Work In Progress. Please Do Not Cite without Permission

Transcription

1 Voter Mobilization through Friends and Family: A Large Scale Field Experiment on the Social Priming about Political Participation Christopher B. Mann University of Miami Department of Political Science cmann@mail.as.miami.edu Casey A. Klofstad University of Miami Department of Political Science klofstad@gmail.com Work In Progress Please Do Not Cite without Permission The authors express our deep gratitude to our partner organization for their cooperation in conducting this experiment and for their commitment to innovation and research about voter mobilization. We thank Kevin Arceneaux, David Nickerson, and Costas Panagopoulos for helpful comments on the manuscript. A previous version of this paper was presented at the 2011 Annual Meeting of the American Political Science Association..

2 Abstract Whether voting should be viewed as an individualistic act or a social act is a fundamental debate in political science. We propose a novel theory of the social aspect of voter participation: emphasizing the social aspect of voting can increase the efficacy of voter mobilization efforts, but this effect is undermined when the social scope becomes too large. To test this proposition we use a new form of social pressure we call social priming. We induce subjects activate considerations of social pressure, rather than pressuring subjects from the outside by a third party, as in existing research. We conducted a field experiment among 198,285 registered voters in Illinois, Michigan, New York, and Pennsylvania during the 2010 General Election using live phone calls to encourage voting. Subjects were encouraged to vote themselves and then asked to remind zero, one, three, or five of their friends and family to vote. There was no effect on turnout from the script with no request to remind friends and family, a small (but insignificant) positive effect for the script requesting a reminder to one person, and a significant increase when asking subjects to remind three people. Asking subjects to remind five of their friends and family to vote had no discernible effect on voter turnout, suggesting that the relationship between social priming and turnout is curvilinear. This variation indicates that reminding others to vote evokes social considerations rather than just internal self-prophecy or commitment mechanisms. Further analysis indicates that the scope of the social prime is the mechanism since the curvilinear effect on turnout mirrors a curvilinear pattern in agreement with the request to remind others to vote. That is, the effect of social priming is conditional on agreeing to the reminder task. These results highlight the need to consider the limits to the social aspect of voting for increasing participation. 2

3 Introduction Whether voting should be viewed as an individualistic act or a social act is one of the fundamental debates in political science, with roots in the discipline s canon (e.g. Michigan School vs. Columbia School ) and extending through much contemporary scholarship. Rather than attempt to settle this lively debate, we argue that it is more appropriate to re-phrase the question: how does varying the scope of the social aspect of voting impact individual decisions to cast a ballot? In response to this question, we propose a novel theory of the social aspect of voter participation: while emphasizing the social aspect of voting can increase the efficacy of voter mobilization efforts, this effect is undermined when the social scope becomes too large. To test this proposition we use a new form of social pressure that we refer to as social priming. Extant studies of social pressure and voter turnout prompt respondents to believe that their turnout is being monitored from the outside by a third party. In contrast, we encourage subjects to activate social pressure within themselves by asking them to remind their friends and family to vote. We find support for our theory in a field experiment that asked registered voters to remind zero, one, three, or five of their friends and family to vote. The treatments were delivered using live phone calls to encourage voter turnout among 198,285 registered voters in Illinois, Michigan, New York, and Pennsylvania during the 2010 General Election. The pattern of results shows increasing effects as the scope of this request expanded: Our treatment without a request to remind friends and family had no effect. Asking subjects to remind one person had a positive (but insignificant) effect. Asking subjects to remind three of their friends and family to vote had a larger and statistically significant effect on voter turnout. Asking subjects to remind five of their friends and family to vote had no discernible effect on voter turnout. That is, we find a 3

4 curvilinear relationship between the scope of social priming and voter turnout. Based on further analysis, we suggest that these results are explained by the willingness of respondents to agree to the request to mobilize friends and family, which follows the same curvilinear pattern. More specifically, the effect of social priming is conditional on agreeing to the reminder task. These results provide deeper insight into how priming citizens to think about voting as a social activity can increase turnout, as well as suggesting that going too far with the scope of social priming will undermine its effect. Individual and Social Antecedents of Voter Turnout At first blush, the act of voting seems highly individualistic. From the standard of one person, one vote in the US Constitution 1 to the fact that votes are cast alone in the privacy of a voting booth, voting appears to be a solitary activity. Perhaps as a consequence of this perception, much of the scholarship on voter turnout, beginning with the Michigan School of voting behavior (e.g., Campbell et al. 1960), has focused on individual-level antecedents such as strength of partisanship, interest in politics, education, income, and the like. Research in this individualistic tradition often downplays the influence of social-level factors (Zuckerman 2004). In fact, the founders of the Michigan School went as far as to say, By and large we shall consider external conditions as exogenous to our theoretical system (Campbell et al. 1960, p. 27). On the other hand, the earliest studies of voting behavior established that the act of voting is firmly embedded in social context (e.g. Lazarsfeld, Berelson, and Gaudet 1944; Berelson, Lazarsfeld, and McPhee 1954). More recently, scholars have become concerned that an 1 This principle is the Constitutional standard for creating legislative districts in the United States under the US Supreme Court s decision in Reynolds v. Sims, 377 U.S. 533 (1964). 4

5 unraveling of the social fabric has undermined political participation in the United States and other advanced democracies (e.g., Putnam 2000; Rosenstone and Hansen 1993). For example, Rosenstone and Hansen (1993) point to the decline of mass popular political parties and the web of personal relationships they facilitated as a major cause of the decline in voter turnout in the United States. Even more recently, scholarship on political behavior has shifted focus from the individual to his or her social network (e.g., see Zuckerman (2004) for a review of this literature). Research in this tradition shows that routine social interactions, such as casual conversations about politics with family, friends, and co-workers have significant impacts on political activity (e.g., Huckfeldt et al. 2004; Klofstad 2011; Klofstad 2009; Klofstad 2007; McClurg 2003; Sokhey and McClurg n.d.). Social Priming and Voting From these two rival traditions it is clear that voting is both an individual and social act. While a voter casts a ballot as an individual, the act of voting is embedded in a rich milieu of social interactions, pressures, and incentives. In line with this logic, recent field experiments have successfully leveraged the social aspect of voting to increase turnout (Gerber, Green, and Larimer 2008; Gerber, Green, and Larimer 2010; Gerber & Green 2010; Davenport et al. 2010; Davenport 2010; Mann 2010; Panagopoulos 2010; Panagopoulos Forthcoming). The mechanism underlying voter mobilization treatments in these extant social pressure studies is monitoring and an implied threat that individuals actions are public knowledge (e.g., Cialdini and Goldstein 2004; Cialdini and Trost 1998; Lerner and Tetlock 1999). Within the context of elections, it is rare that one vote will determine the outcome. Therefore, a rational citizen should feel that the time and energy necessary to vote exceeds policy gains from the 5

6 election outcome, intrinsic psychological benefits of fulfilling one s civic duty (e.g., Downs 1957), and extrinsic social benefits of complying with the norm of voting (Gerber, Green, and Larimer 2008). Social pressure treatments seek to alter the extrinsic social benefits. The way elections are administered in the United States (and many other advanced democracies) makes it unlikely that others will learn that an individual failed to vote. Therefore, it is unlikely that the abstaining citizen be sanctioned for violating the norm of civic duty (e.g., see Harbaugh (1996) on over-reporting of voter turnout). However, the calculus of social sanction and reward can be altered if those around us become aware of our actions (Funk 2009; Knack and Kropf 1998; Knack 1992; Harbaugh 1996; Schultz 1999; Cialdini and Goldstein 2004; Schultz et al. 2007; Whatley, Webster, Smith and Rhodes 1999; Posner and Rasmusen 1999; Rind and Benjamin 1994). The social pressure treatments in the recent line of field experiments provide the citizen with his or her vote history as signal that his or her voting behavior is observable by third parties, and therefore more likely to be subject to social sanction for failing to vote. We take a new approach to encouraging subjects to view voting as a social act. Our social priming treatments prompt individuals to create their own vortex of social pressure by asking their friends and family to vote, rather than signaling the threat of external monitoring by a third party as in the extant social pressure studies described above. We focus on social priming because it is a naturally occurring mechanism to create social pressure. Everyday conversations about politics, elections, campaigns, and/or voting include the type of interpersonal encouragement to vote that we seek to prompt with our treatments. Thus our theory and experimental results provide insight about the natural political world as well as assessing our treatments as a potential mobilization tactic by civic and political organizations. 6

7 Social priming is also useful because the social scope can be varied by changing the number of friends and family members in our treatments. The possibility of varying the social scope facilitates investigation of how voters perceive the social aspect of voting. The research described above indicates that leveraging the social aspect of voting will increase turnout. However when the social scope of voting becomes too large, the social rewards and penalties of voting collapse under their own weight and any positive effects on turnout are undermined. We first discuss how social priming should increase turnout, before turning to the limits of social priming below. Reminding someone else to vote seems to be an externally focused activity, but it has important consequences for the citizen doing the reminding. Reminding friends and family members about voting invites them to pay attention to the voting behavior of the citizen doing the reminding. This invitation to be monitored is likely to enhance the initiator s perception that he is being monitored and therefore his perception of the likelihood of social reward for voting or social sanction for non-voting. To be clear, social priming works through the perceptions of the subject. Since a citizen s perception of monitoring is what influences his behavior, whether or not this monitoring actually occurs is of relatively little concern. By prompting subjects to ask their friends and family to vote, we influence them to perceive that their actions are being monitored by their friends and family, regardless of whether they actually are. Moreover, the people whom a citizen attempts to mobilize are likely to be social intimates within the immediate social network (e.g., friends and family) whose opinions the citizen values. Monitoring by social intimates is likely to generate greater perceptions of extrinsic social rewards for voting or sanctions for failing to vote, than monitoring by a third 7

8 party as in the extant social pressure studies. Since each additional person that a citizen reminds to vote increases the potential for extrinsic social reward for voting, or sanction for failing to vote, the citizen s probability of voting should increase with the number of people he or she reminds to vote. Alternatively, when a citizen reminds someone else to vote he/she is making a tacit promise to vote as well. Making this type of commitment is a self-prophecy that triggers an intrinsic psychological impetus to make behavior consistent with promises (Greenwald, et al. 1987; Sherman 1980; Spangenberg & Greenwald 1999; Sprott, Spangenberg & Fisher. 2003; Mann 2005; Smith, Gerber and Orlich 2003; Michelson, Bedolla and McConnell 2009; Nickerson and Rogers 2010). The intrinsic desire to make actions comply with implied promises is a self-referential, individual consideration and therefore not likely to correlate with the number of persons to whom the promise. If this is the primary mechanism at work in response to our social priming treatments we would expect to see an increase in turnout from inducing citizens to remind others to vote, but the effect would not be conditional on the number of friends and family the respondent is asked to mobilize. If the effect on turnout is conditional on the number of friends and family a citizen is asked to remind others to vote (as we see in our experiment), then extrinsic social benefits are playing a more substantial role in shaping turnout than intrinsic pressures to comply with promises made to others (since the latter not conditional on the number of people a citizen is asked to mobilized). The Limits of Social Priming 8

9 While extant research suggests that our social priming treatments should increase voter turnout, it is important to also consider that citizens may be sensitive to the scope of social priming. Certainly a request to mobilize 1,000 friends to vote would be seen as preposterous, and not just because only super-users of online social networks could claim to have so many friends. Even at much lower quantities, a request to mobilize others might undermine the effect of social priming on voter turnout. Empirical evidence of this phenomenon is suggested by the results of a field experiment by Ha & Karlan (2009) in which a conversational phone call encouraging voting increased turnout, but adding a request that the respondent mobilize a neighbor appears to undermine the effect. First, requesting that someone mobilize additional people makes voting a more burdensome task. Each mobilization attempt requires time and energy that is lost for other activities, costs that increase with the number of people a citizen is asked to mobilize. For example, the costs associated with asking an immediate family member or close friend to vote is very low because the citizen already converses with these individuals regularly. 2 Conversely, as the number of people one is asked to mobilize increases, the effort and opportunity cost is likely to increase because the citizen will have to deliberately seek people out (e.g. make time for an unusual visit or phone call) who are farther out on the periphery of their social network. Second, confronting dissonant points of view about politics may discourage voter turnout (e.g., Mutz 2006; but also see Huckfeldt et al. 2004). As the number of people one is asked to mobilize expands, the odds of encountering individuals with disagreeable viewpoints increases. A scope of social priming that requires a citizen to interact with people who are indifferent or, worse, hostile to their own political beliefs is likely to make the social aspect of voting 2 Mobilization is also more effective (i.e., less costly) if there is a social bond between the dyad. For example, Brady, Schlozman & Verba (1999) show that mobilization to participate in political activities is 28 percent more effective if the target was already acquainted with the mobilizer. 9

10 emotionally unpleasant. Even if the scope of social priming is fairly small, a citizen may feel it requires them to attempt mobilize people with whom they do not regularly discuss politics, and the potential for encountering conflicting viewpoints make this an uncomfortable prospect. As a corollary, if the scope of social priming raises the possibility of reminding people with whom a citizen disagrees about political ideology, partisanship, or other political values, encouraging these political opponents to vote is not only uncomfortable but it is also politically irrational since it will undermine the prospect of obtaining their own electoral preferences. The potential burdens and discomfort from a larger scope of social priming may lead to a form of what social psychology labels reactance. In social psychology, the tendency for people to push back in response to forceful messages designed to change their behavior has informed theories of reactance (Brehm 1966; Brehm and Brehm 1981). Social psychologists have found evidence of reactance when attempting to reduce alcohol consumption, prevent littering, alter consumer product choice, prevent illegal drug use, improve diet, and eliminate smoking (see Burgoon, Alvaro, Grandpre, and Voulodakis 2002 for a review). There is an important parallel between voter turnout and the public health concerns central to the reactance theory literature: Both voting and healthy lifestyles are known to be socially desirable, and yet many people choose to deviate from these social norms. As such, the studies of reactance to public health campaigns are an especially useful analogy for reactance in response to GOTV campaigns. Reactance theory posits that individuals with a negative response to a message may ignore it (Burgoon, Alvaro, Grandpre, and Voulodakis 2002; Stewart and Martin 1994) or perform the opposite of the behavior advocated (Worchel and Brehm 1970; Ringold 2002; Schultz 1999). Further, more forceful messages about changing behavior are more likely to elicit reactance (Miller, Lane, Deatrick, Young, and Potts 2007; Dillard and Shen 2005; Albarracin, 10

11 Cohen, and Kumkale 2003; Grandpre, Alvaro, Burgoon, Miller, and Hall 2003). If expanding the social scope of voting makes citizens perceive the social priming treatment as more forceful then they are more likely to ignore it or act contrary to the voter mobilization intent. The first manifestation of reactance will be an increased likelihood that the citizen asked to remind others to vote will refuse to agree to the social priming request during the voter mobilization call. More subjects will respond that they will not remind others to vote, or perhaps agree to remind others but reject the (large) number in the social priming request. We believe that this refusal to agree to the request is a signal of a more important underlying reactance that will undermine the mobilization intent of our phone call treatments. In short, as the scope of social priming shifts a citizen s perception of voting from easy and rewarding to burdensome and unpleasant, his or her motivation to vote will be undermined (or, at least not enhanced). All else equal, the expected increase in turnout due to social priming is the sum of the perceived social benefits of voting and the burdens of completing the mobilization task. That is, we expect that the effects of social priming on voter turnout will be non-linear:the expected net increase in turnout due to social priming follows a curve that initially rises as the scope of social priming increases, but then turns downward as the burden of asking more and more people to vote increases. Research Design We administered a field experiment using phone calls to encourage turnout among registered voters in Illinois, Michigan, New York, and Pennsylvania during the weekend prior to Election Day for 2010 General Election (N=198,285). The calls were conducted by a commercial call center in Orlando, FL under the supervision of a political consulting firm specializing in 11

12 voter contact calls. The calls began on the Friday prior to Election Day (October 29, 2010) and continued through Sunday (October 31, 2010) with most of the completed calls on Saturday, October 30, These types of phone calls are a staple of voter mobilization efforts. Research using field experiments has shown that live phone calls with a conversational tone can significantly increase voter participation (e.g. Nickerson 2007; Nickerson 2006; Nickerson and Rogers 2010; Ha and Karlan 2009; Michelson, García Bedolla, and McConnell 2009), but calls with an impersonal delivery are not effective (e.g. Gerber and Green 2005; Gerber and Green 2000). 4 A meta-analysis of 28 field experiments determined that turnout increased by an average of 0.55 percentage points when a registered voter was successfully reached by a standard commercial phone bank (Green and Gerber 2008). We randomly assigned registered voters to 5 conditions: a) a control group (i.e. no attempted phone calls), b) encouraging the individual to turnout with no social priming request [ No Request script], c) encouraging the individual to turnout and social priming to remind one friend or family member to vote [ One Person script], d) encouraging the individual to turnout and social priming to remind three friends or family members to vote [ Three People script], and e) encouraging the individual to turnout and social priming to remind five friends or family members to vote [ Five People script]. Table 1 demonstrates that the random assignment produced balanced control and treatment groups across age, gender, race, past voting, and geography. (Copies of the scripts are in the Appendix). The No Request script is the basis for each of the scripts in the field experiment. The core of the mobilization effect from the No Request script was expected to come from asking 4 Based on past research on the importance of an authentic chatty character to the calls (Nickerson 2007), the representatives at the call center were instructed to make the script as personable and chatty as possible. Thus, minor deviations from the script wording were acceptable to make the calls sound authentic. The authors monitored the calls to ensure that these variations were trivial. 12

13 registered voters about when they planned to vote on Election Day and how they planned to get to the polls. Nickerson & Rogers (2010) found that prompting registered voters to make a plan for voting (i.e., implementation intentions) makes conversational phone calls more effective than conventional get out the vote appeals to civic duty or requests to pledge to vote (i.e. selfprophecy of voting intention). In addition, the script asked respondents to pledge to fill out the entire ballot for all candidates and referendums to reduce roll-off for lower salience contests (Mann 2011). The final part of the script requested an address from the respondent to enable our partner organization to send a vote reminder (e.g. Dale and Strauss 2007; Malhotra et al. 2011). This collection effort was a near total failure (only 28 of 32,643 respondents provided s) so we do not address it here. For the three social priming scripts, an element was added to the No Request script following the vote implementation intention interrogatories and prior to the pledge to fill out the entire ballot. This third interrogatory of the respondent said, [w]e re also asking you to remind your friends and family to vote. Can we count on you remind one other person [three other people/five other people] to vote on Tuesday? We choose zero, one, three, and five people to provide a range of social priming sufficient to begin to detect the pattern of effects on turnout described above. We skipped two and four people in order to improve the power of the experiment to detect differences (i.e. we created 4 larger treatments groups rather than 6 smaller treatment groups). We were uncertain whether the One Person script would provide an increase above the No Request script. One person seemed like a relatively weak social prime, since voters were likely to think first of the most socially intimate people (i.e. spouses, partners, close friends, etc) and these highly socially intimate people are unlikely to significantly alter their judgments of the subject based on voting 13

14 or non-voting. Three people seemed likely to show robust effects from social priming one way or the other. We extended the range of social priming out to five people to get further leverage on the trajectory of effects from increasing the scope of social priming. The experimental subjects were selected by our partner, a non-partisan 501(c)3 charitable organization whose mission includes increasing voter participation. The selected registered voters met all of the following criteria (see Figure 1): (1) Registered voters in Illinois, Michgan, Pennsylvania, and New York with a strong match between phone listings and voter file records according to Catalist LLC, a consumer data firm specializing in information on registered voters. 5 (2) Registered voters with a predicted probability of voting between 30% and 70% based on a predictive voter turnout model provided by Catalist LLC. This criterion was based on previous research that voter mobilization contacts have maximum impact for registered voters with a chance of turning out (Green and Gerber 2008 p. 174; Arceneaux and Nickerson 2009; Niven 2004; Hillygus 2005; Parry et al. 2008). (3) Registered voters expected to trust information about political issues from our partner organization based on a proprietary microtargeting model. (4) Any phone number associated with more than one targeted individual was excluded to simplify the process of reaching the correct registered voter in a household, avoid an obvious violation of the stable unit treatment value assumption (Sinclair, McConnell, and Green 2010), and eliminate the complexities of analyzing cluster randomized assignment (Arceneaux 2005; Green and Vavreck 2007; Nickerson 2008). (5) Registered voters who had requested an 5 In practice, the strong match usually means a match of address and full name. Medium and weak match phone numbers include records that match only on address and last name, address only, etc. The standard practice of our partner organization, based on extensive experience with voter contact phone calls, was to use only strong match phone numbers. 14

15 absentee/mail ballot or cast early in person ballot prior to the October 27 th, 2010 were also excluded from the experiment. 6 Our partner organization s selection criteria draw attention to the question of external validity. The results from all field experiments are necessarily specific to the context in which they are conducted, and ours is no different. The opportunity to conduct this experiment in a partnership makes it more realistic in terms of what civic organizations actually do, but this means our subjects are not perfectly representative of registered voters (Hersh and Schaffner 2011; Gerber and Green Forthcoming). Nonetheless, Table 1 shows that our study population is sufficiently diverse to provide valuable insights about voting behavior. 7 In order to measure the effects of our experiment on voter turnout, all subjects in each of the five conditions were matched to official public records of individual level voter turnout following the election. These public records allow identical and unbiased measurement of the dependent variable voter turnout for all subjects across control and treatment groups. A treatment can only impact the subjects to whom it is successfully delivered. However, it is impossible to contact every individual assigned to a treatment condition (e.g., if a selected subject does not answer the phone). Thus, field experiments conventionally report two estimates of the treatment effects results when it is possible to record successful contact, such as with phone calls: the intent to treat [ITT] effect and the average treatment effect among the treated [ATT]. First, we briefly report the intent to treat [ITT] effects. 8 The ITT analysis estimates the 6 The exclusions of absentee, mail, and early in-person voters were based on data obtained from local election officials by Catalist LLC. 7 The larger N of the No Request condition is simply due to arrangement with our partner organization. No Request was the standard script for their voter mobilization program. They allocated ~75,000 records for our script variations, which we divided evenly among the remaining three treatments. 8 We report the intent to treat [ITT] effects, in part, to facilitate our own and future comparisons of our results with the results of other voter mobilization field experiments. Field experiments involving attempted contact with mail, 15

16 effects of each script by calculating the turnout rate among registered voters assigned to each condition, then comparing the rate in each treatment condition to the rate in the control group. Then we focus on the differences between the treatments among citizens who were successfully contacted. In order to estimate the effect of successful delivery of the script, the ATT analysis adjusts for the inability to contact every registered voter assigned to a treatment by dividing the ITT effect by the contact rate. 9 While our method for estimating the effects of our interventions follows that of other field experiments, we make a small break with convention by utilizing a two-tailed test of statistical significance in our hypothesis testing. A one-tailed significance test is appropriate for most field experiments because there is no theoretical expectation that the treatment will reduce turnout. However, our theory anticipates that expanding the scope of social priming too far will undermine its effect, including the possibility of a negative impact on turnout. The decline in effect on turnout in Ha and Karlan s (2009) results suggest that our effects on turnout can go either way with even the minimal One Person social priming request. Given these theoretical expectations and empirical precedent, it is appropriate to use the stricter standard of a two-tailed test of statistical significance. 10 Moreover, use of a two-tailed test makes our standards for hypothesis testing more conservative than using a one-tailed test. 11 After the calling commenced, we became aware of a deviation from the instructions to the supervising firm and call center. The calling for each of the four scripts was treated broadcast media, or other channels in which it is impossible to know whether a voter received the contact can report only the ITT effect. 9 In regression analysis of the results, the ATT is calculated by using random assignment to the five conditions as an instrument for successful delivery of the script (Gerber and Green 2000; Angrist, Imbens, and Rubin 1996). 10 See Mann (2010) for another two-sided theoretical expectation about voter mobilization treatments requiring a two-tailed significance test. 11 Strictly speaking, we have no theoretical expectation of a negative effect for the No Request treatment in the ITT analysis. However, a two-tailed test is a more conservative interpretation of the results so we use a two-tailed test for the No Request treatment for clarity of presentation and interpretation. 16

17 identically with regard to caller training, caller monitoring, repeated attempts, random order of dialing records, and other procedures. Calling for each script was supposed to begin on Friday. However, calling began for the No Request treatment on Friday, for the One Person and Three People scripts on Saturday morning, and the for the Five People script on Saturday at midday. We conducted several robustness checks to be sure that the deviations from the intended protocol do not influence our results (see Supplemental Information). We report the unadjusted results of the full dataset because the results are substantively identical when accounting for the deviation in the commencement of calling. In particular, Table 2 shows the four treatment conditions are well balanced on observable covariates among subjects successfully contacted, which suggests the deviation from calling the scripts in parallel did not have a major impact on reaching subjects. Results Agreement with Social Priming Request The call center representatives recorded whether the respondent agreed to the request to remind the specific number of people in the script, volunteered partial agreement that they would remind at least one friend or family member without agreeing to the specific number in the script, indicated indecision or equivocation (i.e. said maybe or don t know), disagreed (i.e. said no or refused to answer), or terminated the call. 12 Figure 3 shows an inverted U-shaped pattern for agreement with the social priming request. The One Person script receives 82.8% agreement to remind one person to vote. The Three People script generates a significantly higher rate of agreement: 86.9% agreement to remind three other people to vote (difference of proportions: 12 For the One Person script, full agreement and partial agreement were treated as equivalent since both were agreement to remind one person. 17

18 p<0.01), and an additional 3.1% agreement to mobilize at least one other person. The Five People script sees a precipitous decline in agreement to the social priming request: Barely more than half of the respondents (53.7%) agree to mobilize five other people. The partial agreement rises to 9.6% for the Five Person script, but this accounts for only about 1/3 rd of the decline in agreement between Three People and Five People. The differences between the One Person and Three People scripts in other response categories appears consistent with the supposition that subjects might not consider the One Person request carefully. More than half of the increase in agreement with the Three Person script comes the 3.9 percentage point decline in equivocal responses ( Maybe/Don t Know ) relative to the One Person script. The responses to the Five Person script indicate that it elicited strong feelings of resistance and rejection from some subjects. While most of the decline in agreement with the Five Person script is accounted for in the 19.2% of respondents coded No, the other negative response codes provide a sense of a stronger reaction to the request. Refusals to answer and terminations of the call are indistinguishable between the One Person and Three People scripts (difference of proportions: refusals, p=0.21; terminations, p=0.21). For the Five Person script, refusals rise by about 50% to 3.6% (difference of proportions: p<0.01) and terminations more than triple to 9.1% (p<0.01). Increasing the request from reminding three people to reminding five people appears to have generated a three-fold increase in the share of respondents willing to violate social norms by abruptly terminating the phone conversation. This increase in the norm-violating behavior of hanging up the phone suggests that some subjects are not just declining the request, they are forcefully rejecting a request they perceive to be too large. This forceful rejection is consistent with reactance to the treatment that will 18

19 undermine the subject s motivation to vote. We now turn to see if this pattern of agreement with the request to remind friends and family members to vote also appears in the effect of social priming on turnout. Voter Turnout Table 4 reports the turnout in each of the five randomly assigned conditions and the difference between the control group and each of the treatment conditions. Table 5 reports the intent to treat effects from regressing turnout in the 2010 General election on the assignment to each script. Model 1 does not include covariates and therefore replicates the findings from Table 4. Model 2 includes covariates for age, gender, race, past voting history and state (see Supplemental Information for full results). Since the treatments were randomly assigned the covariates are not necessary for an unbiased estimate of the effects, but including covariates reduces the standard error of the estimated treatment effect (i.e. improves the efficiency of the estimates). As expected given the balanced random assignment on these covariates in Table 1, there is negligible difference in the estimates of the effects when the covariates are included in Model 2. Since the variation in the social priming can have no effect unless the voter is successfully contacted, Table 6 presents the average treatment on treated effects [ATT]. The discussion below focuses on the estimates in Model 2 of Table 6 because they are (slightly) more precise and account for successful contact. The results appear to follow an inverted U-shaped pattern as the scope of social priming increases. The inverted U-shaped pattern begins with the No Request script which has no effect on turnout in this experiment. The One Person script appears to increase turnout by 0.63 percentage points, but this is not statistically significant (p=0.25). The expanded social priming 19

20 of the Three People script generates a 1.53 percentage point increase in turnout (p<0.01). However, further expansion of the scope of social priming in the Five People script causes the effect to decline to at least zero. In fact, the decline is so sharp the point estimate of the effect is negative (-0.2 percentage points), but not distinguishable from zero (p=0.71). Since we are interested in the effects of the scripts relative to one another as well as the conventional significance test relative to a null hypothesis of no effect, it is important to note that the pair-wise differences between the effects of the Three People script and each of the other scripts are robust (No Request script: p<0.01, Five People script: p=0.01, and even the positive estimated effect of the One Person Script has a promising trend at p=0.19). Thus, the nonlinearity in the effects as the scope of social priming increases seems unlikely to be random chance. The curvilinear pattern is consistent with our theory above that turnout will initially increase with the scope of social priming, but reverse when the scope becomes too large. The curvilinear pattern in turnout is also consistent with the pattern of agreement with the social priming request described above, which suggests that the perception of the scope of social priming is a mediator for the effect of the calls. In the context of the extant literature, the increase in turnout when progressing from No Request to One Person to Three People is consistent with previous research that finds emphasizing the social aspect of voting will increase turnout. However, the decline from the Three People script to the Five People script is a novel finding. Cost Per Vote We are concerned with the results of different degrees of social priming rather than whether phone calls are effective relative to other voter mobilization techniques, which is well 20

21 established (for a review, see Green and Gerber 2008). However, in addition to estimating treatment effects on voter turnout, it can be useful to calculate the cost per net vote generated by the treatment in order to compare effects across different treatments. The cost per net vote for the Three People treatment is $91 per vote. If we accept the point estimate for the One Person treatment (despite failing to reach conventional levels of statistical significance), it has a cost per net vote of $240. The cost per net vote for the No Request treatment approaches infinity since it has no effect. Based on a possibly negative effect on turnout, the cost per net vote for the Five People treatment would then be something less cost effective than infinity. These costs are considerably higher than the results reported in some previous experiments (e.g. $52, Ha & Karlan 2009; or $29, Nickerson 2007), but well within the variation seen across past studies in different elections, using different scripts, etc (Gerber & Green 2008). The high cost per net vote may be due to a difficult context for mobilization, a sub-par phone bank, or some combination of these and other factors. Nevertheless, the low performance does raise the possibility that call quality might have been low. If so, it seems likely that higher quality calls might increase the impact of all scripts while retaining the same curvilinear pattern of differences. It is possible that superior call quality could sell more citizens on agreeing with the Five Person script, but it would likely be no more than a delay in the decline of the effect. Discussion Our findings suggest that citizens view casting a ballot as a social activity, but only within limits. When the pledge to remind others to vote is carried out, citizens create their own vortex of social pressure increases their perception of the extrinsic social rewards for casting a ballot. Social priming of one person appears to have some (albeit insignificant) effect, but 21

22 certainly does not maximize the impact of emphasizing the social aspect of voting. The expansion to three people increases the likelihood that the citizen will be meaningfully sanctioned non-voting or rewarded voting and therefore increases the effect on turnout. In contrast, citizens appear to resist social priming when the scope of family and friends to remind about voting becomes too large. The drop in agreement with the request to remind five friends and family members to vote appears to be more than simply declining the request: the tripling of call terminations suggests that a form of reactance against the entire mobilization contact was occurring. The parallel pattern of response to the social priming request and effect on turnout suggests that reactance against a large social priming request eliminates any positive effect of leveraging the social aspect of voting with social priming. In the discussion of why social priming would increase turnout, we proposed that compliance with the social priming request might operate, at least in part, through a selfprophecy or commitment mechanism. That is, a citizen who reminds someone else to vote is implicitly predicting and/or committing to vote him/herself. The evidence in our experiment can neither confirm nor reject whether this mechanism is present. However, the minimal effect from the One Person script and the variation across the three social priming scripts does allow us to reject the hypothesis that social priming is merely another form of self-prophecy or commitment influencing turnout. To create the variation in effects here, social priming must activate mechanisms beyond an internal dialogue about consistency with predictions and promises. Our findings also reconcile an apparent contradiction among previous experiments. The line of social pressure field experiments begun by Gerber, Green and Larimer (2008) shows that emphasizing the social aspect of voting increases turnout. However, in a 2004 field experiment on voter mobilization phone calls by Ha and Karlan (2009), adding a request to mobilize a 22

23 neighbor to vote appeared to (unexpectedly) reduce the impact of voter mobilization phone calls. Although the direct request in Ha and Karlan s script was to mobilize one neighbor, it followed an indeterminately large scope request to think of people in your building or on your block that you could encourage to vote (p.366). Our theory and findings provide a richer explanation of the effects of social priming than Ha & Karlan offered when confronted with this unexpected finding. 15 First, our results show that adding social priming improves the efficacy of mobilization phone calls, as they expected and as seen in more recent social pressure field experiments. Second, we explain why a social scope that is too large can reverse the positive impact of social pressure on voting and demonstrate that this occurs with a carefully designed experiment. The results of our social priming intervention support our theory that there is a sweet spot of social priming for voter mobilization. Nonetheless, further replication and extensions are needed to flesh out the potential and limits of social priming for increasing turnout. The results from every experiment are specific to their time and place, so replication is the best way to build confidence that the results are generalizable. We see several aspects of our design that should be addressed in future replications. First, replication among other types of citizens in important. Conducting this experiment in partnership with an actual organization s efforts to increase mobilization make it more realistic in representing the types of voters that such organizations seek to mobilize. On the other hand, this targeting of voters makes it difficult to explore some theoretically interesting questions. For example, we targeted mid to low voting propensity registered voters because they are generally the most responsive to mobilization interventions. However, citizens with a higher voting propensity might be more interested in social rewards and sanctions about voting (i.e. 15 Confronted with their unexpected result, Ha & Karlan (2009) postulated that adding a request to mobilize a neighbor to their phone script seems to undermine the effectiveness of phone mobilization, possibly by shifting the attention of recipients from a vote yourself message to a get your neighbors to vote component (p.362). 23

24 reputational costs and benefits). Higher propensity voters might also be willing to agree to larger social priming requests, thus pushing out the deflection point for the decline in positive effect on turnout. Second, we chose to have fewer groups across a broader range of social priming in our experiment to maximize the power of the experiment, but using one person differences in the treatment (One Person, Two People, Three People, etc) would provide fuller detail for the pattern we observe. Third, we could learn about the impact of different types of social interaction by using terms other than friends and family. Less socially intimate references such as neighbors, social identity groups, or even strangers seem likely to elicit different agreement with (and reactance against) a social priming request. Fourth, it would be valuable to test the delivery of social priming requests through more personal communication (e.g. face-to-face canvassing) and less personal communication (e.g. mail and ). Less personal delivery seems likely to reduce agreement, but might also decrease reactance since the communication is less psychologically forceful. Conversely, faceto-face delivery of the social priming request might increase agreement due to satisficing and other social mechanisms, but it isn t clear whether this higher initial agreement would translate into larger effects on turnout if it is not genuine. Fifth, in thinking about how to strengthen the effect of social priming, it would also be useful to investigate whether reactance to social pressure interventions can be assuaged using restoration of freedom techniques designed to frame the blandishment as a choice rather than a demand (e.g., Miller et al. 2007; Mann 2010). 24

25 Conclusion We conducted a field experiment among 198,285 registered voters that explored the effects of varying the scope of a new form of social pressure we call social priming : registered voters were asked to remind one, three, or five of their friends and family to vote, or no reminder request was made. As our theory of decreasing returns from expanding the social scope of voting suggested, we find an inverted U-shaped curvilinear relationship between the scope of social priming and voter turnout rate: the request to remind three people significantly increased the rate of voter turnout, while there was no effect from the treatment with no voting reminder request, a small but not statistically significant effect for reminding one person, and possibly a negative (but not statistically significant) effect for voting reminders to five people. These results support the proposition that there is a sweet spot when it comes to leveraging the social aspects of voting; too small a social scope failed to impact voting behavior, while too large a scope undermines the value of social priming. It is also important to note that the results of our experiment have implications well beyond the study of voter mobilization tactics. The insights about the self-created vortex of social pressure from asking others to vote have value because this phenomenon shapes voter turnout without intervention by outside groups. Instead, these experimental results suggest that a disposition to talk about voting, particularly encouraging others to vote, makes these talkers more likely to vote. Political talk is not merely a by-product of being a likely voter, political talk is a contributor to propensity to vote. We have developed a new theory of social pressure within the context of voter turnout, and tested it with a new form of social pressure intervention, social priming. The curvilinear inverted U-shaped pattern of results in this experiment highlights the need to consider the limits 25

26 of the social aspect of voting for increasing participation. While it is clear from recent research that emphasizing the social aspects of voting can increase turnout, presenting voting as an activity with too large a social scope appears to backfire by undermining the intended mobilization effect. As in many aspects of human behavior, there can be too much of a good thing when it comes to emphasizing the social aspects of voting. 26

27 Figure 1: Selection of universe for experiment (by partner organization) 27