Social and Political Identity

The tradition of social identity theory and intergroup relations was established by Tajfel (Reference Tajfel1974) and further developed by Tajfel et al. (Reference Tajfel1979). This theory argues that a person's social category (such as class, religion or ethnicity) and the feelings that person has towards that category provide a ‘self-definition that is a part of the self-concept. People have a repertoire of such discrete category memberships that vary in relative overall importance in the self-concept’ (Hogg, Terry and White Reference Hogg, Terry and White1995, 259). This original perspective of social identity is complemented by the notion of ‘self-categorization’ (Turner et al. Reference Turner1987). The self-categorization element of social identity is the cognitive process by which a person highlights differences from out-group individuals and similarities to in-group individuals. In other words, group identity emerges as a feeling of similarity to the ‘typical member’ or prototypical group member (Huddy Reference Huddy, Sears, Levy and Huddy2013).

Membership in a social group is a necessary but not sufficient condition for the emergence of social identity, and social identity is a necessary but not sufficient condition for political identity. Huddy (Reference Huddy, Sears, Levy and Huddy2013, 739) defines political identity as ‘a social identity with political relevance’. Following this definition, for group membership to become a political identity, two elements must be present: first, the presence of a salient social identity is required (Huddy Reference Huddy2001, Reference Huddy, Sears, Levy and Huddy2013); and, secondly, this identity must rest on ‘political content’ (Huddy Reference Huddy, Sears, Levy and Huddy2013, 739).

One implication of this definition is that, as Egan (Reference Egan2020) has argued, these identities are not merely antecedent to political attitudes. Rather, political contexts can affect them, especially through social sorting (Mason Reference Mason2015; Mason Reference Mason2018). The aim of this study is to operationalize an analysis of this phenomenon by identifying which social identities are more relevant in terms of perceived political commonalities.

When social identities are measured directly, comparisons of the relative salience of these identities tend to be difficult. This is partly a methodological issue, as many studies of competing social identities rely on direct one-item questions, without measuring relative strength (see, for example, Evans and Mellon Reference Evans and Mellon2016). Social identity measurements in nationally representative surveys have been criticized for, among other reasons, not adequately measuring the intensity of identification (Huddy Reference Huddy, Sears, Levy and Huddy2013; Wong Reference Wong2010), thereby making comparisons between different identities virtually impossible. On the other hand, more complex multi-item measurements, with identity scales, are typically designed for specific social groups, making comparisons across multiple dimensions of identity hard. For example, even the widely used multigroup ethnic identity measure (MEIM), developed by Phinney (Reference Phinney1992), is designed for comparisons only among individuals and groups in ethnic terms.

Data and Methods

This article presents a novel measurement strategy for assessing the relative salience of social identities for perceived political commonalities. Respondents were presented with two randomly selected profiles of fellow citizens, characterized with ten demographic categories frequently considered as determinants of voting behaviour. Respondents were then asked to assess, in terms of politics, with which of the two profiles they perceived they had more in common. An example of how this appeared to respondents can be seen in Figure 1.

Figure 1.

Survey prompt with example profiles.

The format of previous studies on the political relevance of social identities encourages respondents to think of demographic characteristics one at a time. In contrast, this design requires respondents to evaluate each attribute in the context of many at once, which means that there is an implicit trade-off between the different attributes. Additionally, because the task requires respondents to evaluate profiles with several attributes simultaneously, there is less risk of social desirability bias or conflating identity with sympathy. The fact that the prompt explicitly asks for political commonalities, rather than general closeness, comes from the definition of political identity as a ‘social identity with political relevance’ (Huddy Reference Huddy, Sears, Levy and Huddy2013, 739).

The profiles of characteristics presented to respondents were profiles of real people who were randomly selected from the respondents of the 2017 post-election BES. The randomization was carried out in a probability proportional to size (PPS) manner, with the probability of sampling proportional to the 2017 general election turnout weights. This meant that the profiles presented to respondents followed the distribution of British voters. This is important because it ensures not only that the distribution of individual characteristics is realistic, but also that the (joint) distribution of combinations of characteristics is realistic. If an arbitrary distribution had been used, the relative magnitude of the coefficients on different similarities could have reflected those arbitrary aspects of how different characteristics were put together in the experiment. The experiment, as designed, asks the question of how people choose which of two people is closer to them politically, where those two people are sampled from Great Britain's voting population, rather than from a distribution made up by the experimenter.

This method is closely related to traditional conjoint experiments (Bansak et al. Reference Bansak, Druckman and Green2021); however, instead of a distribution where profile attributes are independently randomized, I have randomly selected complete profiles following their actual distribution in the population. Since the estimates derived from conjoint experiments are obtained by averaging over the treatment distribution, using this external benchmark for the distribution helps with the external validity of estimates but implies that I cannot use the non-parametric identification of the causal effects of individual attributes, introducing model dependence into the analysis of the data (De la Cuesta, Egami and Imai Reference De La Cuesta, Egami and Imai2022).

The profiles' characteristics included in the study were: gender, age, religion, region, home status, education, annual household income, subjective class and subjective family class. The possible levels for each characteristic are detailed in Table 1 in the Online Appendix. The experiment was fielded by YouGov between June and August 2019. The prompt was presented five times per respondent, with different profiles each time. The order in which the attributes were listed was randomized per respondent. A total of 1,656 respondents from Great Britain were recruited for the experiment (8,280 responses). In the analyses, I use sample weights provided by YouGov that make the data representative of the British population on standard demographic and past vote variables.Footnote ¹

To analyse the results of the experiment, information on the respondents' self-categorization is combined with that of the profiles presented to them and their choices. The analysis seeks to assess the probability of a respondent choosing one profile if that profile shares an attribute with the respondent and the other alternative profile does not.

To perform this analysis, two dummy variables, m _Aij and m _Bij, are created to reflect whether the respondent's self-categorization on task i matches profile A on attribute j and whether it matches profile B on the same attribute j (two dummy values per iteration per attribute). The difference between the two dummy variables is the explanatory variable of interest, d _ij = m _Aij − m_Bij. If, for task i, the two profiles present the same levels as the respondent on attribute j, or neither of them do, d _ij will be 0. If the matching attribute is only 1 in the first person (‘Person A’), then d _ij will have the value of positive 1. If the matching attribute is only 1 in the second person (‘Person B’), then d _ij will have the value of −1. The choice of the respondent for task i, that is, the outcome of interest, c _i, is then coded in an equivalent way, with 1 meaning the respondent chose Person A, −1 meaning that the respondent chose Person B and 0 meaning the person chose ‘not sure’. The reason the data are coded this way is that this means that matches with A and B are treated symmetrically and each coefficient describes the effect of moving from no match to a match for a single characteristic, holding the other attributes and the other profile constant.

With these variables, the following subsections will examine how important every d _ij is for every attribute j in explaining the respondents' choice c _i. I will then examine if this relative importance of the attributes varies by respondents' past political behaviour (electoral choice in the 2017 general election and the EU referendum).