Ideology and the left-right continuum

One of the most enduring frameworks for categorizing people based on their political beliefs is the left-right spectrum, named after the seating arrangement in the National Assembly during the French Revolution (those who wanted change were seated to the left; those who favoured tradition, to the right). It has been argued that this categorization is a necessary and useful high-level abstraction and therefore that, “if the left-right distinction did not exist, scholars of ideology would need to invent its equivalent” [3] (p. 5). There is evidence that majorities of citizens across many different countries are comfortable locating their ideological commitments along a simple, left vs. right interval scale [11] and describe the distinction between the two poles in similar terms [12].

Some scholars have therefore argued that this left-right distinction is a durable feature of human nature and point to the fact that individuals who want to conserve institutions and maintain valued traditions, and those who wish to reform and change society, are recognisable in the politics of societies extending as far back as the Ancient Greeks [13]. It has been further argued that preference for these two poles reflects psychological characteristics such as (on the right) epistemic and existential needs that lead to cognitive rigidity [3,6]. Nonetheless, the characterization of this distinction has been the subject of controversy. Some scholars, for example, have argued that the tendency to construe ideology in terms of two poles reflects the didactic, adversarial nature of politics, so that the idea of a spectrum has persisted while its defining features may have changed over time, organized around attitudes towards traditional institutions in the past but, increasingly, as multiparty democracies have become the global norm, organized around attitudes towards inequality [14]. These two conceptions of what it means to be ‘left’ or ‘right’ may not be entirely conflicting as, historically, hierarchies and traditions have been construed as impediments to those who seek social and economic justice [15].

Nonetheless, the conventional way of construing the left-right spectrum faces three empirical challenges. The first is that factor analytic studies have rarely revealed a single dimension of belief corresponding to the spectrum. For example, some researchers have reported models in which there are separate dimensions reflecting social and economic conservatism versus liberalism [9,10]. Although typically highly correlated in the advanced economies of the west, when considered as predictors of preferences for specific policies these dimensions may sometimes align together but at other times may be in conflict [16]. A further complication is that, in the nations of Eastern Europe which, until 1989, lived under communist rule, socially conservative attitudes tend to be associated with left-wing economic attitudes [17] presumably because, in those countries prior to the fall of the Berlin Wall, redistributive, egalitarian policies had been the official doctrine of the governing elites and therefore represent the status quo.

In probably the largest cross-national study of political attitudes undertaken to date Malka and colleagues [18] used data from 325,804 people in 99 nations, collected as part of the World Attitude Survey. The researchers found that, globally, it was more common for socially conservative and economically liberal attitudes to be negatively correlated, albeit often to a small degree, than for them to be positively correlated, as in Western industrialised nations. Hence, the social and economic dimensions may have to be weighted differently when discriminating between right- and left-wing political parties depending on which countries or contexts are considered [19].

A second challenge is that, since the earliest days of research on political attitudes [20,21] it has been known that many citizens routinely lack coherent or ideologically constrained political attitudes [22] and instead rely on cognitive shortcuts when making their voting decisions [23,24]. For example, they may vote on the basis of a loosely felt sense of identifying with a particular party, or based on retrospective judgments of their well-being so that incumbents lose votes from voters who, for whatever reason, feel that they have not been flourishing [25].

Finally, there is some evidence that people who categorise themselves at either pole of left-right spectrum may be in some ways more similar to each other than to people who are non-ideological. For example, people who identify as belonging to either of the political extremes, compared to those who locate themselves in the centre of the spectrum, are much more likely to report certainty that their political beliefs are correct [26], are more likely to say that their beliefs are superior to those of other people, and are more likely to maintain their political beliefs over long periods of time [27]. Moreover, people who describe themselves as left or right-wing seem to be less influenced by contextual information than non-extremists (if asked to guess, say, the population of a large city like Chicago their estimates are less influenced by prior hints) [28].

Political ideologies as belief systems

A seminal contribution to understanding the organization of political ideologies, which contrasted with the factor analytic approaches just reviewed, was made by the American social scientist Philip Converse [21]. His analysis (which, he argued, applied not only to ideologies but to any organized set of beliefs) began by defining a belief system as a network or “configuration of ideas and attitudes in which the elements are bound together by some form of constraint or functional interdependence” (p.207).

Such a system has static constraints as evidenced by the fact that one attitude can very often be predicted given knowledge of others, but Converse also recognised the dynamic constraint that, over time, occurred when changes in some elements within the system would necessitate adjustment to others. Sometimes these constraints are logical (e.g., it is illogical to insist that government expenditure be increased, revenues are decreased, and the budget be balanced at the same time). However, the observed pattern of association between beliefs within a system often cannot be explained in this way (e.g., there is no logical reason why, in countries in which social and economic attitudes are aligned, people who support more spending for the armed forces also, typically, support the use of capital punishment). In these circumstances, constraints might be due to some hidden psychological processes that influence both kinds of beliefs (presumably authoritarian traits in the example just given) or the consequence of the way that ideas have been packaged together prior to dissemination by influential belief entrepreneurs (e.g., political theorists and their activist followers).

An innovative concept introduced by Converse was centrality (e.g., see pp.208–209): some beliefs within a system are more central than others, in the sense that a perturbation of these beliefs will require more radical adjustment of the system (imagine a Marxist deciding that class struggle is a relatively minor force in human history), compared with others that are more peripheral. Thus, Converse’s insight was that the structure of belief systems would depend on the position of an individual within a ‘vertical hierarchy’ down which political information flows, with opinion-forming political elites at the apex, activists and committed party members below them, and ‘mass publics’ at the base. Elites and activists often spend many years discussing and refining their belief systems, so that the individual elements are understood in detail, the relationships between them tightly specified, and contradictions within the system either eliminated or reasoned away with exculpatory hypotheses. However, ordinary citizens may be much less reflective about their political commitments and, as a consequence their belief systems are much less constrained.

Converse supported this account with detailed analyses of qualitative data obtained from large numbers of American voters in the 1950s and early 1960s. His analyses identified five strata of voters: ideologues (estimated to be 3.5% of voters); near-ideologues who could relate policies to abstract political concepts (12%); those whose political commitments were aligned with specific group interests (45%); those who evaluated parties by temporal association (e.g., in the UK, someone might associate the Labour Party with the 2008 financial crisis) or because of gratitude for or dislike of specific policies (22%); and, finally, those who had no policy interests whatsoever but were simply loyal to a particular party or liked a particular candidate (17%). Importantly, in this scheme, only a fraction of voters (less than 20%) structured their political commitments primarily according to any kind of political philosophy. In these voters, compared with those in the third to fifth tiers, attitudes towards policies were highly correlated, and these more ideologically constrained citizens were much less likely to change their vote between one election and another.

Network models of political belief systems

Although Converse’s account has been considered a seminal contribution to the analysis of ideologies, the factor analytic methods which have been available to investigate the structure of political beliefs have been ill-suited for testing and elaborating his model. Nonetheless, some more recent observations made by political scientists are consistent with his original theory. For example, Feldman and Johnston’s [10] observation that attitudes towards social and economic conservative policies are most highly correlated at both ends of the political spectrum but much less so at the centre is entirely consistent with Converse’s original theory. Similarly, Baldassari and Goldberg [29] used a novel technique, relational class analysis, to identify groups of American voters who were similar in their organization of attitudes towards economic issues, civil rights, morality and foreign policy (the correlations between the issues were similar, even though their position with respect to each of these domains might be different) and identified three classes: ideologues (33%) whose beliefs corresponded to the traditional left-right dimension, alternatives (40%) whose attitudes towards economic issues contrasted with their attitudes towards social issues (right economic while left social or vice versa), and agnostics (27%) whose beliefs in these different domains were only weakly associated with each other. The researchers found that the same pattern could be found over twenty years of the American National Election Survey (even though different questions were used and different participants were recruited in the different waves).

Network theory methodologies, which have their origins in physics [30] and have recently been used in the study of psychiatric disorders [31], are uniquely suited to pursuing Converse’s agenda further, so much so that the coincidental alignment of these methods with his theoretical approach seems remarkable. These methodologies abandon the assumption that beliefs (or psychiatric symptoms) necessarily cluster together into recognizable ideologies (syndromes) because of common underlying causes such as specific cognitive traits (or underlying disease processes) but instead assume that the relationships between them emerge as a consequence of the way they can influence each other [31]. In respect to belief systems, this is easiest to see when there is some kind of logical relationship between the constituent propositions. For example, support for tax cuts implies support for limiting benefits and not for spending more money on defence, so that changing one of these attitudes will likely lead to changes in the others. However, there is no assumption that the relationships between attitudes are always logical in this way.

New analytical techniques allow the visual representation of networks, revealing putative interrelationships between the elements. The nodes in these visualisations correspond to the elements that have inter-relationships which we are trying to understand – in our case, attitudes towards particular political policies and ideas (in psychopathology they would be symptoms). The ‘edges’ or lines connecting the elements in the network represent the magnitude of the influence between them [32]. For example, within a specific network it may be that some nodes (attitudes, say towards multiculturalism and gay rights) are highly associated with each other (so that people who favour one will usually favour the other) whereas others (say between benefits for poor and international cooperation) are less so. Once networks have been specified, it is possible to identify communities of nodes that are highly interconnected and therefore ideologically constrained (changing one of the attitudes would very likely require changes in the others). Various centrality statistics are also available to estimate which nodes are most influential within the network as a whole by virtue of their connections with other nodes (in network analyses, the concept of ‘centrality’ exactly corresponds to Converse’s use of the term more than sixty years ago). When networks are estimated for different subpopulations, network comparison statistics are available to test each network’s overall connectivity, as well as the strength of individual edges for different groups. Hence, it is possible to determine whether different groups of citizens differ in the extent to which their networks are ideologically constrained.

Although the majority of psychological applications of network theory have thus far been in the field of mental health, these analytical techniques have been used to address psychological and social problems. For example, Causal Attitude Network (CAN) models, in which behavioural outcomes are seen as products of the interactions between a range of attitudinal and emotional factors connected in a network, have been used to explain the notoriously difficult to predict relationship between attitudes and actions. Dalege et al. [33] used a CAN model to study voters’ decisions to support specific candidates in US elections between 1980 and 2012. Attitudes were found to better predict voting behaviour when the elements in the network were highly interconnected (e.g., during the 2012 election of Obama) than when the relationship between the elements was weak (e.g., the election of G.H.W. Bush in 1992). Moreover, in each election the most central node (e.g., the belief that Clinton would show strong leadership in 1992) was the strongest single predictor of voting behaviour. In a subsequent study, the same researchers [34] showed that American voters who reported being highly interested in politics had more highly connected networks than those who did not, and that connectivity of their networks predicted not only voting behaviour but stability of their attitudes over time.

Similarly, Boutyline [35] constructed network models of American voters’ attitudes during the 2008 Presidential Election, including ideology (self-identification as liberal vs conservative) within the networks. The structure of the networks did not vary substantially between demographic groups, and ideology was the central node, a finding that is consistent with previous demonstrations that many ordinary people do organise their political beliefs according to the left-right distinction [11,12]. Voters who were classified as ‘high information’ (e.g., better educated) had more connected networks than those classified as ‘low information’ (e.g., less educated), as predicted by Converse’s account. Similarly, Brandt and his colleagues [36] studied panel data from New Zealand voters collected between 2008 and 2015, including symbolic (support for particular parties or identification as left vs right) and operational (policy) beliefs in their analyses, revealing that, across multiple years, symbolic nodes were more central in the networks.

In sum, previous research has examined the relationships between ideological and political variables by including them all as elements in a single network. Here, however, we attempt to assess how ideology moderates the relations between policy preferences across a range of different issues. To achieve this goal, it is necessary to examine variations in political attitudes among those whose ideological commitments are determined independently of the elements included in their networks. In the present study, we thus report a psychometric network analysis of policy preferences in a representative sample of the UK public in late 2021. Converse’s model [21] leads us to predict that networks of policy attitudes should be most highly connected in those with either strong right- or left-wing commitments. We tested this prediction by comparing those who described themselves as aligned with the left or right with those who described themselves as belonging to the centre of the left-right political spectrum. We expected the networks to be not only highly interconnected at the extreme ends (left and right) but to also differ in other local and global features. Therefore, we also compared the networks in several other aspects: centrality and predictability indices, as well as their global community (or factor) structure. Finally, as our dataset included childrearing questions which have previously been shown to predict authoritarian predispositions [37–39], we used the opportunity to assess the external validity of our network community structure against this measure.

Sample

Data come from the sixth wave of the Covid-19 Psychological Research Consortium longitudinal survey, which tracks the effect of the coronavirus pandemic on the mental health, attitudes, and behaviour of the UK population [40]. Data for the first wave of the survey were collected in March, 2020, when strict lockdown measures were announced. Quota sampling was used to recruit 2,025 adults (i.e., those over the age of 18 years old) stratified by age, sex, and household income to be representative of the UK population; validation checks conducted at that time found that the sample was also broadly representative on many other variables, including diversity of political allegiances as judged by self-reported voting in the 2016 referendum on Britain’s exit from the European Union, and also in the 2019 UK General Election. Subsequent waves re-contacted participants from earlier waves and then recruited further respondents to replace those lost to panel attrition. As a consequence, subsequent waves were slightly less representative than the first (but still quite diverse).

Wave 6 data were collected between 6^th of August and 28^th of September 2021 and consisted of 1,643 people who had completed one or more earlier survey waves, and an additional 415 newly recruited participants (n = 2,058). Females (n = 1,069) were slightly over-represented compared to males (983; 4 transgender, 2 prefer not to say). Participants came from across the four constituent administrations of the UK (1,702 from England, 205 from Scotland, 108 from Wales and 43 from Northern Ireland; these numbers conform closely to the proportions of the UK population reported by the UK Office of National Statistics of 84.4%, 8.05%, 4.5% and 2.8% respectively [41]). 42.7 percent were university graduates (as opposed to 33.8% with a graduate-level qualification according to the 2021 UK census of England and Wales). When asked about their social class, 42.3% described themselves as working class, 39.6% as middle class, 1.3% as upper class, and 16.9% said they did not belong to a class. Of the 69.7 percent who reported voting in the 2016 referendum on Britain leaving the European Union, 41.8 percent reported voting leave, so remain voters were over-represented (58.2 percent); no data on voting in the 2019 general election was available from this wave. Ethical approval for the study was granted by a blinded for review process (University of Sheffield ethics ref no. 033759). The full dataset used in the current study is publicly available at the Open Science Framework: https://osf.io/v2zur/.

Measures

There was no missing data because our survey software required responses for every item before moving on.

Political attitudes.

Participants reported their preferences toward eighteen political issues, using a scale modelled on Wilson’s [42] questionnaire measuring right- vs left-wing attitudes. The issues were chosen to reflect potential social and economic dimensions of ideology that are relevant to current times; these were: Support for “the death penalty”; “spending money on the armed forces”; “multiculturalism”; “stiff jail terms for criminals”; “euthanasia (physician assisted suicide)”; “gay rights”; “higher benefits for the poor”; “immigration”; “legalised abortion”; “lower taxes to promote business”; “international government”; “rehabilitation of offenders”; “traditional family values”; “monogamy”; “trans rights”; “Brexit”; “public demonstrations such as ‘taking the knee’ to acknowledge discrimination against ethnic minorities”; and, finally, “international aid.” Support for each policy was recorded on a five-point Likert scale (1 = “strongly disagree”; 5 = “strongly agree”). The validity of the items is established by their relationship with self-identification on the left-right spectrum, shown in the results, Table 1. It is important to note that the network psychometric approach taken in this study does not make the assumption that all of these items form a scale; nonetheless, Cronbach’s alpha for the items (with items rescored so that high scores always reflect the choices made by self-defined conservatives as shown in Table 1) was.83.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics and item analyses for individual policy attitudes.

https://doi.org/10.1371/journal.pone.0333595.t001

Analysis plan

Our analyses were conducted using the statistical software R (version 4.1.2). Listwise deletion was employed to handle missing values. Participants were divided into three groups based on their self-reported position on the left-right spectrum (ranging from 1 “Left” to 10 “Right”), which was approximately normally distributed. Respondents with scores of 3 or less were assigned to the left-wing group; those with scores of 8 or above were assigned to the right-wing group; and the rest (that is, those who scored 4, 5, 6, or 7) were classified as centrists. Given the ordinal nature of belief responses, the Wilcoxon rank sum test was used to assess (median) differences in their endorsement across the political groups (which were expected for most, if not all, of the political beliefs). In order to correct for multiple testing (here as well as elsewhere), Bonferroni correction was employed.

General comparisons

Table 1 gives descriptive statistics for the 18 political attitudes, as well as their differences across the ideological groups. From this table, we can see that those on the left and centre differed significantly across nearly all political beliefs. Conversely, fewer differences amongst policy preferences were identified between the right and the centre; however, there were still notable (median) differences for the following issues: P1 (Death Penalty); P2 (Spending money on Army); P7 (Higher Benefits for the Poor); P10 (Lower Taxes); P13 (Traditional Family Values); P14 (Monogamy); P16 (Brexit) (all p values < .001). The left and right differed significantly across all political beliefs, further indicating that the participants’ self-assessment of their ideological orientation mapped onto common definitions of constraint in our data. As a further validation of the left-right self-placement scale we created a sum score for conservativism from the 18 individual policy items (reverse scoring where appropriate so that high scores always indicated right-wing policy preferences); as expected, the correlation with self-placement scale scores was highly significant, r = .504, p < .001.

Comparison of political networks

The three political belief networks (Left, Centre, and Right) are displayed in Fig 1A. Stability and accuracy analyses (reported in the supporting materials) suggested that our network parameters were reliably estimated. Moreover, a complete re-analysis using random and equally-sized sub-samples and alternative estimation procedures replicated the current patterns, adding to their robustness.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Gaussian network models of political beliefs.

Section (A) shows models comprising political beliefs for each ideological group: Left, Centre, Right. Blue vs. red lines (or edges) represent positive vs. negative partial correlations, respectively, with thicker lines indicating stronger correlations. The predictability (variance explained) in each node is displayed around its circumference. Section (B) shows line graph displaying the strength centrality (i.e., extent of absolute connectivity level) for each node (variable) in the three networks. Section (C) shows Global Strength comparison results, revealing significantly higher global strength estimates for the Left and Right networks, compared to the Centrist network.

https://doi.org/10.1371/journal.pone.0333595.g001

Community structure.

The community structure of each network was examined using the walktrap and spinglass algorithms. These procedures revealed the exact same community structure in all three networks (see Fig 2). The first community reflected political attitudes that correspond to classic conceptions of right-wing authoritarianism [58,59]. The second community reflected a willingness to cooperate with, help and accept others even if different (e.g., P3, “Multiculturalism”; P8 “Immigration”); while it was more difficult to find a succinct name for this community we suggest that it reflects altruism and cooperativeness. The third and final community also reflected personal freedom and included issues that commonly align with social liberalism (e.g., P6, “Gay rights”; P15, “Trans rights”; P9, “Legalized abortion”); it was thus labelled as personal liberalism. Although our initial analyses pointed toward a two-factor structure (with the last two communities collapsing into a single liberal domain), the current three-factor solution was deemed more appropriate given its high prevalence in a supplementary bootstrapping procedure (1,000 iterations; 72.5% frequency in bootstrapped distribution of replica networks) (see supporting information).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Community structure of the entire sample.

Three communities are found in all political networks: namely, RWA = Right-wing authoritarianism, ALT = Altruism and cooperativeness, LIB = Personal liberalism.

https://doi.org/10.1371/journal.pone.0333595.g002

Factor scores on these three communities (which reflect the position of an individual on each latent variable) revealed that all three distinguished individuals from the different ideological groups in the expected way: For example, people on the left scored significantly higher on both altruism and cooperativeness and personal liberalism (compared to those on the centre and right); while those on the right scored significantly higher on right-wing authoritarianism (also, compared to the centre and left; see Table 2). A structural equation model (RMSEA = .07, CFI = 0.91, TLI = 0.90) showed that, as expected, child-rearing practices were a robust predictor of right-wing authoritarianism (β = 0.44, p < .001); and a strong negative predictor of altruism and cooperativeness (β = −0.32, p < .001) and personal liberalism (β = −0.31, p < .001).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Political group differences in community (factor) scores.

https://doi.org/10.1371/journal.pone.0333595.t002

Node predictability.

The amount of explained variation in each node is outlined in Table 3 and is also graphically represented by the charts around nodes in Fig 1A. Several similarities and differences were detected in the predictability indices among the networks for those on the Left, Centre, and Right. For instance, P6 (“Gay Rights”) and P15 (“Trans Rights”) scored similarly high across the three networks. The predictability indices that differentiated the networks between the Left and Right included P7 (“Higher Benefits for the Poor”), P9 (“Legalized Abortion”), and P16 (“Brexit”), which were higher in the left network; and P5 (“Euthanasia”), P10 (“Lower Taxes”), and P11 (“International Government”), which were higher in the right network. Notably, the nodes that exhibited the highest levels of predictability within the Centre network scored similarly high in the networks for those on the Left and Right; these include: P3 (“Multiculturalism”), P4 (“Stiff Jail Terms”), and P14 (“Monogamy”).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Predictability indices of particular nodes across the three political belief networks.

https://doi.org/10.1371/journal.pone.0333595.t003

Limitations

The current study should be viewed in light of certain limitations. Although we used a large, diverse sample that was fairly representative of the UK population, some of the political constructs included (e.g., Brexit) were unique to the UK political landscape. There were more items that related to social conservatism than to economic conservatism, which may also limit the generalizability of our findings. It is possible, for example, that a different structure would have emerged if we had included more economic items, perhaps with a specific community relating to economic beliefs, and perhaps reducing the influence of social items, for example relating to gay and trans rights. Given recent observations of cross-cultural differences in the relationship between social and economic conservatism [18] it is also possible that different communities of beliefs would be found in different regions of the world. Similarly, the survey was conducted during the COVID-19 pandemic at a time of great social turbulence, which may also limit generalizability to different periods of time. Although we doubt whether cross-cultural differences would be found with respect to our primary observation that those who locate themselves at the political extremes have more closely connected networks of beliefs (because we think this is a general property of human belief systems) it would certainly be desirable to replicate this work in other geographical locations.

It is also worth noting that the procedure we employed here to categorise individuals into the left, centre, and right is more data-driven than theoretical. However, as already noted, this is standard practice in political surveys and research in political psychology and there is good evidence that most people can assign themselves along the left-right political spectrum [11] and describe the distinction between the two poles in similar terms [12]. Moreover, our between-group comparisons show that these self-categorisations reflect actual differences in policy preferences, and there was the expected correlation between overall support for right-wing policies and self-placement on the left-right scale. Finally, our sensitivity analyses show that our main findings were unaffected when we narrowed the definition of the centre group and widened the definitions of the left and right groups.

We must also acknowledge that the childrearing items we used to validate our community structure had marginal reliability. Our structural equation model found robust relationships with each of the communities. As we would expect the poor reliability of these items to attenuate these relationships, we are reassured by their strength and by the fact that the model fit indices were satisfactory. Nonetheless, this finding should be treated with caution and future studies might consider a wider range of measures to validate observed community structures of political beliefs.

Finally, much network psychometric research has employed the Ising model and has included beliefs along with other variables of similar nature (e.g., feelings). However, in this study, we have modelled political beliefs in a Gaussian manner because Ising models only accommodate binary data. Albeit not a limitation in and of itself, it is worth noting that, in practice, the two models often produce similar results [67].