How voter mobilization spread in households and families. The use of short text messages on cell phones to boost turnout

Transcription

1 How voter mobilization spread in households and families The use of short text messages on cell phones to boost turnout January 16, 2015 Yosef Bhatti**, Jens Olav Dahlgaard*, Jonas H. Hansen* & Kasper M. Hansen* This project is primarily funded by the by the Danish Council for Independent Research (grant no ). The project has also received funding from the Danish Youth Council. *University of Copenhagen, Department of Political Science, Øster Farimagsgade 5, DK-1353 Copenhagen K, Denmark, **KORA Danish Institute for Local and Regional Government Research, Købmagergade 22, 1150 Copenhagen K. & Web: A previous version was presented at NOPSA2014, August in Gothenburg, Sweden.

2 Introduction Large amount of research suggests that citizens' voting behavior is related with the voting behavior of their social environment (Campbell et al. 1960; Lazarsfeld et al. 1968; Zuckerman 2005). This has been shown to be the case for married couples (Wolfinger & Rosenstone 1980; Stoker & Jennings 1995), members of the same communities (Anderson 2009), peers discussing politics with each other (La Due Lake & Huckfeldt 1998), and people living within the same neighborhoods (Cho et al. 2006; Fieldhouse & Cutts 2008) or households (Fieldhouse & Cutts 2012). Furthermore, several studies demonstrate a high degree of transmission of political behavior between parents and their adult children (Jennings et al. 2009; Fieldhouse & Cutts 2012). It has also been suggested that the strongest variable to predict whether a potential first time voter actually goes to the polls is whether her parents vote (Bhatti & Hansen 2012). Thus, parents play a highly important role with regard to children becoming habitual voters or not (Plutzer 2002), either due to socialization, genetics or because parents and child often share household when the young adult becomes eligible to vote. A central concern raised about the relationship within the social network is that many of the studies, regardless of their high degree of sophistication, are of observational character. There is evidence that the political behavior of citizens' who share social environments are related, but it is difficult determining whether there is a causal effect from a given person in a social network to another person because of the potential sorting into networks. The fact that persons in a social network tend to reside in homogenous contexts could be a common cause of turnout, too, which could inflate the actual effect of the network. Recent research have tried to overcome this problem by applying regression discontinuity designs for widowed citizens (Hobbs et al. 2013) and for coresidents (Bhatti & Hansen 2013). These studies indicate that individuals are affected by the death of their spouses (Bhatti & Hansen 2012b). They also show that there is a large degree of congruence 1

3 in voting behavior for two persons living together, and this effect is not likely only to be due to long-term socialization or self-selection. Furthermore, evidence from natural randomizations at a US college suggest that peers indeed have a causal influence on the degree of political participation of each other (Klofstad 2007). In other words, there is reason to believe that the decision whether to vote or not is a social one. The most convincing evidence of the spillover effects of turnout stems from field experiments. Recently, get-out-the-vote-studies have been designed to examine how treatment effects spread across individuals in households and in friendship networks (Nickerson 2008; Sinclair et al. 2012); (Bond et al. 2012). These studies have the advantage that by design they should be unaffected by confounders. Our study uses a get-out-the-vote (GOTV) experiment with short-text-messages distributed on cell phones in the 2013 Danish municipal election to study transmission of treatment effects within households and within families, conditioned on whether they share a household. We thereby aim to contribute to the literature on the influence of social networks on turnout in three ways. First, most observational and experimental studies focus exclusively on social effects within households. This is indeed the most likely place to find spillover effects as members of a household normally interact in a daily basis and are confronted with each other s decision to vote or abstain. However, we need more knowledge regarding if and to which degree treatment effects travel to different type of relations (Bond et al. 2012). One unresolved question is whether such treatment effects also travel to family members, even though they do not live together. In other words, the analysis, examine spillover effects to family members conditioned on whether or not they share the same household. 2

4 Secondly, the few existing studies on travelling treatment effects are indeed useful, but often use treatments that are very difficult or costly for a campaign to apply. For instance, the Facebook study by Bond et al. (2012) demands access to manipulate the Facebook start page the individual meets on Election Day. This is not possible for most organizations. The seminal study by Nickerson (2008) is highly work-intensive as it applies door-to-door canvassing. Sinclair et al. (2012) use of social pressure mailings is less costly, but is impossible to carry out in many countries where turnout is not a matter of open public record. In this study we examine spillover effects of short-text-messages on cell phones. Besides being easy to administer, short-text messages are also interesting because there are some indications that spillover effects can be different across mobilization methods (Sinclair et al. 2012). Thirdly, the existing experimental studies are all from the US. An important question is whether spillover effects are the same, less or higher in other countries and elections. In the existing studies, the baseline turnout has been below 40 percent and most often even below 25 percent. In this article, we test whether spillover effects can be tracked in a Danish municipality election, where the overall turnout was 71.9 percent. As such, there are fewer citizens to mobilize directly and indirectly by applying GOTV-tactics. On top of this, there is a different political culture and media system that stimulates higher degree of political awareness with regard to e.g. discussing political issues and political interest in Denmark (and other countries) compared to US. One indication of this is that door-to-door canvassing seemingly does not have an effect on turnout in Denmark (Bhatti et al. 2014a; Bhatti et al. 2014c), whereas this tool has been quite successful in the US (Green et al. 2013). In the next section, the theoretical underpinnings of the study are unfolded. Then the context and data are described before going into detail about the design of the study, a field experiment utilizing encouragements to vote via short-text-messages to cellphones. In the analysis, 3

5 the main findings of the empirical analysis of the spillover effects are reported within households and within families. Finally, the findings are discussed. Theory the social act of turning out Scholars have had interest in the relationship between social networks and voting for decades. It was early noted that individuals vote in groups and that interpersonal relationships are important for political behavior (Campbell et al. 1960; Lazarsfeld et al. 1968). Several studies have documented a strong relationship between the behavior of the individual s network and her tendency to vote or abstain in elections (Stoker & Jennings 1995; Cutts & Fieldhouse 2009; Fieldhouse & Cutts 2012; Bhatti & Hansen 2012a). There are at least three inter-related potential reasons why the people surrounding citizens in their daily life influence their turnout decision causally - social pressure, low cost information and the companion effect. Firstly, people evaluate their own behavior in relation to their social groups. The norms of what is seen as proper and correct behavior is in this way contingent on how it is perceived by the people around them (Zuckerman 2005). This logic is applied in several GOTVstudies using treatments with a social pressure appeal. For instance, letting people know that their turnout behavior will be exposed to their neighbors significantly increase the turnout of the receivers because it arguably is a social norm to vote and thereby perform one s civic duty. The risk of having your lack of norm compliance exposed to the neighbors has a positive impact on turnout (Gerber et al. 2008). 4

6 A second, related, reason why citizens' are influenced by their social environment may be that the information coming from household members, family members, friends and so on are so-called 'low cost information' (Klofstad 2007). Citizens receive large amounts of information and appeals every day, but might be too busy to cope with it all. A potential shortcut is to take more notice of information coming from one's social network, which people intuitively trust. When deciding whether to vote or not, citizens in this way will be more influenced by the arguments and behavior from their friends and family than by information from impersonal campaigns. Therefore, when a citizen actually goes to the polls due to a given treatment, it will very likely have an extra spillover effect if the citizen talks with people in her network about it. A third, likewise related, reason is a more practical one. Most voters choose to vote at a polling station on Election Day 1. Hence, whether their household members vote or not may influence individuals. If one lives with a voter, it may be difficult to abstain as one is confronted with the household members decision to vote. In addition, it may be more convenient to vote when others go to the polls with you (e.g. sharing transportation costs). In fact, exploratory studies indicate that individuals often vote together with other members of their household, sometimes labelled the companion effect (Fieldhouse & Cutts 2012). One challenge when studying social networks is that individuals self-select into their networks. Homophily in marriages and friendships is a well-known fact of the social world (McPherson et al. 2001). Among other things individuals seem to sort based on political characteristics (Alford et al. 2011). Furthermore, some of the congruence in behavior with our families may be due to genetic similarities or other background characteristics rather than our continuous interactions (Fowler et al. 2008). Finally, turnout can be influenced by contextual factors. As social networks tend to share contextual factors their impact is to some extent clustered percent of the voters voted at a polling station on Election Day at the Danish municipal elections

7 on the network. Even with rigorous control variables it can be difficult to preclude that sorting into relationships or common background characteristics within families can inflate our estimates of the importance of social networks. Two strands of literature have tried to cope with this potential problem. First, a couple of recent studies have applied natural experimental designs (e.g. regression discontinuity designs) to study the effect of social networks (Klofstad 2007; Hobbs et al. 2013; Bhatti & Hansen 2013). Secondly, field experiments have been used to investigate how treatment effects travel from the treatment group to the network of the treatment group. Traditionally, GOTV-research has focused on studying the direct effect of GOTV-campaigns on turnout behavior. That is, what is the effect of a given treatment (e.g. direct mail) on turnout? Evidence is derived from comparing the turnout for a treatment group that receives the treatment with a control group, which does not receive any treatment. However, a couple of recent studies have exploited the fact that when a treatment is administered to a single person in a network without direct effect on others, we can measure the spillover effect on her social network by estimating the difference in turnout between the network of the treatment and the network of the control group (Nickerson 2008; Sinclair et al. 2012; Bond et al. 2012). Nickerson's (2008) seminal study provides strong evidence with regard to the contagious effect of voting behavior within households. In his study, a door-to-door campaign was targeted to reach one member of a two-person household, which in itself had a positive treatment effect. More importantly, however, by comparing indirectly treated individuals with indirectly treated individuals in a placebo-treatment group Nickerson documents a large indirect effect on the other person in a household. For every 100 citizens mobilized to vote, an estimated 60 extra citizens in these persons' households were mobilized. There is, thus, apparently a potential for substantial spillover effect of mobilization campaigns within households. 6

8 In another GOTV-study, Bond et al. (2012) document a substantial and statistically significant spillover effects from a Facebook vote appeal. The study shows that there is a small, direct treatment effect of receiving an online GOTV-treatment. Of more interest to this article, it also shows that there is a contagious effect of the campaign. For each person who was directly mobilized by the campaign message, more close friends to the receivers where mobilized. In fact, the indirect effect of the campaign was larger than the direct effect. An important lesson is that even when the direct effect sizes are small, the true total effect might be larger. However, it demands a relatively large sample size to be able to discover such spillover effects when the direct effect is small. Finally, Sinclair et al. (2012) investigate the spillover effect of social pressure mailings within households and neighborhoods. They find some evidence of spillover effects within household though only slightly above half the magnitude of what Nickerson (2008) finds, while there is no evidence of spillovers within neighborhoods defined by zip codes. Though a couple of high quality studies now exist, there is a need for further studies investigating spillover effects for different groups, using different interventions, and in different contexts. Especially, there is a need for studies investigating spillover effects from digital mobilization methods and within families, including those not sharing a household. Many of the causal mechanisms described above are strongest within households. Therefore it is possible that the spillovers are modest across even closely related family members if they do not live together. As indicated by Sinclair et al. (2012) it is also interesting to investigate whether the large spillover effects are exclusive for door-to-door interventions or extends to all types of interventions. 7

9 Context and Data The high salience Danish municipality elections with actual voting files We fielded the experiment prior to the municipality elections in Denmark November 19, The elections took place simultaneously in all of the 98 Danish municipalities. At municipal elections, each municipality is a constituency where between 9 and 55 mandates are distributed proportionally among multiple parties. There are eight national parties along with few local lists that run for election in each municipality. All Danish citizens and citizens from EU, Norway or Iceland who are 18 or older on Election Day and have permanent residence in the municipality are eligible to vote. Immigrants from non-eu countries are eligible to vote after three years permanent residency in the realm. This is a more lenient eligibility requirement than in national elections where only Danish citizens, 18 or older with permanent residence in the realm are eligible to vote. Voters do not need to register to vote all automatically receive a polling card by postal mail. Arguably, voters perceive the elections as second order compared to the Danish Parliament Elections, but turnout is still around 70 percent with substantial variation across the municipalities (Bhatti et al. 2014d). In the 2013 elections, the average turnout rate was 71.9 percent. The capital, Copenhagen, saw the lowest turnout with 61.2 percent and the small island municipality of Fanoe had the highest rate of 85.4 percent. Furthermore, turnout varies substantially between age groups. In the 2013 municipal elections, the turnout for the year olds where 54.5 percent, while the turnout for the year olds where 77.6 percent (Bhatti & Hansen 2010: 18). Even though turnout is higher in the national elections, high turnout rates demonstrate that the municipality elections are still salient in the mind of the voters. The municipalities do play a central role in the everyday of most citizens. They are the main welfare provider with responsibilities including essential areas such as public schools, childcare, eldercare, and labor 8

10 market counseling. In total, the municipality spending is more than a quarter of the Danish GDP and the municipalities collect taxes directly from its citizens. In Denmark, researchers can be granted access to comprehensive records on the population via Statistics Denmark, the official statistics bureau. The records are based on several individual level registers that can be matched in anonymous form using a scrambled version of the social security number automatically assigned to all Danes and all immigrants with permanent residency. The records include information such as date of birth, gender, ongoing and finished education, residence, ethnicity and income and are considered very accurate. One thing that is usually not registered in the public records in Denmark is voter turnout. However, the voter files contain a unique individual identifier, which makes it possible to match turnout to other information in the public records. We were allowed to do this prior to the elections by contacting each of the 98 municipalities and ask them to send the voting files to us in order to merge them with the public records. Administrative errors at a few polling places after the elections led to deletion of a minor part of the voter files. For that reason, we end up with a sample that is slightly smaller than the population (4.36 million voters or percent of the eligible citizens). 2 One main advantage with the register data is that individuals can be linked within households or families. Therefore, we can investigate how treatments travel within households and families, for instance from children to their parents. This leaves us with highly reliable data including validated turnout and the possibility of linking individuals within formal networks. 2 Our full voter file contains information on 4,362,156 voters. In comparison the official figures from Statistics Denmark are 4,409,251. In other words, our data size is percent of the entire population size. Besides the few missing voter files there are few discrepancies between our file and the official figures in most municipalities. Some of these minor differences (most often less that 0.1 percent) might be voters who deceased, moved out of the country, moved from one municipality to another close to Election Day. 9

11 Research Design A field experiment using text messages In order to study spillover effects in social networks we analyze an experiment that was conducted as part of an investigation of the direct impact of text messages (SMS) on turnout (Bhatti et al. 2014b). The direct treatment effect in the experiment was 1.82 percentage points on average (Table 1). Compared to two US studies, this is a smaller intent-to-treat effect than Dale and Strauss finds (2009) using warm messaging (a message with prior consent from receiver), but a larger intent-totreat effect than Malhotra et. al. (2011) finds using cold messaging (a message without prior consent from the receiver). The Danish law allows mass-distribution of short text messages to cell phones without consent from the receiver as long as it is not done for marketing purposes. Non-partisan text messages with the aim of increasing turnout are considered as such. We cooperated with The Danish Youth Council, an umbrella organization for roughly 70 Danish Youth organizations including scouts, church, students, environmental and political organizations. It functions as an interest organization for the Danish youth in general and, among other things, focus on increasing young citizens participation in civil society and democracy. We restricted this experiment to potential voters aged 22 to 29 on Election Day. To run the experiment we needed cell phone numbers that subsequently could be matched with data from the official voter records. A private company matched names and addresses of individuals with as many cell phone numbers as possible using public phone books. The cell phone number enriched data was then merged back on the official records percent of the population was successfully enriched 3. We restricted the sample to households with only 1 enriched 3 Almost all Danes uses cellphones. 93 percent of all Danes have used a cellphone within the last three months (Danmarks Statistik 2013). For Danes under 35 years, this figure is 98 percent. So in theory, one could reach almost all Danes via short text messages. However, many phone numbers are not linked with sufficient information to merge them with the public records. 10

12 individual in the age group 4. We did this as the number of households with multiple members aged years were quite small, and we wanted to utilize these households in an experiment described elsewhere (Bhatti et.al., forthcoming). We ended up with a sample of almost 50,000 voters. We used this sample as the basis of our randomization into treatment and control groups. Our goal with the original experiment was not only to test the effect of a text message reminder/encouragement to vote, but also to test if the timing of the message matters. To investigate this we assigned 27,500 to one of several treatment conditions. Beginning one week before the election and ending the day before the election, we sent 2,500 voters a text message each day with no overlap between those who received text messages on the different days. Thus, we treated 17,500 over seven days prior to the election. Of this group, 7,500 received an additional text message at 3 PM on Election Day with an additional reminder to vote. We randomly drew an equal amount of voters to each of the treatment arms. On Election Day, we treated another 10,000 voters who we had not treated previously. Every hour beginning at 10 AM and ending at 7 PM we sent 1,000 previously untreated voters a text message with an encouragement to vote 5. The polling stations opened at 9 AM and closed at 8 PM. The design allowed us to explore if there was an overall effect on the 27,500 we treated and if the effect of the treatment is conditioned on the timing or receiving an additional reminder. All were assigned at random. We were able to match 96.4 percent of the individuals in the treatment and control groups to the voter lists (see Table A1 of the appendix for descriptive statistics for the sample). Therefore, the final size of the treatment group is 4 In the original experiment we also included a few individuals who lived in larger households, as it for political purposes was a requirement that all groups of young adults were treated. In all analyses in this paper we exclude all households where more than one individual was phone number enriched (345 individuals) to ensure that only one person from each household was included in the experiment. We furthermore exclude one person with no valid household ID. 5 The text messages where delivered with high precision. All text messages in the days up to the election where delivered on the exact time (11 am). On Election Day, the largest delay in delivery was 14 minutes from the agreed time, which was due to large pressure on the phone network. 11

13 26,873 and the control group size is 20,973. The sources and impact of the attrition was the same in both groups and therefore it does not lead to bias in our estimate. The experiment applied cold text messages (Malhotra et al. 2011) meaning that the recipients had not signed up or indicated that they would like to receive text messages from the campaign before they received a text message. Since the focus of the study was to test the effect of text messages and their timing, we abstained from varying the message more than necessarily. The main inspiration for the content of the text messages were drawn from Dale and Strauss' (2009) study, where they effectively used a civic duty argument in an short text message with an intent-totreat effect of 3.3 percentage points 6. Small adjustments were made to the Dale and Strauss-message in order to fit within a limit of 150 characters and to fit better with normal use of Danish language. Research indicates that effects do not vary greatly by message content while framing the norms used to frame the argument can be efficient (Green et al. 2013: 36-37). Therefore, we tried to formulate messages that framed the norm of voting. As our goal was to make the strongest possible treatment, we added the recipient s first name in the message based on the belief that this would help recipients perceive it as warmer. We did not implement a test of this in the design. The first name where drawn from the public phone records. A message sent out on Election Day to an imaginary voter named Alan would be: Hi Alan. This is a friendly reminder of the Election on Tuesday November 19. Democracy needs you so remember to vote! The message we sent out on Election Day would be: 6 Dale and Strauss (2009) applied warm messaging, meaning that people earlier had signed up to receive messages from the organization. The exact wording in their civic duty message was: A friendly reminder that TOMORROW's Election Day. Democracy depends on citizens like you-so please vote! -PIRG/TxtVoter.org (Dale & Strauss 2009, p. 796). 12

14 Hi Alan. Thank you for voting in the municipality election. If you haven t voted yet, you can make it until 8 PM. Both of the original text messages were in Danish 7. The sender of the messages was stem.dk (the direct English translation would be vote.dk ). This is a webpage sponsored by several organizations including The Danish Youth Council, promoting youth turnout and the organizations mobilization campaigns. Analysis spillover effects within families and households In Table 1, we show the turnout rates in the treatment and control groups and the estimated treatment effect. As the purpose of the paper is to look at spillover effects, we only look at the overall effect from the experiment (the entire treatment group versus the entire control group) and do not distinguish between the various sub-treatments. Table 1: Direct treatment effects (percent). Treatment group turnout N=26,873 Control group turnout N=20,973 Estimated treatment effect 1.82* (0.45) Standard error in (). *p<0.05 (one-sided test). Sample years olds on Election Day where each household only include one eligible year old with a valid phone number. We test for significance with randomization inference clustered on household level with 100,000 iteration using the R-package ri (Aronow & Samii 2012). See appendix A2 for descriptive statistics for the treatment and control groups. 7 The original wording in the two messages in Danish was: 1) "Hej [first name]. Dette er en venlig påmindelse om valget tirsdag d. 19. november. Demokratiet har brug for dig, så husk at stemme!" 2) "Hej [first name]. Tak fordi du har stemt til Kommunalvalget. Hvis du ikke har stemt endnu, kan du nå det frem til kl. 20.". 13

15 Turnout in the treatment group was percent in the treatment group; a 1.82 percentage point increase over the control group turnout of percent. The effect is statistically significant, and even though it is not of the size of many door-to-door canvassing experiments or strong social pressure treatments, it is still a sizeable effect considering the relatively low cost of the treatment administration and the high turnout of the control group. In order to investigate the social effects of voting we match each individual in the treatment and control condition with his or her household members and family members. For the 47,846 individuals in the treatment group and control group we were able to identify 30,223 housemates, 43,512 mothers and 40,938 fathers. Below we examine the spillover effect on these individuals. To investigate the spillover effect we compare the treatment effects on the target population (direct treatment effect) with the treatment effects on their various household and family members (spillover effect). The sample size is smaller than in Table 1, as we only look at those who share a household with others. Single person households are excluded. 14

16 Table 2: Direct treatment effects and spillover effects for household members (percent). Direct effect ALL 1 other in household Treatment group turnout Control group turnout Direct treatment effect 63.00, N=12, , N=10,212 Spillover effect on other house members Treatment group turnout Control group turnout N=9, N=7,555 2 others in household N=2, N= or more others in household N= N= * (0.64) 0.93 (0.73) 4.59* (1.40) (3.62) 69.12, N=16, , N=13, N=9, N=7,555 76,97 N=5, N=4, N=1, N=1,281 Spillover effect 1.43* (0.68) 0.28 (0.73) 1.55 (1.06) 8.00 (3.58) Spillover percentage Standard error in (). *p<0.05 (one-sided test). Sample min. two persons in household. We test for significance with randomization inference clustered on household level with 100,000 iteration using the R-package ri (Aronow & Samii 2012). See appendix A3 for descriptive statistics for the treatment and control groups. The spillover effect is remarkably large. When we look at all household members, the spillover effect is 1.43 percentage points (significant, p<0.05, one-tailed). 8 It is highly interesting that significant and substantial spillover effects can be found given that the turnout in the control group (68 percent) of the network is much higher than in previous studies. Thus, not only can voters be mobilized in high turnout elections, the effect can be spread to their high turnout house members. In columns 3-5 we break down the results by household size. This is necessary in order to calculate the spillover percentage as larger households weigh more in the spillover groups compared to the treatment/control groups when we look at individual level effects for all households simultaneously. Therefore comparing the spillover effect of 1.43 with the direct 8 Some of the overall spillover effect seems to be driven by a very large spillover margin for the large households group (last column). The group is quite small and there is no direct treatment effect for this group. The results for this group are difficult to interpret, but one should still be aware, that it influences the overall spillover effect. 15

17 treatment effect of 1.81 could be misleading. When there are one or two other individuals in the household we get spillover percentages of 30 (0.28/0.93=0.30) and 34 (1.55/4.59=0.34) which is remarkably close to the results found by Sinclair et al. (2012). In other words, for every 100 citizens mobilized to vote by the text messages, about 30 out of 100 citizens in these persons' households were also mobilized to vote 9. Our spillover percentages are around half the size of those found by Nickerson (2008). There are two likely explanations for this. The first explanation could be that the mobilization method matters for the percentage of the direct effect transmitted. In that respect one would perhaps expect short-text-messages to be less noticed by housemates compared to door-to-door canvassing and thereby generate less intra-household discussion. Another explanation could be a ceiling effect due to the very high turnout in the spillover group in the present study. The last column indicates that we cannot see a direct effect for persons who share household with more than 2 others. However, those who share household with the treated have substantially higher turnout rate than those who share household with the control group. Since the direct effect is estimated to be negative and of an absolute size considerably smaller than for the directly treated we cannot obtain a meaningful spillover effect here. So far, our results have largely supported the results of the previous literature but in a different context and with an impersonal mobilization tool. One interesting question is whether the spillover effects extend beyond the household and if so whether it does to the same extent as above. As noted previously we can identify family members even if they are not a part of the household. This can now be utilized to analyze the importance of sharing a household on the spillover effect. In Table 3 we look at the spillover effects on parents. We distinguish between young voters who share 9 We inspected the voters residing in large households. A few, extremely large households drive the effect. Many of these are abnormalities or likely errors in the data. In the next iteration of the paper we will probably restrict the analyses to smaller household and thereby exclude households that seem incredible. As we have been awaiting data for a replication study from the European Parliament election we have not implemented this change, yet. 16

18 a household with their mother or father and young voters who do not live with their parents. For each of the groups we calculate separate direct treatment effects to take into account that the direct effect could differ in the various sub-samples. We also include spillovers to both parents. Table 3: Direct treatment effects and spillovers for parents conditioned on whether parents are part of household (percent). Direct effect Treatment group turnout Lives with both parent N=1,229 Control group turnout N=1,573 Lives mother, not father 48.2 N= N=546 Lives with father, not mother 42.0 N= N=293 Lives parents 64,66 N=17, ,443 Direct treatment effect 2.32 (1.88) 3.20 (3.09) (4.39) 1.40 (0.55) without Spillover effect on mother Spillover effect on father Spillover percentage on mother Spillover percentage on father 0.97 (1.38) (2.69) (3.97) 0.27 (0.43) 0.98 (1.41) 1.51 (2.95) (4.05) (0.44) Standard error in (). *p<0.05 (one-sided test). We test for significance with randomization inference with 100,000 iteration using the R-package ri (Aronow & Samii 2012). NOTE: Control group turnout for parents are around 80 percent. In the first column we depict the spillover effects on parents who both share a household with a person from the treatment or control group. We are somewhat limited by the small numbers in our population living with their parents. Even though the spillover effects do not reach statistical significance (p = 0.24 for both mothers and fathers) the spillover effects are remarkably similar; 17

19 around 40 percent of the original effect. Figure 1 visualizes this point. Moreover, the effects are remarkable given that there may be a ceiling effect as parents in general have very high turnout in the control group (about 10 percentage points higher than housemates in general). The results are theoretically interesting as they indicate that children can influence their parents and not just the other way around. In the final column, we look at spillover effects to parents where neither share household with a voter in the population. We find no evidence of spillover effects on close family members not sharing a household. The spillover effects are close to zero and precisely estimated due to the many voters not living at home. For completion, column 2 and 3 include voters who live with one parent and not the other. These are small groups and the estimates are too noisy to make any claims based on them. While the point estimate and confidence intervals for voters living with both their parents do not preclude that there could be a positive spillover effect, the results supports the idea that sharing a household is an important factor for the inter-personal influence on voter turnout even for close relatives. 18

20 Figure 1: Direct and spillover effect across households. Note: In figure A1 of the appendix we make the same analysis conditioned on gender of the treated. Discussion In this paper, we have presented evidence for substantial spillover effects from sending text messages to individuals on their households' turnout. This is in line with other recent studies that have documented spillover effects within households from door-to-door campaigns (Nickerson 2008), direct mailings (Sinclair 2012), and between friends on Facebook (Bond et al. 2012). Our study shows a direct treatment effect of 1.81 percentage points from sending a civic duty text message to a random sample of year old individuals in Denmark. Of even 19

21 more interest, the study documents a spillover effect on 1.43 percentage points for the entire sample that live together with the treated individual. The effect sizes are quite remarkable as the baseline turnout is high (control group turnout for the direct effect is percent and percent for the household spillover effect). The study's context is therefore a hard test for detecting direct treatment effects and spillover effects. Finding it here makes it more convincing that spillover effects from get-out-the-vote campaigns is not just a U.S.-phenomenon, but stretches to a broader much context. Furthermore, it is an relevant to highlight that the spillover effects are found using a new campaign tool compared to past spillover studies (cf. Sinclair et al. 2012). Spillover effects have now been detected using door-to-door canvassing, Facebook appeals, direct mailings and text messages, which is a wide variety of campaign tools. However, it should be noted that when taking into account household size we find spillover percentages of only about 30 percent of the direct treatment effect which is lower than the 60 percent found by Nickerson (2008). It could be that spillover effects are a general phenomenon but are larger when examining treatments that are more likely to be directly noticed by other household members. An important aspect of our study was to extend the research on spillover effects outside the household. This was possible by exploiting the links between individuals in the data from Statistics Denmark. The question was whether a text message to an individual could have an effect on the turnout for family members who do not share a household with the treated individuals. The results indicate that this do not seem to be the case. At least, there was no solid evidence of spillover effects on either mothers or fathers who do not share household with the treated individual, though the coefficient for mothers was in the expected direction. It could be an indication that sharing household is essential for spillover effects to set in though it should of course be noted that living together can also be a consequence of self-selection. One should also remember that the turnout for the control group in the spillover group was approximately 82 percent, which 20

22 makes it quite hard to increase turnout for this group. Future studies in lower turnout elections and with even higher sample sizes could explore the question of spillover effects outside households further. The results are interesting in their own right, but they are also an important reminder to campaigners. When spillover effects exists, GOTV-evaluations will underestimate the total effect of a treatment by ignoring that citizens interact with one another (Nickerson 2008). When there is a spillover effect from the treated to the household member, friends, neighbors, colleagues and so on, the total treatment effect will be higher than the one found by only studying the direct effect. In other words, getting one extra person to vote might start a "turnout cascade" (Fowler 2005). Not discovering such an effect has consequences as we risk underestimating the cost effectiveness of a given treatment. Furthermore, campaigners end up with a knowledge deficit, which is crucial when they need to decide whether they should use the same type of treatment in upcoming elections. Our study has one more important implication for campaign practitioners. Text messages are cheaper and easier to distribute to a large part of the population than most other mobilization tools, which makes it an appealing campaign form to non-partisan organizations. However, text messages have the distinct feature of going directly to the treated individual, whereas e.g. door-to-door campaigns and mailers are more likely than text messages to be noticed by other household members when they reach the targeted individuals (other household members are probably more likely to notice a person at the door or a letter in the common mailbox than a text to a personal mobile phone). One could expect that it would be difficult to create spillover for text messages for this reason, but this was not the case. Having demonstrated that there are spillover effects from text messages makes these an even more appealing campaign tool in future elections. 21

23 It is important to stress that we might not have detected the entire spillover effect from the text messages. In the current study, we are solely focusing on directly related family members and cohabitants. One could for instance imagine that the recipients forward text messages to friends or discussed among classmates. This could cause a larger spillover effect than we are able to detect. In fact, there is a risk that the messages have been forwarded to persons in the control group. If that happened the treatment and spillover effects could be underestimated in this study. Unfortunately, there is no way of actually detecting to which degree this has happened. It should of course be noted that this is a single study with quite large spillover effect sizes. There is a need for more research on spillover effects with variation in campaign tools and contexts. The results from this study are promising with respect to transferring the existing results to new context and new mobilization methods. Hopefully, researchers will in the future continue down the road of investigating spillover effects of GOTV-campaigns in order to be better able to grasp the total effect of these campaigns. 22

24 References Alford, J. R., Hatemi, P. K., Hibbing, J. R., Martin, N. G. & Eaves, L. J The politics of mate choice. The Journal of Politics, 73(02): Anderson, M Beyond Membership: A Sense of Community and Political Behavior. Political Behavior, 31(4): Aronow, P. M. & Samii, C ri: R package for performing randomization-based inference for experiments. Computer software manual). Retrieved from (R package, version 0.9). Bhatti, Y., Dahlgaard, J. O., Hansen, J. H. & Hansen, K. M. 2014a. The (lack of) effect of door-to-door canvassing in Denmark, in progress. Bhatti, Y., Dahlgaard, J. O., Hansen, J. H. & Hansen, K. M. 2014b. The effect of short text messages on turnout and the importance of timing, in progress. Bhatti, Y., Dahlgaard, J. O., Hansen, J. H. & Hansen, K. M. 2014c. Kan man øge valgdeltagelsen? Analyse af mobiliseringstiltag ved kommunalvalget den 19. november 2013 CVAP Working Paper Series, CVAP WP 3/2014, 2014(3): Bhatti, Y., Dahlgaard, J. O., Hansen, J. H. & Hansen, K. M. 2014d. Hvem stemte og hvem blev hjemme? Valgdeltagelsen ved kommunalvalget 19. november Beskrivende analyser af valgdeltagelsen baseret på registerdata. CVAP Working Paper Series, CVAP WP 2/ (2): Bhatti, Y. & Hansen, K. M Valgdeltagelsen ved kommunalvalget 17. november Beskrivende analyser af valgdeltagelsen baseret på registerdata. Arbejdspapir Københavns Universitet, Institut for Statskundskab, 2010(3). Bhatti, Y. & Hansen, K. M. 2012a. Leaving the Nest and the Social Act of Voting: Turnout among First-Time Voters. Journal of Elections, Public Opinion & Parties, 22(4): Bhatti, Y. & Hansen, K. M. 2012b. Retiring from Voting: Turnout among Senior Voters. Journal of Elections, Public Opinion & Parties, 22(4): Bhatti, Y. & Hansen, K. M The effect of co-residence on turnout. MPSA Annual Conference Chicago. Bond, R. M., Fariss, C. J., Jones, J. J., Kramer, A. D., Marlow, C., Settle, J. E. & Fowler, J. H A 61-millionperson experiment in social influence and political mobilization. Nature, 489: Campbell, A., Converse, P. E., Miller, W. E. & Stokes, D. E The American Voter, Chicago: The University of Chicago Press. Cho, W. K. T., Gimpel, J. G. & Dyck, J. J Residential Concentration, Political Socialization, and Voter Turnout. Journal of Politics, 68(1): Cutts, D. & Fieldhouse, E What small spatial scales are relevant as electoral contexts for individual voters? The importance of the household on turnout at the 2001 general election. American Journal of Political Science, 53(3): Dale, A. & Strauss, A Don't Forget to Vote: Text Message Reminders as a Mobilization Tool. American Journal of Political Science, 53(4): Fieldhouse, E. & Cutts, D Diversity, density and turnout: The effect of neighbourhood ethno-religious composition on voter turnout in Britain. Political Geography, 27(5): Fieldhouse, E. & Cutts, D The companion effect: household and local context and the turnout of young people. The Journal of Politics, 74(3): Fowler, J. H Turnout in a small world. pp In: Zuckerman, A. S. (ed.) The Social Logic Of Politics: Personal Networks As Contexts For Political Behavior Temple University Press, Philadelphia. Fowler, J. H., Baker, L. A. & Dawes, C. T Genetic Variation in Political Participation. American Political Science Review, 102(2): Gerber, A. S., Green, D. P. & Larimer, C. W Social Pressure and Voter Turnout: Evidence from a Large- Scale Field Experiment. American Political Science Review, 102(1):

25 Green, D. P., Mcgrath, M. C. & Aronow, P. M Field experiments and the study of voter turnout. Journal of Elections, Public Opinion & Parties, 23(1): Hobbs, W. R., Christakis, N. A. & Fowler, J. H Widowhood Effects in Voter Participation. American Journal of Political Science. Jennings, M. K., Stoker, L. & Bowers, J Politics across generations: Family transmission reexamined. The Journal of Politics, 71(3): Klofstad, C. A Talk Leads to Recruitment: How Discussions about Politics and Current Events Increase Civic Participation. Political Research Quarterly, 60(2): La Due Lake, R. & Huckfeldt, R Social Capital, Social Networks, and Political Participation. Political Psychology, 19(3): Lazarsfeld, P. F., Berelson, B. & Gaudet, H The People's Choice: How the Voter Makes Up His Mind in a Presidential Campaign: Columbia University Press. Malhotra, N., Michelson, M. R., Rogers, T. & Valenzuela, A. A Text messages as mobilization tools: The conditional effect of habitual voting and election salience. American Politics Research, 39(4): McPherson, M., Smith-Lovin, L. & Cook, J. M Birds of a feather: Homophily in social networks. Annual review of sociology: Nickerson, D. W Is Voting Contagious? Evidence from Two Field Experiments. American Political Science Review, 102(1): Plutzer, E Becoming a habitual voter: Inertia, resources, and growth in young adulthood. American political science review, 96(01): Sinclair, B., Mcconnell, M. & Green, D. P Detecting Spillover Effects: Design and Analysis of Multilevel Experiments. American Journal of Political Science, 56(4): Stoker, L. & Jennings, M. K Life-cycle transitions and political participation: The case of marriage. American political science review, 89(2): Wolfinger, R. E. & Rosenstone, S. J Who votes. New Haven, CT: Yale University Press. Zuckerman, A. S The Social Logic Of Politics: Personal Networks As Contexts For Political Behavior: Temple University Press. 24

26 Appendix Table A1: Descriptive statistics for the study population compared to all year olds. All year olds Study population N Mean Std.dev. N Mean Std.dev. Voted 520, , Age 520, , Male 520, , st or 2nd gen. immigrant 520, , Lives with father 513, , Lives with mother 517, , Household size 520, , Income in DKK 510, , ,933 47, , ,177 In education 520, , Table A2: Descriptive statistics for the treatment and the control group. Treatment group Control group N Mean Std.dev. N Mean Std.dev. Voted 26, , Age 26, , Male 26, , st or 2nd gen. immigrant 26, , Lives with father 26, , Lives with mother 26, , Household size 26, , Income in DKK 26, , ,712 20, , ,842 In education 26, , A logistic regression (with standard errors clustered at the household level) predicting treatment status with the variables above (except voted) is insignificant, p=