Citizen Engagement and Voting Behavior in Publicly Funded Elections WORKING PAPER: Please do not cite without permission of author Michael G. Miller, Ph.D. Assistant Professor Department of Political Science Institute for Legal, Legislative, and Policy Studies University of Illinois, Spring eld
[email protected] November 27, 2011
Abstract Recent work in political science has demonstrated a link between the structural factors of elections and mass political participation.
Yet, the relationship between campaign nance systems and the
political behavior of campaigns and voters has not been fully considered. I employ original survey and interview data from candidates in eighteen states, and demonstrate that the acceptance of full public election subsidies such as those in Clean Elections systems provides candidates with time exibility facilitating higher levels of direct interaction with citizens.
No such eect is present for
candidates accepting smaller, partial subsidies. I then exploit a natural experiment to demonstrate that ballot roll-o is lower in state legislative contests when candidates accept full funding, indicating that voters are more likely to register a preference in those races.
Speci cally, the results of four
dierence-in-dierences models show that the presence of a publicly funded candidate diminishes ballot roll-o by about 2 percentage points.
Acknowledgments:
This research was supported by a National Science Foundation Dissertation Improvement
Grant [SES-0819060] as well as internal grants from the Department of Government and the Department of American Studies at Cornell University, and has appeared in previous versions at meetings of the American and Midwest Political Science Associations. I thank Walter R. Mebane Jr., Peter Enns, Suzanne Mettler, Theodore J. Lowi, Peter Francia, Conor Dowling, Dan Butler, Dino Christenson, and Seth Masket for comments on this and former iterations. Kristin Sullivan in the Connecticut O ce of Legislative Research went out of her way to track down election data for me from that state. Finally, I acknowledge an invaluable team of research assistants for their data collection eorts: Rebecca Dittrich, Benjamin Gitlin, Thomas Hudson, Christopher Martin, Rob Morrissey, and Zachary Newkirk. All mistakes, oversights, and omissions belong solely to me.
1
The candidates that I knew this last election that (funded) traditionally were having fundraisers two or three times a week, while (publicly funded candidates) were going out knocking on doors. That, I think, is a big dierence in how you spend your time. In an evening after work, I can knock on fty to seventy doors of people who will actually go to the polls for me, as opposed to that candidate who has to go out and raise and spend two, three hours with lobbyists who often don't even live in their district. Yeah, they're going to get the money, but I'm the one going out and meeting the voters. 1
-Arizona Legislator
How does the introduction of full public election funding change the behavior of American candidates and voters?
During the last decade, this question has gained salience as direct public election 2
funding programs have become both more common and more generous.
In return for compliance with
state-mandated spending limits, legislative candidates in six states who participated in optional public funding systems during the 2008 election received direct subsidy payments intended to replace private campaign contributions. Candidates in three statesHawaii, Minnesota, and Wisconsinwere eligible to accept partial payments comprising less than half of the mandated spending limit; the remainder of their funding came from private sources.
However, in the fully funded (or Clean Elections) states
of Arizona, Connecticut, and Maine, subsidies were approximately equivalent to spending ceilings and participating candidates raised almost no money from private entities during the general election (For a tabular description of direct public funding programs, see Appendix 3). Recent work in political science has demonstrated a link between participation and structural factors such as the complexity of registration requirements (Wattenberg 2002) and redistricting (Hayes and McKee 2009).
Yet, the relationship between campaign nance systems and the political behavior of
campaigns and voters has not been fully considered. In this article I show that the acceptance of full funding provides candidates with time exibility su ciently powerful to facilitate higher levels of direct interaction with citizens, and that this heightened engagement translates to more people voting in those races. My analysis links changes in campaign nance law to mass voting behavior, building on previous work nding that structural alterations have rami cations for political participation in the United States (e.g., Hayes and McKee 2009) and beyond (e.g., Franklin 2004). Accordingly, I argue that the evaluative framework used to judge campaign nance policy in general, and public election funding speci cally, should be broadened beyond a discussion of nancial numbers and electoral competition to problems of political engagement and voting behavior.
1I
derive this and all subsequent quotes from in-person interviews conducted with Arizona candidates and legislators in
January of 2007.
2 While
I focus on legislative elections here, direct full funding is becoming more common in the elections for other state
o ces as well. New Mexico implemented full funding for its Public Regulation Commission candidates in 2003, and North Carolina did so for judicial candidates in 2004. New Jersey is presently operating a pilot program in its legislative elections. Albuquerque and Portland have recently passed full funding laws for municipal contests and at least a dozen other cities are considering them.
2
My rst action is to examine data collected from a survey of lower house legislative candidates in eighteen states elded during the 2008 campaign.
I construct four public interaction indices from
the number of weekly hours candidates reported devoting to eld activity, phoning, preparing mailings, meeting with interest groups, media relations, speaking engagements, and electronic campaigning. I nd that compared to those accepting only private funding, candidates who accept full subsidiesand raise no private moneyspend a higher percentage of their time engaged in public interaction activities, but that no eect is present for candidates in partially funded systems. My second task is to analyze voter roll-o trends in Arizona, Connecticut, and Maine state legislative races both before and after the implementation of public funding. I nd that in both house and senate elections for all three states, mean roll-o is lower when at least one candidate ran with public funding in the rst election for which public funding was available; an examination of the mean dierence in samedistrict roll-o trends before and after implementation yields the same conclusions. I substantiate these ndings with the calculation of dierence-in-dierences in roll-o from Maine and Connecticut state house elections, using a multivariate regression model that rules out potential confounding factors. Speci cally, I nd that ballot roll-o in treated districts is lower by about 1.5 percentage points in Maine and 2 points in Connecticut. In tandem, the ndings I present in this paper suggest that public election funding aects the political behavior of candidates and voters alike, in a manner that has not been previously described. The paper proceeds as follows:
In the following two sections, I describe the theoretical basis for
hypotheses predicting that full public funding alters candidate and voter behavior, respectively. Next, I explain the research design employed to gauge the eect of public funding on campaign time, followed by the results of that design. I then detail the methods used to examine the relationship between full funding and voting behavior, and the resultant ndings. I close with a discussion of the results and suggestions for future research.
1 Public Funding, Public Interaction, and Voting Behavior Campaign eld operations attempt to sway voters with a combination of media and direct solicitation, and the preponderance of the evidence to date suggests that genuine, personal appeals are more successful at stimulating turnout. Phone calls designed to get out the vote have little eect when they are delivered from large professional phone banks in both a non-partisan (Gerber and Green 2000; Gerber and Green 2001; McNulty 2005) and partisan (Panagopoulos 2009) context. The ineectiveness of professional phone banks is likely due to the low-quality of the message delivered, since their solicitations are often hurried
3
and impersonal (Nickerson 2007). These ndings are consistent with a growing number of mobilization experiments that have found welltargeted, more intimate messages are particularly eective mobilization tools. For instance, a positive eect on turnout has been demonstrated when phone solicitations come from volunteers able to more eectively engage voters (Nickerson, Friedrichs, and King 2006; Nickerson 2006; Nickerson 2005; Ramirez 2005; Wong 2005). Face-to-face canvassing techniques appear to be the most e cacious voter mobilization tools, particularly when delivered on-time to targeted populations (Kramer 1970; Gerber and Green 2000; Niven 2001a; Niven 2001b; Niven 2002; Green, Gerber, and Nickerson 2003; Michelson 2003; Bennion 2005; Hillygus 2005; Nickerson, Friedrichs, and King 2006; Parry et al. 2008). In sum, when it comes to mobilizing support, the most productive campaigns will most likely be those able to make personal connections with voters. The good news for many candidates is that direct communication with voters is probably a preferred activity; the returns from high quality interactions are perceptible, while activities such as canvassing and speeches conform to popular conceptions of what a campaign looks like.
However, despite the
evidence that political campaigns are crucial forces of voter engagement in American elections, the rise of professionalized campaigns increasingly dependent upon expensive media operations has forced many candidates to choose between two distinct campaigns: one for money, and another for votes (Herrnson 2004).
Given the explosion in campaign spending since the mid-1970s, it appears that the former is
winning the battle for candidates' attention (see:
Malbin, Ornstein, and Mann 2008), as campaigns
increasingly opt for mass-media tactics at the expense of retail politics (Schier 2000, 124). In short, candidates in most American elections face a maximization problem. Time is a nite resource; few candidates enter politics so that they can spend countless hours raising money, but fundraising is necessary in order to pay for other campaign activities. Raising funds can be a signi cant challenge that requires a tenacious dedication, but for all the eort expended to raise funds, campaign spending itself does not determine victory (Brown 2011). In terms of expected return on the investment of their
time,
an hour spent raising money therefore imparts an unclear return that for most candidates is worse than the utility derived from an hour spent with voters.
3
So, even while they feel intense pressure to raise
funds, I assume that most candidates will prefer to maximize the amount of time they spend interacting with citizens.
3 The
act of fundraising certainly conveys positive information about the candidate, but in a well-targeted funding plan,
this information is likely to be distributed to individuals and groups who were already predisposed to support the candidate. Moreover, unlike get-out-the-vote eorts, fundraising is not con ned to district boundaries, and a candidate might solicit donors who do not have the ability to cast a vote in the election.
4
From the perspective of candidates, the availability of public election funding has vast potential to resolve this tension. Public funding programs present candidates with a choice between a traditional campaign in which they must devote a signi cant amount of time to fundraising, and one in which the state allocates a single, very large contribution at the outset. The appeal of public funding for a candidate is that acceptance of this subsidy reduces or eliminates the amount of additional money that she must raise to conduct her campaign.
Put another way, I argue here that one understudied eect of public
funding is its potential to alter the tasks that candidates must complete. If candidates no longer need to raise money from private sources, they will logically devote signi cantly less eort to fundraising, which in turn allows them to recover a time opportunity cost by focusing more eort on direct voter engagement. That said, not all funding programs are equally likely to alter candidate behavior.
In the partial
funding programs of Hawaii, Minnesota, and Wisconsin, subsidies typically amount to less than 25% of the average cost of a contested race. However, in the fully funded Clean Elections programs in Arizona, Connecticut and Maine, publicly funded candidates can expect their payments to comprise more than 90% of average campaign expenses on average, and candidates are barred from further fundraising once they 4
accept public funding.
Thus, the candidate experience in a partially funded system is much the same
as in a traditional, privately nanced one: Challengers in particular must still persuade skeptical private donors to contribute, and so they must devote substantial time to fundraising.
Perhaps due to these
conditions, previous analysis has found little competitive change in states oering partial subsidies (Jones and Borris 1985; Mayer and Wood 1995; Malbin and Gais 1998, 136; but see: Donnay and Ramsden, 1995). In contrast, full funding programs appear to stimulate competition in legislative elections (General Accounting O ce 2010, 35; Mayer, Werner, and Williams 2006; Werner and Mayer 2007; Malhotra 2008). Despite a clear potential for Clean Elections programs in particular to aect candidate behavior, little is known about the manner in which public funding changes candidates' ability to mobilize voters. Francia and Herrnson (2003) con rmed that fully funded candidates spend less time raising money, which should not be surprising given that they are legally precluded from this activity. However, Francia and Herrnson did not examine how public funding aects candidates' voter mobilization activities.
As noted above,
I assume that candidates are utility maximizing actors, and should be expected to pursue a strategy that will result in the highest number of votes. Compared even to partially nanced challengers, those accepting full subsidies, who simultaneously enjoy nancial parity with their opponents and the absence 5
of fundraising as a necessary campaign task, are likely to wage a more visible, eld-oriented campaign.
4I
base
this
claim
on
data
obtained
http://www.followthemoney.org
5 The
recent Supreme Court decision in
from
the
National
Institute
on
Money
in
State
Politics,
available
at
McComish v. Bennett struck down the matching funds trigger provisions, and
will eectively eliminate the guarantee of nancial parity unless the Clean Elections programs described here are amended.
5
If this theory is true, then I should nd support for two hypotheses:
Partially funded candidates will devote about the same percentage of their weekly time to public interaction activities as candidates who fund only with private sources. H1:
Fully funded candidates will devote more of their weekly time to public interaction activities than candidates who fund only with private sources. H2:
The natural following question is how public funding might also change mass voting behavior. Rosenstone and Hansen (1993, 218) noted that the media-focused campaign has come at the expense of mass mobilization, and is partially responsible for the decline in voter turnout since the 1960s. If full public funding heightens interaction between the public and candidates, it seems reasonable to expect more people to vote in elections contested by fully funded candidates.
Yet in the context of relatively low-
information state legislative elections, the classic expressions of turnout (the number of votes cast divided by the number of voting-age or voting-eligible citizens) are not the ideal place to look for altered voting patterns, because it is unrealistic to assume that down-ticket races alone can aect the number of citizens who turn out to vote. The presence of public funding in a state legislative election is likely to raise public awareness about that contest. Even so, few people are likely to base their decision on whether to vote
at all on the goings-
on of a state legislative election, and this is particularly true in a presidential election year. Studies of campaign nance laws generally (Primo and Milyo 2006) and Clean Elections speci cally (Miller and Panagopoulos 2011) have yielded little reason to conclude that such policies improve mean levels of e cacy among the electorate. Moreover, Milyo, Primo, and Jacobsmeier (2011) report no relationship between public funding and turnout. That said, it is premature to conclude that public funding has no eects on mass voting behavior. Enhanced engagement between legislative candidates and the public may not drive the turnout decision, but if it increases instances of high-quality interaction between candidates and voters, then it should also heighten general levels of political information and/or election salience among the mass electorate. This, in turn, should raise the likelihood that citizens will cast ballots in the legislative election
after
voting
for higher-pro le o ces. In other words, the presence of a fully-funded candidate in a state house race may not compel a citizen to visit a polling place, but it voters
should diminish voter roll-o, which occurs when
who have already turned out stop marking their ballots for lower o ces.
Roll-o is a di cult phenomenon to reconcile with the classic rational choice voting models (e.g. Downs 1957; Riker and Ordeshook 1968) which predict that citizens will vote when the cost of doing so approaches zero, as it does when a voter is already in the booth. However, when voters know little
6
or nothing about the candidates in a given race, the odds that they will make an incorrect decision increase (Lau and Redlawsk 1997). Most partisan voters in low-information elections can use party labels as reliable cues, even if they know little else about the candidates, but Feddersen and Pesendorfer (1996) demonstrate that when weak partisan voters face a great deal of uncertainty, they have an incentive to abstain from voting and delegate the choice to more informed citizens. Thus, in low-visibility elections (such as those for a state legislature), information is a crucial commodity. Indeed, much of the existing political science literature has approached roll-o as a problem to be solved by raising mass awareness of the candidates and issues in a given contest. Many of these studies have been conducted to better understand a well-documented racial gap, with African-American voters signi cantly more likely to roll-o (Feig 2007; Kimball and Kropf 2005; Herron and Sekhon 2001; Brady et al. 2001; Knack and Kropf 2003; Vanderleeuw and Engstrom 1987). Previous research has also found that African-Americans appear to be more likely to vote in a given contest when black candidates run against white candidates (Vanderleeuw and Utter 1993; Engstrom and Caridas 1991), particularly when black candidates make focused mobilization eorts (Vanderleeuw and Liu 2002). These ndings underscore the importance of information and salience in reducing roll-o in lowvisibility elections. Wattenberg, McAllister, and Salvanto (2000) found that voters approach ballots in much the same way as a test, abstaining when they lack su cient knowledge to make a clear decision. Information in professional, media-heavy campaigns is easier to obtain; yet the attention that top campaigns receive makes it even more di cult for less-visible candidates to provide voters with many of the useful heuristics they use to make a decision, such as incumbency, occupational background, or endorsements (e.g., Lau and Redlawsk 2001; McDermott 2005). Particularly in low-visibility elections such as those for most state assemblies, higher levels of interaction between candidates and the public hold great promise to increase the probability that voters will express a preference in a given contest if they have already turned out to vote. This leads to a third hypothesis:
The presence of a fully funded candidate in a state legislative election signi cantly reduces voter roll-o. H3:
In the following sections, I test the three hypotheses described above.
Due to the nature of the
questions, the appropriate design and methodological concerns are substantially dierent for Hypotheses 1 and 2 than for Hypothesis 3. As such, in the next section, I detail the research design for Hypotheses 1 and 2, followed by a discussion of the results.
Next, I describe the methodology employed to test
Hypothesis 3, and the outcome of that analysis.
I close with a discussion of the ndings for all three
hypotheses.
7
2 Research Design: Public Interaction 2.1 Candidate Survey To determine how public funding aects the manner in which candidates use their time, I collected survey data from the major party lower house candidate populations in eighteen states during the 2008 election, including all six states oering ubiquitous full (Arizona, Connecticut, and Maine) and partial (Hawaii, Minnesota, and Wisconsin) public nancing to legislative candidates. The survey instruments solicited responses to questions regarding candidate attitudes toward their campaign, the electorate, and their competition. Each candidate was also asked to quantify the amount of time he or she personally devoted to various tasks in ten areas, including fundraising, public speeches, eld activity, electronic campaigning, 6
media relations, research, strategy, phoning voters, sending mailings, and the courting of interest groups.
Where applicable, I supplement the survey data with the testimony of candidates collected during inperson interviews in the wake of the 2006 Arizona legislative election, as well as demographic and political characteristics of each legislative district obtained from Lilley
et al.
(2007). The survey instrument is
available upon request. To maximize the possibility of meaningful comparison between publicly and privately (or traditionally) nanced candidates, I also surveyed the candidate populations of twelve additional states with no available public nancing for legislative candidates. Criteria for selection as a comparison state include: the average cost of a legislative campaign, proximity within Squire's (2007) index of legislative professionalization, average district population, chamber size, electoral time-line, and where possible, geographical location as a proxy for regional culture dierences. Thus, I add Alaska, Colorado, Delaware, Iowa, Michigan, Missouri, Montana, New Mexico, Ohio, Rhode Island, Vermont, and West Virginia to the sampling frame, which in total contained 2,971 candidates. Response rates in surveys of elite candidate populations tend to be low, often less than forty percent (e.g., Francia & Herrnson 2003; Howell 1982).
I made multiple contacts to combat this issue, devot-
ing particular attention to fully funded candidates, who comprised roughly 15% of the sampling frame. Candidates received the rst contact, including prepaid return envelopes and links to an identical online version, during the rst week of October. I mailed reminder postcards in mid-November and sent additional monthly electronic invitations to available addresses of non-respondents in mid-October, November, 7
and December.
6 The
At that time, I assessed the response rate of each state, and re-sent full survey packages
question was worded as follows: DURING THE FIRST WEEK OF OCTOBER, what is your best estimate of how
many hours you, yourself, spent engaged in the following activities? Please complete the table below, listing your NUMBER OF HOURS, and NOT A PERCENTAGE OF TIME. If you accepted public funding, do not include time spent qualifying for public money as part of the fundraising category.
7 Electronic
mail addresses were obtained for approximately 60% of the overall candidate population.
8
to non-respondents in both fully funded and low-responding states. I made nal contact with remaining non-respondents in those states by phone in mid-December.
Thus, candidates in the sampling frame
received up to eight contacts, and there was a higher probability of more contact for publicly funded 8
candidates (especially those accepting full funding).
I received 1,022 responses overall, for a response rate of 34.4%. As noted above, this rate is consistent with previous surveys of elite candidate populations (Francia & Herrnson 2003; Howell 1982).
State
response rates ranged between 23.7% in Rhode Island and 49.5% in Arizona. Appendix 1 contains stateby-state response rates, as well as basic characteristics of both the sample and candidate population in the survey frame.
The sample contains a higher percentage of fully funded candidates than the
population, which is resultant of a concentrated eort to illicit responses from those candidates. Women and Democrats also comprise a higher proportion of the sample than the population, which should be expected given the high response rates of fully funded candidates in tandem with existing evidence that both Democrats and women are more likely to accept full public funding in state house elections (General 9
Accounting O ce 2010; Werner and Mayer 2007).
2.2 Matching Design While the resulting dataset is promising, a comparison of the behavior of publicly and privately funded candidates poses some challenges.
For instance, candidate funding status is not likely to be random.
The afore-mentioned dierential participation among candidates of opposite parties or genders described in General Accounting O ce (2010) and Werner and Mayer (2007), as well as the correspondent overrepresentation of women and Democrats in the survey sample, serve as a reminder that individual and/or district characteristics likely aect candidates' public funding status. As such, the covariates are unlikely to be balanced between publicly and privately funded candidate groups because the receipt of public election funding necessarily implies that those groups are drawn from dierent populations. Since imbalance would make inference about public funding di cult using testing of means or even multivariate regression (see: Sekhon 2009), I seek improved balance between the groups' covariates in order to support more reliable conclusions. A necessary condition for such balance is that observable characteristics have the same distribution in the traditionally and publicly funded groups. To achieve this balance, I follow the language of experiments in designating a treatment group of publicly funded candidates and a control group of candidates who
8 There
are no apparent signi cant dierences on the outcome measures between responses collected via mail, Internet,
and phone (two-tailed tests,
9 To
α = .05).
be clear, the eort to over-sample candidates who accepted Clean Elections subsidies in Arizona, Connecticut, and
Maine almost certainly also raised the proportion of women and Democrats in the sample. The matching strategy described below is one way to address this issue in analysis.
9
raised money exclusively from private sources. I then pre-process the survey data with a genetic matching algorithm intended to improve the balance of the observed covariates between treatment and control groups (On pre-processing, see Ho see:
et al.
2007a; 2007b.
On genetic matching and search algorithms,
Sekhon 2008; Sekhon 2006; Diamond and Sekhon 2006; Mebane and Sekhon 1998; Sekhon and
Mebane 1998). The algorithm seeks optimal balance via the construction of matched pairs consisting of one candidate who participated in public funding implemented the alert and one who did not, parsing candidates of one condition whose covariates do not align well with at least one candidate in the opposite condition. Pre-processing in this fashion generally results in a smaller dataset, albeit one in which the observed covariates are more similar between the two groups (Ho described in Ho
et al.
et al.
2007a). The software package
(2007b) supplies post-matching weights that can be used in subsequent analysis of
a parsed dataset. I proceed with an important caveat.
As described above, I expect that candidates in partial and
full public election funding systems experience markedly dierent incentives and political conditions. Accordingly, I perform separate matches for the two groups, re ecting my belief that treatment is likely to have disparate eects on the behavior of partially and fully funded candidates, since the former continue raising money throughout the campaign while the latter do not. Thus, I construct a separate dataset for each treated group, consisting of all candidates who were subjected to that particular treatment and all candidates who were not.
To be clear, the pool of candidates in the control condition is the same
for fully and partially funded candidates, but it is not possible for a fully funded candidate to be paired with a partially funded candidate. The samples from partially and fully funded states contain 96 and 156 publicly funded candidates, respectively, and in both cases the algorithm nds matches for all treated 10
candidates.
10 I
delete list-wise 161 cases in which the candidate was not opposed by a major party challenger, 16 cases in which
candidates did not complete the time component of the survey, 42 cases in which candidates' reported campaign time exceeded the number of hours in one week, and 5 cases in which candidates did not report their name or district, thus making information about their races irretrievable.
Logistic regressions indicate that problematic and missing data are
distributed randomly, and I am con dent that the exclusion of these cases does not bias the substantive ndings.
10
Table 1: Mean Values of Observable Covariates and Balance Assessment, Full Funding Means: Unmatched Data
Means: Matched Data
Treated
Control
Treated
Control
Mean Di
eQQ Med
Percent Improvement eQQ Mean
eQQ Max
Candidate:
Challenger
0.38
0.42
0.38
0.37
52.69
0.00
100.00
100.00
Incumbent
0.35
0.33
0.35
0.37
6.11
0.00
100.00
100.00 0.00
Open Seat
0.27
0.25
0.27
0.27
100.00
0.00
48.79
Male
0.59
0.67
0.59
0.60
85.54
0.00
89.03
0.00
Democrat
0.64
0.56
0.64
0.64
100.00
0.00
45.13
0.00
African-American
0.01
0.02
0.01
0.01
100.00
0.00
100.00
100.00
Experience
0.63
0.52
0.63
0.63
100.00
0.00
40.25
0.00
Employed Part-Time
0.28
0.24
0.28
0.28
85.91
0.00
34.16
0.00
Employed Full Time
0.46
0.47
0.46
0.60
-761.34
0.00
-412.12
0.00
26.50
District:
District Population
40,374
43,634
40,374
37,780
20.44
60.38
44.61
Perc. Urban
64.00
64.39
64.00
65.71
-337.36
42.84
30.74
8.80
Perc. Black
1.95
3.35
1.95
1.94
99.73
53.92
74.33
90.14
6.01
5.49
6.01
4.87
-120.40
-9.21
-17.42
51.09
Household Income
70,004
60,712
70,004
61,696
10.58
15.66
26.39
-8.55
Perc. Coll. Degree
21.40
18.90
21.40
20.11
48.56
44.44
30.59
0.00
2.49
2.53
2.49
2.53
3.62
-50.00
-20.65
50.82
Perc. Hispanic
Household Sz. Multi-Member Dist. Party Vote, 2006 3rd Party Candidate
0.17
0.13
0.17
0.19
37.88
0.00
74.39
0.00
52.18
49.21
52.18
51.86
89.19
33.47
31.15
18.79
0.10
0.33
0.10
0.12
92.14
0.00
87.87
0.00
I allow the matching algorithms to operate on 19 covariates that were unchanged between the point at which the treated candidates opted into public funding and the point at which they were observed. These covariates include measures of candidate employment status, political experience, gender, partisan a liation, and race, as well as characteristics of the candidate's election and demographic traits of the candidate's district. In addition, I include the propensity score (the probability of receiving treatment given a set of observed covariates) in the algorithm, calculated with a logistic regression model (see: Diamond and Sekhon 2006). desired measures.
Ideally, the matching algorithm should improve balance on most of the
11
Tables 1 and 2 contain mean levels of relevant covariates for the full and partial funding groups, respectively, both before and after conducting the matches, as well as the post-matching percent improvement on a number of balance measures. The four leftmost columns of Table 1 depict sample mean dierences between fully funded and privately funded groups, both before and after matching. Entries in the rightmost columns re ect the percent improvement in this dierence, as well as the percent improvement in empirical quantile measures. Taken together, the entries in Table 1 raise con dence that matching results in a better-balanced dataset for fully funded candidates. The absolute value of the mean dierence is smaller after matching for all but three covariates: the indicator for candidates working full-
11 I
attempted more than 20 matches for both full and partial candidate groups, and use the ones that provided superior
overall balance in each case.
11
12
time, the Hispanic population percentage of the district, and the district's urbanity.
The empirical
quantile measures also show marked improvement for most covariates post-matching. The entries in Table 2 yield similar conclusions about the post-matching balance for partially funded candidates, even though the covariates for partially funded candidates are quite similar to the privately nanced group even before conducting the match. Like the fully funded group, the post-matching mean dierence for partially funded candidates is larger for three covariates: the indicators for incumbent and full time employment, as well as the percentage of the district population with a college degree. However, the post-matching dierence in means on all three of these covariates is only about one percentage point. Like those from the matching exercise for fully funded candidates, the empirical quantile measures are also improved post-matching for most covariates in the partial funding match. In short, for both the fully funded and partially funded candidate groups, I am con dent that matching improves balance.
Table 2: Mean Values of Observable Covariates and Balance Assessment, Partial Funding Means: Unmatched Data
Means: Matched Data
Treated
Control
Treated
Control
Mean Di
eQQ Med
Percent Improvement eQQ Mean
eQQ Max
Candidate:
Challenger
0.54
0.42
0.54
0.54
100.00
0.00
26.04
0.00
Incumbent
0.32
0.33
0.32
0.33
-42.90
0.00
-62.71
0.00
Open Seat
0.14
0.25
0.14
0.13
90.72
0.00
40.83
0.00
Male
0.72
0.67
0.72
0.72
100.00
0.00
34.92
0.00 0.00
Democrat
0.64
0.56
0.64
0.65
86.35
0.00
76.76
African-American
0.02
0.02
0.02
0.02
100.00
0.00
-Inf
-Inf
Experience
0.49
0.52
0.49
0.51
38.49
0.00
-62.71
0.00
Employed Part-Time
0.29
0.24
0.29
0.29
100.00
0.00
34.92
0.00
Employed Full Time
0.47
0.47
0.47
0.45
-414.98
0.00
-Inf
-Inf
District:
District Population
38,557
43,634
38,557
39,977
72.04
30.65
27.92
45.60
Perc. Urban
67.00
64.39
67.00
65.28
34.04
-153.96
-103.35
-9.26
Perc. Black
2.85
3.35
2.85
2.68
64.97
81.26
32.94
4.62
Perc. Hispanic
2.56
5.49
2.56
2.47
96.99
55.39
90.65
80.30
Household Income
67,837
60,712
67,837
66,549
81.91
64.19
43.41
-66.69
Perc. Coll. Degree
19.18
18.90
19.18
18.13
-273.19
-120.00
-101.14
-83.33
Household Sz.
2.58
2.53
2.58
2.56
67.97
33.33
34.97
81.44
Multi-Member Dist.
0.00
0.13
0.00
0.00
100.00
0.00
100.00
100.00
47.34
49.21
47.34
46.13
35.22
42.44
10.87
12.50
0.99
0.33
0.99
0.98
98.43
100.00
97.46
0.00
Party Vote, 2006 3rd Party Candidate
12 While
the mean dierence for the latter two covariates remains fairly small post-matching, despite showing worse
balance, the dierence in mean probability of full-time employment is much larger after conducting the match. I return to this point below.
12
2.3 Model Speci cation As noted above, the matching exercises yield separate datasets: one for partially funded candidates and one for fully funded candidates. To determine the eect of accepting public funding on campaign time, I t four separate OLS regression models to each dataset, using sampling weights obtained from the matching exercises (see: Ho
et al.
2007b).
For all models, I seek the average treatment eect on the
treated, or the eect of accepting public funding on the public interaction activities of candidates who accepted public funding. Since the campaign activities that comprise direct public interaction are open to some interpretation, I construct four related outcome variables using additive indices of time measured in raw weekly hours or fractions thereof that candidates devoted to various campaign activities; the dependent variable for each model is the percentage of overall weekly campaign time that each candidate devoted to that index of activities. In other words, the outcome measures incorporate a range of measurements for the percentage of campaign time that candidates devote to interacting withand persuadingvoters. The dependent variable for Model 1 is the percentage of weekly campaign time that each candidate devoted solely to eld activity (canvassing, sign-posting, and other so-called retail politics).
Thus, the outcome variable in Model
1 represents the most direct measure of face-to-face interaction between candidates and voters.
The
outcome measure in Model 2 captures the opposite extreme; it is the percentage of time that candidates spent on all activities
except fundraising.13
The dependent variables in Models 3 and 4 are intermediate
measures of public interaction. In Model 3, the dependent variable is calculated from an additive index of eld activity, electronic campaigning, media relations, public speaking, compiling mailings, phone calls, and interest group meetings, since all of these activities can be plausibly construed to be related to direct public interaction with voters. By contrast, sta meetings and research are left out of the index in Model 3 (in addition to fundraising) since they can be viewed as housekeeping activities that occur out of the view of voters. The dependent variable in Model 4 subtracts mail preparation from the index used in Model 3, since the time a candidate spends on mail might be interpreted as indirect interaction with voters. If the eect of public funding is substantively similar across the measures, then con dence in the robustness of the ndings should be higher. The independent variable of interest in all models is a dichotomous indicator coded 1 if the candidate accepted public funding and 0 otherwise. I also include several additional control variables to account for variation in public interaction activities between candidates, even though matching substantially
13 I
exclude fundraising from all public interaction indices. While the candidate must interact with the public for fundrais-
ing purposes, the fundraising audience is comprised of a narrow sector of the electorate with well-known, favorable preferences (see: Wilcox 2001). In other cases, funds may be solicited from individuals who live outside of the candidate's district or state, negating any potential electoral bene t from fundraising activities.
13
improved balance on most covariates. Speci cally, I add a dummy variable coded 1 if the candidate has been previously elected to a public o ce, since experienced candidates are more likely to have an existing base of support and could conceivably need to spend less eort fundraising than political neophytes. I also include an indicator coded 1 if the candidate is a man.
Previous scholarship has found that
women state legislators are more likely to seek close interaction with constituents than men (Epstein, Niemi, and Powell 2005), and it seems reasonable to conclude that women may exhibit similar tendencies as candidates.
Finally, I add the district population and its mean household income (both in tens of
thousands). Candidates in more populous districts may nd a door-to-door campaign daunting, opting for a media-centric strategy that requires more money (and more fundraising) to execute. Similarly, those in lower-income districts might have to look harder to raise the requisite funds, thus increasing the eort they devote to fundraising. If the analysis supports Hypotheses 1 and 2, then I anticipate that the models of partially funded candidates will return a statistically insigni cant coe cient on the partial public funding indicator and a positive, statistically signi cant coe cient for that variable in the models of fully funded Clean Elections candidates.
In the following section, I detail the results of the models described here and assess the
empirical support for these hypotheses.
3 Findings: Public Interaction 3.1 Summary Statistics Before presenting the results of the models, I rst examine summary statistics derived from the candidate survey. Figure 1 depicts the mean percentage of time that candidates reported devoting to fundraising, eld activities (canvassing, sign posting, etc.), and other campaign tasks, arranged by state and candidate funding status.
14
As noted above, eld work is the narrowest indicator of public interaction behavior
employed in my analysis. The six leftmost bars in Figure 1 re ect the time allocations of participating and non-participating candidates in the Clean Elections states of Arizona, Connecticut, and Maine, which oer full public subsidies.
In all three states, one-tailed tests indicate that candidates in the publicly 15
funded group reported devoting a signi cantly lower percentage of their campaign time to fundraising.
Publicly funded candidates in those three states also reported higher mean levels of eld activity, although only in Connecticut is the dierence statistically signi cant (one-tailed test,
14 Candidates
p=.0002).
That said, the
are not separated into publicly and privately funded groups in Minnesota and Hawaii because only 1
respondent participated in Hawaii's public funding program, while only 2 respondents opted out in Minnesota. In those states, I report overall mean percentages only.
15 P-values
for Arizona, Connecticut, and Maine are .0000, .0189, and .0000, respectively. Two-tailed tests yield the same
conclusions.
14
basic picture depicted in the fully-funded states is one of candidates spending less time raising funds, allocating the remainder of their campaign time dierently than traditionally nanced candidates.
The same cannot be said for partially funded candidates in Wisconsin, which is the only state employing partial public funding for which respondent numbers are su ciently large to facilitate comparison. Wisconsin candidates who took partial subsidies do report slightly lower and higher mean levels of fundraising and eld activity, respectively, but in neither case is the dierence in means signi cant at
α = .05.
Moreover, the mean percentage of fundraising time for publicly funded candidates in Wisconsin,
and in Minnesota, where all respondents but two accepted subsidies, is comparable to that in many other states in which no public funding is available. At very least, the patterns evident in Figure 1 oer no reason to reject Hypotheses 1 or 2.
3.2 Model Results I now report the results of the regression models described in the previous section. The coe cients for all models are contained in Table 3.
Each model employs a separate dependent variable, re ecting a
dierent construction of a public interaction index measuring the percentage of time that candidates devoted to various campaign activities. exercise described above.
All models utilize sample weights derived from the matching
I report OLS regression coe cients and robust standard errors clustered by
state to account for any non-random error variance that may be present due to the multi-state composition of the candidate pools.
15
3.2.1
Partially Funded Candidates
I begin with an examination of the models for partially funded candidates. The most noticeable trend for the models of partially funded candidates is that they do not seem to explain very much about public interaction propensity, with
R2
measures ranging from .01 to .06. Furthermore, in terms of explaining
signi cant changes in public interaction behavior, most of the control variables perform poorly in the models of partially funded candidates. Candidate experience, measured as a dummy variable coded 1 when the candidate had previously been elected to a political o ce, demonstrates a large positive eect in Model 1, but does not achieve signi cance for any of the other models, and is negatively signed for Model 2. As anticipated, the coe cient for the dummy variable coded 1 for a male candidate is signed negatively for all models, but is not statistically signi cant for any of the models.
The coe cient for
district population is also negatively signed for three of the four models, whereas the coe cient for average household income in three of the models is positively signed. However, as noted, none of these relationships are signi cant at
α = .05.
The same can be said of the covariate of interest, which is a dummy variable coded 1 if a candidate accepted partial public funding.
For all four dependent variables, the coe cient for partially funded
candidates is positively signed, substantively small, and does not indicate a statistically signi cant relationship between participation in partial public funding programs and the percentage of time devoted to public interaction. The performance of the control variables in the models utilizing weighted matched pairs is disappointing. However, the substantive nding of all four models regarding the partial public funding indicator is consistent with the additional model speci cation of unmatched data in Appendix 2, which I include as a robustness check. In sum, it seems safe to conclude that partial public funding does not stimulate direct voter mobilization eorts. This nding is consistent with the expectations of Hypothesis 1 above, and is not surprising considering that the nancial realities partially funded candidates face typically require them to raise money from private sources throughout the election.
3.2.2
Fully Funded Candidates
The picture in fully funded elections is markedly dierent. As the entries in Table 3 indicate, the control variables possess more explanatory power in the models of fully funded candidates. For all four models, the coe cients for the indicator variables for male candidates and those with previous elected experience display a negative sign.
However, for the most part, candidate gender and political experience are
not signi cant predictors of public interaction activities; the coe cient for the male dummy variable is statistically signi cant only in Model 3. In contrast, the coe cients for district population and income
16
display a negative, statistically signi cant (α
= .05)
correlation with public interaction percentages; in
most instances, candidates appear less likely to pursue public interaction activities in more populous 16
and more wealthy districts.
The inverse relationship between district income and public interaction is
contrary to the expectations described in the previous section; one possible explanation is that candidates nd more opportunities for fundraising events in wealthier districts.
Table 3: OLS Regression Coe cients: Candidate Public Interaction Activities
Partial Funding
Can. Accepted Full Public Funding
Can. Accepted Partial Public Funding
Can. Previously Held Elected O ce
Candidate is Male
Dist. Population (Ten Thousands)
Dist. HH Income (Ten Thousands)
Constant
Full Funding
(1)
(2)
(3)
(4)
(1)
(2)
(3)
(4)
-
-
-
-
11.673*
11.704*
12.064*
11.625*
(2.361)
(2.074)
(2.032)
(2.118)
0.609
1.769
1.628
0.501
-
-
-
-
(3.082)
(1.908)
(1.433)
(2.435)
10.213*
-0.407
3.265
3.507
-2.266
-1.027
-1.756
-2.997
(3.980)
(2.271)
(2.405)
(2.995)
(2.555)
(1.448)
(2.140)
(2.483)
-0.627
-0.440
-1.309
-2.527
-0.299
-2.789
-4.298*
-2.817
(5.940)
(1.629)
(2.784)
(2.998)
(2.419)
(1.474)
(1.684)
(3.216) -0.378*
-0.803
0.091
-0.772
-0.440
-1.121*
-0.259*
-0.351*
(1.663)
(0.704)
(0.651)
(0.820)
(0.123)
(0.078)
(0.138)
(0.107)
0.604
-0.080
0.197
0.886
-0.770*
-0.266
-0.678*
-0.591*
(1.693)
(0.310)
(0.721)
(0.598)
(0.277)
(0.180)
(0.141)
(0.159)
35.942*
90.084*
78.113*
68.528*
48.329*
92.536*
85.158*
77.319* (3.042)
(15.134)
(3.681)
(5.547)
(6.380)
(3.868)
(2.885)
(2.843)
N
155
155
155
155
279
279
279
279
R2
0.06
0.01
0.04
0.04
0.15
0.31
0.19
0.14
RMSE
21.86
9.82
13.55
14.85
19.99
9.09
14.0
16.28
F -Statistic
16.70
1.72
17.09
25.38
25.73
41.29
27.26
16.05
*p<.05. Robust standard errors in parentheses, clustered by state. Sampling weights derived from genetic matching. Dependent variables are the percentage of time (ranging from 0 to 100) devoted to various public interaction indices. 1: Percentage of time devoted to eld activities (canvassing, sign posting, etc.). 2. Percentage of time devoted to all campaign activities except fundraising. 3. Percentage of time devoted to all campaign activities except fundraising, research, and sta meetings. 4. Percentage of time devoted to all campaign activities except fundraising, research, sta meetings, and mail preparation.
The coe cients for the indicator variable for a candidate accepting full public funding are all positively signed, substantively large, and statistically signi cant. Moreover, the eect size of accepting full funding is similar for all four constructions of the public interaction variable: about 11.5 percentage points. In other words, holding other model covariates constant, candidates in Arizona, Maine, and Connecticut who accept full public funding devote signi cantly more time to public interaction, and this eect remains
16 As
noted above, the full funding dataset demonstrates poor balance on the covariate denoting whether a candidate
was employed full-time during the campaign. While that variable is theoretically relevant in a model of raw hours that a candidate spent on the campaign, it is not immediately clear that it should be included in the present models, in which the dependent variable is the percentage of campaign time devoted to public interaction (regardless of how much time a candidate spent on the campaign). The same is true for the indicator variable for Democrats, who responded to the survey at disproportionately higher rates. Alternate speci cations including these two variables are available upon request, but their inclusion adds no explanatory power to the model, and changes neither the size nor the signi cance of the public funding indicators.
17
stable across various constructions of public interaction indices.
In contrast to candidates accepting
partial funding, who display no change in their public interaction behavior, participation in full funding programs appear to alter the manner in which candidates conduct their campaign. To sum, the subsidy size matters. Participation in partial public funding systems has no eect on the way that candidates use their time, but the acceptance of full funding leads to higher levels of interaction between candidates and the public. With an eect size of 11.5 percentage points, a fully funded candidate entering a legislative race in the rst week of June who campaigns at the mean level of total campaign hours (46.8) reported by survey respondents would devote over 1,000 additional hours to public interaction during the 22 week eort. The shift toward higher levels of interaction with voters therefore has great potential to impart broader eects on mass political behavior; speci cally, on voters' propensity to cast ballots in state legislative elections in which public funding is present.
4 Research Design: Public Funding and Voting Behavior 4.1 Data Considerations I now turn to Hypothesis 3, which posits that roll-o is lower for the legislative elections in districts where at least one candidate accepted full funding.
Compared to Hypotheses 1 and 2, which engage
questions about candidate behavior, an evaluation of Hypothesis 3 requires substantially dierent data and methodological choices; as such, patterns in voter roll-o cannot be properly examined with data from the candidate survey described above. One reason for this stems from the fact that the presence of 18 states in the candidate survey sample frame leads to a problem of data availability; some states 17
simply do not make precinct returns readily available.
Moreover, structural issues correspondent with
a cross-state comparison of roll-o, such as multimember districts in some states but not others, present suboptimal conditions for inference about the true eect of Clean Elections on mass voting behavior.
18
Since the candidate survey includes weighted pairs matched across states (and district types), these data seem less than ideal for examining roll-o. Furthermore, since most of the information required to analyze district-level roll-o is publicly available, I believe that the bene ts from an analysis of all legislative elections in a given state outweigh those from the addition of a few additional covariates derived from a sample of respondents.
To that end,
I compile precinct-level vote totals in all three Clean Elections states for both president and the state
17 Indeed, in Colorado, precinct returns do not seem to exist at all for this election. 18 The implications of the former are obvious, but the latter problem causes concern
since a small percentage of voters in
multimember districts may be unaware that they are able to vote for more than one candidate. The resulting undervotes would bias roll-o upward in states with multimember districts, confounding attempts to compare roll-o between singleand multi-member races.
18
legislative o ce, in all precincts in which voters cast ballots for only one representative or senator. I then aggregate these votes to a of measure district-level roll o, which I express as:
L R = 100 1 − P Where
R
is ballot roll-o,
race in a given district, and
L P
is the total number of all votes cast for the applicable state legislative is the total number of all votes for president in that district.
Lower
roll-o values are indicative of a higher percentage of voters who cast ballots in both the presidential and legislative races.
4.2 Dierence-in-Dierences For the purposes of inference, I exploit an opportunity created by the rules that govern candidate participation in public funding programs. In
Buckley v. Valeo (1976), the United States Supreme Court equated
campaign spending to political speech, rendering mandated spending limits unconstitutional. Because distribution of public money is conditional on acceptance of spending limits, candidate participation is optional, and not all districts are contested by a publicly funded candidate in a given year. Since whether a voter will interact with a fully funded legislative candidate is determined solely by the district in which that voter lives, I analyze roll-o as a natural experiment, dividing districts into treatment and control groups. I assign districts to the treatment group if at least one candidate accepted public funding during the election when it became available and to the control group if all candidates raised money solely from private sources in that year. I rst perform two tests for mean dierence in roll-o levels between the treatment and control group in the house and senate elections of each state, using one tailed
t-tests
in each case. In the rst, I test
for dierences in mean roll-o between treated and control districts in the rst election for which Clean Elections subsidies were available.
The decision to restrict analysis to a single election is guided by
practical considerations. Clean Elections became eective for the 2000 election in Arizona and Maine, and for the 2008 campaign in Connecticut; in the case of the latter, 2008 is the only year for which roll-o 19
data are presently available.
In both Arizona and Maine, candidate participation rates have become so
ubiquitous since 2000 that fewer than ve districts would be in the control condition in either state after that year. The second test compares same-district roll-o dierences between the last election nanced solely
19 2008
is the only presidential election year since Connecticut has enacted Clean Elections.
19
with private money and the rst one during which full public subsidies were available. By same-district roll-o dierences, I mean the roll-o percentage from the last presidential election year before Clean Elections (Election 0) subtracted from the roll-o percentage in the same district in the rst election after 20
implementation (Election 1).
This approach facilitates a district-level panel analysis, as each election
before and after public funding had a contested presidential race at the top of the ballot, and there was 21
no intervening redistricting in any case.
Since my aim is to compare this dierence in roll-o through time between treatment and control districts; my analysis seeks the dierence in dierences.
This approach imparts two major bene ts.
First, the examination of dierences within districts excludes district factors that are time-invariant (or relatively so), such as racial composition or political preferences, as potential confounders. Second, the logic of the dierence-in-dierences design supports a causal framework, since I assume that the roll-o trend in the control group (districts with no publicly funded candidates) approximates the mean of what would have occurred in the absence of public funding.
This assumption establishes a counterfactual
baseline for comparison of roll-o trends between the treatment and control groups. While the results of the means testing is informative for preliminary evaluation of Hypothesis 3, two factors complicate inference about the relationship between Clean Elections and roll-o. The rst is the relatively high potential for confounding variables to bias the comparison between groups. For instance, incumbent name recognition or overall spending may also entice voters to cast ballots, which would reduce roll-o. Thus, changes in other variables through time may obscure the true eect of Clean Elections in a comparison of roll-o trends. Second, with the exception of the lower houses in Connecticut and Maine, none of the legislative houses in the Clean Elections states are comprised of more than 36 districts.
Arizona's legislature is
composed of 30 districts in each house, with 2 members from each district in its House of Representatives; there are 35 members of the Maine Senate and 36 in the Connecticut Senate. Assignment to the treatment or control condition further reduces these numbers. For example, in the rst cycle with Clean Elections, there were 3 treated districts in Connecticut Senate, and 7 in the Arizona House. To be clear, I believe that an examination of the mean dierence-in-dierences in roll-o across all states and legislative houses is informative, particularly if such an examination nds consistent patterns. However, I make no
20 Accordingly,
for each district I compare the dierence in roll-o between 1996 and 2000 in Arizona and Maine, and
between 2004 and 2008 in Connecticut.
21 Gubernatorial
elections such as those in 1998 and 2006 generally draw fewer voters; citizens who do vote when no
presidential candidate is on the ballot may be more knowledgeable or committed, and so roll-o can not be compared between gubernatorial and presidential election years. A stand-alone analysis of gubernatorial roll-o may seem possible in Maine, which is unique among states in that it redistricts for the third election of every decade. However, gubernatorial candidates were also eligible to accept public funding in 2002, when the public fund released over $1.2 million to the contest for governor. I therefore do not analyze roll-o for the 1998 and 2002 Maine elections.
20
claims that such an evaluation is de nitive in all instances, due to concerns about omitted variables and micronumerosity. Elections to the lower legislative house of both Connecticut and Maine oer a potential solution to this problem. Both houses have 151 single-member lower house districts, and the relatively large number of cases in elections to those bodies creates the possibility of calculating dierence-in-dierences within a multivariate analysis. Since they alone hold realistic opportunities for such a design, I construct OLS models of same-district roll-o dierence only for Connecticut and Maine state house elections.
The
dependent variable (dierence in mean levels of roll-o in districts) is described above. The independent variable of interest for all models is a dichotomous indicator coded 1 if the election was contested by at least one publicly funded candidate; the coe cient for this variable represents the dierence-in-dierence estimate. Compared to basic means testing, the main advantage of this approach is that theoretically relevant covariates can be held constant, isolating the eect of the presence of a Clean Elections candidate on roll-o dierences. The process of group assignment is straightforward in Maine, where at least one candidate accepted 22
public funding in 58 of 151 legislative districts during the 2000 general election.
I calculate roll-o only
from precincts in which ballots were cast for a single legislative district; approximately 160 precincts were therefore excluded, but well over 500 remain. In 19 state house districts, ballots were counted only from precincts that overlap multiple districts; as such, no usable roll-o gure can be calculated from those districts, leaving 132 Maine House districts in the original sample. Fifty one of those districts are in the treatment condition. The construction of a su ciently large control group is more di cult in Connecticut, where 87 districts saw contested elections in 2008. However, at least one major-party candidate accepted public funding in all but 2 of those districts, yielding a control group too small to support statistical inference. Following Card and Krueger (1994), who examined the eect of minimum wage changes on employment in New Jersey fast-food establishments with a control group of Eastern Pennsylvania restaurants, I address this issue by constructing a control group of General Assembly districts from Rhode Island. The Rhode Island House is half the size of the Connecticut House. However, the electoral environment and political climate of Connecticut and Rhode Island are similar in a number of important areas, including the average spending in a legislative election, proximity within Squire's (2007) index of legislative professionalization, average district population, electoral time-line, and active minor parties.
Furthermore, I expect their
geographic proximity to mitigate any issues correspondent with cross-state comparisons, such as political
22 I
obtained
precinct-level
election
returns
from
http://www.maine.gov/sos/cec/elec/
21
the
website
of
the
Maine
Secretary
of
State:
culture or regional trends.
23
Since public funding is not available in Rhode Island, I include all Rhode
Island assembly districts as potential controls in the Connecticut dataset.
24
As noted, I expect that factors other than full public funding will aect ballot roll-o trends.
For
instance, expensive contests spawn more media advertising and direct mail, which should raise awareness of the candidates among the electorate. The presence of either a third-party candidate or an incumbent in a race might also eect roll-o, since incumbent name recognition provides a decision heuristic for some voters (see: Lau and Redlawsk 2001) while others may feel that minor parties are more closely aligned with their views. Finally, the percentage of African-American voters in the electorate has been shown to be positively correlated with roll-o in a number of state and municipal elections (see: Vanderleeuw and Engstrom 1987; Darcy and Schneider 1989; Vanderleeuw and Utter 1993; Vanderleeuw and Liu 2002; Feig 2007). Particularly for the Connecticut model, which includes data from a 2008 presidential contest that saw the election of an African-American president, the racial composition of a district might aect roll-o trends despite being relatively invariant through time. Probably the most important determinant of changes in roll-o are uctuations in election contestedness by candidates of major parties. The absence of a major party challenge is almost certain to increase the proportion of voters who roll o, since partisan voters may prefer making no choice to casting a ballot for a candidate of their non-preferred party. Moreover, the contestedness of a race in Election 1 (after public funding implementation) is likely to serve as a major determinant of a candidate's public funding status; unopposed candidates will likely perceive little bene t to expending the eort necessary to qualify for Clean Elections funds (rendering theirs a control district). To illustrate, of the 35 unopposed Maine candidates for which I obtained roll-o data in 2000, only 6 accepted public subsidies, while at least one 25
Maine candidate ran with Clean Elections funds in 46% of districts with contested elections in that year.
It stands to reason that in addition to being less likely to accept public funding, unopposed candidates are also likely to devote much less time to voter interaction in a campaign that they are certain to win. Since treatment occurs in a contested election in Election 1, I believe that uncontested elections that occurred in the rst cycle after public funding implementation do not provide a reliable counterfactual, and I exclude them from the analysis. Following these considerations, I calculate panel dierence-in-dierences with two OLS model speci cations from data in Maine as well as Connecticut and Rhode Island to investigate the eect of public
23 Paired
t-tests con rm that there are no signi cant dierences between the model covariates from Connecticut and
Rhode Island districts in their 2004 condition, which improves con dence in Rhode Island as a control state.
24 I
obtained precinct election returns from the websites of the Maine and Rhode Island Secretaries of State, and a public
information request submitted to the Connecticut O ce of Legislative Research.
25 Participation
rates are substantially higher among unopposed candidates in Connecticut, likely due to a lower quali -
cation threshold.
22
funding on voter roll-o between Election 1 (when public funding was available) and Election 0 (when all elections were privately nanced). In Speci cation 1 I model data from all districts in which Election 1 was contested by two major party candidates, holding the contestedness in Election 0 constant. Speci cation 2 restricts the analysis to districts in which elections in
both years were contested, and therefore
requires no control for changes to the number of major-party candidates. The model speci cations are:
∆Y D = δ + β 1X 1 + β 2∆X 2D + β 3∆X 3D + β 4∆X 4D + β 5∆X 5D + β 6X 6D + Where:
∆Y D X1
is the dierence between roll-o in Election 1 and Election 0 for District
D
is a dummy variable coded 1 for treatment (i.e., at least one candidate in the district accepted public
funding) and 0 for control
∆X 2D
is the in ation-adjusted dierence in total money raised in District
D
between Election 1 and
Election 0, in thousands of dollars
∆X 3D D
is the dierence in whether the election was contested by two major party candidates in District
between Election 1 and Election 0 (Speci cation 1 only)
∆X 4D
26
is the dierence in the number of minor party candidates on the ballot in District
D
between
Election 1 and Election 0
∆X 5D is the dierence in the number of incumbents running in District D between Election 1 and Election 0
27
X 6D
is the percentage of African-Americans living in the district.
Thus, the
β1
coe cient in all models is the dierence-in-dierences estimator of same-district roll-o
changes by treatment condition, holding funding, contestedness, minor party involvement, incumbency, and racial composition constant. I obtained summary nancial information from the website of the National Institute on Money in State Politics,
28
census information on Maine's legislative districts from
Barone, Lilley, and DeFranco (1998), and demographic information in Connecticut and Rhode Island from Lilley
et al.
(2007).
To sum, I calculate same-district roll-o trends in Election 1, as well as mean same-district dierencein-dierences between Election 1 and Election 0, in elections to both legislative houses in all three Clean Elections states.
I also calculate dierence-in-dierences estimators with multivariate OLS regression
models in the lower house elections of both Connecticut and Maine. The latter approach allows me to hold potential confounding variables constant. If Hypothesis 3 is correct, I expect to see lower roll-o rates on average in districts contested by at least one fully funded candidate. Moreover, I expect the rate
26 I exclude all districts in which there was no contested race in Election 1. This 27 Since both states employ single-member districts, values of this variable range 28 http://www.followthemoney.org
23
variable therefore ranges from -1 to 0. from -1 to 1.
of change in roll-o between Election 0 and Election 1 to be smaller in treated districts, relative to those in which all candidates raised money from traditional sources. Finally, I expect these ndings to hold up when subjected to scrutiny in a multivariate context.
5 Findings: Public Funding and Voting Behavior In this section, I rst describe the results of a comparison of means between roll-o levels in the rst legislative election after Clean Elections implementation in all three states: 2000 for Arizona and Maine, and 2008 for Connecticut. I then compare dierence-in-dierences of same-district roll-o trends in the elections to both legislative houses in all six states. Finally, I report the coe cients and standard errors from multivariate regression models run in those elections in Connecticut and Maine house elections, which supply dierence-in-dierences estimates of roll-o holding relevant covariates constant.
5.1 Summary Comparisons Figure 2 depicts mean levels of district roll-o in both state house and state senate elections in Arizona, Connecticut, and Maine in the rst election after the implementation of Clean Elections. The means are portrayed by a district's treatment condition; treated districts are those in which at least one candidate ran with Clean Elections subsidies.
The relative pattern is the same for all six elections: mean levels
of district-level roll-o are lower in those that saw a fully funded campaign. The largest dierences are evident in Arizona and Connecticut house elections, in which roll-o in treated districts was lower by about 29
twelve and fourteen percentage points, respectively. statistically signi cant at
α = .05
One-tailed tests indicate that these dierences are
p
p
30
in Connecticut ( =.0000), but not in Arizona ( =.0727).
Signi cant
p
dierences are observed in the 2000 Maine House elections ( =.0075), in which treated districts saw rollo levels about four points lower than control districts (3.5% to 7.5%, respectively). While the mean level is lower in treated districts, roll-o rates in Maine Senate races were separated by about one percentage
p
point, and levels in the two groups were statistically indistinguishable ( =.115). However, the ten-point dierences apparent between treated and control groups in Arizona and Connecticut Senate races are
p
both statistically signi cant ones ( =.01 and
29 Figure
p =.038).
2 depicts negative mean roll-o for Arizona House elections because that state employs two-member districts.
Since voters may record up to two choices in those contests, Arizona House elections almost always draw more aggregate votes per precinct than the presidential election.
30 The
small number of districts (30) comprising the Arizona House of Representatives poses challenges to inference that
employs districts as the unit of analysis. Here, there are 21 districts in the treated group but only 7 in the control condition. Precinct-level data could not be obtained for the remaining two districts.
24
To sum, the roll-o rates depicted in Figure 2 show a consistent pattern for all six examined elections: Roll-o is lower when at least one candidate accepted full subsidies, and these dierences are statistically signi cant in four of the six cases.
Mean same-district roll-o dierences from state house and senate elections in the three states are depicted in Figure 3; districts are again portrayed by treatment condition.
Thus, Figure 3 portrays
the dierence between roll-o in 2000 and 1996 in Arizona and Maine, and between 2008 and 2004 in Connecticut. In ve of the six legislative houses, mean roll-o increased for districts in the control condition between Election 0 and Election 1; the lone exception is in elections to the Arizona Senate, where roll-o levels were essentially the same in 2000 as in 1996. Yet, roll-o in treated senate districts of all three states
decreased
over the same period. The dierence between these trends (the dierence-
in-dierences) is certainly supportive of the theoretical arguments advanced above; but it is important to note that the pattern observed in Senate elections supports no rm conclusions. As noted previously, inference in the Senate elections is di cult due to the small size of the groups, and while the p-values from
p
one-tailed tests are relatively low, none of the dierences in the Senate elections of Arizona ( =.0796),
p
p
Connecticut ( =.1244), or Maine ( =.0611) is statistically signi cant at
25
α = .05.
Examination of same-district roll-o dierence-in-dierences yields more con dence in House races. Same district roll-o in Connecticut house elections shows a similar pattern to that apparent in the Senate contests: Roll-o increased by about 4.5 percentage points in the control counties, and decreased by about 6.5 points in the treated counties between 2004 and 2008. Moreover, the dierence in mean
p
roll-o change between the two groups is statistically signi cant ( =.0000). Mean roll-o rose for both the treatment and control groups in both Arizona and Maine house elections between 1996 and 2000, but in both states it rose much more in control districts. In Maine, roll-o increased by 4.5 points for districts in the control condition, compared to a very small increase in the treated districts of about one
p
half of one percentage point; this is a signi cant dierence ( =.0011). In Arizona, same-district roll-o rose 13.5 points in treated districts, but 16 points in control districts. While the dierence in Arizona
p
is not statistically signi cant ( =.4278), it is consistent with a clear pattern apparent in elections to
other o ces:
all
Mean roll-o trends higher in districts where no candidate ran with public funding after its
implementation. The patterns in both Figure 2 and Figure 3 are consistent with Hypothesis 3, which asserts that roll-o will be lower in elections contested by at least one fully funded candidate. However, while the observed patterns in these gures certainly do not discredit that hypothesis, they also do not con rm it.
Micronumerosity confounds means testing in most instances, and even conclusions based on the
relationships that do achieve statistical signi cance cannot de nitely rule out confounding variables such
26
as contestedness as the true cause of changes in roll-o.
I therefore now turn to the results of the
regression models described in the previous section, which are designed to isolate the eect of a fully funded candidate on roll-o in elections where group sizes support statistical inference.
5.2 Model Results Table 4 contains coe cients and robust standard errors (clustered by district) from the OLS regression models used to calculate the dierence-in-dierences estimators for lower house elections in Connecticut 31
and Maine, holding relevant covariates constant.
I report two speci cations for each state, and de-
scribe the models' control variables separately below, since they re ect dierent elections with unique considerations.
5.2.1
Control Variables: Maine
I begin by considering the control variables from the OLS models tted to data from Maine, which re ect roll-o changes in that state from 1996 to 2000.
All coe cients are signed in accordance with
expectations, which should improve con dence in the models.
First, the coe cients for the dierence
in money raised in a district re ect a signi cant, negative relationship between fundraising and roll-o in both models, indicating that roll-o declined when candidates in a given district collectively raised more money in 2000 than in 1996. Speci cally, the model coe cients indicate that each additional $1,000 raised in a district would decrease roll-o by about one-tenth of one percentage point.
This result is
consistent with expectations; more money in the election leads to more advertising, sta, and mailings, all of which help to raise awareness of candidates among the electorate. Given that the Maine House is comprised of single-member districts, the coe cient for the dierence in the number of incumbents can be interpreted as the eect of an incumbent being present in the race. This coe cient is negatively signed in both models, suggesting that factors such as incumbent name recognition may serve as an informational shortcut allowing voters to con dently cast ballots, thereby diminishing roll-o. The incumbency coe cient achieves statistical signi cance in Speci cation 2, but
p
narrowly misses signi cance ( =.058) in Speci cation 1. Nonetheless, the coe cients are substantively similar, depicting a relatively strong incumbency eect: Speci cation 2 at least indicates that the presence of an incumbent reduces ballot roll-o by about three- fths of a percentage point. The coe cients for the percentage of African-American residents in a district are positively signed in
31 While
there is no evidence of heteroskedasticity, I report clustered robust standard errors to adjust for any non-random
error variance. Doing so does not change the sign or signi cance of any of the regression coe cients.
27
both speci cations of Maine data. This coe cient achieves statistical signi cance in Speci cation 1 only
p
( =.045). That coe cient is comparable in magnitude to its counterpart in Speci cation 2, suggesting that in districts with a larger black population, roll-o rates were higher in 2000 than in 1996. However, it is worth noting that racial composition is likely not a major determinant of roll-o in Maine, since the average proportion of African-Americans in Maine's house districts is less than 0.5% (see: Barone, Lilley, and DeFranco 1998). The coe cients for the dierence in third-party candidates are negatively signed in both Maine models. This is consistent with expectations, since the presence of a minor party candidate might create the opportunity for some citizens to more accurately express their preferences, leading them to vote when they may otherwise have abstained.
However, neither coe cient for minor party candidates achieves
statistical signi cance. Most likely, voters whose interests are re ected by third parties will support those candidates when they run, but will otherwise settle for a candidate of one of the major parties.
Table 4: OLS Regression Coe cients, Calculation of Dierence in Dierences Estimators of Voter Roll-o in State House Races, Connecticut and Maine
Connecticut
At Least One Publicly Funded Candidate Running
Di. in Money Raised by Candidates in District (Thousands)
Di. in Major Party Contestedness
(2)
(1)
(2)
-4.277*
-2.076*
-1.794*
-1.530*
(0.852)
(0.556)
(0.500)
(0.465)
-0.019
-0.024*
-0.104*
-0.110*
(0.018)
(0.012)
(0.024)
(0.024)
-21.362*
-
-12.387*
-
(0.978) Di. in Presence of Incumbent
Di. in Number of Minor Party Candidates
Perc. of African Americans in District
Constant
N R
2
RMSE F-Statistic
Maine
(1)
(1.404)
0.671
0.761
-0.544
-0.628*
(0.885)
(0.718)
(0.284)
(0.268)
-0.139
0.343
-0.905
-0.394
(0.493)
(0.523)
(0.719)
(0.583)
0.151*
0.098*
0.797*
0.719
(0.055)
(0.046)
(0.393)
(0.403)
0.769
-0.113
0.978*
0.881*
(0.637)
(0.451)
(0.338)
(0.331)
129
82
97
91
0.877
0.333
0.693
0.306
4.32
2.44
2.31
2.16
114.77
6.36
27.59
9.37
*p<.05. Robust standard errors in parentheses, clustered by legislative district. Dependent variable for Connecticut is the dierence in 2008 roll-o and 2004 roll-o. For Maine, the dierence is between 2000 and 1996. The control group for Connecticut is partially constructed from Rhode Island House Districts. Speci cation (1) ts a model to data from all districts in which elections were contested by two major-party candidates in the rst election after public funding implementation. Speci cation (2) ts the model to all districts in which both elections were contested by two major-party candidates.
28
Speci cation 1 includes an additional variable re ecting changes in the contestedness of the election by major-party candidates.
32
As expected, the coe cient for this variable demonstrates that major-
party contestedness is a powerful attenuator of roll-o; when an election was contested in 2000 (but not 1996), roll-o in that district was reduced by more than 12 percentage points. The relationship between contestedness and roll-o is statistically signi cant and consistent with expectations; it is reasonable to conclude that partisan voters are likely to abstain from voting in elections when their preferred party does not advance a candidate.
5.2.2
Control Variables: Connecticut
Table 4 also contains coe cients and clustered robust standard errors for the models used to calculate dierence-in-dierences in Connecticut state house elections (using those from Rhode Island as a control group). The Connecticut models re ect changes in roll-o from 2004 to 2008. Like those for Maine, the coe cients for dierences in money raised (in thousands of dollars) are negatively signed, indicating that the presence of more money (and correspondent campaign communication) diminishes roll-o. However, while the coe cients for the fundraising variable in Speci cations 1 and 2 are quite similar (-.019 and
p
-.024, respectively) only the one in Speci cation 2 achieves statistical signi cance ( =.042). Racial composition is a signi cant predictor of roll-o trends in both Speci cations 1 and 2, which return coe cients of .151 and .091, respectively.
These coe cients, both of which achieve statistical
signi cance, indicate that a one percentage-point increase in African American residents would raise rollo by about one-tenth of one percentage point from 2004 to 2008. This nding is consistent with previous research that has found higher roll-o in districts with more African-Americans (e.g., Feig 2007; Kimball and Kropf 2005; Herron and Sekhon 2001; Brady et al.
2001; Knack and Kropf 2003; Vanderleeuw
and Engstrom 1987). Particularly given the dynamics at play in the 2008 presidential election, which saw unprecedented turnout from African-American voters (see: McDonald 2008), both the direction and signi cance of these coe cients align with theoretical expectations. Unlike those in the Maine models, and contrary to expectations, the coe cients for dierences in thirdparty candidates are positively signed. However, like in the models for Maine, changes in the number of minor-party candidates on the Connecticut and Rhode Island ballots did not result in signi cantly altered roll-o patterns. As in Maine, the presence of third-party candidates does not appear to aect roll-o in a statistically meaningful way. The coe cients re ecting the dierence in the presence of an incumbent are also signed incorrectly in both models, but once again, neither coe cient is a signi cant
32 Speci cation
2 models data from districts in which elections were contested by major-party candidates in both years.
As such, there is no variance in contestedness.
29
predictor of roll-o changes from 2004 to 2008. While there is a plausible explanation for the failure of third party candidates to aect roll-o trends (third-party voters simply transfer their vote to the closest major party), it is slightly more di cult to account for the failure of incumbency to diminish roll-o. One possible explanation is that change was a major theme of the 2008 presidential campaign, which may have reduced the informational value of incumbency. However, these data do not allow sound conclusions on this question. On the question of dierences in major-party contestedness, Speci cation 1 in Connecticut mirrors the ndings in Maine. The model re ects a strong, signi cant, negative relationship between contestedness and roll-o; in elections that were contested by two major parties in 2008 (but not 2004), roll-o declined by about 21.4 points, holding other covariates constant.
Once again, this nding is consistent with
theoretical expectations, since the presence of two major-party candidates presents most voters with an acceptable choice, reducing the chances that they will roll o.
5.2.3
Treatment Indicators: All Models
I now turn to the treatment dummy variables for all four models, which are the predictors of interest with regard to Hypothesis 3. As is apparent in Table 4, all four coe cients for the treatment dummy variables agree in both sign and statistical signi cance, and indicate that the presence of a fully funded candidate in the legislative race decreases roll-o in both Connecticut and Maine.
For the models of
p
data from Connecticut, Speci cation 1 returns a negative coe cient of about 4 points ( =.0000), while
p
Speci cation 2 returns a negative eect about half that size ( =.0000). The models also re ect reduced roll-o in Maine districts contested by at least one publicly funded candidate. The size of this reduction is 1.79 percentage points in Speci cation 1 and 1.53 percentage points in Speci cation 2 for Maine, holding
p
all other covariates constant. Both of these coe cients are statistically signi cant ( =.0001). In short, all four models hold that roll-o is signi cantly lower in districts where at least one candidate accepted public funding. When coupled with the patterns depicted in Figures 2 and 3, which demonstrate in every case that roll-o was lower in districts that saw fully funded campaigns, this section oers strong evidence that the presence of a publicly-funded candidate can aect mass voting behavior. Speci cally, this section supports Hypothesis 3, which predicts that the presence of a fully funded candidate will signi cantly reduce ballot roll-o, and is consistent with the theoretical narrative that the higher levels of interaction between campaigns and citizens demonstrated above lead to more voters casting ballots in races when a publicly funded candidate is present.
30
6 Conclusion Almost all of the literature on the subject of full public election funding programs like Clean Elections evaluates their ability to improve competition. The results are encouraging on that front (see: Malhotra 2008; Mayer, Werner and Williams 2005), and given that combating stagnant competition is a stated goal of the reform movement, these studies are valuable. However, the ndings I report in this paper show that diminished victory margins are only one of several major eects of full funding, and the most powerful bene t of programs like Clean Elections might be their ability to alter the manner in which candidates use their time.
Simply stated, full funding facilitates changes in basic campaign strategy, enhancing
the ability of campaigns to directly engage voters.
This, in turn raises levels of voter participation in
low-visibility elections contested by fully funded candidates. Yet, these ndings do not allow a de nitive conclusion with regard to the casual mechanism functioning between the exposure to the candidate and the decision to vote. One answer might be that, as Wattenberg, McAllister, and Salvanto (2000) suggests, increased interaction between candidates and voters simply lowers voters' information costs, providing them with a political education su cient to allow a clear choice. Another possible explanation is that when voters meet candidates, the personal connection leads to higher levels of mass political e cacy, which has been shown to be particularly important in reducing roll-o in studies in certain cases (Vanderleeuw and Utter 1993; Vanderleeuw and Liu 2002). The answer is likely that both knowledge
and
e cacy increase as public interaction rises.
That said, future work
should seek to untangle this question. It is worth noting that many fully funded candidates who participated in this study possessed a broad view of their campaign, describing its eects on civic engagement and the availability of political information. Those candidates reported feeling that public money aorded them a unique opportunity to stage a serious public discourse that would not otherwise have occurred to the extent that it did. Fully funded challengers seem to seriously re ect upon their ability to present voters with a choice, and view the dialog they have with voters as a personal contribution to the quality of their state's democracy. This is a unique situation in American politics, and stands in stark contrast to the prevailing notion of selective mobilization eorts that are deployed only in close races (see: Rosenstone and Hansen 1993, 210). As one informant noted, (Accepting public money) was the dierence between (having no candidate and) being able to get out there. . . (to) say the things that needed to be said about making changes in the system and why I needed to be involved in this process. The corresponding potential for broader citizen participation is a worthwhile end with normative implications for American politics.
31
In short,
when public funding provides candidates with the nancial resources to challenge an incumbent, it will likely make elections more competitive. When it causes citizens to cast votes in more elections, public funding will make them more democratic. Since public funding was rst implemented, the question from both an academic and policy perspective has been:
Does public money improve elections? Holistic evaluation of public funding programs is
beyond the scope of any single paper, and while I cannot say that public subsidies render elections better, I demonstrate here that subsidized elections are certainly dierent. Many publicly funded candidates, and particularly those who accept full funding, approach their campaign from a unique strategic framework. Spared from the onerous task of fundraising, fully funded candidates may focus entirely on strategy, message, and voter outreach. A dierence in weekly public interaction time of over ten percentage points is a substantial one, particularly since a typical candidate's time commitment in a long campaign season approximates that of a full-time job. Fully funded candidates' enhanced mobilization capability translates to at least hundreds, and possibly thousands of high-quality voter contacts that would not have otherwise occurred.
These
personal contacts happen even in districts where the candidates stand no chance of winning, and in districts where full subsidies are present, more citizens vote.
It is hard to argue that these are not
positive changes, and if there is value in enhancing the visibility of political discourse, then full public election funding certainly creates a public good. Future studies of public funding and other campaign nance regulations should not only go further in examining the links between public funding programs and mass knowledge and e cacy, but also in accounting for the fact that these policies likely have additional, unstudied eects on voter turnout and other forms of political participation.
32
Appendix 1: Response Rates and Sample Characteristics Candidate
Response
Population
Rate
Rhode Island
114
23.7%
Michigan
207
24.6%
Ohio
175
26.9% 28.2%
State
Iowa
170
West Virginia
151
28.5%
Missouri
244
29.1% 31.7%
Vermont
240
Minnesota
266
35.3%
Hawaii
78
35.9%
New Mexico
101
36.6%
Connecticut
239
38.5%
Maine
288
38.9%
Montana
183
40.4%
Delaware
64
40.6%
Wisconsin
166
41.0%
Alaska
69
42.0%
Colorado
121
42.1%
Arizona
97
49.5%
Characteristics of Candidate Population and Sample Population
Sample
Total
2,971
1,022
Incumbents
46.7%
39.7%
Challengers
34.9%
37.8%
Dem
54.1%
61.3% 33.4%
Women
28.2%
Full PF
15.8%
20.4%
Part PF
9.3%
10.4%
33
Appendix 2: Alternate Speci cations With Unmatched Data OLS Regression Coe cients: Eect of Accepting Public Funding on Candidate's Public Interaction Activities, Unmatched Data Dummy: Candidate Accepted Partial Public Funding Dummy: Candidate Accepted Full Public Funding Dummy: Candidate Previously Held Elected O ce
(1)
(2)
0.828
0.973
(1.218)
(0.985)
9.485*
9.779*
(2.382)
(1.551)
-
-0.357 (1.036)
Dummy: Candidate is Incumbent
-
-0.872 (1.141)
District Population (Thousands)
-
-0.517*
Household Income in District (Thousands)
-
-0.541*
77.928*
84.251*
(1.108)
(1.760)
(0.166) (0.159) Constant N
p
R2
801
801
0.065
0.093
RMSE
14.60
14.4
8.70
22.34
F -Statistic
* <.05. Robust standard errors in parentheses, clustered by state. Dependent variable is the percentage of campaign time devoted to public interaction activities, ranging from 0 to 100.
34
Appendix 3: Public Funding in the States Table A2: Summary of Public Funding Regulations, 2008 Election
Quali cation Spending Limit Max. Subsidy Max. Match
Quali cation Spending Limit Max. Subsidy Max. Match
Hawaii
Minnesota
Wisconsin
raise $1,500
$1,500 in amounts < $50
raise $1,725 in amounts < $100
app. $32,000*
$31,400*
$17,250*
15% of spending limit
50% of spending limit
45% of spending limit
NA
NA
NA
Arizona
Connecticut
Maine
220 $5 donations
$5,000 from 150 in-district contributors
50 $5 donations
$35,673**
$41,000**
$6,148**
$31,673
$35,000
$5,648
up to 3X subsidy
up to 2X subsidy
up to 2X subsidy
*Hawaii's spending limit $1.40 for each voter in a district. Minnesota and Wisconsin spending limits apply only if all candidates in a given race accept them. **In Arizona, Connecticut, and Maine, candidates are allowed to raise $5,000, $6,000, and $500 prior to qualifying. Once they accept public funding, they may raise no additional money.
35
References [1] Arceneaux, Kevin, Alan S.Gerber, and Donald P.Green. 2006. Comparing Experimental and Matching Methods Using a Large-Scale Voter Mobilization Experiment.
Political Analysis 14(1):
[2] Barone, Michael, William Lilley III, and Laurence J. Defranco. 1998.
Demographics. Washington D.C.:
37.
State Legislative Elections: Voting Patterns and
Congressional Quarterly.
[3] Bennion, Elizabeth A. 2005. Caught in the Ground Wars: Mobilizing Voters during a Competitive Congressional Campaign.
The Annals of the American Academy of Political and Social Science 601(1):
123-141.
Counting All the Votes: The Performance of Voting Technology in the United States. Berkeley, CA: University of California.
[4] Brady, Henry E., Justin Buchler, Matt Jarvis, and John McNulty. 2001.
[5] Brown, Adam R. 2011. Does Money Buy Votes? The Case of Self-Financed Gubernatorial Candidates, 1998-2008. Typescript.
[6] Card, David, and Alan B. Krueger. 1994. Minimum Wages and Employment: A Case Study of the Fast-Food Industry in New Jersey and Pennsylvania.
American Economic Review 84(4):
772-793.
[7] Darcy, R., and Anne Schneider. 1989. Confusing Ballots, Roll-O, and the Black Vote.
Quarterly 42(3):
The Western Political
347-364.
[8] Diamond, Alexis, and Jasjeet S. Sekhon. 2006. Genetic Matching for Estimating Causal Eects: A General Multivariate Matching Method for Achieving Balance in Observational Studies. Available at: http://sekhon. berkeley. edu/papers/GenMatch. pdf [Accessed March 16, 2009].
[9] Donnay, Patrick D, and Graham P Ramsden. 1995. Public Financing of Legislative Elections: Lessons from Minnesota.
Legislative Studies Quarterly 20(3):
[10] Downs, Anthony. 1957.
351-364.
An Economic Theory of Democracy. New York:
Harper and Brothers.
[11] Engstrorm, Richard L., and Victoria M. Caridas. 1991. Voting for Judges: Race and Roll-O in Judicial Elections. In
Political Participation and American Democracy, W. Crotty, ed. Westport, CT: Greenwood.
[12] Epstein, Michael J., Richard G. Niemi, and Linda W. Powell. 2005. Do Women and Men State Legislators Dier? In
Women and Elective O ce, Sue Thomas and Clyde Wilcox, eds. New York:
Oxford University Press.
[13] Feddersen, Timothy J., and Wolfgang Pesendorfer. 1996. The Swing Voter's Curse.
The American Economic Review
86(3): 408-424.
[14] Feig, Douglas C. 2007. Race, Roll-O, and the Straight-Ticket Option.
Politics and Policy 35(3):
548-568.
[15] Francia, Peter L., and Paul S. Herrnson. 2003. The Impact of Public Finance Laws on Fundraising in State Legislative Elections.
American Politics Research 31(5):
520.
36
[16] Franklin, Mark. 2004.
1945. New York:
Voter Turnout and the Dynamics of Electoral Competition in Established Democracies since
Cambridge University Press.
Campaign Finance Reform: Experiences of Two States That Oer Full Public Funding for Political Candidates. Washington, D.C.: United States Government.
[17] General Accounting O ce. 2010.
[18] Gerber, Alan S, and Donald P Green. 2000. The Eects of Canvassing, Telephone Calls, and Direct Mail on Voter Turnout: A Field Experiment.
The American Political Science Review 94(3):
653-663.
[19] Gerber, Alan S., and Donald P. Green. 2001. Do Phone Calls Increase Voter Turnout?: A Field Experiment.
Opinion Quarterly 65(1):
Public
75-85.
[20] Green, Donald P, Alan S Gerber, and David W Nickerson. 2003. Getting out the Vote in Local Elections: Results from Six Door-to-Door Canvassing Experiments.
The Journal of Politics 65(4):
[21] Hayes, Danny, and Seth C. McKee. The Participatory Eects of Redistricting.
Science 53(4):
1083-1096.
The American Journal of Political
1006-1023.
[22] Herrnson, Paul S. 2004.
Congressional Elections: Campaigning at Home and in Washington. 4th Edition. Washington,
D.C.: CQ Press.
[23] Herron, Michael C., and Jasjeet S. Sekhon. 2001. Overvoting and Representation: An Examination of Overnoted Presidential Ballots in Broward and Miami-Dade Counties.
Electoral Studies 22(1):
21-47.
[24] Hillygus, D. Sunshine. 2005. Campaign Eects and the Dynamics of Turnout Intention in Election 2000..
of Politics 67(1):
Journal
50-68.
[25] Ho, Daniel E., Kosuke Imai, Gary King, and Elizabeth A. Stuart. 2007(a). Matching as Nonparametric Preprocessing for Reducing Model Dependence in Parametric Causal Inference.
Political Analysis 15(1):199-236.
[26] Ho, Daniel E., Kosuke Imai, Gary King, and Elizabeth A. Stuart. 2007(b). Matchit: Nonparametric Preprocessing for Parametric Causal Inference.
Journal of Statistical Software. Available at:
[27] Howell, Susan E. 1982. Campaign Activities and State Election Outcomes.
http://gking.harvard.edu/matchit/.
Political Behavior 4(4):
401-417.
[28] Jones, Ruth S, and Thomas J Borris. 1985. Strategic Contributing in Legislative Campaigns: The Case of Minnesota.
Legislative Studies Quarterly 10(1):
89-105.
[29] Kimball, David C., and Martha Kropf. 2005. Ballot Design and Unrecorded Votes on Paper-Based Ballots.
Opinion Quarterly 69(4):
Public
508-39.
[30] Knack, Stephen, and Martha Kropf. 2003. Roll-O at the Top of the Ballot: Intentional Undervoting in American Presidential Elections.
Politics and Policy 31(4):
575-594.
[31] Kramer, Gerald H. 1970. The Eects of Precinct-Level Canvassing on Voter Behavior. 34(4): 560-572.
37
The Public Opinion Quarterly
[32] Lau, Richard R., and David P. Redlawsk. 2001. Advantages and Disadvantages of Cognitive Heuristics in Political Decision Making.
American Journal of Political Science 45(4):
951-971.
[33] Lilley, William, Laurence J. DeFranco, Mark F. Bernstein, and Kari L. Ramsby. 2007.
Elections, 3rd Ed. Washington, D.C.:
[34] Malbin, Michael J., and Thomas Gais. 1998.
American States. Washington, D.C.:
Almanac of State Legislative
CQ Press.
The Day After Reform: Sobering Campaign Finance Lessons from the
Brookings Institution Press.
[35] Malbin, Michael J., Norman J. Ornstein and Thomas J. Mann. 2008.
Vital Statistics on Congress, 2008. Washington,
D.C.: Brookings Institution Press.
[36] Malhotra, Neil. 2008. The Impact of Public Financing on Electoral Competition: Evidence from Arizona and Maine.
State Politics & Policy Quarterly 8(3):
263.
[37] Mayer, Kenneth R, Timothy Werner, and Amanda Williams. 2006. Do Public Funding Programs Enhance Electoral Competition?. In
The Marketplace of Democracy: Electoral Competition and American Politics.
Michael P.
McDonald and John Samples, eds., Washington, D.C.: Brookings Institution Press.
[38] Mayer, Kenneth R, and John M Wood. 1995. The Impact of Public Financing on Electoral Competitiveness: Evidence from Wisconsin, 1964-1990.
Legislative Studies Quarterly 20(1):
69-88.
[39] McDermott, Monika. 2005. Candidate Occupations and Voter Information Shortcuts.
The Journal of Politics 67(1):
201-219.
[40] McDonald, Michael P. 2008. The Return of the Voter: Voter Turnout in the 2008 Presidential Election.
The Forum
6(4). Available at: http://www.bepress.com/forum/vol6/iss4/art4
[41] McNulty, John E. 2005. Phone-Based GOTVWhat's on the Line? Field Experiments with Varied Partisan Components, 2002-2003.
The Annals of the American Academy of Political and Social Science 601(1):
41-65.
[42] Mebane, Walter R., and Jasjeet S. Sekhon. 1998. Genetic Optimization Using Derivatives (GENOUD). Available at: http://sekhon.polisci.berkeley.edu/rgenoud/. [Accessed March 16, 2009].
[43] Michelson, Melissa R. 2003. Getting Out the Latino Vote: How Door-to-Door Canvassing In uences Voter Turnout in Rural Central California.
Political Behavior 25(3):
247-263.
[44] Miller, Michael G., and Costas Panagopoulos. 2011. Public Financing, Attitudes Toward Government and Politics, and E cacy.
Public Financing in American Elections. Costas Panagopoulos (ed.) Temple University Press.
Milyo, Jerey, David M. Primo, and Matthew L. Jacobsmeier. 2011. Does Public Funding of State Election Campaigns Increase Voter Turnout?
Public Financing in American Elections.
Costas Panagopoulos (ed.) Temple University
Press.
[45] [46] Nickerson, David W. 2005. Partisan Mobilization Using Volunteer Phone Banks and Door Hangers.
the American Academy of Political and Social Science 601(1):
The Annals of
10-27.
[47] Nickerson, David W. 2007. Quality Is Job One: Professional and Volunteer Voter Mobilization Calls.
Journal of Political Science 51(2):
269-282.
38
American
[48] Nickerson, David W. 2006. Volunteer Phone Calls Can Increase Turnout: Evidence From Eight Field Experiments.
American Politics Research 34(3):
271-292.
[49] Nickerson, David W., Ryan D. Friedrichs, and David C. King. 2006. Partisan Mobilization Campaigns in the Field: Results from a Statewide Turnout Experiment in Michigan.
Political Research Quarterly 59(1):
85-97.
[50] Niven, David. 2001. The Limits of Mobilization: Turnout Evidence from State House Primaries.
Political Behavior
23(4): 335-350.
[51] Niven, David. 2002. The Mobilization Calendar:
American Politics Research 30(3):
The Time-Dependent Eects of Personal Contact on Turnout.
307-322.
[52] Panagopoulos, Costas. 2009. Partisan and Nonpartisan Message Content and Voter Mobilization: Field Experimental Evidence.
Political Research Quarterly 62(1):
70-76.
[53] Parry, Janine, Jay Barth, Martha Kropf, and E. Terrence Jones. 2008. Mobilizing the Seldom Voter: Contact and Eects in High-Pro le Elections.
Political Behavior 30(1):
Campaign
97-113.
[54] Primo, David and Jerey Milyo 2006. Campaign Finance Laws and Political E cacy: Evidence from the States.
Election Law Journal 5 (1):
2339.
[55] Ramirez, Ricardo. 2005. Giving Voice to Latino Voters: Nonpartisan Mobilization Eort.
A Field Experiment on the Eectiveness of a National
The Annals of the American Academy of Political and Social Science 601(1):
66-
84.
[56] Riker, William H., and Peter C. Ordeshook. 1968. A Theory of the Calculus of Voting.
Science Review 62(1):
The American Political
25-42.
[57] Rosenstone, Steven J. and John Marc Hansen. 1993.
Mobilization, Participation, and Democracy in America.
New
York: Macmillan.
[58] Schier, Steven E. 2000.
By Invitation Only: The Rise of Exclusive Politics in the United States. Pittsburgh:
University
of Pittsburgh Press.
[59] Sekhon, Jasjeet S. 2006. Alternative Balance Metrics in Matching Methods for Causal Inference. Available at: http://sekhon.berkeley.edu/papers/SekhonBalanceMetrics.pdf [Accessed March 16, 2009].
[60] Sekhon, Jasjeet S. 2008. Multivariate and Propensity Score Matching Software with Automated Balance Optimization: The Matching package for R.. Available at: http://sekhon.berkeley.edu/papers/MatchingJSS.pdf [Accessed March 16, 2009].
[61] Sekhon,
Jasjeet S. 2009. Opiates for the Matches:
Matching Methods for Causal Inference.
Available at:
sekhon.berkeley.edu/papers/opiates.pdf [Accessed March 16, 2009].
[62] Sekhon, Jasjeet S., and Walter R. Mebane. 1998. Genetic Optimization Using Derivatives.
Political Analysis 7(1):
187-210.
[63] Smith, Jerey, and Petra Todd. 2005. Does Matching Overcome LaLonde's Critique of Nonexperimental Estimators?.
Journal of Econometrics 125(1-2):
305-353.
39
[64] Squire, Peverill. 2007. Measuring State Legislative Professionalism: The Squire Index Revisited.
Policy Quarterly 7(2):
State Politics &
211.
[65] Vanderleeuw, James M., and Glenn H. Utter. 1993. Voter Roll-O and the Electoral Context: A Test of Two Theses.
Social Science Quarterly 74(3):
664-673.
[66] Vanderleeuw, James M., and Richard L. Engstrom. 1987. Race, Referendums, and Roll-O.
The Journal of Politics
49(4): 1081-1092.
[67] Vanderleeuw, James M., and Baodong Liu. 2002. Political Empowerment, Mobilization, and Black Voter Roll-O.
Urban Aairs Review 37(3):
[68] Wattenberg, Martin. 2002.
380-396.
Where Have All the Voters Gone?
Cambridge, MA: Harvard University Press.
[69] Wattenberg, Martin P., Ian McAllister, and Anthony Salvanto. 2000. How Voting is Like Taking an SAT Test: An Analysis of American Voter roll-o.
American Politics Research. 28(2):
234-250.
[70] Werner, Timothy, and Kenneth R. Mayer. 2007. Public Election Funding, Competition, and Candidate Gender.
Political Science and Politics 40:
PS:
661-667.
[71] Wilcox, Clyde. 2001. Contributing as Political Participation. In
A User's Guide to Campaign Finance Reform.
Gerald Lubenow, ed. New York: Rowman and Little eld.
[72] Wong, Janelle S. 2005. Mobilizing Asian American Voters:
Academy of Political and Social Science 601(1):
102-114.
40
A Field Experiment.
The Annals of the American
