![Sample our Politics & International Relations journals, sign in here to start your FREE access for 14 days]({"type":"image" , "src" : "/sda/5947/TOC-Subject-Sample-2021-Politics-International-Relations.jpg"})
Abstract
Current research on coalition formation is plagued by two serious problems. First, we cannot predict more than about one-third of the Western European governments, and, second, we do not have a good understanding of the causal mechanisms that explain the effects found in large-n coalition studies. This article illustrates that by combining statistical and case study analyses we can solve these problems. Since statistical analyses are well equipped for measuring and isolating effects, we argue that a coalition study should start with such an analysis. Predictions made in this analysis are then used to select cases. In order to study the mechanisms underlying effects found in large-n coalition studies, we argue for selecting cases that are predicted, and then applying the method of process verification. In order to find new explanatory variables, we argue for selecting cases that are deviant, and then applying the method of process induction. Substantive results of our analysis for coalition theory point to the importance of party strategies based on parties' past experiences, which aim at curtailing present and future costs of competing and governing with other parties.
Acknowledgements
Earlier versions of this paper were presented at the ECPR General Conference in Marburg, September 2003 and at the Annual Meeting of the American Political Science Association in Chicago, September 2004. The authors wish to thank Ted Hopf, Martin Lundin, Jan Teorell and Kåre Vernby for helpful comments on previous versions of this paper.
Notes
1. This is why De Winter et al. (Citation2002) advocate re-appraisal of the inductive approach towards coalition theory in their call for a ‘theoretically informed inductive approach’.
2. Hence, even if the existing model was deductively derived and the new variable(s) can be integrated in a formal model, the theory arrived at will owe to both a deductive and an inductive approach (see discussion in Morton Citation1999: chapter 3; Camerer Citation1997 and for coalition studies, see Dumont Citation2006).
3. Other research disciplines may of course be plagued by other problems.
4. We focus here on some of the most important hypotheses that can be derived from the coalition literature. We have also chosen to focus on many of the hypotheses evaluated by Martin and Stevenson (Citation2001) in order to be able to compare our results with their findings.
5. Here we stick to the operationalisation provided by Martin and Stevenson (Citation2001) in order to generate comparable results. For a better operationalisation of this concept, see Dumont (Citation2006).
6. We count a change of cabinet whenever: 1) general elections are held, 2) there is a change in the party composition of the cabinet, 3) there is a change of Prime Minister, and we only consider minority situations (formation opportunities where no single party holds a majority of seats).
7. We have a much larger number of potential governments than Martin and Stevenson (Citation2001), as their full sample consisted of 33,000 potential governments while their number of formation opportunities (220) was higher than ours. This is due to the fact that we work on an updated data set – the last opportunities they analysed dated back to the mid-1980s – and the fact that political systems since the mid-1980s have been much more fragmented than in the pre-1970s period, causing an exponential increase of potential governments.
8. Updated for post-1998 government formations by the authors.
9. The Laver and Hunt survey (Citation1992) asks country experts to place the parties on a number of scales ranging between 1 and 20, and the Huber and Inglehart survey (Citation1995) asks country experts to place the parties on what they consider to be a left–right scale ranging between 1 and 10. We have standardised the scales to vary between 0 and 1. To come to terms with the fact that we had missing values for some parties, we have performed a small survey with country experts. We still have some missing values for some parties however. A drawback with using these data instead of manifesto data (the main available alternative when we are studying post-war Western European governments), is that the expert survey data do not vary over time, that is, we have to assume that the parties' policy positions have stayed stable between 1970 and 2000. There are several reasons for using expert surveys instead of manifesto data. One reason is that party manifestos are likely to reflect the image that parties want to give to the voters rather than their ‘true’ policy positions, a view that has recently been analysed by Pelizzo (Citation2003: 87). Parties may also anticipate future coalitions when writing their policy programme, which makes manifesto data unsuitable to use when studying coalitions (see Bäck Citation2003; Dumont Citation2002, Citation2006; Dumont and Bäck Citation2006). Moreover, if the main advantage of using party manifesto data is to provide measures for each election year, the empirical exercise provided by Budge and Klingemann (Citation2001) does not convince us to use these more dynamic data, as they show that parties' locations remain fairly stable and they rarely leapfrog each other on a left–right scale.
10. The conditional logit model is nonlinear, which means that the unstandardised coefficients cannot be directly interpreted as what happens with the probability that a potential coalition is chosen when we alter a choice-specific variable which we are ultimately interested in. There is however no straightforward way to calculate and present changes in predicted probabilities for a conditional logit model with such a high number of alternatives (the number of potential governments vary between 7 and 16,383 in the formation opportunities present in the sample).
11. Note that as predictions are almost all unique equilibriums in models 2 and 3, the efficiency rates presented for these models roughly equal their prediction rates.
12. Since the level of the highest predicted probability varies greatly between formation opportunities (hence the absolute differences automatically become small when dealing with low predicted probabilities), we standardise the differences by dividing them with the highest predicted probability in each case. The procedure for calculating the differences can be described by this equation: Prhigh – Practual/Prhigh×100. This gives us a measure that varies between 0 and 100.
13. We could of course use this type of criterion to select other cases, for example, we could be particularly interested in studying cases where some specific theory fails to predict the outcome. We could also be interested in studying some particular deviations, such as minority governments, or oversized governments, which would mean that we should select outcomes that exhibit such features.
14. With the strategy adopted in this paper, that is, selecting cases to be intensively studied on the basis of large-n tests of a number of classic theories, we anticipate to some extent this requirement.
15. In the field of coalition formation research, explanatory variables may come from deductive theoretical models, from inductive work conducted by country specialists, or even from attempts at explanations or justifications by participants of government formation themselves. In the latter case, it is quite clear that an in-depth study of the case is necessary to take into account the motivations of participants for proposing their own ‘theory’, and that the investigator must nevertheless translate the participants' ‘stories’ in theoretical terms.
16. The two illustrative case studies below build on in-depth studies by Andeweg (Citation2002) and Dumont (Citation2002), who used newspaper reports, radio/TV broadcasts, internal party documents and performed interviews with negotiators and observers.
17. If this was the mechanism at work, the office and/or policy-based models would rightly predict the formation of the first of the two consecutive governments in case such size and policy features disappear in the second formation opportunity (transmitted effects), or both of the consecutive governments if these features remained (spurious positive incumbency effect), in our large-n study.
18. Assuming strong policy-seeking motivations of political actors, the incumbent administration may also represent for some opposition parties a better outcome in terms of policy distance from own preferences than what an alternative would offer, thereby inducing these parties to break attempts at forming an alternative government to which they would be invited.
19. In the other countries studied here, the ‘burden of proof’ lies in the opposition camp, as only a no confidence vote voted by a majority in parliament can unseat the government in place.
20. Some would perhaps argue that negotiations do not have to fail between other parties for a reversion point mechanism to be at work, and that the fact that parties consider bargaining with other potential coalition partners could qualify as evidence for the idea of a reversion. We contend that this is not what the creators of the reversion outcome idea had in mind, and we thus argue that this idea is only supported if we find that negotiations including other parties have occurred and failed.
21. Following Gerring (Citation2007), this case is not only a ‘typical’ (or ‘predicted’, that is, lying on the regression line) case but rather a ‘crucial’ case of the ‘least-likely’ type (all causal factors except the factor of interest – here the incumbency dummy variable – point away from a predicted outcome, which nonetheless occurs). The intensive study of such a crucial case may confirm the causal effect of incumbency and probe the causal mechanism at work.
22. Informateur Zalm (VVD) explained that the three parties had started out and lived through events without damaging relations amongst cabinet ministers. A feeling of loyalty to the purple formula developed between them, so he felt that there was no reason to bury it (Andeweg Citation1998).
23. Chamber of Representatives and Senate (federal level); Flemish, Walloon and Brussels-capital parliaments (regional level); French-speaking and German-speaking parliaments (community level).
24. Coalition talks would then involve the Liberals, Socialists, Greens and the Flemish nationalist Volksunie in order to reach a majority in the Flemish parliament.
25. The same type of argument may also hold for the largest party, which would lose its ‘dominance’ (or, if the party was not dominant in the first place, a substantial amount of voting power), as was the case of the CVP in Flanders in the case analysed.
26. This hypothesis is testable in a large-n setting distinguishing the stage of formateur selection from the stage of the choice of full government composition by this formateur, see Bäck and Dumont Citation2006.
27. If the former largest party is located in the centre of the political spectrum, which is likely if it stayed in power for a long time, the probability of the formation of an unconnected coalition becomes high. Belgium (1999) but also the Netherlands (1994) and Finland (1995) fit nicely with this mechanism. In each of these three countries, these unconnected coalitions excluding the former dominant party were re-formed for a second term (in Belgium, the Greens were however ousted in 2003).
28. Golder (Citation2006) shows that public pre-electoral agreements are more likely the closer the policy positions of its potential partners, but she does not investigate the majority/opposition hypothesis.