Strategic Incentives, Issue Proximity and Party Support in Europe

Abstract

The Issue Yield model predicts that parties will choose specific issues to emphasise, based on the joint assessment of electoral risks (how divisive is an issue within the party support base) and electoral opportunities (how widely supported is the same issue outside the party). According to this model, issues with high yield are those that combine a high affinity with the existing party base, together with a high potential to reach new voters. In previous work, the model showed a remarkable ability to explain aggregate issue importance as reported by party supporters, as well as issue emphasis in party manifestos. This paper tests the implications at the individual level by comparing a conventional model where issue salience is determined from manifesto data with a revised model where issue salience is determined by issue yield. The empirical findings show that issue yield is a more effective criterion than manifesto emphasis for identifying the issues most closely associated with party support in the minds of voters.

Acknowledgements

Mark Franklin is grateful for the support of the College of Arts and Social Sciences, Australian National University, which provided him with ideal working conditions during September 2011.

Notes

 1. Albeit with mixed results, given that the cornerstone of Downsian theory is precisely that only the synthesising of political conflict into a single dimension (Black 1958; Downs 1957) can avoid the inherent disequilibria and decision cycles produced by issue multidimensionality (Arrow 1951; Riker 1982). As a result, most of the literature has concentrated on determining the conditions upon which the median voter theorem can be extended to multiple dimensions – conditions that are in general very restrictive (Ansolabehere and Snyder 2000; Davis and Hinich 1966; McKelvey 1986; Plott 1967).

 2. This example should clarify the decoupling between the potential for a trade-off and the empirical presence of a difference of opinion regarding policy alternatives. Even in a country where 100 per cent of voters agreed on income redistribution, a potential trade-off would still exist on the issue: but, so long as the trade-off does not show up in public discourse, we would expect competition to take place on that issue in valence rather than in positional terms.

 3. The 25 per cent disagreeing on the policy confirms the presence of a proper distribution of preferences.

 4. We do not suppose that party leaders necessarily spend their time poring over polling figures trying to determine the issues on which their supporters are agreed. Politicians have other ways of gauging public opinion. Astute party leaders presumably have well-honed abilities to evaluate the potential of certain issues without necessarily requiring survey data. It is we who need survey data to understand the strategies that they eventually adopt.

 5. These insights are not entirely new, as they are present – though part of a slightly different theoretical framework – in both the saliency theory (Budge and Farlie 1983) and the issue ownership (Petrocik 1996) approaches.

 6. Defined as the share of respondents lying on the ‘agree’ side of the corresponding response item.

 7. Compared to the default expectation that agreement and disagreement within each party is the same as in the whole electorate.

 8. If we define p as the share of respondents that support a party, i as the share of respondents that support an issue, and f as the share of respondents that jointly support both, the formula for issue yield is (see De Sio 2010; De Sio and Weber 2011).

 9. Of course, it might be possible to convert opponents of the issue, both within and beyond a party, but this will produce a different proportion of supporters within and beyond the party and give rise to a different value of L.

10. The main quantities for computing issue yield indicators (party support, issue support, joint party–issue support) were derived from vote intention in national elections (‘And if there was a general election tomorrow, which party would you vote for?’) and from standard, five-point Likert response scales (strongly agree, agree, neither agree nor disagree, disagree, strongly disagree) on 12 issues presented in Appendix Table A1. These latter scales were dichotomised to calculate the measure of issue yield. Neutral values were coded 0.5, resulting in a conservative estimate of bridge issues because (dis)agreement is shrunk toward the midpoint.

11. The possibility for parties to exploit either the positive or the negative side of an issue has been taken into account in the analysis by separately computing issue yield for each side of each issue, then assuming that each party would exploit the side with the higher yield.

12. Otherwise, all predictors (being respondent-level) would have the same value across all the within-respondent observations: thus PTV differences across party-level observations for the same respondent could not be explained.

13. Control variables were included at the party–respondent level based on socio-demographics at the respondent level, and on respondents’ party evaluations (which party did they feel closest to and which party did they think would perform best in terms of the most important problem facing their country). The specific procedure for doing so is based on multivariate regressions (run separately for each party) of a party's PTV on the specific predictors: predicted values (y-hats) are then centred on their means and saved as scores for use in the later analysis as party–respondent-specific predictors (van der Eijk and Franklin 1996; van der Eijk et al. 2006).

14. A fourth version was computed as a robustness check (see below, note 32), but its detailed results are not reported.

15. With scores −1, 0, +1 corresponding to disagreement, neutral position, agreement.

16. In case of lack of any mention, the ‘neutral’ code is assigned.

17. Simplifying the five-point into a three-point scale is needed to render the voter and party scales comparable.

18. We also give a score of 0 when both are neutral, thus expressing the fact that this is a weaker form of agreement than when both are on the same side.

19. Actually, the difference between the emphases given to the two sides of the issue. For comparison, a model is provided in which issues are not weighted by manifesto emphasis.

20. Analogously to the manifesto emphasis score, this is based on the difference between issue yields for the positive or negative side of the issue.

21. As in the manifesto version, by the difference in yield between the positive and negative side of the issue.

22. With responses nested within respondents who are nested within countries. Random intercepts are specified at the country and respondent level. Since the model contains no party-level variables, we do not specify a party level in this model (see Table 2 for models that include a party level).

23. Education, union membership, marital status, employment status, social class (worker), family income, religion, and religiosity (church attendance).

24. Political interest and campaign interest.

25. Whether a party was considered best able to deal with the most important problem facing the country (the identity of the problem is not specified here).

26. Though one benefit of a stacked dataset is that it can include party-specific variables, we include no such variables in this model. We do have an expectation, detailed earlier, that party size and its interactions with yield-weighted issues will play a role in explaining party support. Specifically we expect support for small parties to be affected more strongly than support for large parties. However, in this initial test we have 12 measures of issue yield, and interactions between each of these and party size would produce more variables at the party level than we can handle in a multi-level model. Once we have established the ubiquity of the relationship we theorise, we can simplify the model so as to be able to include party size.

27. Obtained by squaring the correlation between observed values and values predicted by the model.

28. In this table we control for ‘Party best for most important problem’, which arguably already incorporates some effects of issue yield. We do this to provide the most conservative available test of those effects. Appendix Table A2 shows Models B and C without the Most Important Problem variable. Those models explain considerably less variance but show even greater gains for the Issue Yield model in comparison with a model employing salience-weighted affinities.

29. In Figure 1 we compare models B and C because this is the more critical comparison: models B and C have fewer significant differences than Models A and C. It is noteworthy that, with two exceptions, for issues where model C performs significantly better than Model B it also performs significantly better than Model A (it performs significantly better than Model A on an additional issue as well); but the two exceptions cancel out. For Q60, Model C performs significantly better than Model B but not than Model A, whereas for Q63 it is the other way around.

30. Effects for yield-weighted issue affinities do not indicate the importance of the issue. An issue with low yield, whose low yield is correctly anticipated, will get as high a coefficient as a correctly placed high yield issue. But this expectation breaks down for issues that promise no yield at all. An effect of zero for a yield-weighted issue implies that the issue concerned is not a relevant basis for distinguishing between parties. That being the case, the yield-based effect will be zero, which would quite often be less than an issue effect calculated on a different basis.

31. Identification of the actual mechanism responsible for the superior performance of the Issue Yield model must wait on future research.

32. As a robustness check we created a fourth set of issue affinities for the 12 issues by generating y-hat predictors for each issue, in just the same way as for demographics and political interest. A y-hat affinity is, as already explained, ‘tuned’ to the dependent variable and can be regarded as the best affinity measure that can be gleaned without the benefit of any theoretical basis for linking voter preferences to parties. The overall predictive power of a model employing y-hat issue affinities is given by an R ² of 0.367: more than the R ² for Models A or B, but less than the R ² for Model C. So yield-weighted issue affinities apparently work better even than a measure that posits a basis for these affinities derived purely from data-fitting (with y-hats, a correspondence such as the negative effect of Q65 in Models A and B, that makes no substantive sense, counts as ‘correct’ prediction).

33. See note 30.

34. Estimation was performed using R's lme4 package. The quantities of interest for Figure 2 were simulated using Clarify (Imai et al. 2007; King et al. 2000). Because it is difficult (perhaps logically impossible) to simulate quantities of interest from a hierarchical model that contains random coefficients, we did not include random coefficients in any of the models in Table 2. However, a separate analysis (not shown but available on request) indicates that the standard error of the interaction term is only 0.005. So the effect of 0.5 shown for the interaction of party*size in Model E varies by no more than 0.01 (with 95 per cent probability) across countries. The effect of yield-weighted size itself varies rather more across countries, but by no more than 0.15 with 95 per cent probability. These variations are consistent with the standard errors for the same effects seen in Model E, suggesting no need to explicitly model these country-level variations. Moreover, the AIC is hardly reduced and BIC actually increases when random slopes are introduced into the analysis, supporting our choice of models to present in Table 2.

35. Because y-hat affinities are tuned to the dependent variable, in the absence of collinearity with other independent variables their effect would be 1.0. However, their indicated effect depends on the extent to which their uncontrolled effect is shared with other variables – the precise point of concern in this analysis. Model D tells us that about a third of the effect of party size is shared with other variables, probably mainly the Most Important Problem variable (because one-third is the extent to which 0.65 is less than 1.0). Model E shows no change in this situation, but Model F sees the bulk of the remaining effect of issue yield transferred to its interaction with party size. Substantively identical findings are obtained if, instead of b coefficients, we calculate standardised beta coefficients in these models, as explained earlier. But this would make it computationally challenging for us to plot the interaction of interest.

36. Indeed, voters adopt an emphasis on different issues that accords better with party needs than we would see if we adopted the blindly empiricist procedure of weighting each issue according to its observed covariation with party support (see note 32). The subtleties of the Issue Yield model do better in this respect than data-fitting.