Conclusiones

employed and those who are not in the workforce for any other reason. The latter group is mostly retired people but also includes stay-at-home parents and disabled people, among oth- ers. Unsurprisingly, a majority of respondents (50.2%) indicated that they were employed, with only a small number (8.0%) claiming to be unemployed. The remaining 41.8 percent are not in the workforce. All of these variables are coded as dummy variables indicating each of the non-reference cases.

2.3

Method of analysis

This thesis uses a stacked dataset design, which was strongly influenced by van der Brug, van der Eijk and Franklin (2007, 40-46), although the methods used are different. The motivation for this design is that the key variables of interest exist at the measurement level, which relates particular individuals to particular parties. For instance, the dependent variable, party support is such a variable. The degree to which individualisupports party j cannot be said to belong completely to either the individual level or the party level—it belongs to both. Since this party×individual, level, which is referred to in this thesis as the measurement level, is where the most important measurements occur, it is appropriate for the unit of observation to be a party–individual pair. This objective was achieved by transforming the EES survey dataset, which was measured at the individual level. Parties are represented by repeated questions— that is, each party-related question is asked once for each party. The dataset was transformed such that each individual appears in the dataset once for each party they were asked about, with only the questions about the relevant party included in each new row. Figure 2.6 illus- trates this process.

The key analytical method used in this thesis is multilevel modelling. Multi-level model- ling is an extension of linear regression and generalised linear modelling that explicitly models variation between groups. This is achieved by giving the model coefficients a probability model of their own (Gelman and Hill 2007, 1). The decision to use multilevel modelling was mo- tivated by the clustered nature of the data. One of the consequences of stacking the dataset as described above is that individual observations have been repeated in the stacked dataset many times. If linear regression were used this would artificially deflate the estimated stand- ard errors but multilevel modelling can account for this sort of structure. In any event, the data used in this thesis is inherently structured. For one thing, there are twenty-five countries in the analysis and to ignore this would be to assume implicitly that there is minimal variance

Figure 2.6: Data stacking process

country

resp_id

age

support_p1

support_p2

support_p3

UK 1 88 0 8 5

UK 2 21 7 3 8

UK 3 40 3 8 0

⇓

country

party_id

resp_id

age

support

UK UK-Lab 1 88 0 UK UK-Lab 2 21 7 UK UK-Lab 3 40 3 UK UK-Con 1 88 8 UK UK-Con 2 21 3 UK UK-Con 3 40 8 UK UK-LD 1 88 5 UK UK-LD 2 21 8 UK UK-LD 3 40 0

In the original dataset, each observation corresponds to an individual and variables relating to particular parties are repeated for each party. In the stacked dataset, each observation corresponds to a party–individual pair, eliminating the need to repeat variables.

between these countries, which is a strong assumption to make. Furthermore, many of the variables that are of interest, including party support, the dependent variable, involve opinion about specific parties. Since there may be variation between a country’s parties as well as variation between countries, there are clusters within clusters.

While ignoring this clustering would lead to deflated standard errors, it must be acknow- ledged that there are other ways of addressing this particular problem. It has been shown that robust standard errors and aggregation can produce the same results as multilevel mod- els under the appropriate conditions (Arceneaux and Nickerson 2009). Nonetheless, each of these other methods is limiting in some specific way. Aggregation cannot be used to estimate individual-level effects and robust standard errors are not recommended for clusters of fewer than about twenty observations (188). The number of parties in each country in this study varies from five to fifteen. Furthermore, multilevel modelling is sufficiently flexible that is possible to include predictors at multiple levels within the same model. This is particularly important for this study, since economic voting theory relates variables that necessarily relate to multiple levels. Incumbency is only meaningful at the party level. Economic performance is measured at the national level. And it is individuals who make vote choices. A key advantage of multilevel modelling is that these variables and their interactions can be included in a single

2.3. METHOD OF ANALYSIS 59 model (Gelman 2007a, 7-8). Of course, it would be possible to aggregate all of these together and estimate everything at the national level. It would also be possible to model separate equations for every political party of interest. Both approaches are problematic however. The former approach, complete-pooling analysis, ignores all variation between groups and the lat- ter, no-pooling analysis, is statistically inefficient. A key strength of multilevel modelling is that it offers a compromise between these extremes (Gelman and Hill 2007, 256). Multilevel models produce both fixed effect and random effect estimates. The fixed effects are those of primary substantive interest given the research questions behind this thesis. The decision to use multilevel modelling is motivated by the need to account for the structure of the data, rather than interest in the random effects specifically. As a result, the discussion of random effects is limited to discussions of variance and sometimes covariance. Individual countries are not normally discussed, as the purpose of this thesis is to gain an understanding of large cross-national trends.

The decision to use multilevel linear regression models distinguishes this study from pre- vious survey-based research into economic voting behaviour. As discussed in the previous section, the stacked dataset design of this thesis was influenced by van der Brug, van der Eijk and Franklin (2007), but they chose not to use multilevel modelling in their study. They reason that the chief advantage of multilevel models is that they avoid the biased standard errors that would result from naive regression modelling and they observe that the this prob- lem can be avoided equally well by using robust standard errors, which they do (47). They also argue against multilevel modelling on the grounds that it is not well equipped to handle the cross-classified data structure that results from the stacked dataset design (48). While there is merit to these arguments, there is still much to be said for multilevel modelling in this context. Although it is true that the use of robust standard errors can avoid the problem of deflated standard error estimates, there are other advantages to using multilevel model- ling, particularly in that they permit greater flexibility in the models. As for the problem of cross-classification, it must be acknowledged that this does add complexity to the models but modern computing power combined with Bayesian estimation techniques makes it possible to estimate these complex models. Duch and Stevenson (2008), on the other hand, do use some multilevel modelling. They adopt two different methods for their analysis, a one-stage and a two-stage strategy. Their one-stage strategy uses multilevel models to estimate all of the para- meters together, whereas their two-stage strategy involves estimating the level of economic voting in national surveys and then using those estimates in cross-national models (94-100).

One of the key differences between their methods and the methods used in this thesis is that they use multinomial logistic regression to predict vote choice, whereas this thesis uses linear regression to predict party support, for the reasons given in the previous chapter.

Most of the independent variables in the models presented in the following chapters have been centred around the grand mean. By contrast to the centring of the dependent variable, the reasons for which have been given earlier, this centring of the independent variables has been done for technical reasons. Specifically, centring in this way markedly increases the con- vergence speed. It is also helpful for variables which are involved in quadratic and interaction terms. One consequence of this centring is that some caution is required when interpreting intercepts and interactions. Because of this and also owing to the complexity of some of the models, model coefficients are typically not interpreted directly in this thesis. Instead, post- estimation simulation is used to derive quantities of more direct interest. The method used is that described in Gelman and Hill (2007, 140–143). Because the predictions resulting from this method take into account the uncertainty of many predictors, the error bands around them can be deceptively wide. In particular, it is very often the case that two predictive intervals overlap even though there is a significant difference between the actual predictions. In order to minimise confusion, predictive intervals are not normally shown in any plots but instead the question of significant difference is tested directly and discussed in the text. Standard errors are of course quoted in the text for most predictions as well. The full model coefficient tables can be found in Appendix B. There are multiple methods for computing p-values for multilevel models. The p-values shown in the coefficient tables are based on Satterthwaite estimates of the degrees of freedom, although these values play no part in the analysis in this thesis. The PseudoR2 reported for these models is based on Xu (2003).

One challenge that arises with any use of survey data is that of missing data. This has been dealt with using listwise deletion. Owing to the stacked dataset design, a respondent does not have to be excluded using this method simply because he or she has not answered a question about a particular party. Only the row corresponding to that party–individual pair is removed, while rows are included for each party that the individual did answer questions about. This means that the only individuals that had to be excluded completely were those who refused to answer any party related question—and those voters leave very little basis to impute missing values—and those who did not answer questions at the individual level. As people typically did not refuse to answer the demographic questions posed, the only problematic question was prospective economic assessment. The proportion of respondents who declined

2.3. METHOD OF ANALYSIS 61 to answer this question was 13.3 percent in 2004, 4.1 percent in 2009 and 5.4 percent in 2014. Other techniques for managing missing values were also considered. Multiple imputation in particular has much to be said for it (Rubin 1978; King et al. 2001, 50) but that approach did not prove viable here because multiple imputation techniques and software have not yet been developed for all of the multilevel models that are used in this thesis. Ultimately, the structure of the data was considered a more important issue than missing data and that is what led to the decision to use listwise deletion and multilevel modelling.

The survey response rate has also been given some consideration. Response rates in the EES surveys typically ranged from 60–80 percent for the face-to-face mode but were lower for the telephone mode, sometimes below 20 percent. One approach to the issue of potential bias resulting from low response rates is post-stratification survey weighting. Survey weights have been included in the EES survey data to correct for non-response bias. These weights were computed using a raking procedure on the variables of age, sex, region, education and, in some countries, whether or not the household has a fixed phone line. In principle, weighting in this way allows for the more accurate estimation of population parameters from the survey data by reducing the effect of non-response bias. These weights are not however used in the multilevel models in this thesis. Instead, non-response bias is minimised by including all relevant demographic variables in the models as controls. Since the weights are conditioned on variables that are already being controlled for in the models, they add no extra information. There has long been a controversy in the literature about the relative merits of weighted and unweighted least squares estimators for linear regression (DuMouchel and Duncan 1983, 535) and the relative merits of model-based and sampling-based approaches to the problem of unit non-response continue to be debated today (Gelman 2007a, 2007b; Bell and Cohen 2007). The decision to use a model-based approach is motivated by the fact that the parameters of most interest are regression coefficients rather than simple means or proportions. There is also the problem that weighted estimators simply have not been developed for certain cross- classified multilevel models, so the sampling-based approach would severely limit the types of analysis that could be undertaken using these methods.

The statistical analysis described in this thesis has been performed using the statistical pro- gramming language R (R Core Team 2016). In addition to the core language and its libraries, the analysis was supported by the lme4(Bates et al. 2015),lmerTest(Kuznetsova, Brockhoff and Christensen 2016) and ordinal (Christensen 2015) packages. Most of the plots in this thesis were produced using the ggplot2package (Wickham 2009). Finally, thearm(Gelman

and Su 2015),dplyr(Wickham and Francois 2015) andRSQLite(Wickham, James and Falcon 2014) packages were all used heavily for utility purposes.

2.4

Conclusion

This chapter has introduced the dataset and the methods that will be used throughout the rest of the thesis, as well as discussing how the key variables have been measured. This study will undertake a multilevel analysis of survey responses collected in twenty-five European Union member states. The primary data source is the pooled responses from the 2004, 2009 and 2014 waves of the European Election Studies surveys. These waves correspond to time points before, during and after the Great Recession. This survey data is supplemented with contextual data from other sources. The key dependent variable is party support, the degree to which a voter states that he or she is likely to vote for a particular party in the future. Important independent variables are the survey year, spatial distance between party and voter, party identification, incumbency and prospective economic assessment.

The next chapter will introduce the basic models used with various extensions throughout this thesis. Owing to the complexity of the complete model, several simpler models are presen- ted first, examining the first hypothesis from various angles. These simpler models are easier to interpret but only the complete model takes advantage of all of the available data, which provides a clearer overall picture. The intention behind this approach is to use the simpler models to illustrate various aspects of economic voting behaviour and then use the complete model to obtain definitive estimates. Later chapters of the thesis extend or alter these models in various ways, so as to test the remaining hypotheses.

Chapter 3

Voting in a time of crisis: how the Great Recession

affected the economic vote

After the global financial crisis of 2007–08, most developed countries slipped into recession, in an event that has become known as the Great Recession. This was the worst event of its kind since the Great Depression of the 1930s and a number of governments suffered catastrophic electoral defeats in the following years. The example of Ireland was introduced earlier, where the formerly dominant party Fianna Fáil was reduced to approximately half of its previous vote in 2012 (Marsh and Mikhaylov 2012, 478). Such results accord with the established theory of economic voting, which predicts that poor economic conditions will lead to voters turning against their governments at the ballot box. On the other hand, this theory was almost entirely developed using evidence relating to less turbulent economic conditions, what might be described as the ordinary boom and bust cycle of the economy. Whether or not voters respond to a severe transnational crisis in the same way as a typical recession is not yet clear. This chapter explores this question by comparing voters’ party suport levels in 2004, well before the crisis, to those in 2009, at the height of the first wave of the Great Recession, and in 2014, after the initial shock had subsided.

In the immediate aftermath of the Great Recession, some political commentators argued that the situation would benefit the Left, since they are traditionally critics of the economic system which produced the crisis (Bartels 2012, 44) but these expectations have not been borne out. There is little evidence of an ideological shift in OECD countries as a result of the crisis (Bartels 2014). Commenting on the United States, Bartels (2013, 70), observes that:

The truth of the matter is that ideological mandates are exceedingly rare in Amer- ican politics, even in times of economic crisis. Indeed, what may be most striking about the politics of the Great Recession is how ordinary they look. In bad times, as in good times, ordinary citizens have a stubborn tendency to judge politicians

and policies not on the basis of ideology or economic doctrine, but of perceived success or failure.

Economic voting theory offers a more plausible account of electoral behaviour during the crisis. In an analysis of aggregate data from twenty-eight OECD countries, Bartels (2014, 188–194) shows that governments generally received an increased vote when GDP growth was positive and a decreased vote when negative. Kriesi (2014, 305–315) similarly finds a relationship in European countries between incumbent vote share and the economic indicators, particularly inflation in Central and Eastern Europe and unemployment in Western Europe. Kenworthy and Owens (2011, 212–216) examined US voter attitudes in survey data since the 1970s and found that voters do tend to lose confidence in whichever party is governing at the time but they also found that this effect was actually quite weak during the Great Recession. Others have also found the electoral response to the crisis to be weaker than might be expected (for example, Kriesi 2012).1

These results suggest that the electoral response to the Great Recession was an economic voting one, rather than one motivated by a deep ideological shift among voters, but they still leave questions unanswered. The key question concerning this chapter is: was the economic vote stronger during the Great Recession than at other times? Given that the recession was far deeper, this seems likely, but some of the studies just mentioned have found clues that the opposite might actually be the case. On the other hand, such findings might also be artefacts of the particular methods used by those studies, since there has not yet been enough research done to know how different approaches affect the results found.

This chapter has two key purposes. First, it describes how a multilevel model was construc-