The Longitudinal Impact of a GOTV Campaign: a Local Election Follow up to a Constituency

s voting habit forming? The longitudinal impact of a GOTV campaign in the UK

David Cutts, Edward Fieldhouse and Peter John

University of Manchester

Abstract

This paper examines whether voters mobilised by a GOTV campaign continue to turn out in a subsequent election. It thereby tests the extent to which they gain the habit of voting (Gerber et al 2003b). The research examined electoral registers in the 2006 English local elections in the constituency of Wythenshawe and Sale East. This was the site of John and Brannan’s (2008) GOTV experiment, which had compared the impact of telephoning and door-to-door canvassing in the 2005 General Election. Descriptive statistics show that the downstream effect of the treatment was quite small at three per cent in 2006. But the habit effect was large. Voting in 2005 raised the probability of voting in 2006 by more than half. Regression analysis confirms the size of the habit effect. The paper discusses the limitations of sample size in making statistical inferences about the prevalence of the habit effect in one location.

Introduction

Voting is a habit, or so it is commonly assumed (e.g. Verba and Nie, 1972: 148; Franklin 2004: 203-205). Consider Plutzer’s statement that there is a “longstanding agreement that voting behaviour is habitual” (2002, 42). But, with some significant exceptions (Green and Shachar, 2000; Gerber, et al, 2003b; Plutzer 2002, Fowler, 2006), there has been little research on the phenomenon. Moreover, behavioural models of turnout question whether habit exists at all (Bendor, Diermeier and Ting, 2003). Individuals instead engage in casual voting. However, an alternative behavioural model which assumes a different method of reinforcement and inhibition challenges these findings (Fowler, 2006). The conclusion to draw from the current state of knowledge is that habitual voting is highly plausible, but there is just not enough evidence for it, especially outside the USA. So in this paper we begin to plug the gap by analysing data from a two-stage randomised field experiment. The results provide new evidence of the habit of voting in the British context.

There are three reasons why individual voting is likely to correlated across elections (Gerber et al, 2003b). First, voter turnout is influenced by factors that tend to remain stable over time, such as civic duty, strong partisanship and interest in politics (Campbell et al, 1960; Finkel, 1985). Second, parties manage and target campaign effort in specific areas - as well as to particular groups within the social structure - to enhance mobilisation (Huckfeldt and Sprague, 1992; 1995). Voters who have a longstanding record of participation are more likely to be contacted and targeted than non-voters. These same people will be mobilised each election, whose vote then correlates more across elections than had there been no campaign. Third, the act of voting itself may be habit forming. The more often people vote, the more likely they regard going to the polling station as “what a person like me does on Election Day”. Voting may reinforce the belief people have that they are civic–minded and fully engaged. According to this account, there is an underlying disposition to vote, but the act of voting gives expression to this preference. The underlying beliefs in civic action are periodically renewed each time an election takes place. And once they locked into being civic, so inertia sets in. Even without trying very hard, they differ in their behaviour to their non-civic counterparts. Inertia, however, is not symmetrical. It has much less of an effect on non-voting (Plutzer, 2002). Psychological studies of “foot in the door” techniques seem to reinforce this as current participation makes future participation more likely (Freedman and Fraser, 1966; Cialdini, Trost and Newsom, 1995). Getting people to repeat their behaviour is based on strengthening self-perception. This is the idea that individuals often determine their attitudes from observing their own behaviour (Beamer et al 1983, Burger and Caldwell 2003). A complementary perspective is that voting reduces the subsequent cost of voting. For once voters have been to the polls, they gain local knowledge, such as the location of the polling booth, which they retain for use in subsequent elections. And due to the asymmetry of inertia, gains in participation are harder to reverse than they were to acquire (Plutzer 2002: 43).

As a result of these psychological and cost factors, we expect change in the vote, such that induced by a Get Out the Vote (GOTV) campaign, to have a lasting effect. However, there remains the possibility that the increased propensity to vote might be temporary. After the initial interest wears off, many citizens may return to their long-established patterns of behavior. This would suggest habit is weak. Mobilisation may either have an effect for one election only or there is not much of a spill-over from one to the other.

However, before such a hypothesis can be substantiated, it is important to ascertain whether voting in one election affects the probability of voting in a subsequent one. Whilst a small number of experimental analyses have been carried out in the U.S., no experimental design has been used to detect habitual voting formation in other contexts. By applying a longitudinal panel design to a GOTV experiment we are able to directly measure habit or downstream effects for a panel of voters (Gerber et al. 2003b). By artificially stimulating voting in one election we are able to gauge the effect of habit by observing whether there is any downstream increase in the propensity to vote in the experimental treatment group. Here we build on the first randomized GOTV experiment in the UK, conducted by John and Brannan (2008). Previously, GOTV experiments have mainly been carried out in the United States (Arceneaux, 2005; Arceneaux and Nickerson 2006; Bennion, 2005; Gerber and Green, 2000a; 2000b; 2001; 2005; Gerber et al 2003a; Green, 2004; Green et al 2003; Green and Gerber, 2004; Green and Shachar, 2005, McNulty, 2005; Nickerson, 2006; 2007a; 2007b; Ramirez, 2005; Trivedi, 2005; Wong, 2005). But they have recently been implemented in other contexts (John and Brannan, 2008; Guan and Green, 2006).

The Manchester 2005 General Election GOTV campaign study combined two interventions in the same sampling design. This allowed for a direct comparison between the two methods of stimulating the vote. In this study, we recorded the voting behaviour of the same sample of electors that had been the subject of the GOTV experiment twelve months earlier. This two-stage experimental design offers a more robust assessment of the habit of voting than can be achieved using multivariate regression techniques. It is the first such study outside of the U.S (Gerber et al, 2003b). We derive a two-stage instrumental variable regression model to determine the effect of habit on voting. The results indicate a moderately significant habit effect after controlling for the underlying propensity of individuals to vote and a range of socio-economic variables, despite the limitations of a small sample size. The findings suggest that there may be long-lasting effects of voter mobilisation among these voters.

Experimental Design

The Original 2005 GOTV Experiment

The original experiment took place in the Wythenshawe and Sale East parliamentary constituency at the UK 2005 General Election. In the months before, John and Brannan (2008) randomly selected a sample of telephone accessible persons on the electoral register. A total of 6,900 individuals were drawn from this pool and then randomised into three groups of 2,300. One group received a visit from the GOTV field-force encouraging them to vote. The other got the same message from telephone call from a survey company. The third group was the control, which was left alone (see John and Brannan, 2008 for more details). Following a register check, 1,549 individuals could not be found on the electoral register (name or property was absent from electoral records). For the 2005 experiment, this decreased the randomly assigned sample size to 5,340 individuals, of which 1,764 were randomly assigned to the canvass group; 1,781 to the telephone group; and 1,795 to the control group. John and Brannan (2008) also had to remove postal voters as it was not possible to examine how they voted, as well as non-eligible voters that appeared in the sample.[1] This impacted on the sample size of the treatment and control groups. It may have contributed to the relatively borderline statistical effect of the original results.

The issue of sample size becomes critical in the detection of habit. The likelihood of a reduction in the effect of the intervention over time may limit the ability to make a statistical inference. Furthermore, it is only possible to detect a habit effect amongst the additional number of electors who were persuaded to vote in the original experiment. Even a large habit effect may not be statistically significant without a very large sample and a large treatment effect in the original study. As John and Brannan (2008) show, both interventions in the 2005 experiment raised turnout by 6.7 per cent for canvassing and 7.3 per cent for telephoning. Both interventions were statistically significant at respectively p= .035 and p=.038. The original results showed that the more impersonal telephoning had about the same impact as the more personal door-to-door canvass. Whereas the UK results replicated the magnitude of the face-to-face form of mobilisation found in the US studies, the telephoning results were about twice the average effect size of across the Atlantic.

So do both interventions have a long lasting effect? The argument is that voting should continue into the future because the newly mobilised voters gain the habit of voting. This persists irrespective of the type of mobilisation (Gerber et al 2003b). Here we provide a robust assessment of the habit of voting by examining whether the mobilised Wythenshawe and Sale East voters continue to vote on average more than those in the control group. Further to Gerber et al’s research, we show that the estimated habit effect is consistent across different types of treatment – telephoning and canvassing.

The 2006 Follow-Up Study: Descriptive Findings

This research is based on an examination of marked electoral registers of the same sample of voters from the 2006 local elections, which happened almost exactly a year later than the General Election. The key question is whether voters mobilized in 2005 continued to be more likely to vote than the control group in 2006. It is important to acknowledge that local elections are a different form of electoral contest than a General Election. Average turnout is lower in local contests because of their second order nature. However, we are primarily interested in the relative difference in treatment effects (which determines the effect of habit (δ) as shown in figure 1 later). By moving to a lower turnout election, we are shifting the overall level of voting. Yet the relative size of the effect should not be affected.[2] But the numbers reduce because of the need to filter out new postal voters, new cases of ineligibility, people who have deceased, and those that have moved or have removed their names from the register. Following the 2006 electoral register check, the overall sample size was 3,419 individuals, of which 1,104 had been randomly assigned to the canvass group, 1,158 to the telephone group, and 1,157 to the control.

Before the 2005 GOTV campaign, we can assume that those in the control and treatment group have equal underlying (or unmeasured) propensities to vote (by virtue of randomisation).[3] Following the interventions (canvassing or telephone), we expect that those electors in the treatment group were more likely to vote than electors in the control group. But the original study showed that the GOTV campaign persuaded at least some of those electors with a relatively low propensity to vote to participate in 2005. At the 2006 local elections we might expect these electors to revert back to non-voting, if, indeed, the impact of the GOTV campaign was temporary. Alternatively, did those who started out with a lower propensity to vote get into the habit of voting, following the GOTV campaign? It is possible that the GOTV campaign only had a temporary effect, in which case we might expect those electors in the treatment group who voted in 2005 to be less inclined to vote again in 2006. We might also expect that this decline would be higher in the treatment groups (both interventions) than in the control group. The null hypothesis for the experiment is that, following the GOTV intervention, voting in 2005 had no effect on turnout in 2006.

As noted earlier, both personal and telephone canvassing had statistically significant effects on voter turnout in 2005 (John and Brannan, 2008). Further descriptive evidence of this finding is shown in Table 1. In 2005, turnout among the control group was 51.5 per cent as compared to 55.0 per cent and 55.1 per cent in the canvass and telephone groups respectively. This 3.5 percentage point gap understates the effect of both personal and telephone canvassing as only 53.7 per cent and 47.7 per cent respectively, were actually contacted. When recalculated to take account of the actual contact rate, we derive the statistically significant 6.7 percentage point personal canvassing effect, and the 7.3 percentage point telephone effect (John and Brannan, 2008). While both personal canvassing and telephone canvassing increased voter turnout in 2005, the question is whether voters were more likely to return to the polling booths in the local elections exactly one year on. As expected, turnout was much lower in the 2006 local election. In our sample, 31.3 per cent of those in our sample voted in 2006. Yet Table 1 shows that turnout rates were higher in both treatment groups than in the control group. While 29.8 per cent of the control group voted in 2006, this figure increased to 31.4 per cent for those who received personal canvassing in 2005, and 32.7 per cent who were in the telephone group.

Table 1 about here

Table 2 shows a more detailed analysis of the effects of personal canvassing on 2006 voter turnout. After taking account of those in this treatment group that were actually contacted (contact rate is 54.4 per cent), we replicate the Green et al’s (2000) method in order to calculate the turnout differential for the canvass group. This is the difference between the percentage voting in the canvass group and the percentage voting in the contact group. Thus, from the 2006 results, we can produce an estimated effect of personal contact on voter turnout by dividing the intent to treat estimate by the contact rate. This is reported below in Table 2. Even though the 2005 canvass intervention had a positive effect on 2006 voter turnout of three percent, the difference was not statistically significant. The size of the habit effect (δ) can be estimated as the ratio of the treatment effect in 2006 to that in 2005 (3.0 compared to 6.7 in the original study), giving an estimated habit effect of 0.45 (see Gerber et al, 2003b). This means, in other things being equal, voting in 2005 increased the probability of voting in 2006 by almost 50 per cent.

Table 2 about here

We use a similar procedure to estimate the effect of telephone canvass intervention in 2005 on voter turnout in 2006 (see table 3 below). As described above, the treatment effect is 5.9 per cent, which is higher than personal canvassing and just outside the .05 significance level. Our initial findings therefore suggest that electors in the treatment groups (both interventions) were more likely to vote in both elections than those in the control group. Voting overall fell across the board at the local election of 2006 as might be expected. However, there is a larger drop off in the two treatment groups than in the control group. This infers that, for some people, the GOTV experiment only had a temporary effect. However, despite the passage of time, there was still a sizeable (albeit marginally significant) treatment effect in 2006. This suggests some people developed the habit of voting. The habit effect estimated from the telephone intervention is rather larger at δ = 0.81. However, we should bear in mind that the treatment effects these estimates of δ are based on have large sampling errors and in the case of 2006 are not significant at a 95 per cent confidence level. That said, this does suggest a rather large habit effect, but re-emphasises the point that with a moderately sized sample and initial treatment effect, quite large estimates of δ may not reach statistical significance. We also pooled the data to assess the overall treatment and habit effect on the combined sample (including the canvass and telephone treatment groups). The overall treatment effect for 2006 was 4.4 per cent and it was just outside the 0.05 significance level (0.08). The overall habit effect (δ) is fairly large at 0.63.

Table 3 about here

Up to this point, we have only examined the relationship between the dependent variable and both personal and telephone canvassing. However, our specific aim is to examine the effects of voting in 2005 on electoral participation in 2006. In order to achieve this goal, we need to provide a reliable estimate of the habit effect using the treatments in a model that takes account of unobserved heterogeneity in voting behaviour.

Modelling the Habit Effect