Measuring issue salience divergence
Issue salience divergence is the degree to which parties in a party system emphasize issues to different extents.
Sigelman and Buell (2004) propose an index for the US case, but this results only in dyadic comparisons. This works well in a two-party system where both parties receive identical scores. In multi-party systems, scores for individual parties and party-systems as a whole are necessary to compare across cases and time periods. In this section, I walk through my operationalization that explicitly connects party-dyads to party-election and country-election level scores, followed by a concrete example for the generation of my measure in the UK’s 1979 election. Additional descriptive statistics for issue salience divergence beyond those discussed later in this paper can be found in
Supplementary Appendix B.
I conceptualize issue salience divergence as a spectrum. At the convergent pole, parties issue emphases are identical. There is no distinction between parties in terms of how they allocate their attention across issues. At the divergent pole, parties’ issue emphases are not just different, but opposites. At the center of the spectrum is the midpoint where parties’ issue salience distributions are unrelated. Therefore, I argue that a correlation-based operationalization matches the concept as it captures both the strength and direction of the relationship between party dyads. To do this, I need a data source for parties’ issue salience across relevant domains and a way to scale these correlations to the relevant level of analysis.
I measure issue salience using MARPOR Project data. This data source takes each quasi-sentence from party manifestos and categorizes them into mutually exclusive topics. These data provide a consistent procedure for measuring this manifestation of the party agenda across countries and elections. The data also represent an unmediated depiction of the parties’ issue emphases. While no data set is perfect (
Gemenis, 2013), the MARPOR project’s geographic and temporal inclusiveness, as well as its focus on issue attention, makes it an ideal window into the issues that parties want to make important in elections.
While previous studies have also used manifestos as a data source (
Dolezal et al., 2014;
Sigelman and Buell, 2004: 658), results using manifesto data may not perfectly generalize to other data sources. As longer documents, election programs discuss a wide range of issues and are released prior to the actual campaign, thereby incentivizing parties to cover a broad range of issues. Other indicators of the political agenda—such as campaign advertisements (
Banda, 2013,
2015;
Lipsitz, 2013), press statements (
Seeberg, 2020a), and candidate speeches (
Sigelman and Buell, 2004)—are more dynamic and targeted representations of the political agenda and may, therefore, present more divergent issue emphases. Because the goal of this paper is to investigate the trends and determinants of issue salience divergence across a long period of time and many countries, I am unable to assess the degree to which the results would differ using another data source. There is evidence that parties’ issue attention in manifestos and press releases are closely related, but parties emphasize their owned issues more strongly in press releases than in manifestos (
Tresch et al., 2018: 37–38). Therefore, I proceed with manifestos as one valid—though conservative—indicator parties’ issue agendas and issue salience divergence.
I begin by calculating ISD
it, issue salience divergence for a single party
i in election
t, using the formula below. I first take the Manifesto Project’s coding scheme, which includes some directional categories such as ‘pro-welfare state’ along with non-directional categories like ‘peace’, and condense them into 15 policy domains (
Volkens and Merz, 2018).
2 Critically, this categorization scheme allows me to calculate parties’ positions and salience for each category. Combining the positive and negative categories removes the positional component of the data yielding the
salience of each domain, which I measure as the percent of the manifesto in that area.
I then take the correlation between the salience of these issue domains for each party dyad in an election, which I denote as
Ci,j,t where
i refers to the party of interest and
j refers to a different party in election
t. Strong positive correlations in this context indicate that the dyad’s issue emphases are quite similar. Negative correlations indicate greater differentiation between the parties in terms of emphasis allocation. I then calculate ISD
it by taking the average of the correlation for the dyads including the party of interest, weighted by the vote share of the other parties,
vj,t,. This value is then multiplied by negative one, so higher values indicate divergence and lower values indicate convergence This operationalization also treats large parties as more influential, which is appropriate as these are the most important parties in the system. Weighting by party vote share, in this case, follows the same logic as in the calculation of the effective number of electoral parties or ideological polarization, both of which weight parties by size (
Dalton, 2008;
Laakso and Taagepera, 1979). To arrive at a party-system level measure, I take the average of these party-level values.
Concrete example: United Kingdom parliamentary election 1979
Figure 1 plots issue salience for the 15 issue categories used in this paper for each of the parties included in the MARPOR data from the 1979 election in the United Kingdom. First, I calculate the correlations between the issue saliences of each party dyad in the election. Because there were only three parties in this election, I calculate correlations for the Labour-Conservative, Labour-Liberal, and Liberal-Conservative dyads. The dashed identity line indicates where all points would fall if both parties in a dyad had identical issue salience distributions. When the party on the y-axis emphasizes an issue more than the party on the x-axis, the point ends up above the line. When the party on the x-axis emphasizes an issue more, it falls below the line.
As can be seen in
Figure 1, the Labour-Conservative and Labour-Liberal correlations are around 0.5, which is more divergent than the global average, but this is not surprising given the acrimonious nature of the 1979 election. In the Labour-Conservative dyad for example, the Conservative party emphasizes the Role of the State (free vs controlled markets) much more than the Labour Party, which fits with Margaret Thatcher’s ideology. The Conservatives also emphasize unions more than Labour, which reflects the state of affairs following the Winter of Discontent and the public’s souring towards unions at that time. Labour puts more emphasis on the welfare state and environment than the Conservatives.
The Liberal-Conservative dyad is particularly interesting in this election. The Liberal Party was the only party to emphasize Civil Rights (including democracy and law and order). This reflects the party’s emphasis on electoral reform including the voting system, the relationship between parliament and government, citizen’s rights, and reduction in crime. For perhaps obvious reasons, the two major parties in the system have much less to say on these issues, particularly electoral and governmental reform. The Liberal Party also called for an “environmental perspective” across all areas of policy. The differences between these priorities and those of the Conservative party are stark, leading to a low correlation of 0.097.
With these correlations, calculating issue salience divergence for a party in this election is the average of the correlations for each dyad including it. This average is weighted by the vote share of the other party in those dyads. I then multiply that average by −1, so higher values indicate more divergent issue emphases. This is just convention. The averages could be left as they are and be treated as issue salience convergence rather than divergence. I choose to multiply by −1 so that higher values mean more differentiation as is the case with standard measures of polarization and fragmentation. In this example, the issue salience divergence of the Conservative Party comes to −0.400, the Liberal Party −0.303, and the Labour Party −0.522. As the dyadic values indicate, the Liberal Party is the most divergent party in this election, and the Labour Party is the least. A score for the election is the simple average of these values because the party-level scores are already weighted by vote share. Using this method, the UK’s 1979 election has an issue salience divergence of −0.409, which is substantially more divergent than either the second 1974 election (−0.771) or the following election in 1983 (−0.834).
Independent variables
I argue that issue salience divergence varies due to political institutions, the decline of cleavages, and mainstream versus challenger status. I operationalize electoral systems using the tripartite division of party systems into majoritarian, mixed, and proportional from the fourth version of the
Democratic Electoral Systems dataset (
Bormann and Golder’s, Forthcoming). The majoritarian systems include both plurality and absolute majority systems, which should incentivize direct issue engagement to achieve pluralities and majorities. The proportional and mixed cases, however, provide weaker incentives to court diverse voters via broad appeals and should result in more distinctive issue emphases.
3
H2 above argues that issue salience divergence declines over time due to the weakening of cleavages and the transition to catch-all parties. To reflect this, I measure time at the party-level with the party’s age, the difference between the year of the first election that a party enters the data, and the current election. Although measuring party age in this way will be collinear with the year in across-time analyses, there is still variation within elections as some parties will be older than others. I also include squared and cubed values to allow for the possibility of non-linear effects.
Lastly, I use the Manifesto Project’s classification of political parties into party families. I then use
Spoon and Klüver’s (2019) categorization of these parties as either mainstream or challenger parties, which considers Christian democrats, conservatives, liberals, and social democrats to be mainstream.
4 All other party families are considered challengers. While this operationalization is time-invariant, it does distinguish challenger parties as those seeking to introduce or increase the salience of a specific set of issues on the political agenda. These parties are also more policy-seeking than office-seeking, so they will maintain their core set of issues despite competitive pressures to expand their appeals.
Controls
In addition to the independent variables of interest, I also include indicators for executive type (presidential vs. semi-presidential vs. parliamentary) and region (Western vs. Eastern Europe) to account for the possibility that post-communist trajectories have led to different degrees of issue salience divergence (
Vachudova and Hooghe, 2009). I do not make explicit predictions for the effects of either of these variables. I also include the number of years since 1945 in models where it would not be perfectly colinear with party age or age of democracy.
I also include ideological polarization on economic and non-economic issues as well as the effective number of electoral parties (ENEP) as election-level variables (
Laakso and Taagepera, 1979;
Sartori, 2016). I control for party system factors to account for the possibility that issue salience divergence is a reflection of the number of parties and degree of positional disagreement in the system. Note that although correlations between issue salience divergence and polarization (economic and non-economic) are significant and positive, they are low in magnitude (0.197 and 0.104 respectively). The correlation between issue salience divergence and the effective number of parties is not statistically significant. These values suggest that they may share determinants but are not mere reflections of each other.
Following previous literature, I calculate the standard deviation of party positions on economic and non-economic issues weighted by party vote share (
Gunderson, 2021;
Hazan, 1997;
Kim et al., 2010). I calculate the party positions using data from the Manifesto Project because it covers all the cases included in the analysis. See Table 1 in the
Supplementary Appendix for the division of categories into economic and non-economic.
Modeling
The operationalization of issue salience divergence forms a panel data set where each observation is a party-election. In the 30 European countries from 1945 to 2021 included in this paper, I have a total of 381 unique parties in 426 national elections that enter my analysis. As shown in the descriptive statistics, party observations can be aggregated to the country-election level. Therefore, this dataset contains both rich temporal and substantial cross-unit variation. Given that my hypotheses laid out above refer to both differences of level and change as well as having implications at multiple levels of analysis (party-election vs country-election), I use several modeling strategies to capture the variation of interest.
At the country-election level, I utilize OLS estimate models with all cases pooled, country fixed effects, and first differenced. The pooled analysis will act as a baseline model with which to compare the other two. Its results blend temporal and cross-unit variation. The country-election fixed effect model controls for between-country variation, so its results speak to the extent that the independent variables are related to temporal variation in issue salience divergence. The first-difference model takes this one step further by narrowing the effect of the independent variables to only changes between adjacent elections. The extent that the results from the fixed effect and first-difference models differ from the results in the pooled model indicates whether the pooled results refer primarily to differences across time or between countries.
5
To better reflect the multilevel structure of the party-level models, I use multilevel models with random intercepts by party and country. This allows me to include variables referring to the parties themselves and contextual variables that may influence party behavior. The inter-class correlation coefficients indicate that both random intercepts warrant inclusion in the model and explain a substantial proportion of the overall variation. I also leverage the higher number of cases at the party level to test the robustness of results from the overall sample by running separate regressions for cases in Western and Eastern Europe. This split sample analysis clarifies the extent to which parties from either region drives the aggregate results.