Knock Knock : Do personal and impersonal party campaigning activities increase voter turnout? Evidence from a UK-based partisan GOTV field experiment

Transcription

1 Knock Knock : Do personal and impersonal party campaigning activities increase voter turnout? Evidence from a UK-based partisan GOTV field experiment Joshua Townsley * Draft, August Keywords: Campaigns; GOTV; Field experiment; Turnout; Canvassing; Leaflets; Voting Abstract Experimental evidence from non-partisan Get Out The Vote (GOTV) studies suggest that faceto-face canvassing have powerful effects on turnout compared to impersonal methods such as leaflets. The decline in face-to-face party contact has even been linked to declining turnout. However, it remains unclear how far these findings apply to partisan campaign activities. There is also uncertainty around the generalisability of American GOTV experiments to a European context. This paper reports the results of a partisan field experiment conducted with the Liberal Democrats during the 2017 local elections in England, testing the effect of a) a leaflet and b) a canvass visit after a leaflet. I find that exposure to partisan campaigning increases turnout by 3.7 percentage points. In contrast to existing non-partisan experiments, I find that impersonal contact boosts turnout (by 4.5 percentage points), and that the additional effect of canvassing is minimal. Postal voters, meanwhile, are unaffected by treatment. This paper contributes to our understanding of partisan GOTV efforts, and adds to the emerging body of partisan GOTV studies in Europe that questions the generalisability of non-partisan, US-based GOTV studies to a European context. * PhD Candidate in Politics, Department of Politics and International Relations, University of Kent, j.townsley@kent.ac.uk,

2 1. Introduction There is an ongoing discussion over how much political participation is impacted by party campaign efforts, and which forms of campaign contact can increase voter turnout the most. Given declining levels of voter turnout across democratic countries (Lijphart, 1997), this is a particularly pertinent issue. The wealth of evidence gained from the use of field experiments have shed light on the effectiveness of various Get Out The Vote (GOTV) methods in terms of increasing participation. Based on non-partisan civic GOTV experiments, conducted largely in the United States, the consensus reigns that personal, face-to-face interactions are the most powerful way to mobilise voter turnout (Green and Gerber, 2008; Green et al, 2013). However, there are questions over the extent to which these findings generalise to a) a European context, and b) partisan campaigning. The majority of existing GOTV studies examine the impact of non-partisan campaign exposure, rather than partisan campaign efforts, and focus predominantly on the United States (Bhatti et al, 2016; 2017; Foos and de Rooij, 2017; John, 2017). This is an important distinction, as the majority of local campaigning activities in Europe are conducted by political parties (Nyman, 2017). To more clearly understand the impact of certain campaign activities on turnout, it is important to examine the partisan activities that voters are regularly exposed to at election time. This is a problem for field experiments, which are vulnerable to criticism from an external validity standpoint (Mutz, 2011). The geographical imbalance in the location of most GOTV experiments restricts our ability to generalise existing findings to other countries, and has led to calls for more field experiments in Europe (Nielsen, 2014; Bhatti et al, 2016). This study adds evidence to the emerging body of European GOTV studies that test the generalisability of US-based studies to a European, partisan context. This represents a clear gap in the existing literature that this study seeks to address. I report the results of a partisan field experiment conducted in the United Kingdom local elections in The study has two purposes. Firstly, to test whether partisan campaign contact stimulates political participation. Secondly, to examine whether the key findings of existing non-partisan GOTV research that personal contact has the most powerful effect on turnout generalises to partisan campaign activities. The experiment tests the impact on voter turnout of exposure to partisan leaflets (or

3 mailers in US terminology), and the effect that a face-to-face canvass visit has in addition to receiving a leaflet. By showing that that exposure to partisan campaigning even on behalf of a minor party locally boost turnout by around 4 percentage points, this paper contributes to our understanding of the effects of party campaign activity on voter turnout. I also find that while canvassing + a leaflet has a greater effect on voter turnout, exposure to just a leaflet can also boost participation by itself. These findings are important in that they question the extent to which results from non-partisan, US-based GOTV studies are applicable to European party campaigns. The second contribution of this study is that by replicating the experiment among postal voter households and non-postal voter households, I can compare the effect of party campaign contact between the two groups. This represents a novel contribution of this experiment, as postal voters are usually excluded from the experiment population altogether. To the best of my knowledge, my experiment represents one of the few GOTV studies to replicate treatment and compare effects between postal voters and those registered to vote in-person (see Arceneaux et al, 2012 for an aggregate-level experiment). In line with Arceneaux et al (2012), I find that campaign contact boosts turnout among in-person voters, but has negligible effect on postal voters. This is evidence that lower information and lower propensity voters are more likely to respond to information provided by partisan campaigns than high propensity voters. This suggests that, in accordance with previous research on who is most affected by GOTV efforts (Arecenaux and Nickerson, 2009), voters that are highly predisposed to turnout are less likely to be influenced by additional campaign efforts. The paper proceeds as follows. I first discuss existing research on the effects of GOTV campaigns, and the place of this study in relation to existing scholarly work. I then outline my theoretical expectations regarding why we would expect party campaigning to mobilise voter turnout at all, and my hypotheses. The following section then outlines the experiment s research design, describing the randomisation process and experimental setting. I then present and discuss the results and their implications. 2. Partisan and Nonpartisan GOTV Experiments

4 The extent to which various efforts by parties and campaigners can effectively increase political participation has long been a topic of great academic debate. While it is now largely agreed that local campaigns do have an effect on electoral outcomes, the topic of which campaigning methods boost turnout the most has been subject to intensive experimental research over recent decades. While not the first experiments used in this field (see for example Gosnell, 1927; Eldersveld, 1956; and Miller, Bositis and Baer, 1981), Gerber and Green initiated this surge of field experiments at the turn of the millennium (2000). Their first study found that non-partisan personal contact increased turnout by 8.7 percentage points compared to the control group (Gerber and Green, 2000: 658), while the effects of telephone calls and direct mail were minimal. The field of GOTV experiments has since burgeoned, with the effects of leaflets, letters, phone calls, canvassing, texts, and s all being tested using randomised control trials (for an up-to-date summary, see Gerber and Green, 2017). The majority of this now vast body of GOTV research strongly suggests that face-to-face campaign contact has the most powerful effects on voter turnout at election time. The effect of personal campaign contact trumps the impact of impersonal methods. Table 1 summarises the precision-weighted effect on turnout of the most commonly-used GOTV tactics based on Green et al s (2013) meta-analysis review of existing GOTV studies. The effect of face-to-face canvassing is around 2.5 percentage points, making it the most powerful GOTV tactic. As the tactics become less personal, their effect size gets smaller. In contrast to canvassing, leaflets and direct mail have only minimal impact on voter turnout. Table 1 Effect of Common GOTV Tactics (Green, McGrath and Aronow, 2013) GOTV Tactic 2 Effect (95% CIs) Based on No. of Studies Face-to-Face Canvassing (1.817, 3.255) 71 2 The effect of leaflets and direct mail are the Intent to Treat (ITT) effect, which represents the difference in turnout between those randomly assigned to treatment conditions compared to those assigned to the control conditions. The effects of faceto-face canvassing and phone calls are the Complier Average Causal Effects (CACE), which represents takes into account the fact that not all those assigned to receive certain forms of treatment are successfully treated. This is known as one-sided noncompliance among subjects, and CACE represents a common method used to interpret experimental treatment effects when this is an issue. It is calculated by dividing the ITT by the successful contact rate (Green et al, 2013).

5 Volunteer Phone Calls (1.298, 2.575) 37 Commercial Phone Calls (0.504, 1.456) 25 Leaflets/Direct Mail (0.078, 0.247) 110 conventional mailings 3 The number of GOTV field experiments that have been carried out in recent decades have established a clear consensus that personal campaign contact trumps all other GOTV tactics. Leading scholars in this field even suggest that the decline in voter turnout is linked to the decline in levels of face-to-face contact between parties and voters (Green and Gerber, 2000). However, the vast majority of studies that have formed this consensus are based in the United States (Bhatti et al, 2016). This has created a sharp imbalance in the literature between US and European based studies. It is not clear, therefore, to what extent the power of personal consensus generalises to other locations. There is, however, an emerging body of Europeanbased GOTV studies that are seeking to rectify this. Bhatti et al s (2016) meta-study of 9 canvassing experiments across six European countries summarises these new studies. They find that canvassing in Europe has a much lower impact than in American studies a precisionweighted effect of 0.78 percentage points, compared to in the United States. While they find canvassing to be less effective in Europe, the limited number of experiments conducted in Europe means there is still considerable uncertainty around the effects of European-based GOTV experiments. They call for further GOTV experiments to continue to redress the imbalance (Bhatti et al, 2016). The extent to which non-partisan GOTV findings translate to partisan campaigning is also unclear, as partisan studies are rare and thus far present mixed results. McNulty (2005) finds no turnout effect of partisan phone calls, Panagopoulos (2009) finds that party phone calls have the same impact as nonpartisan calls, while Nickerson et al (2006) find that partisan canvassing does boost turnout. Green et al (2013) conclude from their review of leaflets and direct mail that partisan mail has a negative effect (though statistically insignificant) effect on voter turnout, while non-partisan mail has a small positive effect. Recent partisan studies in 3 Conventional mailings refer to leaflets that do not contain any social pressure messaging, which are shown to have stronger effects (see Green et al, 2013 for a summary.

6 Europe also show mixed results. Some experiments appear to confirm the findings from nonpartisan GOTV studies that canvassing increases turnout. A field experiment in Sweden found that party canvassing increases turnout by 3.6 percentage points (Nyman, 2017), while Cantoni and Pons (2016) find an effect size of around 2 percentage points. Experiments that randomise both personal and impersonal partisan treatment at the aggregate level show no effect on voter turnout (Ramiro et al, 2012; Pons, 2014; Pons and Liegey, 2013). In their UK-based field experiment, Foos and John (2016) find that partisan leaflets and canvass visits do not boost overall turnout, though were equally effective at mobilising party supporters. Foos and de Rooij (2017) find that partisan telephone calls boost turnout by 2-3 percentage points. There is no clear consensus on the impact of partisan campaign activities, with especially few studies conducted in a UK or European context. Beyond a handful (Foos and John, 2016; Foos and de Rooij, 2017), the small number of GOTV experiments conducted in the UK, for instance, (John and Brannan, 2008; Cutts et al., 2009; Fieldhouse et al., 2003; 2012) have only tested the impact of nonpartisan interventions. The first contribution of this study, therefore, is to add further evidence to the growing body of partisan, European-based GOTV research. This leads to my main research questions does partisan campaigning increase voter turnout? Does face-to-face canvassing have a larger effect than impersonal leaflets? And do the effect sizes of contact differ among postal voters? 3. Theory and Hypotheses We know that GOTV campaigns can increase voter turnout, but why would we expect partisan campaign contact to increase turnout? First and foremost, exposure to partisan campaigning can provide easily-accessed information to voters about the election, the issues at stake, and the candidate (Aldrich, 1995; Arceneaux and Kolodny, 2009). According to Riker and Ordeshook s calculcus of voting (1968), the decision of whether to vote or not is a function of the probability that the vote is decisive (p), the material benefits of the individual s favoured party winning (B), the expressive benefits of casting one s vote (D), and the costs associated with voting (C), so that: Voting (V) = pb + D C.

7 People vote, therefore, when the costs associated with voting are smaller than the benefits (pb + D > C). Given that most individuals cognitive abilities with regards to political information are limited, they struggle to process political information in a reliable way (Mondak, 1993). Processing political information, especially at a low-saliency election, takes time and requires mental energy. As a result, most individuals tend to have relatively low levels of political information (Page and Shapiro, 1992). Exposure to partisan campaigning can address this by providing information to the voter about the election, a candidate, and the issues at stake, all to their doorstep (Lenz, 2009). This serves to raise the saliency of the election, and reduce the costs associated with gathering this information by one s self and thereby boost turnout (Rosenstone and Hansen, 1993). Given the low-saliency nature of this experimental setting, we would expect this information effect to boost turnout. My first hypothesis is, therefore: H1 Exposure to partisan campaign efforts will increase voter turnout. My expectation is that exposure to campaign contact will have a positive effect on voter turnout. To this end, as contact provides information to voters about the election and reduces costs, we might also expect the effects to be stronger among low-information voters who may face higher costs. I replicate the experimental treatment among postal voters and non-postal voters, in order to compare the effects between voters with a high predisposition to vote (postal voters) and those with a low predisposition (non-postal voters). As postal voters have made the extra effort to register to vote by post, their turnout is usually far higher than those who are registered to vote only in-person. For postal voters, the costs associated with voting are already far lower given that their ballot papers are delivered to their door, to be filled out and sent back at a time convenient for them. The provision of extra information about the election would likely have less of an impact on their propensity to vote, given that their costs are already low and they are starting from a much higher base of predisposition to vote. Inperson voters, meanwhile, would have a far lower predisposition to vote on election day as the costs of voting are higher with having to walk to the polling station on a particular day. As such, I expect that the effects of exposure to campaign contact to reduce some of the costs of voting will have a larger effect among non-postal voters. This expectation is also in line with

8 existing research into how the effects of GOTV efforts differ between different voters. Arceneaux and Nickerson (2009) find that GOTV campaigns are more effective among voters closer to their threshold of indifference between voting and abstaining. As postal voters are far more likely to turnout, they can be characterised as being further away from their indifference threshold even at low-saliency elections. My second hypothesis is, therefore: H2 the effect of campaign contact on voter turnout will be lower among postal voters than among non-postal voters. The most dominant finding of existing non-partisan GOTV research is that personal contact is a more powerful stimulant of political participation than impersonal contact (Green et al, 2013). But why is this the case? Green and Gerber argue that personal contact increases the social connectedness between voters and the voting process. Mobilising a voter, they argue, is akin to inviting them to a social occasion (Green and Gerber, 2004: 92), and that personal contact works because a campaigner s willingness to devote time and energy to talk personally to an individual signals the importance of participating in the electoral process and voting (2015: 38). Alternative explanations suggest that it is the noticability of a campaign intervention that drives its effect (Dale and Strauss, 2009). While originally applied to explain the surprising effect of text messages on turnout, this explanation can also readily be applied to canvassing. Having a party activist appear at your door serves as a very noticeable reminder that there is an election approaching and that one s vote is sought after. The mechanism that drives the effect of personal interventions is still unclear, however (Broockman and Kalla, 2016), but the bulk of non-partisan GOTV studies provide strong reason to expect canvassing to be more powerful than leaflets. The effect sizes of canvassing can be as high as almost 10 percentage points (Gerber and Green, 2000; Nickerson, 2008) in some cases. In their review of 71 canvassing studies, Green et al (2013) reveal a weighted average effect of closer to 2.5 percentage points. Leaflets meanwhile tend to show minimal effects. In a review of 31 advocacy mailings, Green et al find that mailings have an estimated effect of percentage points on voter turnout. While recent European experiments provide mixed evidence (Foos and John, 2016; Pons, 2016), the overwhelming bulk of literature gives us plenty of reason to expect canvassing to have stronger effects than impersonal contact. I anticipate, therefore, that this will also apply to the effect of partisan canvass visits. Therefore:

9 H3 Those exposed to a canvass visit in addition to a leaflet will vote at a higher rate than those exposed to a leaflet only. The paper will now move on to outlining the research design of my partisan field experiment through which I test these three hypotheses. 4. Research Design I conduct a randomised field experiment with the Liberal Democrats in the English local elections in Field experiments offer several methodological advantages to a survey or observational-based research design that justify its use. Firstly, experiments can identify causation and make causal claims more robustly than observational-based approaches (John, 2017). Random assignment controlled by the researcher overcomes the problem of selection bias that can exist in survey-based campaign studies (Green et al, 2013). The researcher can also control which subjects are exposed to treatment, allowing their effects to be more accurately measured, and manipulate the treatment, allowing me to compare the effects of canvass visits to receiving a leaflet only. The outcome variable in this case, voting can be more reliably measured with official voter records rather than with a survey, which can suffer from false recall or social desirability bias and the over-reporting of voting. Given the causal nature of my research questions, a field experiment provides the most suitable means through which to test my hypotheses. Finally, field experiments also benefit from being conducted in realistic conditions. During an election, this is particularly helpful, as a myriad of factors can shape voting behaviour on polling day. As opposed to a laboratory-based study, field experiments are conducted in real-life settings surrounded by the normal dynamics of an actual election. By conducting a field experiment in real-life conditions, internal validity is improved. 3.1 Study Population My study population comprised registered postal voters and non-postal voters in the ward (known as an electoral division for county councils) of Thedwastre North in Suffolk, England. Taken from the Liberal Democrats voter database prior to the beginning of the local campaign period, the study population totalled 6,525 registered voters. At the 2013 County Council

10 elections, turnout in Thedwastre North was 31.3%. This compares to a turnout of 31% across County Council elections nationwide on that day (Electoral Commission, 2013). This suggests that the area is not atypical of the rest of England in terms of its underlying level of turnout prior to any effect of this study s campaign exposure. The division is, however, dominated by the Conservative party, for whom the division is a safe seat Clustered Random Assignment All voters in the division were first clustered into their households. In order to compare the effect of treatment between postal voters and non-postal voters, households were then separated between those containing at least one postal voter, and those containing no postal voters at all 5. Each type of household was then randomly assigned to either a control group, or one of two treatment groups. This process is summarised in Figure 1 6. The first treatment group was assigned to receive an A3 Liberal Democrat leaflet, while the second group received the same A3 leaflet followed by a canvass visit from a volunteer. The households assigned to the control group were not exposed to any Liberal Democrat campaign treatment. The treatments were replicated between the households containing postal voters and those containing no postal voters. In each case, a random sample of households were exposed to either a) a leaflet, or b) a leaflet followed by a doorstep visit from a party activist. In essence, this random assignment is designed to examine what additional effect a canvass visit has on turnout on top of a regular party leaflet. 4 Figure A1 in the Appendix shows the location of Thedwastre North within Suffolk County Council, and Table A1 shows the result from the previous local election in Postal voter households are defined as such if they contain at least one postal voter. As such, some households contain a mix of registered postal voters and registered non-postal voters, and therefore, not all registered voters in the postal voter treatment arms are postal voters. A note of these numbers: within the canvass group, 396 of the 515 registered voters were postal voters; within the letter group, 410 of the 540 are postal voters; within the control group, 206 of the 269 individuals were postal voters. 6 Several villages were excluded from the experiment as they were subject to a pilot experiment prior to the experiment. To avoid any contamination of multiple treatments voters in these villages were removed from both the treatment and control groups (i.e. removed from the population altogether).

11 Figure 1 Field Experiment Research Design Process Registered Voters N =6,528 (3,295 households) Non-Postal Voter Households N =5,202 (2,659 households) Postal Voter Households N =1,323 (636 households) Canvass and Leaflet Leaflet Control Canvass and Leaflet Leaflet Control N=1,095 (580 households) N=1,132 (592 households) N =2,975 (1,487 households) N =515 (250 households) N =539 (250 households) N =269 (136 households) Random assignment is clustered at the household level rather than at the individual level ignoring the household. This is for both practical and methodological reasons. Namely, that when canvassing I have no control over who opens the door when I knock. If one individual in a household is assigned to the treatment condition while another in the same household is assigned to the control condition, it is impossible to ensure that only the former opens the door when I am canvassing. If the latter opens the door, they will be, to a limited extent, incorrectly exposed to campaign treatment. Similarly, leaflets that are delivered through the letterbox could be picked up, or seen, by anyone within the household. Who this will be exactly is beyond the researcher s control. There is strong evidence that campaign contact effects spill over to others in the household who are not themselves directly exposed to campaigning (Nickerson, 2008). As there is a need to avoid unintended exposure to campaign treatment among those assigned to control conditions, campaign contact is therefore administered at the household level.

12 This decision also has empirical ramifications regarding what it is that the experiment is measuring. Namely, does the cluster-randomised design examine the turnout effect of direct exposure to canvassing or leafleting treatment? Or does it measure the effect of living in a household in which at least one member of the household is exposed to campaign treatment? This design can only answer the second of these questions. However, this is based on research that shows that the effect of campaign exposure spills over to others in the household (see, for instance, Nickerson, 2008). This is supported by a body of research that shows that living with a voter increases one s probability of voting to 90 per cent, compared to 10 per cent for someone living with a non-voter (Cutts and Fieldhouse, 2009). As such, I am confident that the research design examines the effect on individual turnout of exposure to campaign treatment conducted at the household level. 3.3 Treatments All treatments were partisan in nature, and carried out by the researcher or a volunteer. Treatments were kept as authentic (i.e. as close to what a typical local campaign would do anyway), as possible. Several weeks were spent touring the Thedwastre North area with the Liberal Democrat candidate to collate photographs and stories to populate the leaflet. The leaflet was printed in full colour A3, and contained the legally-required party imprint, as well as Liberal Democrat branding. Figures 2 and 3 show images of the front and back of the leaflet delivered. Figure 2 Lib Dem Leaflet 1

13 Figure 3 Lib Dem Leaflet 2

14 As Figures 2 and 3 show, the leaflet contained numerous local policy issues and manifesto commitments covering libraries, transport, and schools, all of which are under the remit of the local County Council for which the election was being held. The leaflet also introduces the candidate, Jon James, and his local connections to the area, including several photographs of him. The leaflet was designed to provide colourful, eye-catching information about the candidate and the issues at stake at the election, in accordance with the theory that information reduces costs for voters. Postal voter households within the leaflet treatment

15 group received the leaflet between the 4 th and 14 th April, to coincide with the period just before they received their postal votes. Non-postal voter households received their treatment a week later, coinciding with the period leading up to polling day. In other words, the treatment for both postal and non-postal voters was carried out during the two-week period leading up to their being able to cast their ballots. It is possible that the differential timing of the treatment could mean that turnout-inducing events might impact one component of the experiment and not the other. However, given that both the postal voter and non-postal voter components of the experiment had their own control groups, and the experiment is designed to measure the difference in turnout between the treatment groups and their respective control groups, the effect of any such events would not affect the credibility of the experiment. Those households assigned to the leaflet + canvass visit group received their leaflets at the same time as those in the leaflet only group. However, these households were then visited by a party volunteer several days after their leaflets were delivered. The boxes below present the canvass scripts delivered to both postal voter (Box 1) and non-postal voter households (Box 2) 7. Box 1 - Canvass script for postal voter households Hello, my name is, I am a volunteer for the local Liberal Democrats. I m just calling round to remind you that the Suffolk County Council elections are coming up, and that it s important that you have your say in how our area is run. The local Lib Dem candidate, Jon James, will be a strong voice for our area, and will fight to improve local transport and health services. According to our records, you are registered for a Postal Vote. Your Postal Vote papers should be arriving over the next few days, so just keep an eye out for those, then you can vote, send it off, and you ve had your say. 7 Canvassers were instructed to give the script, but not to stick rigidly to the script and nothing more. They were encouraged to be as friendly and chatty as they liked. The reception among voters was relatively mixed, with many being happy to chat to the canvasser, and many wishing to end the interaction promptly.

16 Box 2 Canvass script for non-postal voter households Hello, my name is, I am a volunteer for the local Liberal Democrats. I m just calling round to remind you that the Suffolk County Council elections are coming up, and that it s important that you have your say in how our area is run. The local Lib Dem candidate, Jon James, will be a strong voice for our area, and will fight to improve local transport and health services. 3.4 Contact Rates and Noncompliance Table 2 summarises the final sample sizes and contact rates. All of the impersonal treatment was carried out successfully, as all those households randomly assigned to receive a leaflet did so. As such, the leaflet contact rate across the canvass and leaflet, and the leaflet only groups were 100%. In total, 500 households containing at least one postal voter received a leaflet. Likewise, all 1,172 non-postal voter households assigned to receive a leaflet during the campaign did so. However, one-sided noncompliance was an issue among those living in households assigned to receive a canvass visit in addition to the leaflet. It must be noted that a voter is considered to be successfully canvassed if they live in a household in which at least one person opened the door and spoke to a canvasser. In such cases, all individuals in the household are considered treated. Table 2 Sample size and contact rates by treatment group Registered Voters Of which were exposed to Leaflets Leaflet Contact Rate Canvass Contacts achieved Canvass Contact Rate Postal Voter Households Canvass + Leaflet % % Letter only % - - Control Non-Postal Voter Households Canvass + Leaflet 1,095 1, % % Leaflet only 1,132 1, % - - Control 2, Total N 6, postal voter households, and 580 non-postal voter households were assigned to receive a canvass visit in addition to the leaflet. The successful canvass contact rate was 28%. In most

17 cases, several canvass attempts were made, all within 3 days of the leaflet delivery. The successful canvass rate is slightly lower than other similar studies. For instance, Foos and John s UK-based experiment had a successful canvass rate of 40% (2016). This is likely to have been due to a number of factors. Much of the campaigning fell over the Easter period, when many people were away on holiday, which hindered our success rate. Also, much of the doorknocking was carried out by two men in their mid-20s, which may have contributed to many people, particularly elderly voters, opting not to open their doors to strangers. Finally, as a very rural area, the door-knocking campaign posed logistic issues, requiring significant travel times to small hamlets in order to door-knock only a handful of randomly-selected households. This is of course far less efficient than a campaign that covers every household on a street rather than a pre-determined random sample of them. Such factors made it difficult to secure a higher success rate for the doorstep visits. Nevertheless, 28% is not an altogether out of the ordinary contact rate for an experiment of this kind. Altogether, 3,248 individuals lived in households that were, through personal or impersonal means, contacted successfully by the campaign once (leaflet or letter only) or twice (those assigned to receive a canvass visit in addition to the letter or leaflet). This represents 52.5% of the experiment population. 3.6 Balance Test and Data Assumptions Randomisation ensures that observable and unobservable characteristics of the experimental subjects should be balanced between treatment and control groups (Gerber and Green, 2012). This is because all subjects have a known and equal probability of being assigned to treatment conditions. Table 3 presents the balance of available individual-level characteristics across the treatment and control groups in the non-postal voter and postal voter components of the experiment. This data was obtained through the local Liberal Democrats voter ID database. Table 3 Balance of Pre-Treatment Covariates between Treatment and Control Groups Postal Voter Experiment Control Leaflet Only Canvass + Leaflet Non-Postal Voter Experiment Control Leaflet Only Canvass + Leaflet Lib Dem (20) (45) (34) (332) (123) (111) Women (135) (284) (281) (1,537) (569) (556)

18 Voted in (131) (242) (230) (753) (264) (257) Single Occupancy 14.9 (40) 13.2 (71) 14.4 (74) 12.2 (362) 13.0 (147) 13.2 (145) Age (21) (57) (38) (229) (76) (84) Age (11) (14) (13) (77) (26) (23) Age Under (21) 6.7 (36) 8.0 (41) 7.7 (229) 8.5 (96) 5.8 (63) Voted in 2009 most recent turnout data available in Thedwastre North. While more recent turnout data would be better, the balance between the groups is crucial. As Table 3 shows, the covariates do not vary appreciably between assignment groups. To verify this statistically, I follow the procedure outlined by Gerber and Green to test whether imbalances are larger than one would expect from chance alone (2012 :109). I run a multinomial logistic regression model for both components of the experiment to test whether assignment to treatment groups is significantly related to available individual-level covariate data. The results of the regression are reported in Tables A2 and A3 in the Appendix, and show that all covariates are statistically insignificant predictors of assignment to treatment conditions, affirming the randomisation process. The pre-treatment covariate data is also used in the subsequent analysis, and includes sex, party support based on previous canvass analysis, ward, registered postal voters, previous turnout, proportion living in single occupancy households, and age group. The random allocation to treatment and control conditions was conducted by the researcher. The random nature of assignment underpins the causal claims the study is able to make (Gerber and Green, 2012). Throughout the paper, I also assume one-sided noncompliance, noninterference, and excludability. These assumptions will be briefly outlined below. Noncompliance occurs when not all of the subjects assigned to receive treatment were actually treated. This is an issue for the canvass treatment, which I take into account in the following section in which I present the results. I use the standard procedure for dealing with one-sided

19 noncompliance as outlined in Gerber and Green (2012). Crucially however, I assume that this noncompliance is one-sided, meaning that no subjects in the control group were treated. This is fulfilled in so far as I know that no leaflets were delivered to, nor canvass visits attempted on, houses that were assigned to the control condition. It would be impossible, therefore, for subjects living in households assigned to the control condition to be directly treated. Of course, it is possible that subjects living in neighbouring households may be exposed to treatment indirectly through interaction with their neighbours. The fact that this cannot be ruled out simply means that I can only claim that outcome effects are a result of living in a household exposed to treatment. I also assume non-interference, which means that every subject s probability of being assigned to treatment is unaffected by other subjects assignment. Excludability is also assumed, which means that the outcome is affected by treatment and not by assignment to treatment. The main threat to these assumptions is that treated subjects then discussed their treatment with other subjects, encouraging them to vote (Nyman, 2017). I reduce the extent to which this can happen by cluster randomising at the household level, so that there is no risk of subjects assigned to treatment living in the same house as subjects assigned to control conditions. Because people sharing a house together interact on a daily basis, this would violate the non-interference and excludability assumptions. 4. Analysis and Results I focus on one dependent variable: the percentage-point difference in voter turnout at the 2017 local election between those assigned to the control group and those assigned to each treatment group. The data was obtained from the local authority, who released official individual-level turnout records shortly after the election. These were then matched to my experimental records to calculate the rate of turnout across the control and treatment groups. The main results of the experiment are shown in Table 4. The table includes the results for the full sample, and for the postal voter households and non-postal voters separately. From left to right, the table shows the number of subjects and the turnout in the control group. The next columns show the estimated effects of exposure to (1) Lib Dem campaign, (2) a leaflet, and (3) a canvass visit after a leaflet. The estimated effects are in relation to the control group. The table presents the turnout rate of each group, the intent-to-treat (ITT) effect, which is calculated by comparing the turnout between the treatment group and the control group, and

20 the covariate-adjusted ITT effect. All of the models are estimated using regression in order to obtain robust standard errors clustered on households, which are presented in parentheses. This is important, as there is strong turnout correlation within households, and not correcting for this would deflate the standard errors substantially. If the assumptions discussed in the previous section hold, the differences in the turnout rates can be interpreted as the causal effect of the treatments. These are known as the Intent-To- Treat (ITT) effects, and are calculated by comparing the turnout rate (%) between those assigned to the control and treatment groups. By comparing the turnout difference between the groups, regardless of whether or not the entire treatment group was successfully treated, the ITT reveals the effect of being assigned to a treatment group compared to a control group, and avoids the spuriousness associated with a turnout comparison of those successfully contacted and those not contacted (Gerber and Green, 2000; 2015). Because we cannot know for sure whether all of those individuals living in a treated household actually read the leaflet, the ITT effect can be considered a conservative measure of the treatment effects of leaflets. The covariate-adjusted effects are calculated by controlling for individual-level variables in a multivariate regression framework. The estimates have the advantages of correcting for any minor imbalances in the assignment process, while also increasing the precision of the estimates. Adjusting for covariates still produces unbiased effect estimates (Gerber and Green, 2012: 109). I include in the models the covariates that were available from the local party voter database, including sex, party support (based on previous canvass analysis), ward, postal voter registration, previous turnout, whether the individual lives in a single occupancy household, and age group. The covariate-adjusted estimates produce near-identical results, which testifies to the quality of the randomisation process, with the only difference being that the covariateadjusted estimates are more precise. I focus on the covariate-adjusted estimates due to their greater level of precision, but I present both for robustness and transparency. In order to illustrate the effects visually, I also present the predicted marginal effects of the treatments in Figure The effects are covariate-adjusted.

21 Table 4 Experiment results (1) Lib Dem Campaign (2) Leaflet Only (3) Canvass + Leaflet N Control Turnout Turnout Effect (SE) Covariate adjusted (SE) Turnout Effect (SE) Covariateadjusted Effect (SE) Turnout Effect (SE) Covariateadjusted (SE) Full sample 6, * (1.6) 3.7* (1.5) (1.9) 2.8 (1.8) - 4.6* (1.9) 4.6* (1.8) Postal Voter Households 1, % 68.9% -1.3 (4.2) -2.6 (3.9) 66.9% -3.3 (4.7) -4.8 (4.3) 71.1% 0.8 (4.6) -0.3 (4.3) Non-postal voters 5, % 29.3% 4.4** (1.7) 4.8** (1.6) 28.8% 3.9+ (2.1) 4.5* (2.0) 29.9% 4.9* (2.1) 5.1* (2.0) Robust standard errors (in parentheses) clustered on households. All tests two-tailed. ***p<.001, **p<.01, *p<.05, +p< Effects of the Liberal Democrat Campaign All effect sizes that are statistically significant at conventional levels of at least 95% confidence are highlighted in bold for ease of interpretation. Looking first at Model 1, it is clear that subjects exposed to the Liberal Democrat campaign in any form were 3.7 percentage points more likely to vote, significant at the 95% level. The experiment clearly confirms my first hypothesis that exposure to partisan campaigning boosts voter turnout. In this case, across postal and non-postal voter households, I find that the Liberal Democrat campaign increased participation by 3.7 percentage points. Turning now to comparing the effect sizes of the replicated treatment among postal and nonpostal voter households. As outlined in the research design section, I split households within the Thedwastre North Electoral Division between those containing at least one postal voter, and those containing none at all, and replicated the treatment and control conditions among each in order to compare if postal voters react differently to treatment. While both sets of treatment were conducted at the same election, they were carried out at different times, with leaflets delivered and canvass visits being conducted around two weeks prior to the rest of the population. The differing timescales are on account of the fact that postal voters receive their ballot papers through the post several weeks before polling day. However, given the treatment is replicated across both, we can compare the effects.

22 The experiment clearly shows that postal voters reacted very differently to the Lib Dem campaign than non-postal voters. Among non-postal voters, the effect of exposure to the campaign (Model 1) was to boost turnout by 4.8 percentage points, significant at the 99% level. Unsurprisingly, turnout was substantially higher among postal voter households. Among these voters, the campaign had a negative, but statistically insignificant, effect on voter turnout. This provides strong evidence confirming my second hypothesis that high-propensity voters who are more likely to turnout anyway are less influenced by additional campaign exposure. Postal voters turnout at a high rate, and the potential for further increases as a result of exposure to party campaigning is lower. Figure 4 Covariate-Adjusted Predicted ITT Effects of Campaign Contact on Voter Turnout, with 95% Confidence Intervals 4.2 Effects of the Personal and Impersonal Treatments Finally, we can examine the effect of the two treatments on voter turnout. Model 2 shows the effect of the leaflet compared to the control group, while Model 3 shows the effect of the canvass visit and leaflet treatment combination. Across all voters, the leaflet increased turnout

23 by 2.8 percentage points, though this is statistically insignificant at conventional levels. The canvass visit and leaflet combination boosted turnout by 4.6 percentage points, significant at the 95% level. This appears to suggest that while canvass visits have strong effects on voter turnout, leaflets are ineffective. However, looking at non-postal voters only, it is clear that the leaflet did have an effect on turnout. Among non-postal voters, the leaflet increased turnout by 4.5 percentage points compared to the control group, while the canvass visit + leaflet increased turnout by 5.1 percentage points, both significant at the 95% level. Again, among postal voter households each treatment had negative and statistically insignificant effects. Across the pooled groups, the difference in turnout between those assigned to canvassing + treatment and those assigned to leaflets only is 1.78%, representing an intent to treat effect of 1.78% for canvassing, though this is statistically insignificant. However, the ITT effect can underestimate the extent of the effect of exposure to the treatment due to noncompliance among the canvass group. The ITT does not take into account the successful contact rate among those assigned to receive a canvass visit. I follow the procedure laid out in Gerber and Green (2012) and utilise a second measure, the Complier Average Causal Effect (CACE). The CACE takes the successful contact (or complier ) rate among the canvass group into account by dividing the ITT (the difference in turnout between the canvass + leaflet and leaflet only groups) by the proportion of the canvass group successfully reached 9. In so doing, it estimates 9 CACE is laid out at length in Gerber and Green, Consider the following model: Y = a + trx + u Y represents the dichotomous outcome indicating whether an individual voted or not (0,1). The coefficient tr is the effect of treatment, while X indicates whether an individual was successfully contacted by the canvasser or not. Given that whether or not an individual was actually contacted (X) is a function of whether or not the individual was assigned to the canvass treatment group, we can introduce the dichotomous Z variable that indicates whether an individual is assigned to the group. Cr denotes the contact rate. Thus: X = crz + e. Estimating the treatment effect (tr), given the contact rate (cr), we can divide the intent-to-treat (ITT) effect by the contact rate (cr), to give us the Complier Average Causal Effect (CACE). Thus: ITT / cr = CACE. This measure is a standard assessment of the effect sizes in GOTV field experiments given issues of noncompliance (i.e. not everyone in the treatment group(s) accepting their assigned treatment). Gerber and Green (2000) note that this is comparable to running a two-stage least squares regression of turnout on campaign contact, with randomisation as an instrumental variable (Gerber and Green, 2012).

24 the effect of exposure to treatment if all those assigned to receive treatment were successfully treated. The ITT of canvassing in comparison to the leaflet only group is 1.78%. However, the successful contact rate for canvassing was 27.7%. As such, the Complier Average Causal Effect (CACE) of canvassing in comparison to the leaflet only group is 6.44%, with a 95% confidence interval of to As the confidence interval captures 0, it is clear that the additional effect of a canvass visit compared to leaflets is not statistically significant. This provides mixed evidence for my third hypothesis that the effect of personal contact would be higher than the effect of impersonal contact. On the one hand, across the full sample the only treatment that is statistically significant is that of the canvass visit in addition to the leaflet. The effect of the leaflet by itself falls short of conventional levels of statistical significance. But breaking the sample down and looking at non-postal voters only shows that both treatments have strong effects on voter turnout. Discussion The effect of exposure to the Liberal Democrat campaign, increasing turnout by 3.7 percentage points, is striking. This is especially so given several factors. Firstly, the strength of the contact was relatively small. The campaign only included one leaflet, albeit a large and colourful leaflet. The canvass visits were also carried out by relatively inexperienced canvassers, as the more seasoned campaigners were campaigning in nearby areas that were more winnable for the party. This relates to the second factor the position of the Liberal Democrats locally. Despite holding the neighbouring division of Thedwastre South, the party is a relatively minor player in Thedwastre North. At the previous election, the party won 4.43% of the vote, while the Conservatives polled almost half of the vote. This makes the effect of the campaign all the more remarkable, given that it was not tapping into a significant base of underlying Liberal Democrat support. It may be that given its relative safety, the division does not usually witness intense party campaigning, making the efforts of my Liberal Democrat campaign more noticeable. As far as I am aware, two other parties had notable campaign efforts in Thedwastre North the Conservatives and the Green Party. Both parties delivered (to me knowledge) one

25 leaflet to most households in the area (see A4 and A5 in the appendix for photographs of these leaflets). I do not have an account of which households the opposing parties targeted. But, given the random assignment, the households exposed to other parties campaigns had an equal chance of being exposed to additional campaign efforts from the Liberal Democrats. The 3.7 percentage point turnout effect can therefore be interpreted as the effect of Lib Dem campaigning only in addition to any campaigning households were exposed to from other parties. It may be that the Lib Dem campaign had such an effect on aggregate turnout by mobilising existing Liberal Democrat party supporters. This is entirely plausible, given existing research showing that party campaigning mobilises supporters and demobilises opposition supporters (Foos and John, 2016). However, given that the Liberal Democrat vote share only increased from 4.4% to 6.6% in this election, it seems unlikely that the 3.7 percentage point effect can be entirely attributed to mobilising Lib Dem supporters. Exposure to the party s campaign did not necessarily, and dramatically, boost the party s position at the polls. This suggests that exposure may have boosted turnout among supporters of rival parties, in contrast to findings from Foos and John s (2016) experiment in the south west of England. Interaction effects with pre-treatment party support data were conducted, but given limited existing canvass data, meaningful comparisons were impossible. The effect of exposure to a party s campaign contact on voter turnout provides evidence that at the individual level, exposure to campaign information does indeed reduce the costs associated with voting and increases turnout even among supporters of opposition parties. The comparative effect of party contact among own supporters and supporters of rival parties is worthy of further experimental research. As per my expectations, the effect of campaign exposure was limited only to those who had a relatively low predisposition to vote, namely, non-postal voters. While postal voters turned out at a rate of 70%, non-postal voters turned out at a rate of 25%. Such voters have low underlying propensity to vote at local elections, and face higher costs associated with casting a ballot on the day of voting itself. When costs are high, it seems that exposure to party campaign contact has an effect at boosting turnout. Postal voters, meanwhile, already vote at far higher rates and were unaffected by additional campaign efforts. Campaigning appears to have strong mobilising effects among voters with a low predisposition to turnout. This is in line with

26 Nyman s (2017) experiment, who found that party canvass visits in Sweden boosted turnout by 3.6 percentage points, driven largely by effect estimates for occasional and first-time voters. It may be that low turnout voters, for various reasons, face higher costs associated with voting, meaning that attempts to reduce such costs reap greater turnout rewards. This suggests that mobilising efforts might reap greater benefits from targeting non-postal voters than they would from postal voters. Or, similarly, that they should dedicate resources to recruiting postal voters. While this study represents one of the few to compare effects between postal and in-person voters, the findings are in line with existing research. Arceneaux et al s aggregate-level experiment in the United States (2012) found that vote-by-mail (i.e. postal voting) precincts were less impacted by GOTV canvassing efforts than precincts using traditional in-person voting. However, the sample size of postal voters, who only comprise of approximately 20% of registered electors, was small. Despite their small numbers, postal voters represent a significant proportion of total votes cast at local elections (Rallings and Thrasher, 2014). Given this, parties invest significant resources into securing their votes. While this study adds to the limited investigation into this, further research into how the effectiveness of GOTV campaigns varies depending on how voters cast their ballots is needed. While exposure to the campaign boosted turnout, I was particularly interested in the relative GOTV effects of personal contact in addition to impersonal methods. The impact of leaflets alone fell short of conventional levels of statistical significance across both experimental components. However, among non-postal voters, the leaflet along boosted turnout by 4.4 percentage points. This effect is particularly large compared to other partisan experiments (e.g. Green et al, 2003; Cardy, 2005). While a canvass visit in addition to the leaflet had a larger effect, and one that was statistically significant across the full experiment sample, the difference between the canvass + leaflet and the leaflet by itself was not significantly different. Even after accounting for the one-sided noncompliance by calculating the CACE, there is no significant difference between the personal and impersonal campaign treatments. Though, it is likely that with a higher successful canvass rate, the effect would be statistically significant. Also, because I did not test the impact of canvassing by itself (i.e. without a leaflet preceding it), I cannot directly compare the two, and it is therefore unclear whether leaflets are doing all of the work of the effect. Nevertheless, among the non-postal voter households a leaflet was

27 enough to significantly and substantially increase turnout, while the canvass visit added only a little to that. The existing consensus on GOTV experiments suggests that leaflets are ineffective at boosting turnout especially partisan leaflets compared to canvassing. Impersonal methods such as leaflets and direct mail tend to be ineffective at increasing turnout according to most studies (Green et al, 2013). The findings from this experiment, therefore, represent an important qualifier on this consensus partisan leaflets, when presenting bright and eye-catching information, can substantially boost turnout. The fact that this experiment detects a significant effect provides strong evidence that the findings of non-partisan GOTV studies do not generalise perfectly to our understanding of partisan campaigning. This represents an important contribution of this study, which questions the generalisability of the extant body of GOTV literature to partisan, European-based campaigning. This finding is also consistent with the emerging body of European-based GOTV studies, which tend to find weaker non-partisan canvassing effects in Europe (Bhatti et al, 2016). The results are also consistent with Foos and John s UK-based partisan experiment (2016), which finds no difference between canvass and leaflets with regard to increasing turnout of party supporters. It seems that, in the UK at least, the effect of a partisan canvass visit is not as powerful as in the United States. Leaflets, by themselves, can provide significant boosts to turnout. Acknowledgements Big thanks go to Lesley Paris, Oliver Townsley, Ashley Vardey, and Tom Parkington, for their significant efforts helping me carry out the assigned treatment door-knocking and leafleting through rain and shine. I am also enormously grateful to Jon James, for standing as the Thedwastre North Lib Dem candidate for the purposes of my experiment and for his invaluable advice and guidance. Thank you as well to the rest of the Mid Suffolk Liberal Democrats for their support throughout this study, especially Penny Otton, Martin Redbond, Frank Wright, and John Field. I would also like to thank my PhD supervisory team, Dr Ben Seyd and Dr Laura Sudulich at the University of Kent. Special thanks go to Professor David Cutts and Dr Florian Foos for their

28 advice on the experiment and the paper. I would also like to thank the European Field Experiment Summer School teaching team of Dr Alexander Coppock, Dr Eline de Rooij, and Dr Florian Foos for their fantastic course. References Aldrich, John H Why Parties? The Origin and Transformation of Party Politics in America. University of Chicago Press. Arceneaux, Kevin, and Kolodny, Robin Educating the least informed: group indorsements in a grassroots campaign. American Journal of Political Science. 53(4): Arceneaux, Kevin, and David Nickerson Who Is Mobilized to Vote? A Re-Analysis of 11 Field Experiments. American Journal of Political Science. 53(1) :1-16. Arceneaux, Kevin, Thad Kousser, and Megan Mullin Get Out The Vote-by-Mail? A randomized field experiment testing the effect of mobilization in traditional and Vote-by-Mail precincts. Political Research Quarterly. 65(4): Bedolla, Lisa Garcia and Michele Michelson Mobilizing Inclusion: Transforming the Electorate through Get- Out-The-Vote Campaigns. Yale University Press. Bhatti, Yosef., Jens Olav Dahlgaard, Jonas Hansen Hedegaard, and Kasper M Hansen Is Door-to-Door Canvassing Effective in Europe? Evidence from a Meta-Study across Six European Countries. British Journal of Political Science. Online early access. Bhatti, Yosef., Jens Olav Dahlgaard, Jonas Hansen Hedegaard, and Kasper M Hansen. M Can governments use Get Out The Vote letters to solve Europe s turnout crisis? Evidence from a field experiment. West European Politics. Doi: / Broockman, David E., Joshua Kalla Durably reducing transphobia: A field experiment on door-to-door canvassing. Science. 352(6282): Cantoni, Enrico, and Vincent Pons Do Interactions With Candidates Increase Voter Support and Participation? Experimental Evidence from Italy, Working Paper No Cambridge, MA: Harvard Business School.

29 Cutts, David, and Edward Fieldhouse 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, Allison, and Aaron Strauss Don t Forget to Vote: Text Message Reminders as a Mobilization Tool. American Journal of Political Science. 53(4): Electoral Commission (2013). Local Elections May The Electoral Commission, available: Fieldhouse, Edward, David Cutts, Paul Widdop, and Peter John Do Impersonal Mobilisation Methods Work? Evidence from a Nationwide Get-out-the-Vote Experiment in England. Electoral Studies. 32(1): Fieldhouse, Edward, and David Cutts Shared Partisanship, Household Norms and Turnout: Testing a Relational Theory of Electoral Participation. British Journal of Political Science. Foos, Florian Eline de Rooij The role of partisan cues in voter mobilization campaigns: Evidence from a randomized field experiment. 45: Foos, F. and Peter John Parties are no civic charities: Voter contact and the changing partisan composition of the electorate. Political Science Research and Methods. Doi: doi: /psrm Gerber, Alan S Does Campaign Spending Work?: Field Experiments Provide Evidence and Suggest New Theory. American Behavioral Scientist 47(5): Gerber, A. S., and Donald P Green The Effects of Canvassing, Telephone Calls, and Direct Mail on Voter Turnout: A Field Experiment. The American Political Science Review. 94(3): Gerber, Alan S. and Donald P. Green Field Experiments: Design, Analysis and Interpretation. New York: WW Norton. Gerber, A. S., and Donald P Green Field Experiments on Voter Mobilization: An Overview of a Burgeoning Literature. in A. V. Benerjee and E. Du o (eds.), Handbook of Economic Field Experiments. Oxford: Elsevier, Green, Donald, and Alan Gerber Get Out the Vote! How to Increase Voter Turnout. Washington, DC: Brookings Institution Press.

30 Green, Donald, and Alan Gerber Get Out the Vote: How to Increase Voter Turnout. 2nd ed ed. Washington, D.C.: Brookings Institution Press. Green, Donald, Mary C. McGrath, and P. M. Aronow Field Experiments and the Study of Voter Turnout. Journal of Elections, Public Opinion & Parties, 23(1): John, Peter Fields Experiments on Political Behaviour. OXFORD RESEARCH ENCYCLOPEDIA, POLITICS (politics.oxfordre.com). Oxford University Press USA, John, Peter. and Tessa Brannan How different are telephoning and canvassing results from a Get Out The Vote field experiment in the British 2005 general election? British Journal of Political Science. 38(3): Lenz, Gabriel Learning and opinion change, not priming: Reconsidering the priming hypothesis. American Journal of Political Science. 53(4): Lijphart, Arend, Unequal participation: democracy's unresolved dilemma. American Political Science Review. 91(1): McNulty, John E Phone-Based GOTV--What s on the Line? Field Experiments with Varied Partisan Components, The ANNALS of the American Academy of Political and Social Science 601(1): Mondak, Jeffery J., Public opinion and heuristic processing of source cues. Polit. Behav. 15(2): Mutz, Diana C., Population-based Survey Experiments. Princeton University Press, Princeton. Nickerson, David Is Voting Contagious? Evidence from Two Field Experiments. American Political Science Review, 102: Nickerson, David W., Ryan D. Friedrichs, and David C. King Partisan Mobilization Campaigns in the Field: Results from a Statewide Turnout Experiment in Michigan. Political Research Quarterly 59(1): Nielsen, Julie Hassing, Why use experiments in EU studies? Comp. Eur. Polit Nyman, Par Door-to-door canvassing in the European elections: Evidence from a Swedish field experiment. Electoral Studies. 45, pp Panagopoulos, Costas Partisan and Nonpartisan Message Content and Voter Mobilization: Field Experimental Evidence. Political Research Quarterly 62(1):

31 Page, Benjamin I., Shapiro, Robert Y., The Rational Public: Fifty Years of Trends in Americans' Policy Preferences. University of Chicago Press, Chicago. Pons, Vincent. Will a Five-Minute Discussion Change Your Mind? A Countrywide Experiment on Voter Choice in France. Harvard Business School Working Paper, No , January (Revise and resubmit requested, American Economic Review.) Pons, Vincent, and Guillaume Liegey Increasing the Electoral Participation of Immigrants. Experimental Evidence from France. Working Paper: Rallings, Colin and Michael Thrasher Local Elections in England May The Electoral Commission, Electoral-data-report.pdf Ramiro, Luis, Laura Morales and Maria Jiménez-Buedo Effects of Party Mobilization on Electoral Results. an Experimental Study of the 2011 Spanish Local Election. IPSA Working paper. Riker, William H. and Peter C Ordeshook A Theory of the Calculus of Voting. The American Political Science Review. 62(1): Rallings, Colin and Michael Thrasher Local Elections in England, May Electoral Commission. Available: available: data/assets/pdf_file/0003/175062/local-elections data/assets/pdf_file/0003/175062/local-elections-2014-electoraldata-report.pdf. Rosenstone, Steven and John Mark Hansen Mobilization, Participation, and Democracy in America. MacMillan:New York.

32 Appendix Figure A1 Thedwastre North Division (highlighted in light blue) within Suffolk County Council Table A1 Suffolk County Council Thedwastre North Division Result, 2013 Party Votes % votes Conservative UKIP Green Labour Lib Dem Total 2, % turnout

33 Balance Test As per Gerber and Green (2012), I show that almost all pre-treatment covariates do not differ appreciably between assignment groups. In order to statistically check whether imbalances are larger than one would expect from chance alone, I run a regression of the assigned treatment on all covariates, and calculate the F statistic. The Likelihood Ratio (LR) Chi-Square test and its accompanying p value for each model are reported in the tables below. As the randomisation process was carried out separately for the postal voter and non-postal voter components of the experiment, the regressions are presented separately below. Table A2 shows the results in the non-postal voter experiment. Table A3 shows the results in the postal voter experiment. The tables show that the only significant predictor of assignment to the Canvass + Leaflet group is being aged under 35 within the non-postal voter part of the experiment. As per Gerber and Green (2012), I present analyses controlling for all covariates, thereby producing unbiased estimates (2012: 109). Table A2 Non-Postal Voter Experiment Assignment, Balance of Pre-Treatment Covariates assignedgroup Coef. P>z 4 (base outcome) 5 LibDem woman votedin singleoccupancy pvhousehold 0 age age ageunder _cons LibDem woman votedin singleoccupancy pvhousehold 0 age age ageunder _cons

34 N 5,204 LR chi Prob > chi Table A3 Postal Voter Experiment Assignment, Balance of Pre-Treatment Covariates assignedgroup Coef. P>z 1 LibDem woman votedin singleoccupancy pvhousehold 0 age age ageunder _cons (base outcome) 3 LibDem woman votedin singleoccupancy pvhousehold 0 age age ageunder _cons N 1,324 LR chi Prob > chi

35 Table A4 Full Results of Covariate-Adjusted Treatment Effects NPV1 PV1 Pooled1 NPV2 PV2 Pooled2 Lib Dem Campaign 0.048** * (.016) (.039) (.015) Leaflet Only 0.045* (.020) (.043) (.018) Canvass Visit + Leaflet 0.051* * Ward Reference = Badwell Ash (.020) (.043) (.018) Elmswell & Norton *** *** *** *** *** *** (.018) (.039) (.016) (.018) (.038) (.016) Woolpit (.042) (.035) (.041) (.035) Woman (.008) (.021) (.008) (.008) (.020) (.008) Voted in *** 0.146*** 0.236*** 0.269*** 0.150*** 0.245*** (.019) (.031) (.008) (.019) (.030) (.016) Single Occupancy * Party Support (.018) (.036) (.016) (.018) (.036) (.016) Reference = Rival Party Lib Dem * (.039) (.096) (.037) (.039) (.097) (.037) Unknown (.039) (.078) (.031) (.033) (.079) (.031) PV Household *** 0.302*** *** (.022) (.037) (.022) Age *** *** 0.107*** *** (.029) (.050) (.022) (.028) (.050) (.025) Age (.045) (.080) (.040) (.045) (.079) (.040) Age Under (.023) (.055) (.022) (.023) (.054) (.022) constant 0.289*** 0.854*** 0.332*** 0.289*** 0.862*** 0.331*** (.036) (.089) (.034) (.037) (.086) (.034) N (Households) 5,202 (2,697) 1,323 (676) 6,525 (3,371) Clustered standard errors in parentheses. ***P<.001, **p<.01, *p<.05. 5,202 (2,697) 1,323 (676) 6,525 (3,371)

36 Figure A4 Other Party Leaflets Encountered During Campaign (Conservative)

37

38

39 Figure A5 Other Party Leaflets Encountered During Campaign (Green)

40