Brady et al., 1995, p. 274]. Brady et al. [1995] show that there is an inverse relationship between income and free time [Brady et al., 1995, p. 273]. Since learning about new election policy and obtaining a necessary ID requires time, voters with rigid work schedules or less free time may be adversely affected. Barreto et al. [2007] argue that gap in civic skills between white and non-white voters affects turnout in Voter ID states: Voting may be less costly for those with greater levels of political resources such as money, time, English language abilities and education. Therefore, any increases in costs associated with voting should have Driver s License or State Issued ID Card 9

10 the greatest impact on those with the fewest political resources racial and ethnic minorities, the less educated, immigrants, and the less affluent [Barreto et al., 2007, p. 9]. Citizens with less political information, for example, are less likely to know which ID they need to vote, potentially affecting turnout: citizens might not vote because they mistakenly thought that they could not do so if they did not have certain forms of identification [Urbina, 2008]. Overall, the time and resource gaps affect individuals not familiar with the voting system such as young adults, college students and minorities. This leads to the following hypothesis: H1: Differing ID ownership rates and civic skills levels across swaths of the electorate suggest that minorities, renters, the elderly and young adults will be more burdened by the election administration policy changes relative to others. When can Voter ID statutes affect turnout? Voter ID statutes exact costs on voter subgroups who lack suitable forms of identification or familiarity with the electoral system. These ballot security measures may have little impact on voter turnout, however, if campaigns devote resources toward voter mobilization and outreach, interest groups synchronize their efforts with traditional allies and organized interests shift their missions to focus on voter outreach and education campaigns rather than persuasion. Campaigns are dynamic and react swiftly to changes in election administration policies. They have an opportunity to educate their voters about ID requirements and can target their fixed resources based on changes in election laws. In the 2012 race, for example, the Obama campaign has sent teams to address new Voter ID statutes: Field workers for President Obama s campaign fanned out across the country over the weekend in an effort to confront a barrage of new voter identification laws that strategists say threaten the campaign s hopes for registering new voters ahead of the November election [Shear, 2012]. The passage of election laws may cause organized interest groups, along with their traditional allies, to synchronize their efforts. The AFL-CIO, for example, vowed to mount their biggest voter registration and protection efforts ever to counter these laws [Voter ID statutes] [Greenhouse, 2012]. Further, voter ID laws have increased 10

11 interest groups coordination efforts with traditional ideological allies: The federation s [AFL-CIO] leaders said they would work closely with other groups, including the N.A.A.C.P. and the National Council of La Raza, to maximize voter turnout and provide whatever help is needed to enable elderly, disabled and poor Americans to get voter IDs [Greenhouse, 2012]. Finally, these election statues can cause interest groups to reorient their short-term goals to focus on mobilization rather than persuasion. The NAACP, for example, will focus on voter education and outreach ahead of this year s presidential election in the wake of a U.S. Supreme Court ruling on voter identification laws [Haines, 2008]. Organizations have fixed resources, and the decision to allocate these resources to grassroots mobilization rather than paid media such as televised advertising could lessen the impact of election law changes. Grassroots mobilization drives, enhanced campaign efforts and synchronized interest group activity, however, may wane in subsequent election cycles. New Voter ID battlegrounds may arise, for example. If enhanced mobilization and coordination efforts are not maintained, then Voter ID statutes may negatively impact the participation rates of new voters and individuals without suitable forms of identification. Overall, I argue that Voter ID laws will affect voter turnout when elevated mobilization efforts are absent. I argue that Voter Identification Laws will exert more of an impact during midterms contests (relative to presidential elections) and that the effects of the policy change may be most influential multiple election cycles after implementation. H2: Voter ID statutes will be most influential multiple election cycles after their implementation. Since campaign and interest group mobilization efforts are less intense in midterm elections compared with presidential contests, I also predict that Voter ID statutes will reduce turnout disproportionately during these midterm contests and other low interest elections: H3: Voter ID statutes will be influence voter turnout more in midterm elections than in presidential contests. 11

12 A note on effects sizes Despite the above discussion, election administration policies still appear to impact voters on the margins. Most voters have valid photo or non-photo form of identification for their state of residence. Few voters would be turned away from casting a ballot solely based on 2008 state-by-state ID requirements, according to two separate studies. 20 Second, most individuals without proper IDs are unlikely to cast a ballot, regardless of whether they reside in policy or non-policy states, according to Rick Hasen, an election law expert: It s not possible to show, he says, that many people have actually been deterred from voting by these laws. In part, that s because many of the laws are new, and in part it s because many of the people who lack an ID card tend not to be interested in voting in the first place [Firestone, 2012]. Third, a majority of states with ID policies allow voters to cast a provisional ballot, after signing an affidavit, for instance, which could eventually be counted. Many of the stringent laws with limited provisional ballot options have been passed only in recent months or are pending. Fourth, studies of Voter ID statutes do not examine the full effect of requiring an ID versus asking for no ID. The Help America Vote Act established minimum identification standards for first-time voters and absentee voters, so any change in Voter ID policy is relative to those federal minimum standards. Research Design Election laws rarely affect all voters equally. Voter ID statutes may exert a disparate impact on population subgroups that are less likely to own a valid form of identification, have limited time or resources or are concerned election administration policies will not be implemented fairly. In this section, I outline my approach for detecting the impact of Voter Identification statutes among subgroups. Using a difference-in-differences estimator [Ashenfelter and Card, 1985], I compare the change in voter turnout before and after a policy change among specific subgroups 20 One study estimates that.5% of respondents would be prevented from voting [Alvarez et al., 2008] while a 2007 study estimates that one tenth of one percent of voters would be unable to vote because of an ID requirement [Ansolabehere, 2007]. 12

13 such as African Americans with statewide voting trends. Furthermore, I analyze individual level voting patterns both in states implementing a new Voter ID policy and in states where election law remained constant. I compare change in turnout among specific subgroups in these control states with broader statewide change. This additional set of controls creates a difference-in-difference-in-differences (DDD) approach. Suppose State S implements a strict Voter Identification law a few months after the 2004 presidential election. We test whether the statute reduces turnout among subgroups such as college students, who are less likely than other voters to have a suitable form of identification with a current address. Our outcome variable is the percent of eligible adults who cast a ballot before and after the policy change, the treatment group is college students in State S and the control group is all other residents in State S. This difference-in-differences approach tests whether the policy intervention in State S increased or decreased the turnout disparity between college students and statewide turnout. If the turnout rate in a Voter ID state was 50% for both college students and statewide in November 2004, 50% statewide in November 2008 and 45% for college students in 2008, we conclude that Voter ID policies reduced turnout by 5% among college students. Figure 3 on Page 14 uses simulated data to display the first difference-in-differences approach in a state where a Voter ID law takes effect between 2004 and In this state, turnout among college students increases relative to overall statewide turnout from 2004 to This design s main shortcoming is that factors unrelated to State S s new policy may affect the political participation of college students relative to statewide turnout. The Obama campaign s mobilization efforts in 2008, for example, buoyed turnout among college students nationwide. Therefore, a model focusing solely on policy states could produce the spurious result that Voter ID laws boost turnout among college students. We improve the approach by using an additional control: one or more non-policy States S. The approach described above tested whether policy interventions in State S increased or decreased the difference in turnout between college students and statewide turnout. By including control state(s) S, we can examine whether the college student to statewide turnout differential expanded more in policy states S versus non-policy states S. This controls both for factors unrelated to the Voter ID policy that affect turnout among the subgroup nationally, such as Obama mobilization efforts. Further, it controls for features unique to State(s) S that either 13

14 Figure 3: Simulated data Change in turnout in policy states among college students (blue) and statewide (red). increase or decrease turnout statewide, such as a competitive gubernatorial race or a particularly weak economic climate. Figure 4 on Page 15 displays the full difference-in-difference-in-differences approach that also takes into account the relative differential between college students and statewide turnout in areas where election law remains constant. Turnout among college students rises across policy and non-policy states, though it rises much more sharply in non-policy states. In this simulation, we conclude that Voter ID laws decrease the turnout of vital subgroups. 14

15 Figure 4: Simulated data Change in turnout in both policy states (right panel) and non-policy states (left panel) among college students (blue) and statewide (red). This figure highlights how the first difference-in-differences approach, shown in Figure 3, can be misleading if factors unrelated to the policy affect overall voter turnout patterns among a subgroup. Figure 4 takes into account subgroup, temporal and state specific factors. We can compute the OLS estimate for the DDD specification using a series of indicator variables and interactions shown below y = β 0 + β 1 S + β 2 G + β 3 S G + λ 0 T + λ 1 T S + λ 2 T G + λ 3 T S G + u (1) where S (State) is either a policy or non-policy state, G (Group) is either statewide turnout or turnout among the disadvantaged group, T (Time) is pre-intervention or post-intervention and λ 3 is the coefficient of interest from the triple interaction of State, Group and Time. A negative coefficient on λ 3 indicates that state Voter Identification laws decrease turnout among a particular subgroup. We can estimate this model using either individual level data or by aggregating data to the precinct, block group, county or state level. 15

16 The notation below breaks the difference-in-difference-in-differences (DDD) estimator into four components δ = [(ȳ S,G,2 ȳ S, G,2 ) (ȳ S,G,1 ȳ S, G,1 )] [(ȳ S,G,2 ȳ S, G,2 ) (ȳ S,G,1 ȳ S, G,1 )] DDD est. (1) (2) (3) (4) where δ is the estimator, ȳ is mean turnout, a state is either S (policy) or S (nonpolicy), the group is either G (college students) or G (all non-college students), and the time T is either Time 1 (pre-intervention) or Time 2 (post-intervention). The first quantity is the difference in turnout between a voter subgroup (i.e., college students) and the entire state post-intervention in policy states, while the second sum is the same subgroup versus statewide comparison pre-intervention. The third quantity is the difference between the voter subgroup and the statewide total in time 2 in non-policy states S, while the fourth sum is the same subgroup versus statewide comparison pre-intervention in non-policy states. We find the population analog by taking the expected value of the eight groups above. The DDD estimate starts with the changes in average turnout for college students in the treatment state and then nets out the change in mean turnout for the non-college students (i.e., statewide) in the treatment state (2), the change in means for college students in the control state (3) and the change in means for the non-college students in the control state (4). This approach should control for trends in voting among college students across states that have nothing to do with the policy, trends in voting among all residents in the policy-change states possibly caused by policies affecting everyone s voting propensity and trends in voting among all residents in the non-policy states possibly caused by policies affecting everyone s voting propensity in the non-policy states. 21 This basic approach can be extended to study different elections, alternate voter subgroups or other state-level election law interventions. The denominators used to calculate turnout are annual statewide voting age population or citizen voting age population figures for each subgroup. It is worth noting that we are measuring a 100% sample and do not have sampling error in these calculations. 21 For instance, think of non-policy state moving from competitive to non-competitive over the period of interest. 16

17 The potential outcomes framework We must satisfy a series of assumptions to make a causal inference using observational or experimental data. Foremost among these assumptions are the impact of nonrandom assignment and the potential violation of the stable unit treatment value assumption (SUTVA). Non-random assignment. We assume that treatment assignment is independent of outcomes. The adoption pattern of Voter ID statutes first in the South, then the Mountain West and finally in the Southwest and Midwest makes apparent that individuals were not randomly assigned to treatment. The widespread adoption of these policies has been spearheaded by strategic elite behavior among Republican legislators, and it is possible that these officeholders are targeting states with voter subgroups likely to be affected by the policies. For instance, Voter ID laws may negatively impact turnout among African Americans in Georgia more than African Americans in North Carolina. If legislators pass a law in Georgia but not in North Carolina, then the estimated treatment effect will be biased. These Voter ID laws have been driven mainly by the partisan composition of the state legislature and governor s office rather than specific characteristics of the state s residents. The difference-in-difference approach utilized in this study accounts for variation in turnout rates across states and voter subgroups. Previous scholars have used these estimators to address problems associated with non-random assignment [Alvarez et al., 2011]. 22 We also can address this issue by matching on covariates with Catalist s 1% sample, which contains detailed information on more than two million voters. With this approach, the sole difference between units is the assignment to treatment. SUTVA. The stable unit treatment value assumption (SUTVA) assumes that the assignment status of any unit does not affect the potential outcomes of other units. 22 Alvarez et al. [2011], for example, use a difference-in-differences estimator with Current Population Survey data to estimate the impact of Voter ID statutes: Finally, identification requirements are not randomly assigned across states. This is a problem if states with historically lower turnout also tend to adopt stricter identification requirements, then we will have trouble isolating whether the low level of turnout is due to the identification requirement or to other factors that lead a given state to have lower turnout rates. The estimation strategy used exploits the temporal and geographic variability in voter identification requirements to sidestep the problem on non-random assignment. This is referred to as a difference-in-differences estimator and our analysis is built on a generalization of this procedure. In particular, we use a multilevel model also referred to as a random effects model to assess how voter identification requirements affect participation by registered voters, using data from four years of recent CPS Voter Supplement data [Alvarez et al., 2011, p. 10]. 17

18 The broad awareness of Voter ID statutes nationwide, 23 and the widespread adoption of these election laws in the past decade calls into question the validity of this assumption. 24 For example, African Americans in Alabama (a non-treatment state) may see African Americans in Georgia (a Voter ID state) assigned to treatment and then act differently. News of Voter ID laws in other states could mobilize African Americans in neighboring states, which will lead to a biased treatment effect. 25 Unfortunately, the potential violation is increasingly a concern as more and more states adopted the Voter ID policies. Finally, the second part of SUTVA assumes that there is no variation in treatment across groups, that the treatments for all units are comparable. 26 Voter ID statutes are multi-faceted and diverse, and each state has unique minimum and maximum requirements, making it nearly impossible to calculate a single average treatment effect. These concerns are highlighted by [Alvarez et al., 2011] Twenty-one percent of Americans said they heard or read a lot about states putting in place new photo identification requirements for voters, according to a Washington Post poll in The pool was conducted July 18-29, 2012 and based on 2,047 telephone interviews. Twenty-seven percent said they heard or read some, 15% said not much and 36% said they had heard or read nothing at all. 24 The problem posed by interference across units is very similar; if unit i s potential outcome under treatment A depends upon another unit j s assignment status, then there are really multiple (compound) treatments involving A for unit i, each of which involves a different assignment for unit j. Each of these multiple treatments is associated with a corresponding potential outcome. Note that this kind of interference across units does not necessarily present a problem for defining the effect of a single one of these compound treatment As. It just means that asking What is the effect of treatment A? makes no sense it is not a well-posed causal question. http: //blogs.iq.harvard.edu/sss/archives/2006/02/thoughts_on_sut.shtml 25 Alternatively, African Americans in Georgia, an ID state, may see that African Americans in other states were not assigned to treatment and then may be deterred from casting a ballot. 26 SUTVA posits, for each unit and treatment, a single fixed potential outcome, not a distribution of potential outcomes. Thus, if there is a potential outcome for the weak version of treatment A and a different potential outcome for the strong version of treatment A, then one cannot speak of the potential outcome that would have been observed following treatment A: there are in fact two treatments...[a]s long as there is a single version of the control intervention, one could still coherently define causal effects for each unit in terms of the difference between (observed) potential outcomes under heterogeneous treatment interventions and (unobserved) potential outcomes under control Alvarez et al. [2011] address the variation in treatments: [T]here are many methodological problems unique to this data, one of which is the ordinality of voter identification requirements. As is apparent from the listing of the types of regimes, it is not the case that a state either requires identification to vote, or does not. States require many different levels of identification from simply stating one s name to showing a picture identification. This further complicates the question, as we must determine not just one effect but several potentially incremental effects. Second, states 18

19 The next section discusses the series of elections and important constituencies this study examines. Data Assembling a Voter ID Database I utilize a series of sources to assemble a database of Voter ID policies such as the National Conference of State Legislatures, LexisNexis, newspaper reports and the work of previous scholars. The time series policy database facilitates statistical tests examining the impact of changes in Voter ID policy on voter turnout. 28 The National Conference of State Legislatures places ID requirements in four categories: No Voter ID law, Non-photo ID law, Photo ID law and Strict Photo ID law. I follow their methodology Overall, 13 states changed their Voter ID statutes between the 2002 and 2004 election, seven strengthened their laws between 2004 and 2006, two tightened their policies between 2006 and 2008 and three implemented additional Voter ID provisions between November 2008 and November I examine the impact of relative changes in Voter ID policy from one of the four categories to another. Other scholars have examined the impact of eight different types of Voter ID laws Alvarez et al. [2011], and my coding approach can be extended may differ in their implementation of similar requirements. While one state may consider a student identification card or discount club membership card to be valid photo identification, another state may only recognize government-issued photo identification cards [Alvarez et al., 2011, p. 9-10] 28 Assembling a database of Voter ID statutes for each midterm and presidential election since 2002 is an arduous task. States have unique minimum requirements, maximum requirements and policies for handling provisional ballots, multiple courts may render judgment on newly enacted Voter ID policies, and election officials such as the Secretary of State can modify policies in the days leading up to an election. In a report to the Election Assistance Commission, the authors note the complexity of Voter ID policy: We recognize the difficulties in summarizing each state s voter ID requirements. The problem is illustrated by the number of footnotes to Table 1 below. The variety of statutory and regulatory details among the states is complex (Eagleton Institute report to the EAC, p. 20) The strict photo option refers to states that do not allow voters to cast provisional ballots unless they present a photo ID : AZ, MT, SD, ND, HI, CO, AR, LA, AL, FL, TN, SC, MD; : WA, AZ, NM, FL, IN, OH, HI; : MI, GA; : ID, OK, UT. 19

20 to measure both the impact of a change in a relative and absolute sense. 32 Voter File and Census Data My primary elections data source is Catalist, a national voter database containing voter data for 270+ million Americans. 33 The query-able database includes detailed voter histories, along with demographic and commercial data appended to each name in the voter file. With this rich data source, I can compare voting patterns among subgroups within and between states. I assembled a database using voter data from the 2004 presidential election ( ) and 2008 presidential election ( ). I extracted data for the following voter groups from Catalist: African Americans, African Americans by gender, Hispanics, Hispanics by gender, adults across a series of age cohorts, family income, length of residence and owner/renter. I also collected data for the 2006 midterm election ( ) and 2010 midterm election ( ) to compare the impact of Voter ID laws in presidential versus midterm contests, and to determine the impact that Voter ID laws may exert multiple election cycles after implementation. This approach has advantages over common methods of assessing state-level turnout. These aggregations provide precise, accurate outputs of voter turnout without sampling error, compared with state-level estimates from the Current Population Survey s November supplement. Aggregate analyses that highlight counties or states with high proportions of a specific subgroup are extremely vulnerable to ecological inference concerns. Race / Ethnicity. I identified the number of adults who voted in the following racial groups: Asian, Black, Caucasian, Hispanic, Middle Eastern, Native American and nonwhite and accounted for shifts in statewide population size by using annual Census data from the American Community Survey listing total Voting Age 32 Alvarez et al. [2011] have developed an eight point classification scale based on the strictness of the statute and have categorized each state from 2000 through Their scheme includes the following categories, ranging from the least intrusive to the most stringent: 1) voter states name, 2) voter signs name, 3) voter signs and signature match, 4) voter is requested to present proof of identification or registration card, 5) voter must present proof of ID or voter registration card, 6) voter must present proof of identification and signature match, 7) voter is requested to present photo id, and 8) voter is required to present photo id [Alvarez et al., 2011]. 33 I received access to Catalist through Stanford University s Academic Subscription. 20

21 Population and Voting Eligible Population by racial group as a denominator. 34 The racial data in the Catalist database comes primarily from two sources: selfidentified responses on voter files and CPM Ethnics race prediction software. The race variable in the database uses self-identified race in many Southern states, where residents list their racial status when registering to vote. In other states, however, CPM Ethnics assigns a race based on an algorithm including the respondent s first name, middle name, last name, age and characteristics of their Census geography. Given the clustering of Americans by race, the resulting predictions are highly accurate, though, not without error. 35 We will have the highest confidence when using racial breakdowns based on data coming from the South, where voters self-report their ethnicity. Hersh (2011), for example, finds that the race variable included in the Catalist data file is accurate between 91% and 96% of the time in Southern states. 36 Age. Adults in each age group were identified using their birth date. 37 For each election between 2004 and 2010, I have compiled data for the following voting subgroups: adults under 25, 25-34, 35-44, 45-54, 55-64, 65-74, and 85 years of age and older. I accounted for shifts in statewide population size by age cohort by using annual Census data from the American Community Survey Voting Eligible Population was defined as 18 years of age and older and a citizen. 35 A note from CPM Ethnics: In external blind testing against self-reported ethnicity identification, CPM Technologies solutions have shown over 20% more coverage than other established ethnicity appending services. While other providers have software that is identifying 48% of the African Americans in lists, CPM Ethnics software can find over 75% of the African Americans in lists and still maintains an accuracy of over 80%. CPM s algorithms are based upon modern machine learning techniques and are built using tens of millions of samples with known race The exact model Catalist uses to predict race is proprietary, but we can check the quality of the prediction using survey responses that have been matched into the Catalist database. The 2009 Cooperative Congressional Election Study (CCES) was matched into Catalist s database. For the registrants with listed races, 96% of voters self-reported races were the same as the publicly listed races. For the registrants whose races were predicted with confidence, 91% had the same self-reported race as predicted by Catalist s model. Though the match between self-reports and the Catalist data is not in perfect agreement, it is sufficiently accurate that each racial group in the Catalist database can be divided in two... (Hersh 2011 p. 9-10). 37 Nearly all records contain a birth date. Records with missing age, however, were supplemented using a model the number of years the individual has been registered to vote, the age of the head of household and the individual s first name. 38 Ansolabehere and Hersh, 2010 find that 1 in 7 records does not have a listed birthdate, and for many voters who do have a listed birthdate, the date entered is inaccurate. (Ansolabehere and Hersh, 2010, p. 2). The primary inaccuracy, however, is that the voter s birthdate was entered as 21

22 Length of Residence. I identified the length of residence for individuals using household level commercial data. I isolated adults who had lived in their current household for the following periods of time: less than one year, one to five years, six to 10 years, 11 to 20 years and more than 20 years. Owner / Renter. I isolated individuals who were either renting or who owned their residence using household level commercial data. Family Income. I isolated household incomes in the following ranges using household level commercial data: less than $5,000, $5,001 to $12,50, $12,501 to $20,000, $20,001 to $30,000, $30,001 to $40,000, $40,001 to $60,000, $60,001 to $100,000 and over $100,000. Overall, the official tallies and demographic variables reported in the Catalist database are highly accurate. 39 Results and Figures This section isolates groups that scholars have hypothesized will be disparately impacted by Voter ID statutes, such as African Americans, young adults, adults over 65, Hispanics, renters and others. I compare the change in turnout of these subgroups before and after the policy intervention with broader statewide turnout and with turnout among similar groups in control states. I examine the impact of Voter ID statutes separately for presidential and midterm contests. I finally test the long-term impact of legislation by examining changes in turnout among voter subgroups four years after a policy implementation. the first of the month or as January There may be a slight discrepancy between the official vote tally and the number of votes cast in the database due to voter purges: A vote tally from a registration file excludes the votes cast by citizens who were purged from the file since the election. For instance, a person who voted in 2006 but was since removed from the rolls would not be included in the county on the registration list but would have an official ballot counted. This presents a minor problem since it only applies to voters who confirmed with the registrar that they moved (Ansolabehere and Hersh, 2010, p. 15). States vary in the discrepancy between vote tallies and official results, though the discrepancy is less than 5% in most states: The 2008 and 2006 vote history discrepancy rates vary considerably by states. In Oregon, North Carolina, Rhode Island, Delaware, and many other states, discrepancies are at a minimum, representing fewer than 5% of all votes. However, in other states like Mississippi, New York, and Texas, the 2008 discrepancy rate is closer to 10% (Ansolabehere and Hersh, 2010, p. 15) 40 22

23 Difference of means Table 1 displays each component of the difference-in-difference-in-differences estimator. Each component is weighted by the statewide subgroup population. First, we calculate the change in turnout before and after the intervention in policy states across African Americans and statewide. Second, we take the same calculation over the same time period for non-policy states. The difference between these two quantities is the impact of Voter ID statutes on the main voter subgroup. The final estimate, -.005, the DDD estimator, is the impact of Voter ID statutes among African Americans. Table 1: Difference in mean for African Americans, Policy States Non-policy states Voter Turnout Voter Turnout African Americans Statewide Difference DDD estimator Figures The figures below display treatment effects by subgroup for three separate election comparisons: presidential ( ), midterm ( ) and lagged ( ). In each figure, I aggregate turnout among tens of millions of records in the national voter file by the respective subgroup. The primary model specifications compare turnout between the November 2004 and November 2008 presidential elections. During this period, nine states tightened their ID policies and one relaxed its policies. Overall, I find that Voter ID policies exerted a limited impact on turnout during these presidential elections. The second specification compares turnout between the November 2006 and November 2010 midterm elections. During this period, five states tightened their ID policies and one relaxed its policies. These states include Georgia, Idaho, Michigan, Oklahoma and Utah. Mobilization efforts are less intense during midterm campaigns, and 23

24 I find that Voter ID laws are more likely to reduce turnout when such mobilization efforts wane. Finally, state Voter Identification policies may not affect turnout immediately after implementation but only after multiple electoral cycles. Mobilization forces may counteract any immediate, negative turnout effects by expending more effort trying to turnout African Americans and others. In future elections, however, the same rules will be in place but additional mobilization forces may focus on new Voter ID battlegrounds. I focus on the change in turnout at least two elections after the implementation of the voter identification statute. This third group captures the lagged effects among states that changed their policies between 2004 and This section proceeds by analyzing the impact of Voter ID statutes across Americans age, income, length of residence, race and home ownership. Age. Figure 5 on Page 25 displays the treatment effects across eight age cohorts. The x-axis presents the treatment effect in percentage points, the y-axis displays various age cohorts and the panels present three separate election models. The effects in the (left panel) are muddled and suggest that Voter ID laws increased turnout among young adults. Otherwise, however, the effects are in the expected direction the impact of a Voter ID statutes attenuates as we move to older age cohorts. The midterm comparison (center panel) suggests that Voter ID laws reduce turnout among young adults but may lead to a slight increase in turnout among older adults. For example, Voter ID laws reduced turnout among adults under 25 by two percentage points between 2006 and 2010, compared with a three percent increase among adults over 75 in that same time period. The lagged comparison (right panel) looks like a carbon copy of the midterm case, with Voter ID statutes reducing turnout among adults under 25 and adults aged but generally causing increases among older citizens. This figure displays clearly the advantages of using granular aggregate data. While Voter ID policies do not appear to affect overall turnout (mean treatment sizes are around 0), there are heterogeneous effects across age cohorts. 24

25 Figure 5: Voter ID treatment effects by age cohort and election. Income. Figure 6 on Page 26 displays the treatment effects across six household income strata ranging from less than $5,000 to household income exceeding $100,000 annually. Again, the effects in the 2004 to 2008 comparison are unintuitive and suggest that Voter ID statutes increased turnout among poor voters. It is possible that during this cycle campaigns targeted such voters inordinately in Voter ID states. Otherwise, Voter ID laws do not appear to have much of an impact on the change in turnout between 2004 and Among states that changed their policy between 2006 and 2010, Voter ID laws appear to exert a disparate impact among poor voters. Americans in households earning less than $13,000 annually are about three percentage points less likely to cast a ballot in Voter ID states, while statutes have a smaller influence on middle class and upper middle class Americans. In the model detailing turnout between 2004 and 2010, the effects are in the expected direction and accentuated. In states that implemented Voter ID statutes between 25

26 2004 and 2008, turnout is reduced among poor voters by at least five percentage points. Figure 6: Voter ID treatment effects by income strata and election. Length of Residence. Figure 10 on Page 31 displays the impact of Voter ID statutes based on the length of time the current resident has lived in his or her current household. Individuals who have lived in their current residence for less than one year are negatively affected by Voter ID statutes in all three panels, though the effects are most negative among adults who have lived in their household for six to 10 years. Overall, however, the plots (especially the center and right panels) suggest that Voter ID statutes depress turnout among residents who have lived in their current household for a relatively short amount of time. Across the three election comparisons, we never witness declines in turnout among individuals who have lived in their current residence for at least 10 years. 26

27 Figure 7: Voter ID treatment effects by length of residence and election. Race. Figure 8 on Page 28 displays the impact of Voter ID statutes for African Americans, Hispanics, Caucasians and all non-whites. Overall, the effects are quite small; however, the right panel suggests that Voter ID laws cause an approximately 2 percentage point decrease among African Americans while not affecting turnout among white Americans between November 2004 and November Hispanic voter turnout does not appear to be affected in any of the specifications. 27

28 Figure 8: Voter ID treatment effects by race and election. Own / rent. Finally, I aggregated data based on whether respondents were currently renters or owners. Figure 9 on Page 29 shows that Voter ID statutes do not affect turnout for either group for the 2004 or 2008 presidential elections; however, in the midterm comparison and the comparison, there are significant differences between turnout patterns among renters and owners. The data suggests that Voter ID policies decrease midterm turnout among renters by about seven percentage points and decrease turnout between 2004 and 2010 by 10 percentage points, an even larger treatment effect. 28

29 Figure 9: Voter ID treatment effects by home ownership and election. Census Block Group data The previous section analyzes individual level Catalist voter file information. I conduct a separate analysis using aggregate data at the Census block group level. There are more than 200,000 block groups across the country, meaning that the average population for these block groups is approximately 1,500 persons. This secondary analysis presents evidence broadly supportive of the findings in the previous section and introduces a few new measures of interest such as block group level household income. 29

30 Block Group election returns I obtain Block Group elections returns for the Catalist Voter database. Generally, I divide variables such as income, racial composition and home ownership rates into bins for which there are equivalent U.S. Census Bureau measures. Then, I take the sum of the total votes across the four general election cycles. Census block group denominator I download Census block group data using the American Community Survey s five-year block group estimates. I sum the number of individuals living in block groups with certain characteristics. For example, I measure the proportion of African Americans living in each block group, divide the measure into deciles and aggregate the total number of individuals living in each decile. These estimations provide a denominator for the Catalist block group election returns discussed in the previous subsection. We expect that areas concentration with African Americans, lower income individuals and renters will experience turnout declines subsequent to Voter ID statute implementations. On the other hand, block groups populated with white Americans, the wealthy or home owners should experience little to no impact in their political participation rates. Figure?? below displays Americans household income at the block group level in five midwestern states. Block Groups with the highest incomes are shaded dark red, while those with the lowest incomes are shaded blue yellow. I divide block groups into deciles by household income and estimate the average impact of Voter ID statutes across the states. For example, three of these states, Indiana, Michigan and Ohio, had a Voter ID law take effect between 2004 and Table 2 on Page 32 displays the change in turnout based on analyses at the block group level. 41 The top panel of Table 2 displays voting patterns for individuals 41 As a rule, I only display block group categories that contain at least five percent of the American population. This leads me to pool a few results. For example, four percent of Americans live in block groups with fewer than 10% home owners. I sum election returns for all Americans living in block groups with fewer than 30% homeowners. This procedure does not change any substantive takeaways. Similarly, I sum election returns for all Americans who live in block groups with an average household income above $125,000. The Census measure goes up to $250,000, but very few Americans resides in block groups with median household incomes this high. Finally, I sum election returns for Americans living in block group with at least 50% African Americans. Seven in 10 30

31 Figure 10: Block Group Household Income for Five Midwestern States based on the household income in their block group. We expect that areas with lower incomes will experience a decline in turnout relative to areas with wealthier residents. Between 2004 and 2008, turnout in block groups in Voter ID states with average household income under $25,000 annually decreased by 1.5 percentage points. In middle class and upper class block groups, the Voter ID intervention either had a minimal effect or a slight positive impact. Voter ID policies reduce turnout among the lowest two income categories for the midterm election comparison and the third comparison as well, while voter turnout among richer households actually increases in policy states. The middle panel of Table 2 displays turnout patterns based on the proportion of African Americans residing in a decile. We expect that areas with mainly African Americans live in a Census block group with fewer than 10% African Americans, meaning that the remaining nine deciles do not have huge populations. 31