Party Polarization and Mass Partisanship: A Comparative Perspective

For reviews of this research, see Fiorina and Abrams (2008), Hetherington (2009), and Layman et al. (2006).

I use the terms partisanship, party attachments, and party identification interchangeably to refer to an individual’s self-identification with a political party.

I conceive of partisan attachments as lying on a continuum. Each individual has in mind how attached she feels to a party, and as that attachment increases, she gets closer to an arbitrary threshold above which she is willing to tell an interviewer that she identifies with the party. As a result, any increase in the intensity of attachment will also increase the probability of moving from the non-partisan column to the partisan one. I therefore discuss the conceptual effect of party polarization interchangeably as “strengthening” partisanship or “increasing the likelihood of” partisanship. Due to data limitations, I empirically measure only the latter manifestation.

Both sets of theories of partisanship imagine that citizens evaluate (however subconsciously) parties on some salient political dimension. What dimension that is may vary across individuals, allowing party systems to polarize on one dimension without necessarily polarizing along another. The more concrete implication from these theories, then, is that party polarization along a particular dimension strengthens partisanship among individuals for whom that dimension is most salient. Much of my empirical analysis assumes that for most people, the economic left-right dimension matters most—an empirical regularity that is widely documented—but I also examine survey items that do not impose that particular dimension.

This hypothesis may seem to fly in the face of a conventional wisdom in U.S. politics that Americans prefer bipartisanship (Hibbing and Theiss-Morse 2002; Ramirez 2009; but see Harbridge and Malhotra 2011). However, it would not be inconsistent for individuals to both prefer bipartisanship or consensus consciously and for such party cooperation to weaken attachments through the more subconscious process of identity formation.

In theory, this relationship involves mediation: party polarization increases perceived polarization, which in turn increases partisanship. Analyzing mediation effects requires strong modeling assumptions, and better methodologies currently being developed (e.g., Imai et al. 2010) are not yet equipped to handle multilevel data. I therefore conduct what some scholars refer to as implicit mediation analysis by demonstrating the three relationships involved in the theorized causal pathway.

Descriptive statistics and information about survey methodology for these studies are provided in the supplementary appendix.

The items on party positions reference a limited number of parties (six in the first CSES module and nine thereafter), so we might be concerned that they leave out important parties in very fragmented systems. Therefore, I exclude studies in which the parties referenced in the survey together received less than 80 % of the vote in the relevant election. Setting the threshold at 90 % of the vote excludes additional surveys but does not substantively change my results (see supplementary appendix).

I include only those surveys conducted in minimally democratic settings, defined as those receiving a positive Polity score. The countries included are Austria, Brazil, Bulgaria, Canada, Croatia, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Japan, Mexico, Netherlands, New Zealand, Norway, Peru, Philippines, Poland, Portugal, Romania, Russia, Slovakia, Slovenia, Spain, South Korea, Sweden, Switzerland, Taiwan, Turkey, United Kingdom, United States, and Uruguay. I omit surveys conducted in Albania, Australia, and Belgium because the items on either partisanship or party placement were not comparable to the rest of the sample.

In U.S. studies, the latter respondents are typically referred to as leaners, although they behave more like outright partisans than true independents (Keith et al. 1992; Petrocik 2009). Indeed, Barnes et al. (1988) show that leaners do consider themselves “close” to a political party when asked the CSES-style item. Most analyses of partisanship with ANES data employ an ordinal scale in which leaners are considered weak partisans. My ANES results are substantively equivalent using the ordinal scale (see supplementary appendix), but I report results with the dichotomous measure for comparability with the CSES analysis.

The specific wording is, “In politics people sometimes talk of left and right. Where would you place [party name] on a scale from 0 to 10 where 0 means the left and 10 means the right?”

The correlation between respondent- and expert-based party placement is 0.45. Another set of expert surveys conducted by scholars based in Chapel Hill offers similar measures of party placement (Hooghe et al. 2010). Those surveys were conducted in 1999, 2006, and 2010, but only in the subset of European Union member countries. Their advantage is that they rely on the responses of multiple experts per country, making their measures more reliable. Matching each country-year in my CSES dataset to the temporally closest Chapel Hill survey, my results are substantively equivalent (see supplementary appendix). I also employed a measure of economic left-right placement from the Comparative Manifestos Project, which is available for most of the countries in my sample. These placements are based on the number of times particular keywords appear in the preelection manifesto of each party (Budge et al. 2001). Again, my results are substantively similar using this measure (see supplementary appendix).

These vote shares are the absolute percent of the popular vote received by each party in the national lower house elections, as reported by the CSES. In mixed electoral systems, this is the proportional portion of the vote. If lower house elections were not the subject of the election study, I use presidential vote shares. Dropping the presidential cases from the analysis does not substantively change my results (see supplementary appendix).

Since the CSES does not ask respondents/experts to place every party in the system, an unweighted measure using these data will also be highly misleading in cases where a significant number of small parties are omitted.

The placement question in the ANES offers respondents only 7 categories, as compared to the 11 in the CSES. The question states, “We hear a lot of talk these days about liberals and conservatives. I’m going to show you [from 1996 on: Here is] a seven-point scale on which the political views that people might hold are arranged from extremely liberal to extremely conservative. Where would you place the [Democratic/Republican] Party on this scale?” Respondents who did not place themselves on the ideological scale were not asked to place the parties. Beginning in 1984, a follow-up question was included that asked respondents where they would place themselves ideologically “if you had to choose.” This significantly reduced non-response and thereby increased the number of respondents asked to place the parties. Limiting my sample to surveys from 1984 on does not substantively change my results (see supplementary appendix).

The question asked, “During the election campaign, would you say that there were major differences between the [parties/candidates], minor differences, or no differences at all?” My variable distinguishes those who saw “major differences” from the other response categories.

The item was not asked in the 1974, 1978, and 1982 studies.

A preferable measure would more directly capture the respondent’s political knowledge, but comparability and reliability across countries and time pose serious challenges. Still, educational attainment and political knowledge are highly correlated (e.g., Delli Carpini and Keeter 1996; Highton 2009).

I include a squared age term to capture a potential quadratic relationship, whereby partisanship levels off after a certain age (see Converse 1969).

Huber et al. (2005) argue that these two effects can be estimated separately by including the effective numbers of electoral and legislative parties. In my dataset, these variables are correlated at 0.92, which precludes including both.

Since the CSES reports the vote shares of only the parties covered in the survey, they do not always sum to 100 %. I count the missing votes as a residual category and employ the correction proposed by (Taagepera 1997).

The data on party age come from Brader et al. (2013a) for modules 1 and 2 of the CSES. For module 3, I collected my own data on party age.

As already noted, fuller consideration of aggregate variation must also account for the ideological distribution of citizens. But my interest is in the individual-level relationship. I include this figure merely to illustrate the cross-national variation in these data.

In some models, fewer than 25 % of respondents were union respondents, so rather than representing the effect of the interquartile range (which would be zero), the figures simply show the effect of becoming a union member.

Among the countries with more than one survey in the CSES sample, the average change in party polarization is 1.04, roughly half the interquartile range. The average span of time between surveys is 4.5 years, so two decades of continual polarization would roughly double the effect in Fig. 2.

This finding is consistent with Hetherington’s (2001: 627) results from a more limited ANES sample.

The gold standard for identifying a causal relationship is through an experimental manipulation, although experiments often also entail costs to generalizability. I report elsewhere on a survey experiment I conducted (Lupu 2013) and focus here on employing panel survey data to improve causal inference.

Indeed, there is quite a bit of variation on both partisanship and perceived polarization over the course of the panel. The polychoric correlation of partisanship across waves is, on average, 0.55; that of perceived polarization is, on average, 0.45.

Another potential problem with panel survey data is attrition. Across the three reinterviews, the average retention rate was a remarkably high 82 %. Jennings et al. (2009: 783) also note that, “panel status never accounts for over 2 % of the variation in the scores of explicitly political measures.”

Respondents in the PSS were asked to place the two U.S. parties on an ideological dimension only in 1965, so I am unable to use the distance-based measure of perceived polarization.

The same index is used by Highton (2009).

Katz and Mair (1995) offer an explanation of partisan erosion in Western Europe that relies in part on policy convergence, although the micro-level foundations of their theory are quite different from those suggested here.