Being a voter: developing a survey instrument for expressive voting

ABSTRACT It has been argued that expressive motivations are a crucial element in persuading people to vote, overcoming the “paradox of voting” and it has been demonstrated that in experimental situations, subjects behave in a manner consistent with expressive voting. However, there have been limited attempts at operationalizing expressive voting in a survey context and these have been unable to successfully separate the influence of outcome related benefits (instrumental pay-offs) from expressive pay-offs. In this paper, we use a web survey experiment to first identify citizens with a propensity to behave expressively and then assess the validity of the instrument.


Introduction
Why do people vote when the chance that they will affect the outcome is vanishingly small? This is a question that has troubled political scientists for decades. One answer is that people vote to express themselves rather than in the hope of affecting the election result (Brennan and Lomasky 1997). In other words, voting is an act of "being" rather than "doing" as voters gain utility both from expressing their support for a candidate with whom they identity and affirming their status as a supporter through the act of voting (Schuessler 2000). Expressive benefits are not dependent on the outcome of the election so may be thought of as consumption benefits, which are crucial in resolving the paradox of voting (Riker and Ordeshook 1973). Despite their importance in explaining why people vote, there have been few successful attempts to operationalize expressive motives in a survey context. In particular, whilst survey data has been used to capture variation in the expressive pay-offs from voting, for example as measured by proxy with party identification (Fiorina 1976), as far as we are aware, no previous survey research has separated expressive pay-offs from those that depend on the outcome of elections. In this article, we use an experimental design to develop an instrument for measuring the propensity of citizens to vote expressively which is applicable to voters and non-voters alike and does not rely on survey respondents to rationalize their own motivations for voting.
Drawing on theories of expressive voting (Brennan and Buchanan 1984;Brennan and Lomasky 1997;Schuessler 2000), expressive partisanship (Green, Palmquist, and Schickler 2002;Huddy, Mason, and Aarøe 2015) and party identification as a social identity (Huddy 2001;Greene 2004), we assess the extent to which more expressive citizens are more likely to vote than less expressive ones. We show how expressiveness as a trait increases the likelihood of turning out to vote.

Theory
In classic rational chance models of voting the expected benefits from turning out to vote are a product of the difference in the benefits (B) that the voter would receive if her party or candidate won the election multiplied by the probability (P) of her vote being decisive. These expected benefits have to be set against costs (C) of voting which, whilst are usually assumed to be low (Aldrich 1993), are likely to exceed any objective estimate of the expected benefits, due to the extremely low value of P. This leads to the paradox of voting (Riker and Ordeshook 1968).
There have been many attempts to resolve the paradox, the most successful of which have emphasized the intrinsic benefits of voting (D) which are accrued by citizens who vote whether or not their preferred candidate is elected (Downs 1957;Fiorina 1976).
There are two widely discussed intrinsic benefits of voting both of which are primarily expressive in nature; civic duty and the expression of social or political identity (Hillman 2010;Hamlin and Jennings 2011). By expressive, we mean acts that are motivated by their symbolic meaning, not by the benefits that accrue directly as a result of the action. According to some, the notion of civic duty and other normative motivations for voting, such as altruism or ethical voting, reflect expressive considerations (Fiorina 1976;Drinkwater and Jennings 2007;Fowler and Kam 2007;Hamlin and Jennings 2011). That is voters wish to express their support for the democratic process and in doing so affirm their identity as a voter or as a responsible citizen (Hillman 2010). Thus the distinction between voting as duty and voting as an act of expression of social or political identity is not so much the expressive content, but rather that the latter has implications not only for whether or not a citizens votes but also how they vote (Jones and Hudson 2000).
If voting is an expression of identification then identifying with a particular social group (Huddy 2001), or a particular party, candidate or policy position (Brennan and Hamlin 1998) then who one votes for is integral to the expressive returns. Some have argued that expressive behaviour is, by definition, always about identity (Hillman 2010). From this perspective, expressive benefits can be thought of as "the utility or disutility of satisfying or violating one's party allegiance" (Fiorina 1976, 395). As such it has often been likened to the pleasure of cheering for your team at a football match or even in front of the TV (Brennan and Buchanan 1984;Brennan and Lomasky 1997;Green, Palmquist, and Schickler 2002;Huddy, Mason, and Aarøe 2015). There is no realistic possibility that cheering for the team will help secure victory or affect the outcome in any way. The cheering is purely expressive and, unlike an investment in the market place, is decoupled from the material benefits that may be accrued as a result of the investment of time, energy or money. Similarly in most elections, at least where the electorate is large, voting is very unlikely to affect the outcome yet many citizens still vote. The utility gained from voting expressively is not simply about the preference of one candidate or policy over another, but rather the expressive voter prefers to prefer one candidate or policy (Brennan and Lomasky 1997, 32).
But why do citizens prefer to prefer one party or candidate over another? As suggested above the value of expression is closely linked to the affirmation of identity. If the football fan does not identify with the team then there is little incentive to cheer. The theory of expressive partisanship (Huddy, Mason, andAarøe 2015, 2018) likens party identification with a social identity (Tajfel and Turner 1979) insofar as citizens form an emotional attachment with a political party (Huddy 2001;Greene 2004). By voting for that party the citizen re-affirms his membership of the group (for example Democrats) and re-affirms his identity as the kind of person who votes Democrat (Brennan and Hamlin 1998). Voting is, therefore, an act of identification which is primarily about "being" rather than "doing" (Schuessler 2000). It is the symbolic value of supporting a party (or a policy) that drives the expressive voter, rather than the policy outcomes that might result from the election (Brennan and Lomasky 1997).
Whilst the theoretical basis for expressive voting is well established, demonstrating it empirically is more difficult. The crucial problem is differentiating expressive motives from instrumental motives when, more often than not, they coincide. For example, people who have a strong preference for a set of policies are also likely to identity with a party advocating those polices. Brennan and Lomasky (1997, 39) outline how preferences revealed by actual choicesuch as giving to charity or making a purchasemay diverge from preferences expressed though voice, including voting. This occurs because when voting one does not face the full cost of obtaining the private benefits of being seen to support (for example) charity or income distribution (Tullock 1971). This insight has helped researchers to isolate expressive motives for voting in experimental settings, for example by manipulating whether the subject will bear the full cost of supporting redistribution (Carter and Guerette 1992;Etang, Fielding, and Knowles 2016). Because of the weakness of the link between individual choice and outcome, intrinsic or expressive benefits loom large in the voting decision as voters have a strong incentive to vote in a way which is consistent with their selfimage. This makes expressive voters especially open to social pressures over their choices. In other words, elections are not like market places insofar as voters' choices are not decisive, potentially leading to different (and possibly normatively less desirable) outcomes due to the dominance of expressive considerations (Tullock 1971;Brennan and Lomasky 1997;Hamlin and Jennings 2011). Moreover, insofar as social norms are felt more keenly amongst those that strongly identify with a group (Terry and Hogg 1996), the incentive to vote expressively should be related to the strength of partisan identification (Fieldhouse and Cutts 2016).

Survey measures of expressiveness.
Although there is a strong theoretical case for believing expressive motivations dominate electoral choice, measuring expressive voting has proved more problematic. Outside of the political context, social psychologists have developed measures of personal expressiveness, such as the Instrumental and Expressive Behaviour Inventory (Holmbeck and Bale 1988) and the Personally Expressive Activities Questionnaire (Waterman 1993). Neither is ideal for our purposes. The former (the IEBI) contains 92 items and as primarily designed for assessing the validity of measures of masculinity and femininity first proposed by Bem (Bem 1977;Helmreich and Spence 1978) and which have been strongly critiqued from a feminist perspective (Gill et al. 1987). Johnson et al. (1975), for example, argued that expressiveness is not the opposite of instrumentality but a separate independent dimension. The latter (PEAQ) was designed to capture non-hedonistic happiness or eudaimonia and self-realization which have been linked to intrinsic motivation to take part in specific activities (Waterman 2005). However, its applicability to low cost, low return activities such as voting in elections is unclear. Others have developed and demonstrated the validity of measures of individual expressiveness within specific contexts such as the family (Halberstadt et al. 1995). Overall, the existing research in social psychology points towards a distinct expressiveness personality trait, but its measurement has been contextspecific and not suited to identifying a propensity to act in a manner consistent with the theory of expressive voting, or more specifically to carry out costly actions in order to signal group identity (Schuessler 2000;Thorbjørnsen, Pedersen, and Nysveen 2007).
As noted above experimental research has helped establish the existence of expressive voting empirically and provided some clues as to who votes expressively (Fischer 1996;Etang, Fielding, and Knowles 2016), but this has proved difficult to export to the survey context. This limits our ability to examine in detail how expressive voting varies by social and demographic characteristics and perhaps more crucially with political preferences. The failure to develop a reliable survey instrument to measure the expressiveness of voters not only makes our understanding of voter choice and turnout as incomplete but also restricts our understanding of the disconnection between vote choice and policy preferences. This is not to say survey researches have ignored expressive choice. One method is to ask voters directly whether they vote to express their political views (Carlsson and Johansson-Stenman 2010;Krajina and Prochazka 2017). However, this has the limitation for the analysis of voter turnout since, whilst it is possible to ask voters why they voted or intend to vote, it makes little sense to ask this of non-voter. Moreover, it is difficult for people to rationalize their own motives for behaviour which may be unconscious and subject to social desirability bias. A more usual way of dealing with expressive motivations is to equate them with things that correlate with expressive pay-offs from voting. For example, if voting is about affirming political identity, then expressive pay-offs should increase with the strength of partisanship whether measured conventionally or using a social identity scale (Huddy, Mason, and Aarøe 2015). Similarly, we would expect these pay-offs to be highly correlated with party and candidate feeling thermometer scores or policy attractiveness (Jones and Hudson 2000). In attempting to forge a more direct measure others have adopted more explicit indicators of expressive pay-off such as the emotional responses to candidates and parties, including anger, hope, fear, and pride (Kan and Yang 2001). Whilst these are all useful measures of the expressive pay-offs from voting, they are not separable from the instrumental pay-offs associated with voting for the same candidates or parties. In other words, dislike of the policies that a candidate would implement if elected is likely to invoke emotional responses toward the candidate. They are therefore better deployed as a means of validating a measure of the propensity of the citizen to vote expressively rather than as indicators. Drinkwater and Jennings (2007) used the commitment to voting as a duty as an indicator of expressive voting. Whilst we agree that duty may be an expressive concept, the use of duty has a number of drawbacks. First, it fails to identify those that do not vote out of duty but out of a wish to express their identification with a political party, policy or social group. Second, the concept of voting as a duty is bedevilled by the possibility of reverse causation: people who vote, when asked, reason that voting is a duty whilst nonvoters quite reasonably prefer to deny this claim (Matsusaka and Palda 1999). In other words, it is, in part, a post hoc rationalization of voting. Galais and Blais (2016), in a robust defence of civic duty as an exogenous influence on turnout, find that it may also be party caused by voting. In addition, the specific operationalization adopted by Drinkwater and Jennings (2007) is not ideal since it is based on the respondent's view of voting in general, not on their own personal motivations. Like feelings towards parties, strength of partisanship, and emotions, duty may be better deployed as a tool for validating a more direct measure of the propensity to behave expressively than as a measure of expressiveness.
So what might indicate a propensity towards expressive behaviour that could be readily measured in a survey context? Brennan and Lomasky suggested it might be captured by the sending of get-well cards, since sending cards have no possible effect on the health of the patient (Brennan and Lomasky 1997, 33). One possible objection is that the sender does hope to achieve an outcome beyond the warm glow of sending the cardthat is, making the sick person feel better psychologically, if not physically. Moreover, even setting aside this objection, in the age of social media, sending traditional greetings a card is likely to vary demographically. A good measure of the propensity to act expressively should be equally valid for all citizens.
Copeland and Laband (2002) make a different objection: that the sending of get-well cards does not necessarily translate to political expressiveness. They argue that a propensity to general expressiveness may not be an accurate reflection of an individual's propensity to behave expressively when making political decisions (Copeland and Laband 2002, 352). Instead, they advocate using the wearing of a button/displaying a sticker, poster or yard sign in support of candidate, which they show these to be strongly correlated with turning out to vote. Whilst it is true that general expressiveness may not be directly transferable into the political arena, unfortunately, the wearing of buttons (etc.) is clearly correlated with the instrumental motives to vote for a candidate, thus failing the same exclusion restriction as feeling scores or strength of identity. In other words, wearing a button is likely to be related to the dependent variable, vote, though mechanisms other than expressiveness.
In this paper, we develop a survey measure of expressiveness that should not be related to vote choice or turnout through any mechanism other than expressiveness. We use a survey experiment to identify citizens who display a propensity to act expressively in an inconsequential non-political context. By allowing the survey researchers to identify who votes more or less expressively, such a measure would enable the researcher to answer important questions about how expressive voting affects election and policy outcomes as well as how it affects electoral participation.

The instrument
We asked the following questions to a sample of respondents on the British Election Study Internet Panel Wave 11 (Fieldhouse et al. 2017). First to identify relevant objects with which citizens might identify we asked: Do you feel a strong connection with any of the following? If more than one please tick the one you feel the strongest connection to.
As response options, we offered a list of objects which might appeal to different sections of the population. It covers a range of objects so that the results of the experiment are not skewed by the choice of objects. For example, the use of get-well cards as suggested by Brennan and Lomasky (1997) may skew measurement to older voters who are more likely to send greetings cards rather than electronic messages. Similarly, the choice of any single object (e.g. a sports team) might bias the measurement towards a specific group (e.g. young men). The selection of object with which to identify is designed to cover the widest possible socio-economic and demographic range. These are listed in the appendix (Table A1) along with the percentage selecting each option. Political parties are not included in the selection to ensure that the measurement of expressiveness is independent of political identity, since the latter will almost certainly be related to non-expressive (i.e. instrumental) motives to vote.
To check the extent to which respondents would vote as a result of stronger identification rather than expressiveness, we measure the strength of identity with the object, although it is important to note that strong identity might result from being a more expressive person. The question used to test strength of identity with the object is as follows: How much do you agree or disagree with the following statement? I identify with my favourite[object of identity]] Having identified a connection with one of the objects, not surprisingly most agree that they do feel a connection with that object (see appendix Table A2). On a seven-point scale, the mean strength of identity score varied from 5.0 (film star/celebrity) to 6.0 (church or religious group) (see appendix Table A1).
Having established the object and strength of identification, the next step was to design an experimental test capable of detecting whether citizens are prepared to undertake costly actions which allow them to enjoy the benefits of expressing their identification, whilst excluding them from any instrumental benefits (Brennan and Hamlin 1998;Etang, Fielding, and Knowles 2016). This is achieved by randomizing a manipulation of the perceived likelihood of being influential: -Chance of winning is randomized to either "no chance" or "a good chance". -Cost to vote is randomized at nothing, 50 pence, £1 or £2 in order to identify the extent to which the expressive value may be offset by participation costs.
As in the case of electoral turnout, these costs are set fairly low (Aldrich 1993). The purpose randomization of cost is a calibration tool, to identify the optimal amount at which to set the cost in future data collection exercises rather than to attach any specific value to expressive participation. It was anticipated that some cost would discriminate between more and less expressive voters better than no cost, since there is little disincentive to vote if there is zero cost. The object of identity is piped from the previous question.
Imagine Overall 4266 respondents identified an object of identity and provided an answer to the voting question. The sample was fairly evenly split between those who said they would be likely or very likely to vote (50%) and those that would not (35%). The remainder said they were neither likely nor unlikely (11%) or that they did not know (3%). The full distribution is shown in table appendix, Table A3).
The first test of the instrumentality of participation is the extent to which electors respond to the cost of voting. Data are unweighted as the assignment to different costs and chance of winning is randomized and we are interested in the difference between these groups. 1 Figure 1 shows that they do respond to cost. The higher the cost of the vote, the fewer people take part and the probability of voting (on 1-5 scale) is significantly related to cost overall. The most noticeable difference is between zero cost and some cost, though there is a significant fall in turnout when the cost increases from £1 to £2. Of course, if the cost of voting was to keep increasing, at some point nobody would vote. In this sense, the question is not about whether voters are instrumental in that they respond to cost, but how far they respond to the chance of affecting the outcome (P). In the rest of this article, we use the term instrumental narrowly as pertaining to the value of P (affecting the outcome), as opposed to being sensitive to cost. The next step is to compare the level of voting in "Britain's Best Vote" by random assignment of whether the nominated object of identity had any chance of winning. In this experiment, the P term is only measured as a dichotomy ("no chance" versus "a good chance") but theoretically may take any value. 2 The extent to which a larger proportion of electors vote when they are told there is a chance of victory compared to when they do not suggest that some people take the value of P into account when deciding whether to vote. Table 1 shows the per cent of electors who are very or fairly likely to vote according to both the cost and the value of P. We see that only when the cost is high (£2) is there a statistically significant difference in the chance of voting depending on P. This is what we might expect since according to Equation (1) above when P = 0, the probability of voting is simply D-C or the expressive benefits minus the cost. In other words, as C increases V is reduced by a proportionally greater amount if P is small.
Table 1 tells us that whilst subjects were sensitive to the cost of participation, they barely responded to Pthe likelihood of influencing the election outcome. Overall the chance of voting is just over two percentage points higher if there a chance of winning (marginally insignificant), or just under 5 percentage points if the cost is as high as £2 (which is statistically significant). This difference suggests that there are a few electors who primarily voted instrumentally in this context (i.e. are dependent on P). However, the large proportion voting when P = 0 suggests that many more voted expressively, though this depends on the cost. When the cost is non-trivial (£2), the proportion who vote despite having no chance of winning is only 35%.
We can fit the classic turnout model shown in Equation (1) using the information in the experiment since we have manipulated the value of P and C, and if we assume that B is a constant (and small), we are able to estimate the effect of D. It is worth reminding ourselves here that that whilst we do not need to assume B = 0, the experiment was designed to ensure there were no outcome related benefits from voting. That is voters stood to gain nothing if their favourite team (for example) won the competition, except for perhaps some satisfaction at seeing their team win. This reflects the possibility that some outcome related benefits might also be expressive, insofar as  the pleasure of seeing one candidate succeed, when there are no material or policy benefits at stake, can be described as expressive. This is a semantic question that we do not attempt to resolve here. Instead, we can empirically estimate the effect of P (using the variable "winProb" which is assigned a value of 1 or 0) on the probability of voting (V). The value of D is represented by the constant term or the intercept. Table 2 (model 1) gives the coefficients for the regression equation: where "e" is the regression error or the unobserved propensity for each individual to take part. The analysis was repeated as an ordered logit and also using a logit model where the outcome was fairly/very likely to vote. In both cases, the substantive results were identical with no change to the significance of variables. The OLS model is shown for ease of interpretation, and no predictions fell out of the range of the original scale (1-5). Model 1 (Table 2) confirms the large constant term, representing the expressive element (D) and a significant impact of cost (C). The effect of cost equates to a decrease in chance of voting of approximately half a point on the five-point scale for every £1 increase in cost. There is no significant variation by the P term (winProb). From the equivalent logit equation, when there is no chance of affecting the outcome, the predicted probability of being likely to vote falls from 64% to 49% when the cost increases from zero to £1 and to 33% at £2. Overall, the models confirm the findings of the descriptive results: voting in an inconsequential contest is largely driven by expressive motives but these are quite quickly offset by the cost of voting.
Having established the major role of expressive motives in voting in a fictional election, we now extend the model to take into account the extent to which participation depends on other characteristics of the voter. As in the political world, electoral turnout is influenced not only by costs and benefits but by characteristics of the individual. For example, it may be that because of the nature of the electoral context and what is at stake, certain demographic groups might be more or less likely to take part. We, therefore, control for age and gender. Voting is usually also correlated with education and political interest (Smets and Van Ham 2013), and although this is a non-political election where such factors may be less important, we control for political interest to take this into account. 3 Moreover, the expressive benefits to be accrued depend on the extent to which the elector identifies with the group that the candidate represents (Schuessler 2000). In our experiment, this is measured by the variable "identityStrength" (as shown in table A2). Table 2 (model 2) shows that the strength of identity does indeed have a strong influence on the likelihood of voting, as well as age, gender and attention to politics. 4 As we might expect, the cost is still significant but winProb (the PB term) is still insignificant. To formally test whether the effect of the P term is conditional on the cost we add an interaction between winProb and voteCost (model 3). This proves to be insignificant. Finally, in model 4, we test whether the P term is conditional on the strength of identityin other words, are people who identify more strongly more likely to vote despite an increase in cost? The answer is no, as there is no significant interaction between "winProb" and "Identity Strength". Overall, then, we find no evidence of the P term having any significant effect on taking part in this inconsequential election regardless of cost and strength of identity with the object being voted for. This serves to confirm the original aim of the question wordingthat instrumental or outcome related benefits are excluded. The experimental manipulation of P meant that had voters been influenced to vote by outcome related benefits, this would have shown up by a significant coefficient for P, either on its' own or conditional on cost or strength of identity.

Criterion validity
In order to optimize tests of criterion validity the questions were asked to a different subset of respondents in May 2018 (wave 14). This allowed for a simpler version of the instrument to be tested, without randomization, which would be easier to replicate in future surveys and maximizes the usable sample size. The initial question to identify the object of identification was modified very slightly to replace the least popular response, "film star or celebrity", with "company". This made very little difference as it was only selected by 1% of respondents. However, a slightly larger percentage (45%) selected "none of the above" than in the wave 11 experiment (38%).
The main vote question was worded so that the cost to vote was fixed at £2 and the object identified with had "no chance of winning" meaning that all voting can be classed as expressive based on the logic outlined above. This increased the sample size for our analysis of validity since there was no need to discard cases in which instrumental motives might be relevanti.e. where the object of identification had a "good chance" of winning. In addition, BBC Television was changed to "National Television" to make the question transferable to other contexts. This may have had a small effect on the likelihood of voting since the BBC brand conveys prestige not necessarily invoked by 'National Television'. The question was therefore asked as follows with the response options unchanged. Using this version of the question in wave 14, 32% of the sample (n = 3635) that identified with an object said that they would be either very likely or fairly likely to vote compared to 35% in the relevant sub-sample in wave 11 (as shown in Table 1) with a mean score of 2.53 (compared to 2.67 in wave 11).

Convergent validity
Before turning to predictive validity, we make a brief assessment of convergent validity (Drost 2011) by reference to other BES items that we expect to be correlated with a propensity to expressive behaviour. In wave 4 and wave 6, we asked the questions suggested by Kan and Yang (2001): Now we would like to know something about the feelings you have towards each of the parties. Which of these emotions do you feel about each of the parties?
The emotions asked about were anger, hope, fear and pride. We might expect that people who are more expressive as measured by our instrument would be more emotionally expressive about the political parties. Table 3 shows the percent of the sample who expressed each emotion towards any political party according to whether or not they were likely to vote in Britain's Best. As we would expect if both are capturing expressiveness, there is a significant difference in the reporting of pride and hope between those that said they would be likely to vote and those that would not. However, there was no significant difference on scores for anger and fear. This suggests that the expressiveness trait being captured by the instrument is linked with positive rather than negative emotions, which is not entirely surprising given that the design of the instrument was based around the affirmation of support for an object of identity. Whilst this is consistent with classic expressive motivations for voting (Schuessler 2000) it does suggest that the instrument may be less useful for detecting negative expressive motives to vote.
Another test of convergent validity is the association with civic duty which some have used as an indicator of expressive voting (Drinkwater and Jennings 2007). Not only do we find a significant correlation between the instrument and the standard BES measure of civic duty 5 but a regression model (Table 4) confirms that this survives controls for age, gender and political interest. This equates to an average predicted civic duty score of 4.2 on a five-point scale for the most expressive respondents (very likely to vote in Britain's Best) compared to 4.0 for the least expressive. Table 4 also shows the instrument significantly predicts the respondent's propensity to express a liking for parties (defined as their highest party like score). 6 The coefficient equates to an average predicted maximum score of 7.2 for the least likely to vote in the experiment versus 7.7 for the most likely.
We might also expect that more expressive respondents would be more likely to donate money to parties and more likely to express closeness to political parties. With regard to donation, the British Election Study asked whether "during the last 7 days, have you done any of the following" including "Given any money to a political party, organization or cause". As the question only covers a very short time frame, the number of donations is fairly low, but Note: Standard errors in parentheses. Weighted by full wave 14 weights. * P < .05, ** P < .01.
since donations peak, during election campaigns, we use data from wave 13 collected shortly after the 2017 election. Overall 3.3% of our sample made donations and it was significantly higher amongst those who said they would vote in Britain's Best contest (4.3%) compared to those who would not (2.9%). 7 Similarly, we might expect that more expressive voters would display a campaign poster (Copeland and Laband 2002). We find that this is indeed the case with 6.6% of respondents who were likely to vote in Britain's Best contest displayed a poster, compared to 3.3% who were unlikely. 8 Turning to partisanship, we examine the extent to which our measure of expressiveness predicts concurrent party identification (PID) and PID strength. Model 1 (Table 5) shows a logistic regression model predicting whether each respondent either had no PID or weak PID (40% of the sample) compared with strong or fairly strong (60%). 9 The results indicate that lacking a strong PID is negatively related to voting in Britain's Best, controlling for the effects of age, gender and political attention. However, the effect is quite small, equivalent to about a five percentage point difference between someone who is very likely to vote in Britain's Best versus someone who is very unlikely.
These findings are supported by an analysis of the strength of identity (model 2). The dependent variable is an Item Response Theory score based on items designed to capture expressive partisanship (Bankert, Huddy, and Rosema 2017), a higher score representing stronger partisanship. Again controlling for age, gender and political attention, we find that there is a significant relationship between voting in the experiment and the strength of expressive partisanship. This is what we would expect if the instrument is capturing expressiveness, although the effect size is modest, the difference between the most and least likely to vote in Britain's Best being approximately 0.2 on a standardized scale.

Predictive validity
To assess predictive validity (Drost 2011), we examine the extent to which voting in Britain's Best contest predicts voting in the General Election in 2017. This is important because it is the ability to predict turnout in real-world political elections which motivated the search for a measure of expressiveness. We find that those least likely to vote in Britain's Best have a weighted non-voting rate of approximately 19% compared to only 13% amongst those who were most likely. 10 Although we have measured expressiveness independently of any political preferences, the difference could partly be due to a tendency of women, older respondents and politically interested people to vote in the experiment. We, therefore, add in the same basic control variables as in the previous models. The results, shown in Table 6 (model 1), indicate that our measure of expressiveness has a significant impact on the likelihood of voting in the General Election.
Having already established that those voting in the experiment are more likely to strongly identify with a political party, we ask whether the effect of party identification is moderated by expressiveness. The answer is revealed by model 2 and Figure 2 which shows a positive interaction between weak/ no PID and expressiveness. This merits some comment particularly since it confounds the expectation that identifiers should be more driven by expressive motives (Terry and Hogg 1996). One possible explanation for the interaction is a ceiling effect: turnout in the sample is very high (94%) amongst strong party identifiers. This means there is little room for party identifiers to become more likely to vote regardless of how expressive they are. However, for non-identifiers, there is a significant increase in the probability of voting the higher the respondent scores on the expressiveness scale. Another way of interpreting this is that people who strongly identify with a party will normally vote unless there is some practical or contextual reason not to (e.g. due illness or living in a non-competitive seat). This holds whether or not he or she is expressive. For non-identifiers, by contrast, some additional factor is required to motivate turnout. One such motivation is to express a preference for a preferred candidate, with more expressive non-identifiers opting to vote. Moreover, we have already shown above that more expressive respondents are more likely to identify with a party. Combining these findings, we may surmise that expressive citizens are more likely to identify with a party, but when they do not identify with a party, their expressiveness increases their chance of voting compared to less expressive nonidentifiers.

Conclusions
In this article, we have attempted to develop and validate an instrument for measuring the propensity of citizens to vote expressively. Until now survey items used to measure expressiveness have been inextricably linked to the expressive pay-off of voting for a preferred candidate. Given the importance of expressive voting to the theory of turnout overcoming this limitation is a potentially valuable endeavour. It is probably reasonable to conclude that the exercise has been, at least partially, successful. The design of the survey instrument ensured face validity for capturing expressiveness, insofar as there were no outcomes related benefits on offer in the fictional inconsequential election. This was supported by the absence of any empirical evidence of respondents being influenced by the likelihood of affecting the outcome, yet they were prepared to vote in large numbers. Moreover, tests of concurrent and predictive validity demonstrated that the instrument performed as expected across a number of scenarios. The qualifier that the exercise was only partially successful reflects a number of caveats that should be considered in further developmental work. First, the instrument relies on offering a pre-determined list of objects with which respondents might identify. A fair proportion of our sample (around two in five) failed to pick any of the options on offer. A more comprehensive list could be adopted, although it is possible that a failure to identify with any non-political objects may indicate a genuine lack of expressiveness. Indeed additional models not reported here suggest that failure to identify with any of the objects is negatively correlated with turnout and expressive partisanship. 11 Second, by restricting the list of options to non-political objects we cannot identify those people that are very expressive about politics but not about other things. Whilst this may be a common trait amongst political scientists, hopefully, it is not a major problem amongst the population at large. Third, the wording variant for the condition when the contest is competitive could be improved to make clearer that the election is neck-and-neck, rather than the less strong condition that there is a good chance of winning. Nevertheless, in practice, this is not crucial because the instrument can be used in the non-experimental format with only the "no-chance-of-winning" variant, as we have demonstrated that varying the P-term makes a negligible difference. Indeed the validation tests use this version of the question. Similarly, the follow-up validation study and future operationalization of the instrument does not require experimental manipulation of the cost. We have successfully shown that expressive benefits are cost sensitive and for the purpose of maximizing the usefulness of the instrument the cost of voting should be set at the equivalent of approximately £2. This was successfully adopted in the follow-up study.
The fourth and final caveat concerns the substantive findings from the tests of criterion validity. Whilst the empirical validation confirmed expectations and provided criterion validity for the instrument, the magnitude of the impact on turnout was only moderate. For comparison, the effect on turnout was roughly equivalent to a change in political attention from 5 to 7 on a zero-to-ten scale, and between respondents aged 20 and aged 60. There are a number of possible factors that might limit the impact of expressiveness, not least that the instrument is, by definition, non-political. It is therefore hardly surprising that it translates only weakly into the political domain (Copeland and Laband 2002). An alternative (but related) interpretation is that the instrument was successful at measuring propensity to expressiveness, but expressiveness is not an important factor in the voting decision. It is difficult to distinguish between these explanations, although the relatively high level of turnout in the experimental inconsequential election, and the lack of influence of P (the chance of being decisive), does support previous claims that expressive motivations are important. Additionally, the expectation that expressiveness should be more important for party identifiers was not supported, and there was some evidence of the opposite.
Despite these caveats, we hope that this attempt provides some new impetus to the seemingly long-abandoned search for an effective survey measure of the propensity to behave in a manner consistent with the expressive theory of voting. This measure could be used for a large variety of purposes including models of voter turnout since it is applicable to voters and non-voters alike. The availability of such a measure would therefore potentially open up new avenues for research, not only in the study of turnout but in other aspects of political behaviour including political participation, choice and communication.

Notes
1. All analyses of the randomised experiment are reported unweighted, but weighted analyses were also performed and produced very similar results. In the analysis of criterion validity where the dependent variable is subject to non-response bias standard BES full population weights are applied (full-weightW14). These adjust for a number of demographic and political characteristics. 2. The wording of the P clause might also be improved in future applications to encapsulate competitiveness rather than simply a good chance of winning which might imply that victory was assured. 3. We also tested the effect of education which did not add any explanatory power after including political interest (with which it is highly correlated). Because education level is not available for all respondents, we retain political interest rather than education to maintain sample size. 4. The intercept in these models can no longer be interpreted in a meaningful way as the intercept gives an estimate only for those belonging to the reference category (where age is zero). As above we report OLS, but substantively identical results are found modelling as a binary logit or an ordered logit. 5. How much do you agree or disagree with the following statements? It is every citizen's duty to vote in an election (5-point scale, strongly disagree … strongly agree), also measured in wave 14. 6. Respondents are asked how much do you like or dislike each of the main political parties, with responses on a scale from zero to ten (strongly dislike to strongly like). 7. Probability difference > 0 = .01 8. Probability difference > 0 = .00 9. As a percentage of all those included in the experiment weighted by standard BES wave 14 full weight (fullweightW14). Models of party identification and turnout are also weighed to address imbalance in the BES panel in regard to political enragement and voter turnout (Mellon and Prosser 2017). 10. The unweighted per cents are 11% and 9%, respectively. A test of difference in proportions (unweighted) shows the difference is significant at P < .001 11. Models shown in Tables 5 and 6 were re-run with 'no connection type' replacing likelihood of voting in Britain's Best and was statistically significant and negative (P < .01) for turnout and expressive partisanship, and positively correlated with weak/no PID but insignificant. Table A1. Connection type (in order of frequency).

Connection type Percent
Mean identity strength A town or city 17.2 5.7 A sports team 10.9 5.8 A local church or religious group 6.5 6.0 A charity 4.5 5.8 A musician or band 4.3 5.5 A social club or community group 3.9 5.1 A film/TV series 3.9 5.2 A School, college or university 3.4 5.5 A film star or famous celebrity 0.6 5.0 Don't know 6.9 None of the above 37.9 N 7732