What Voters Want: Reactions to Candidate Characteristics in a Survey Experiment
There has been extensive research into the extent to which voters utilise short cuts based on gender and race stereotypes when evaluating candidates, but relatively little is known about how they respond to other background characteristics. We compare the impact of candidates' sex, religion, age, education, occupation and location/residence through a survey experiment in which respondents rate two candidates based on short biographies. We find small differences in the ratings of candidates in response to sex, religion, age and education cues but more sizeable effects are apparent for the candidate's occupation and place of residence. Even once we introduce a control for political party into our experimental scenarios the effect of candidate's place of residence continues to have a sizeable impact on candidate evaluations. Our research suggests that students of electoral behaviour should pay attention to a wider range of candidate cues.
We know relatively little about what socio-demographic characteristics voters value in election candidates – and the extent to which short cuts based on stereotypes matter when it comes to the way candidates are viewed by voters. The literature on candidate effects is large, but it is also partial and geographically skewed. There is a voluminous and sophisticated literature looking at some types of candidate characteristic, of which by far the most common are biological sex and race. But other characteristics are much less studied, and the majority of the literature draws on data from one country, the United States.1
Traditionally, and for good reasons, electoral studies in countries such as the United Kingdom were especially dismissive about the importance of anything that occurred below the level of the national campaign. Elections were seen as national events, in which national campaigns produced nationwide vote swings. A combination of an electorate divided along class lines (Butler and Stokes, 1974) and an electoral system that did not create strong incentives to cultivate a personal vote (Carey and Shugart, 1995) led to talk of local or candidate effects being dismissed as a failure to understand psephological reality. But with the decline in partisan and class alignment (Crewe et al., 1977; Mughan, 2009; Särlvik and Crewe, 1983) and increasing evidence of variations in constituency behaviour, various local or candidate characteristics are coming to be seen as more important.2 Indeed, not only are British elections becoming increasingly localised – with the potential for candidate effects to have a greater impact than in the past – but also the issue of candidate ‘representativeness’ has become a politically live one, with all the major political parties making efforts to improve the diversity of their candidates (Cowley, 2013), with the potential for electoral contests to feature a greater heterogeneity of candidate types in future.
Just as in the US, however, most of the work on candidate characteristics in the UK has focused on a relatively narrow range of characteristics, with sex and race being the most studied. Women candidates in Britain generally now face little discrimination from voters, although selectorates are often not as progressive in their behaviour (Norris and Lovenduski, 1995; Shepherd-Robinson and Lovenduski, 2002). Research has shown that some of the US findings, where in some cases women do seem to be more likely to vote for women candidates, do not apply to the British case (Campbell and Cutts, 2009); and some experimental research indicates that male candidates may be more popular with Conservative voters (Johns and Shephard, 2008). There are also a number of studies that have looked at race/ethnicity and candidate preference in Britain (Fisher et al., 2011; Norris et al., 1992; Saggar, 1998).
One surprisingly understudied topic is that of residency. Despite a growing focus on the local campaign (Denver and Hands, 1997; Denver et al., 2002; Fisher et al., 2007), a widespread acknowledgement within British political parties that being a local candidate can be an asset, some evidence that the number of MPs with local roots is on the increase (Childs and Cowley, 2011; Johnson and Rosenblatt, 2007) and that voters say they value local MPs (Cowley, 2013; Johnson and Rosenblatt, 2007), there has been relatively little examination of whether this can be an electoral asset (Arzheimer and Evans, 2012).3 Other candidate characteristics – such as their education, their occupation, their age and so on – are even more infrequently, if ever, discussed.
Moreover, even those candidate characteristics about which we have some knowledge are rarely compared to others, to measure their relative impact on voters. Even the highly developed literature on the influence of candidate race and sex in the US, for example, does not compare the size of its effects to other significant biographical information about candidates. The analysis in this article therefore reports findings from a survey experiment testing the impact of candidate characteristics in a low-information context in Britain. The experiments examine the relative importance of six different types of candidate cue. These include some of electoral studies' hardy perennials – such as biological sex – but we deliberately examine the effect of a wider range of cues than in extant studies.4 The experiments appear to reveal that many of the social information variables that academics have researched are in fact relatively unimportant, while also revealing sizeable differences triggered by background factors that have previously been largely ignored by researchers.
Experimental methods are becoming increasingly popular in research examining candidate evaluations (Birch and Allen, 2011; Druckman et al., 2006; Huddy and Terkildsen, 1993; Rosenberg and McCafferty, 1987; Sanbonmatsu, 2002). They allow us to separate out confounding effects much more neatly than is possible using ordinary survey methods (Mutz, 2011). Moreover, although some observable candidate characteristics – such as biological sex – can be relatively easily coded up, others (such as ‘localness’ or race) are much more difficult to code definitively. In addition, the fog of electoral war makes it very difficult to know what information voters have actually received about candidates. Survey experiments thus offer the opportunity to model hypothetical races, giving us insights into the priorities of voters that are not possible with conventional survey or observational data.
The creation of electronic survey companies has also made population survey experiments noticeably more affordable and manageable. This is particularly useful when using Britain as a case study. Electoral constituencies in British general elections are relatively small when compared to those in US federal elections and surveys are almost never conducted with sufficient sample size from each constituency to model individual races or to compare them effectively. Furthermore, research has shown few statistically significant differences between models of political choice estimated using reputable electronic survey companies and traditional face-to-face methods (Clarke et al., 2008; Sanders et al., 2007).
We used a survey experiment to create a low-information environment where respondents had to compare two candidates across three traits and choose which one they would prefer to be their representative. The electoral context was initially pared back to one where biographical information about the candidate was the only material available to respondents. We initially sought to give each characteristic the maximum chance of having an impact on preference without introducing another layer of complexity by interacting with political party. As Diana Mutz notes, ‘Needless complexity seldom makes for better experimental research’ (Mutz, 2011, 125). We then introduced political party in the final wave of the experiment.
The core of the study consisted of six split-sample internet surveys. Each survey involved respondents reading two short profiles about hypothetical candidates, and then answering four questions about those candidates. Following Kira Sanbonmatsu, our research design included profiles of two candidates (Sanbonmatsu, 2002), whom we (initially) called John and George:
John Burns is 48 years old, and was born and brought up in your area, before going to university to study for a degree in physics. After university John trained as an accountant, and set up a company ten years ago; it now employs seven people. John has interests in the health service, the environment, and pensions, and is married with three children.
George Mountford is 45 years old; he lives in the constituency and studied business at university. He is a solicitor and runs a busy local practice. George is passionate about education, with two children in local schools and a wife who is a primary school teacher.
John and George are plausible election candidates in a British election; they are both middle-aged men, in professional occupations, and although we alter these profiles throughout the experiment, the biographies remain those of plausible candidates.
Wave 1 of the survey was an attempt to measure the impact of gender on the perceptions of candidates, but also served to establish our baseline comparison for the later surveys. Half of the respondents read the two profiles given above. (In this, and all other waves of the experiment, the order in which respondents received the two profiles was randomised). The other half of respondents saw the same profiles but with John's name changed to Sarah (‘Sarah Burns is 48 years old …’). No other changes were made, the profile remaining in every other way identical. In subsequent waves, we made similarly small changes to one of the profiles in order to test attitudes to other cues. Details of the profile wordings used can be found in the Appendix.
There are obviously a very large number of cues that we could test for, each with an equally large number of variations that could be examined. We selected a mixture of usual suspects from political science – sex and religion – along with four other characteristics that have been more rarely studied: age, occupation, location and education. Election literature in Britain is dominated by the names and pictures of candidates, which enable voters to have a good idea about a candidate's sex and (approximate) age, and to make inferences about their race and in some cases religion.5 In addition, the apparent professionalisation of politics and the narrowing range of previous career experience held by representatives has become part of a familiar critique of politics in Britain and elsewhere (Carnes, 2012; Jun, 2003; Riddell, 1993; Ruostetsaari, 2003; Rush, 1994), with a sense evident in public discourse that the political class has become a self-perpetuating institution that lives off politics rather than for it. We therefore wanted to test whether voters did in fact react differently to candidates based on their occupational background. We selected residency because candidates made frequent mention of it in their election literature, yet it was largely absent from academic research. And we selected education because it is clearly an area where parliament is hugely unrepresentative of the population. Many writers on representation would have no problem with this – as education is often used as a measure of the ‘quality’ of candidates – but we know very little about how voters respond to educational attainment and it may be that they treat it as a proxy for social class. The age of candidates is frequently used in election campaigns to promote or undermine candidates; older candidates are alternately portrayed as frail and vulnerable or wise and experienced and younger candidates are described as either ‘novice’ or dynamic and forward-looking. Again we know relatively little about how voters respond to the age of candidates.
Crucially, election candidates frequently provide information about themselves of the sort given in our hypothetical profiles. To give some illustrative examples, all taken from campaign material from the 2010 British general election: ‘Jack lives in Filton with his wife Lucy, who is a nursery nurse, and their three children all of whom have gone to local South Gloucestershire schools’ (Conservative, Filton and Bradley Stoke); ‘There's something special about representing the area where I was born, grew up, went to school, worked and live … As the only candidate born and bred here I will proudly continue to live and work locally’ (Labour, Nottingham North); ‘I am your local candidate with experience. I live in Goffs Oak with my wife, Fiona, and our three children’ (Conservative, Broxbourne); ‘When I am elected as your Member of Parliament I will make my home in Wavertree’ (Labour, Liverpool Wavertree); ‘As well as nearly five years as a teacher, Nick has spent four years working for pension companies’ (Conservative, Cambridge); ‘Steve has lived in Oxford for over 20 years. He met his wife Julia here and his children, Laura (9) and Alexander (6), were both born in the JR and attend local schools’ (Lib Dem, Oxford East).
In selecting the precise phrasing, we deliberately chose fairly unsubtle cues (on the basis that if effects cannot be found with unsubtle cues they will probably not be found with more nuanced phrasing), while also attempting to avoid extreme cases or caricatures. When choosing to examine the effect of occupation, for example, we examine the impact of a candidate being a local doctor (a profession that enjoys high levels of trust in Britain) compared to having a background as an aspirant politico (a profession that does not).
There are also a very large number of candidate traits that can be examined: the literature on candidate effects is littered with different potentially important traits, including (but not limited to): ‘competence’, ‘experience’, ‘strength’, ‘leadership ability’, ‘effectiveness’, ‘integrity’, ‘honesty’, ‘morality’, ‘trustworthiness’, ‘compassion’, ‘warmth’, ‘approachability’ and ‘likeableness’ (Bartels, 2002; Johns and Shephard, 2008; McDermot, 1998; Miller and Shanks, 1996; Miller et al., 1966; Peterson, 2005; Rosenberg and McCafferty, 1987). We examined the impact of cues on three candidate traits: approachability, experience and effectiveness. Approachability was selected both to tap into feelings of ‘compassion’ (commonly associated with femininity and which we expected could trigger gender stereotypes where they exist) and to be useful when thinking about local versus distant candidates or political versus ‘caring’ professionals; approachability evokes an element of commonality or shared understanding between voter and candidate rather than simply suggesting an agreeable glow emanating from one to the other. We selected experience as a measure of competence rather than strength or leadership ability based on the same logic. Strength is associated with masculinity and would most likely yield gender effects, but when comparing local or distant candidates or political versus ‘caring’ professions strength is unlikely to be a particularly effective trigger. Finally we chose effectiveness because it is a measure of competence which looks at potential outcomes that should be the most important to voters.6
Respondents were asked to compare the hypothetical candidates across these three traits and then to select a preferred candidate:
Without knowing which party they stand for, which of them do you think would be:
More approachable as an MP: John/Neither/George
More experienced as an MP: John/Neither/George
More effective as an MP: John/Neither/George
Which would you prefer as your MP: John/Neither/George.
When designing the research, one of our concerns was that respondents might simply select down the line – preferring the same candidate for everything – and prove unable to distinguish between them on these different traits. The pilot study showed that this was unlikely to be the case and this was then confirmed in the main survey.7
Table 1 shows the respondents' views when shown the initial comparison of John and George, as detailed above. John was preferred to George on all three traits (by a consistent 16 percentage points) and as the preferred candidate (by 21 points). There was also a significant minority of respondents who were unable to choose between the two profiles (ranging from 28 per cent to 39 per cent, depending on the question). Given the research design, one concern we had was that without the heuristic of a party label not enough people would be able to make a choice between the candidates; yet in every case a majority of voters were able to choose, with John emerging as the preferred candidate. We do not know why John was the preferred candidate but given the research design it does not matter. This John/George comparison is used as the baseline against which we measure future small changes in the profile; what matters therefore in what follows is not the absolute standing of the two imaginary candidates, but the relative change from this baseline.
Table 1. John vs. George Initial Comparisons (row %)
|N = 1,444|| || || || |
The first change we made to the profiles was to examine the effect of biological sex, by changing John's name to Sarah but with the rest of the profile remaining constant. There were some obvious differences between how respondents saw Sarah compared to John, despite the fact that all that had been changed was the name serving as a cue for candidate gender. Sarah was seen as noticeably more approachable than John; instead of the 16-percentage-point lead in approachability enjoyed by John, Sarah led George by 28 points. There was, however, also a noticeable drop in how experienced she was perceived to be; whereas John was seen as more experienced than George (by 16 points), Sarah was seen as less experienced than George (trailing by 3 percentage points). Both of these differences were statistically significant. There were also minor but statistically insignificant changes to perceptions of effectiveness and to the preferred candidate. This produced the net effects when compared to the baseline John/George trade-off shown in the first section of Table 2.
Table 2. Net Experimental Effects
|N = 2,864|| || || || |
|Religion|| || || || |
| ||John||George/Daniel/ Mohamed||Neither||John lead over Daniel/Mohamed|
|George becomes Daniel|| || || || |
|N = 2,810|| || || || |
|George becomes Mohamed|| || || || |
|N = 2,777|| || || || |
|Occupation|| || || || |
| ||John||George||Neither||John lead over George|
|George becomes a GP|| || || || |
|N = 2,710|| || || || |
|George becomes a politico|| || || || |
|N = 2,754|| || || || |
|Age|| || || || |
| ||John||George||Neither||John lead over George|
|George becomes younger|| || || || |
|N = 2,874|| || || || |
|George becomes older|| || || || |
|N = 2,874|| || || || |
|Residency|| || || || |
| ||John||George||Neither||John lead over George|
|George arrived two years ago|| || || || |
|N = 2,879|| || || || |
|George lives 120 miles away|| || || || |
|N = 2,856|| || || || |
|Education|| || || || |
| ||John||George||Neither||John lead over George|
|George left school at 18|| || || || |
|N = 2,793|| || || || |
|George has a PhD|| || || || |
|N = 2,760|| || || || |
|George left school at 16||John||George||Neither||John lead over George|
|N = 2,327|| || || || |
Overall, then, changing the candidate's sex – and nothing else – generated a 12-percentage-point increase in their lead on approachability, and a 19-point decrease in their lead on experience but had no statistically significant impact of sex on the candidate's perceived effectiveness or preference for the candidate. In common with extant studies, both in the UK and elsewhere, this appears to confirm the belief that biological sex now has a negligible influence on choice of candidate (Sigelman and Welch, 1984; Sigelman et al., 1995; Trent et al., 2001), but also that gender stereotypes can influence candidate trait evaluations (Johns and Shephard, 2008; King and Matland, 2003; Smith and Fox, 2001). This gives us confidence in the nature of our experiment.
The second wave of the experiment tested changes in the apparent religion of the candidates: first, we examined the effect of a candidate having an apparently Jewish name (‘Daniel Goldstein’), and second that of having an apparently Muslim name (‘Mohamed Lafi’). Again, nothing else in the profiles was altered. As the second section of Table 2 reveals, the effect of making one of our imaginary candidates Jewish was relatively small. There were some differences with the baseline (John's lead increased when George Mountford was instead called Daniel Goldstein on two of the three candidate traits, as well as the preferred candidate), but for the most part these differences were small and statistically insignificant; the only exception was approachability where John's lead extended by a statistically significant 6 percentage points. The religion of candidates mattered more when one of the candidates appeared to be Muslim. This effect, though, was more complicated than might have been expected. John increased his relative lead over George/Mohamed on all three candidate traits (and by a statistically significant amount) but this was as a result of the ratings of both John and Mohamed falling but at a differential rate, with the percentages of respondents preferring neither candidate increasing on every question (by between 7 and 10 points). As will be clear from what follows, this did not happen with any of the other experiments we tested. There was also no change in John's lead as the preferred candidate, with both his ratings and that of Mohamed falling by 5 percentage points, and the proportion saying they preferred neither candidate rising by 10 points (also statistically significant). The effect was that in this case, but in no other set of profiles we tested, the plurality response for the preferred candidate was neither of the profiles. Our suspicion is that in this case some voters had identified, consciously or subconsciously, the presence of a ‘right’ or socially desirable answer; fearing that they might give the ‘wrong’ answer, an increasing number chose neither candidate.
Our third wave tested differences in candidates' preferences and ratings when we altered their occupations. The ‘original’ George was a solicitor, who had studied business at university. In the third wave we changed his occupation to, first, that of a local doctor – known in the UK as a general practitioner, or GP – and then second, a politico. The third section of Table 2 shows that both profile changes produced sizeable differences in the way the candidates were perceived by respondents. The effect of making George a GP was to increase all of his ratings relative to John; the increase in perceived experience and effectiveness was minor (although statistically significant) but there were sizeable (and statistically significant) differences in his perceived approachability and in respondents' overall preference. In the original comparison, John had led George the solicitor by 16 percentage points in terms of approachability; the effect of changing George to a GP was to put George ahead by 15 points, a massive 31-point change in the overall position. There were effects of a similar scale in terms of respondents' preferred candidate: in the original comparison John was preferred to George by 21 points; George the GP, however, became the preferred candidate of respondents with a lead of 6 points, a 27-point change in the overall standing of the two candidates.
Making George a politico had a huge effect on how voters perceived his experience relative to that of John, but less dramatic changes elsewhere. The original John had led by 16 points on experience; but making George have a political background was enough to make him perceived as easily the most experienced of the two candidates, with a massive 49-point change in the overall position. Other changes were less dramatic, although all were statistically significant. In other words, respondents clearly recognised the extra experience that a candidate who had a background in politics would enjoy, but they did not then especially reward it when it came to deciding which candidate they preferred overall.
Something similar was seen in the fourth wave, when we altered George's age. The original George was 45 years of age. To test respondents' judgement about younger candidates, we changed his age to 32 in one sample; to test what would happen if George was older, we changed his age to 60 in the second sample. Both produced very large and statistically significant differences in how respondents perceived the candidates' levels of experience. Making George aged 32 extended John's lead over him on experience by a further 25 points (and it increased his lead on effectiveness by 10 points); making George aged 60 had the opposite effect, reducing John's lead on experience by 26 points, with George perceived as the more experienced of the two candidates. But just as with George the politico, there was little impact in terms of respondents' overall preferences (although, again, these differences were statistically significant). Respondents did, therefore, judge the candidates differently in terms of their experience (and in a way that was intuitively reasonable: perceiving a 32-year-old as less experienced than a 45-year-old, and a 45-year-old as less experienced than a 60-year-old), but they did not then see it as important when determining which candidate they preferred overall.
The fifth wave altered George's residential status, as a means of measuring how important the local nature of the candidate was to respondents. The original George had been identified as living in the constituency. In wave five we changed that, first, to his having moved into the constituency two years previously, and then to living some 120 miles away, but being prepared to move to the constituency if elected. Both had the effect of noticeably increasing John's relative appeal. Even though the original profile had not noted when George had moved into the seat (merely that he lived there), explicitly identifying that he had moved there recently had the effect of increasing John's relative approachability by some 15 points, as well as his experience and effectiveness (by 5 and 8 points, respectively); it also increased John's position as the preferred candidate by some 12 points (all statistically significant effects). This 12-point effect on John's status as the preferred candidate – caused by merely noting that someone had moved into the constituency two years before – is greater than all of those previously discussed, apart from when we made one of the candidates a GP.
The effect was even more dramatic when we made George live outside the constituency. This saw John's lead extend by anywhere from 5 points (on experience) to 33 points (on approachability), and produced a 30-point increase in his position as the preferred candidate. These are huge differences, all statistically significant, and the largest of any of the tests we carried out. It is worth recalling that John's original lead on approachability and effectiveness was some 16 points. Making George live outside the constituency therefore was sufficient to increase John's lead on approachability threefold with his lead on effectiveness doubling. Similarly, John's baseline lead as the preferred candidate was 21 points; changing the profiles so that George lived outside the constituency had the effect of more than doubling John's lead.
The final main wave of the study examined the effect of changes in educational status. For reasons explained in the Appendix, altering a candidate's educational background required more consequential changes to the profiles than any of the other profile changes. The original George had gone to university; in wave six we changed George to, first, having left school at eighteen (in other words, before university, and with no mention in his profile of a later university degree); and then, second, to having gained a PhD. Changing levels of education did produce differences in respondents' verdicts on the candidates, although perhaps not in the expected direction. The less educated version of George was seen as less experienced (John's lead increasing by some 17 points), but in every other way he was seen as a better candidate than the university-educated version. He reduced John's lead on approachability by 35 points (thus becoming the more approachable candidate), and he reduced John's lead as preferred candidate by 23 points (again, thus becoming the preferred candidate overall), with all these differences being statistically significant. By contrast, being even better educated than the original version of George did not especially help. The version of George with a PhD was perceived to be slightly more approachable, but on other traits a higher level of education merely made him seem less experienced and less effective, and produced no difference at all when it came to who was the preferred candidate.
Perhaps because we spend our lives working in universities, the finding that respondents seemed noticeably to prefer a candidate who had not been to university surprised (and depressed) us somewhat. But we were also concerned in case the various consequential changes to the profile – especially changes to George's occupation – were instead the factor that was causing these differences. In the seventh wave of the survey, therefore, we tested the same profile but for a would-be candidate who had left school at sixteen (the minimum leaving age in the UK). This produced the results shown in the final section of Table 2.
The sixteen-year-old school leaver is not seen as such a good candidate as the eighteen-year-old school leaver, but even the candidate who left school at sixteen is preferred to one who emerged from university with a PhD. Although we have concerns about the extent to which we can compare these samples to the original John/George baseline, these three profiles are themselves all directly comparable. All that changes between them is the educational status, and they show clearly that a school leaver at eighteen is preferred to one at sixteen, but both are preferred to one with a PhD.
Having concluded these six main waves of the survey, we used the seventh wave of the study for the big fight, to test the relative effect of two of the socio-information variables that earlier waves had shown to be most powerful: a candidate's residential status and their occupation. In one sample in wave seven, therefore, we offered respondents the choice between John and a version of George who was both a GP (which we had found to be a highly positive cue) but who also lived 120 miles away (which we had found to be highly negative). This produced the figures in Table 3. In both the three individual traits and overall, it had the effect of increasing John's appeal relative to George, with all the differences being statistically significant; in other words, when put up against each other, localism trumped occupation.8
Table 3. The Net Effect of George being a GP but also Living 120 Miles Away
|N = 2,316|| || || || |
It is fairly well established that voters are more likely to ‘use social and political stereotypes to help them make decisions in low-information races’ (McDermot, 1998, p. 914). Thus far these experiments have indeed been fairly ‘low-information’ races. Most obviously, they have lacked any information about the candidates' political party. Some of the cues analysed clearly have a potentially strong pull, but we need to know whether they are strong enough to withstand the competing pull of party.
To test this we ran a final wave of the experiment but this time we divided the sample into Labour voters and leaners and Conservative voters and leaners, stripping out voters for other parties. Labour voters and leaners were shown a variant of the original John biography – but this time with John identified as the Conservative Party candidate. They were then also shown one of two alternative biographies for George, this time identified as the Labour candidate. In the first version George's residency is within the constituency; in the second he is described as living ‘about 120 miles away, although he has said that if elected, he would move to live in your area’. A mirror of this experiment was then applied to Conservative voters and leaners: John was described as the Labour Party candidate, George the Conservative.9
This is therefore a robust test of any residency effect: it takes only those who have said they intend to vote for a party and offers them a choice of a candidate from a rival party. If the effects previously discovered were weak (even if they appeared large), then we would have expected them to vanish when party is reintroduced into the experiment. If, however, the effects are stronger, then we would expect them to remain, at least in part.
As Table 4 shows, the effect of changing a candidate's residential status was (not surprisingly) less when we controlled for party than before, but a significant effect remained. When both candidates were local, respondents overwhelmingly preferred George, with leads of between 30 points (experience) to 44 points (overall preference). But when we changed the profile and made George live 120 miles away, these leads were substantially reduced. George's lead on approachability suffered most (down by 29 percentage points), but – importantly – there was also a significant impact on respondents' overall preference. George was still the preferred candidate but changing his residency was enough to lower his lead as the preferred candidate by 16 percentage points.
Table 4. Effect of Changing Residency with Party Included (row %)
|Both local|| || || || |
|George not local|| || || || |
Party therefore remained a very strong determinant of preference (as we would have expected) – and George remained the preferred candidate over John – but the residency effect was clearly strong enough to withstand a significant part of the pull of party. Altering the residency of the candidates was enough to shift sufficient respondents so that almost one in five respondents preferred the candidate of the opposite party to that which they had said they intended to vote for or were leaning towards, and the 16-point change in the standing of the candidates is the equivalent of an 8 per cent swing, far from trivial.10
Discussion and Conclusion
This study consists of a broad comparison of the impact of six candidate characteristics at the aggregate level. For reasons of space, we make no attempt here to test the extent to which the results vary within subgroups of the population. Future research will dig deeper into responses to the individual characteristics and assess whether there is heterogeneity in the experimental effects.11 Our interest here is in the overall relative impact of the characteristics and while we found that all six of the cues we tested had some statistically significant impact on the way voters perceived candidates, we found particularly large effects with education, occupation (with candidates who had served as a local doctor being highly rated) and residency (with candidates from outside the local area being especially heavily penalised). While we have not managed to test all of these effects up against the competing pull of party, we did test the effect of residency – the experiment that demonstrated the strongest effect – and found that altering the residency of the candidates produced a significant change in the standing of the candidates even when the experiment included the party label.
We are aware of a number of qualifications and caveats. This is, self-evidently, a study of one country, at one time; we would expect different results in other political systems with different political cultures. It seems to us unexceptional, for example, to suspect that the results of our religion experiments might be different if conducted in, say, Israel or Egypt; but we might also expect more subtle differences on other characteristics in other systems, depending on the political culture and norms. Indeed, in any system with half-decent residency qualifications (which are absent from the UK), tests for localness would need to be adapted to make any sense, although the carpetbagger or parachutiste candidate is a common enough phenomenon (Pedersen et al., 2007) for us to be sure that this is not solely a British experience. Similarly, we cannot claim that we have tested the overall impact of any of these cues. For example, we have not tested the overall impact of ‘occupation’, merely two particular occupations – a doctor and a political researcher. To examine the impact of others, we would need to compare a range of other occupations. The same applies to almost all of the cues we tested. These are all areas for further study, beyond the scope of this article.
We are also aware that some of the characteristics we test may interact with one another. When we alter one characteristic, therefore, we cannot be sure that all the effect generated is being produced solely by that characteristic. To take the most obvious example, sex and occupation may well interact (a female solicitor may be perceived differently to a female nurse, but in a distinct way from how a male solicitor may be perceived differently to a male nurse). We also realise that the experiments leave out other candidate characteristics that might influence voters, such as their ideology or issue positions. This is a concern in itself, because we might wish to know about the influence of such matters (although in terms of the external validity of our experiments, it is something we are relatively sanguine about, for reasons we explain further below), but there could also be a potential indirect effect, in which characteristics that we test are used by respondents as proxies for those that we are not testing.12 We do not think that any of the profiles utilised in the experiments are likely to have produced major interaction effects, and doubt that any characteristics are acting as significant proxies, but (again) we would need more research to be certain of this.
Perhaps more significantly, we are also of course aware that the effects we have discovered by utilising an experiment are not the same as measuring the effect of these variables in real-world elections. For one thing, some of the variables might impact on the way candidates behave, in a way that also alters voters' perceptions of them. One reason better educated candidates predominate may well be because they are – all other things being equal – more articulate, and in a real-world electoral contest that might more than compensate for the voter's inherent preference for a less-educated candidate. Moreover, for these variables to make a difference in any real-world election would require at a minimum: (a) for a voter to have some knowledge of the candidates' characteristics; and (b) for there to be at least a perceived difference between candidates. Some of the variables that have been discovered here to have most impact – such as occupation and ‘localism’ – are not as easily discoverable by a voter as, say, the sex or race of their candidate, and/or are more open to manipulation and obfuscation by the parties.13 But that is precisely why we have utilised a survey experiment, to attempt to cut through the fog of electoral war in order to measure the latent impact of these cues.
Yet all of these caveats notwithstanding, it is worth ending by recalling the relative impact of these cues. Our experiment with the sex of the candidate produced a 2-percentage-point difference in the overall preference of voters. Our experiment with religion produced at most a 4-point difference. When it came to age, the largest effect was 6 points. However, when we came to examine the effect of education we managed to produce a 23-point difference in the overall preferred candidate; testing occupation produced a 27-point difference; testing the residential location of the candidate produced an even bigger 30-point difference. In other words, the cues that have been most studied, in Britain and elsewhere, produced relatively small effects. Those cues that have been largely ignored produced much larger effects. The impact of whether a candidate was local or not was fifteen times that of biological sex. Indeed, even once we introduced party into the experiment, we still produced an effect some eight times larger than the impact seen in our original biological sex experiment. Even allowing for the caveats we note above, these are substantial effects and ones that deserve further investigation. The purpose of this article therefore is to attempt to start a debate about the focus of electoral studies when it comes to candidate effects.
Appendix: Appendix: Profile Wording and Survey Details
In wave one, half of respondents saw the John/George profiles as given in the text of the article; the other half saw the exact same profiles but with John's name changed to Sarah (‘Sarah Burns is 48 years old …’). No other changes were made, the profile staying in every other way identical.
The second wave tested attitudes to religion. Half of respondents saw the profiles given above, but with George's name changed to Daniel Goldstein (‘Daniel Goldstein is 48 years old’ …). The second half of the sample saw George's name changed to Mohamed Lafi. As with wave one, only the names were changed; all other aspects of the profile stayed the same as for the original John/George comparison. We did not give any explicit indications of religious belief or religiosity, beyond the candidate's names, which were highly suggestive if not definitive indications of religious background or heritage.
Wave three tested reactions to occupation. Half of respondents saw the John/George profiles given above, but with George's occupation changed to that of a general practitioner doctor:
George Mountford is 45 years old; he lives in the constituency and studied medicine at university. He is a local GP with a special interest in elderly care. George is passionate about education, with two children in local schools and a wife who is a primary school teacher.
The amended text is in italics. The other half of respondents saw a George who was an aspirant politico (again, with the amended text in italics):
George Mountford is 45 years old; he lives in the constituency and studied for a degree in politics. After university George worked for two MPs and became a local councillor. George is passionate about education, with two children in local schools and a wife who is a primary school teacher.
Wave four tested attitudes to the age of the candidates. Half of respondents saw a George who was 32 (instead of 45 in wave one); the other half saw a George who was 60. Only the candidates' ages were changed; all other details remained constant to that in wave one.
Wave five tested whether attitudes to the candidates depended on how ‘local’ they were. Half the respondents saw a George who instead of ‘living in the constituency’ had ‘moved into your consistency two years ago’. The other half saw a George who ‘currently lives about 120 miles away, although has said that if elected, would move to live in your area’. As a consequential change, this second version also required us to lose the reference to George having two children in ‘local schools’; instead he now had ‘two children in school’.
And wave six tested attitudes to the candidates' educational backgrounds. Half saw a George who had left school at eighteen, with no mention of university. The other half saw a George who had gone to university, emerging with a PhD. This question required more widespread consequential changes than in any of the other waves. Because it might not be felt plausible for George to have left school at eighteen and yet to have become a solicitor, we altered his occupation to a more generic one. We also made the same change to the profile of our PhD-endowed George, to allow the two to be compared. They read:
George Mountford is 45 years old; he lives in the constituency. He left school at 18 and now works in insurance. George is passionate about education, with two children in local schools and a wife who is a primary school teacher.
George Mountford is 45 years old; he lives in the constituency. He went to university, gaining a PhD, and now works in insurance. George is passionate about education, with two children in local schools and a wife who is a primary school teacher.
These two profiles are thus slightly less comparable with the baseline George than are the other profiles, but are comparable with each other.
Having carried out initial analysis of these six waves, we then carried out a further seventh wave, designed to check on the relative weight of two of the apparently most powerful influences as well as double check some aspects of the findings. This required the seventh wave to be split into three. To test the relative impact of localism and occupation, we created a profile of a George who was both a GP (considered positive by respondents) but who also lived outside the constituency (negative). All other aspects of the profile remained constant. It read:
George Mountford is 45 years old; he studied medicine at university and is a GP with a special interest in elderly care. He currently lives about 120 miles away, although has said that if elected, would move to live in your area. George is passionate about education, with two children in school and a wife who is a primary school teacher.
As with earlier waves, we tested this against our original John profile. To test whether asking the three questions about traits was influencing the question about overall preference, we split the sample so that a third of respondents to wave seven were shown this profile and asked the original four questions, and one-third were asked only the question about preferred candidate.
The final third of respondents to wave seven were shown an amended education question, in which we lowered George's school leaving age to sixteen (the minimum possible in the UK), while keeping all other attributes the same. This third of respondents were shown the four standard questions.
The eighth wave introduced party labels. Those respondents who said they intended to vote Labour in the next election (or who said they were unsure but who were leaning towards Labour) were shown a profile of a Conservative candidate called John. The profile is very similar to those used in the earlier waves, except that it now incorporates an explicit party label:
John Burns is the Conservative candidate. He is 48 years old, and was born and brought up in your area, before going to university to study for a degree in physics. After university John trained as an accountant, and set up a company ten years ago; it now employs seven people. John has interests in the health service, the environment, and pensions, and is married with three children.
We then randomised the description of the Labour candidate, with approximately half seeing a candidate who lived locally:
George Mountford is the Labour candidate. He is 45 years old; he lives in the constituency and studied business at university. He is a solicitor and runs a busy local practice. George is passionate about education, with two children in local schools and a wife who is a primary school teacher.
The other half saw a Labour candidate who was identical apart from living outside the constituency:
George Mountford is the Labour candidate. He is 45 years old; he currently lives about 120 miles away, although has said that if elected, would move to live in your area, and studied business at university. He is a solicitor and runs a busy local practice. George is passionate about education, with two children in school and a wife who is a primary school teacher.
For those who said they were voting (or leaning) Conservative we simply changed the party labels; that is, they saw a Labour candidate (John) who was local, and then either a Conservative candidate (George) who was local or who lived 120 miles away.
All the survey work was carried out online by YouGov, an online polling company, who sample from their panel of some 360,000 British adults. Members of the panel are recruited using a variety of methods including self-selection, advertising and commercial relationships with websites.14 Waves one to seven were commissioned by the authors; wave eight utilised space on the British Election Study's Continuous Monitoring Survey (for which we are very grateful). The surveys had the following field dates and sample sizes:
- Wave 18.8.11–9.8.11 (N = 2,864)
- Wave 29.8.11–10.8.11 (N = 2,700)
- Wave 310.8.11–11.8.11 (N = 2,577)
- Wave 411.8.11–12.8.11 (N = 2,656)
- Wave 514.8.11–15.8.11 (N = 2,847)
- Wave 615.8.11–16.8.11 (N = 2,665)
- Wave 723.8.11–24.8.11 (N = 2,709)
- Wave 823.5.12–15.6.12 (N = 1,152).
The first seven surveys – which are compared to one another – were thus all carried out within a sixteen-day period; only the eighth wave – which is self-sufficient and is not compared to the earlier work – was conducted separately. As a result of these overall Ns, each split sample in the first six waves was seen by c.1,300 respondents. The split samples in the seventh wave were each seen by c.900 people. The split samples in the eighth wave were seen by between 125 and 210 people. Each experiment was run with a new sample of respondents; any individual respondent was only asked to respond to one trade-off.
All the reported data in this article were weighted by YouGov's standard weighting based on age, gender, social class, region, party identity and the readership of individual newspapers. However, we conducted the analysis on both weighted and unweighted data, and the resulting difference in the effect sizes between the two sets of analysis was minimal; for example the difference in the net effect of George living 120 miles away between the weighted and unweighted data is 1 per cent.15
With many thanks to Peter Allen, Sarah Birch, Sarah Childs, Cees Van Der Eijk, Jocelyn Evans, Stuart Fox, Mathew Humphrey, Edward Page and Tim Smith, as well as the participants at the 2011 EPOP Conference, for their comments on drafts of this article. Thanks also to the British Election Study team for including our pilot items in the BES 2010 rolling survey and to the University of Nottingham's Integrating Global Society group which provided the funding for the survey work. Extra special thanks to Joe Twyman and Anthony Wells of YouGov for their enthusiasm and for all the lovely data. The data used in this manuscript are available for replication purposes upon request.
Space prevents a detailed examination of this literature, but for studies examining the impact of the sex of candidates, see, for example, Cook, 1998; Dolan, 1998; 2001; 2004; Paolino, 1995; Plutzer and Zipp, 1996; Sanbonmatsu, 2002; Sigelman and Welch, 1984; Sigelman et al., 1995; Trent et al., 2001. There is also a considerable body of research, again mainly developed in the United States, assessing the impact of candidate's race on voter preferences (Brouard and Tiberj, 2010; McDermot, 1998; Sigelman and Welch, 1984; Sigelman et al., 1995; Terkildsen, 1992). There is a similarly large literature on the role of Latino candidates (Barreto et al., 2005; Kaufmann, 2003; Stokes-Brown, 2006). There is a growing body of literature on candidates' visual image (Banducci et al., 2008). There is, however, relatively little on, for example, a candidate's occupation or age (for rare examples see McDermot, 2005; Trent et al., 2010).
Although some of these local factors are still party rather than candidate focused, such as Ron Johnston et al.'s extensive work on neighbourhood effects (see, for example, Johnston et al., 2004).
The relative absence of residency from the UK electoral studies literature is even more surprising, if only because it has been studied elsewhere. In the US literature, for example, it dates back to Key's landmark Southern Politics in 1949 (Key, 1949), with a ‘home state’ advantage in presidential races (Dudley and Rapoport, 1989; Garand, 1988; Rice and Macht, 1987a; 1987b). In Australia, there is evidence that voters are more likely to vote for a candidate the longer he or she has lived in the constituency (Studlar and McAllister, 1996). In Ireland, where this effect is well known, there is even evidence that the location of the candidates' house within the constituency can affect their electoral performance (Gallagher, 1980; Gorecki and Marsh, forthcoming; Weeks, 2008).
The literature in this area tends to refer to candidate gender, but if gender is understood to be a social construct and biological sex is understood as a (generally) simple dichotomy assigned at birth, then gender is not a perfect fit with biological sex, and biological sex is the more appropriate term when comparing voters' reactions to male and female candidates (Lovenduski, 2005).
The US research on candidate backgrounds tends to measure responses to race and religious cues separately. This distinction makes sense in the US where levels of attendance at religious services and religious belief are much higher than in Britain and where ‘religiosity’ or the intensity of religious feeling/practice is generally seen as a ‘good thing’ by a large number of voters and can reasonably be included in a list of desirable candidate traits. In Britain politicians often evade questions about their personal religious beliefs, partly from fear of ridicule, but mainly because they believe high levels of religiosity may undermine their chances of electoral success. Yet religious background can still play a role in voters' assessments of candidates. Religion might be significant as an ‘in-group’ measure (where voters seek out candidates from their own religion) or it might signify ‘out-groups’ who are subject to prejudice. There is an obvious overlap between ethnic religious groups and race and it is here that we expect to see the impact of religion on candidate preference in Britain. Over the last 30 or so years simple colour prejudice has declined dramatically in Britain (Ford, 2008) but in its place there has been a rise in ‘Islamophobia’. We therefore chose to use an ethnic-religious cue rather than race or religiosity.
In order to test the validity of our choice of trait measures we conducted a pilot study using the British Election Study's Continuous Monitoring Survey (BES CMS), part of the larger ESRC-funded BES project. The full details are available from: http://bes2009-10.org/. The CMS is a rolling internet survey conducted every month by YouGov on behalf of the BES team, for which researchers can submit proposals for the inclusion of short batteries of questions. We would like to thank the BES team for including our questions in the February 2010 CMS. Working with a single candidate profile (which we manipulated for different samples), we asked respondents to judge a hypothetical candidate's approachability, experience and effectiveness. This pilot work – available on request from the authors – revealed that our trait measures produced meaningful variation in responses.
Of the 19,016 respondents to the seven waves of the experiment there were 10,826 who did not choose the same candidate across all four questions. To be absolutely sure that there was meaningful variation in the trait selections we also ran a logistic regression of traits (coded as dummy variables) on preference. We also examined the relationship in each of the subsamples, and despite some variations, again found that respondents were able to distinguish between the traits and did not simply make selections at random. Details are available on request from the authors.
A further sample from wave seven saw the same profiles tested, but without offering the three questions about candidate traits, to test whether asking about the traits was somehow skewing the overall figures. This produced the same effect – with localism trumping occupation – but was even stronger: John's lead as the preferred candidate extended by some 19 points.
Full details of the wording are provided in the Appendix.
Space constraints preclude an extended discussion of the heterogeneity of the responses, but suffice to note that we found an even stronger effect among Conservative voters (where the lead on overall preference changed by 22 percentage points, or an 11 per cent swing). But even among Labour voters, the shift was some 14 points, or a 7 per cent swing, the sort of swing that most candidates would kill for.
This will include whether there is evidence of respondents favouring candidates with whom they share characteristics: whether, as the identity politics or descriptive representation literature would theorise, voters want a representative who is ‘like them’ (Cutler, 2002).
For example, imagine that we tell you a candidate is a GP but nothing about the candidate's position on issues; you might assume that a GP would be in favour of extra spending on the health service; if you too favour extra spending on the health service you might therefore favour that candidate; however, you would do so because of their assumed issue positions, and not because of their occupation per se; you would perhaps like a tax inspector who favoured extra spending on the health service just as much. In this particular example, we do not think this is a major problem, since (as we observed above) in our GP experiment we also found huge increases in ratings of approachability as well as in the scores for being the preferred candidate, which makes us suspect there is something about the role or job which has proved attractive to respondents, but in general we are aware of the potential for this to be an issue.
More positively, however, this is one reason why we are relatively sanguine about the absence of candidate issue positions from our experiments. In British elections we suspect most voters have even less knowledge of the individual candidate positions on issues – as differentiated from the issue position taken by their party – than they do their candidates' personal characteristics.
YouGov estimate the participation rate of their non-probability internet panels at around 30 to 40 per cent.
See http://research.yougov.co.uk/services/panel-methodology/ for full details.
Rosie Campbell is a Senior Lecturer in Politics at Birkbeck, University of London. She has research interests in voting behaviour, political participation, representation, political careers and gender and politics. Her book Gender and the Vote in Britain was published in 2006 and she has recently published in the British Journal of Political Science, the British Journal of Politics and International Relations, British Politics, Political Quarterly and Political Studies Review. Follow Rosie on Twitter: @Rosiecampb. Rosie Campbell, School of Politics and Sociology, Birkbeck, Malet Street, Bloomsbury, London WC1E 7HX, UK; email: email@example.com
Philip Cowley is Professor of Parliamentary Government at the University of Nottingham. His recent books include The British General Election of 2010 (with Dennis Kavanagh; Palgrave, 2010), Developments in British Politics 9 (edited with Richard Heffernan and Colin Hay; Palgrave, 2011) and The Rebels (Politico's, 2005). Philip Cowley, School of Politics and International Relations, University of Nottingham, University Park, Nottingham NG7 2RD, UK; email: firstname.lastname@example.org