In the eye of the beholder: voters’ perceptions of party policy shifts

1. Note that using other model specifications, e.g. random-effect by individual or by party, yields similar results to the ones presented in the article. See Tables S2 and S3 in the Supplemental data for complete results.

2. The Left Party and the CSU are excluded due to lack of data.

3. Note that in our second set of models, only correct responses about the direction of shifts are coded as 1. Correct responses when there was no shift constitute less than 10% of the cases and therefore do not affect the substantive results and the inferences drawn therefrom. However, to ascertain that this is indeed the case, we also ran a set of corresponding specifications of the multinomial logistic regression models where the dependent variable took three values accounting for the correct responses where there was no shift, and our main results are confirmed. Note also that using the subsample of instances where the party’s position does not change, salience has no effect in this case. Hence, it seems that saliency only plays a role when there are policy shifts (see Table S4 in the Supplemental data).

4. We do not measure the absolute change between the respondent’s current and lagged placement of the party on a given issue for two main reasons. First, the extent of shift is not in itself a useful indicator for testing the hypotheses because the theory does not relate the magnitude of a policy shift to voters’ perception of a policy shift. Secondly, it is commonly known, many policy shifts are small and may well be attributed to measurement error (e.g. Benoit et al. Citation2009). Hence, a variable measured on absolute values would have asymmetric right-skewed proprieties (see Meyer Citation2013: 223). However, since some shifts in policy positions can be so minute as to be imperceptible while dramatic policy shifts, albeit rather rare, are likely to be observed irrespective of the importance voters attach to these policies, we opt to differentiate between noticeable vs. non-noticeable shifts as discussed below.

5. For a discussion on party positions, its various measurements and drawbacks see the Supplemental data.

6. Attempts have been made to improve the scaling of party positions derived from political texts, e.g. by constructing a log ratio scale of the CMP/MARPOR data (Lowe et al. Citation2011). While the merits of different scaling procedures continue to be debated, all transformations come at a cost (Franzmann Citation2013; Meyer Citation2013: 42–3). The most contested CMP/MARPOR scale is the L–R (rile) scale, which we do not use in our analysis. Furthermore, the logit scores correlate with the original L–R (rile) scores at r = 0.94 (Budge and McDonald Citation2012) and this appears to also be the case for the policy sub-scales Lowe et al. (Citation2011) propose.

7. Not surprisingly, parties in Western Europe are quite consistent in their choice of platforms and therefore the magnitude of one-period changes is quite small (Adams et al. Citation2016; Fernandez-Vazquez Citation2014: 1936; König and Luig Citation2009; Saalfeld and Zohlnhöfer Citation2014).

8. The analysis reveals that estimates are similar across cutoff points and become inefficient with cutoff points above 2. Insignificant results for the variable saliency for a cutoff of 2 and above are mostly due to the fact that swift policy shifts, albeit rather rare, are likely to be perceived by voters irrespective of the importance voters attach to these policies.

9. We also re-run our analysis looking at significant intsead of noticeable shifts. Hence we calculated 95% confidence intervals around parties’ policy positions following Benoit et al. (Citation2009) and define as significant a shift that goes beyond these confidence bounds. Additional results are presented in the online supporting material. We thank a reviewer for this suggestion.

10. We set each CMP coding x_i of party j on the 0 to 100 scale to the value [0.1(x_i)+0], which recalibrates these codings to an 11-point scale or to the value [0.06(x_i)+0], which recalibrates to a 7-point scale. Although the theoretical range for the CMP policy positions is 0–100; there are negative values for Europe and immigration in our dataset because the derived varible for these two issues was created by subtracting the values of the oppositional categories per CMP coding (see the online appendix for details), while the negative values for the nuclear variable reflect the substantive meaning of the positional scale in the voter dataset. Thus, for nuclear, Europe and immigration, we set each CMP coding x_i of party j on the ‒100 to +100 scale to the value [0.05(x_i)+5], which recalibrates these codings to an 11-point scale or to the value [0.03 (x_i)+3], which recalibrates to a 7-point scale.

11. The use of open-ended questions is justified on the grounds that they will more accurately reflect the ‘true’ saliency that voters attach to an issue than the closed-ended questions. Indeed, research suggests that when asked questions in the closed-ended format, voters tend to rate all issues at higher levels of importance and minor issues are particularly prone to greater increases (Fournier et al. Citation2003).

12. We acknowledge that assigning EU integration to either the left or the right side of the ideological spectrum is less straightforward than doing so with the other three issue dimensions. Several have even discussed to what extent the EU related to a left/right competition (see Bakker et al. Citation2012 for a review). In Germany and the period under scrutiny in this article, the European project was championed more strongly by the parties of the centre-right than for the centre-left and we hance consider more EU integration as right-wing. However, it has to be stressed that flipping the scale has very little impact on our regression models.

13. We also replicated all our models adding a variable measuring lagged party’s perceived position (t‒1) to control for voters’ long-term perceptions of party position; the resulting estimates support the same conclusions that we report in the article.

14. We have surveyed other European countries which have panel data available, including Sweden, Norway and Italy; however, they either only include the general left–right measurement or some issues are available only in one of the waves of the panel study.