Who Gets into Government? Coalition Formation in European Democracies

Abstract

This article investigates different dynamics in government formation in 16 Western and 10 Central-Eastern European democracies during the post-war period. The study provides the first systematic comparison of determinants of participation in government in the East and the West. Applying mixed effects logit regression models while taking into account missing values in the dataset, the results demonstrate substantial differences between the two regions and show that most of the existing findings about participation in government are driven by Western democracies. Policy-based factors are relevant in Western countries, but no indications were found for these factors in Central-East European democracies where membership of government is mainly the result of electoral gains and losses.

Acknowledgements

Both authors contributed equally to the article and their names are in alphabetical order. An earlier version of this paper was presented at the 5th ECPR General Conference in Potsdam, at the Colloquium on Political Behaviour at the European University Institute in Florence, at the MZES-Colloquium in Mannheim and at the annual meeting of the Swedish Political Science Association. We are grateful for detailed comments on earlier versions of the paper by Robin Best, Hanna Bäck, Marc Debus, Aline Grünwald, Mark Franklin, Johannes Freudenreich, Bjørn Høyland, Kaare Strøm and Till Weber; as well as Laurie Anderson, Armen Hakhverdian, Alexia Katsanidou, Quinton Mayne, Autumn Lockwood Payton and Alyson Price at the Max Weber Programme.

Notes

1. The unit of analysis in our study is an individual party and we aim to investigate parties’ likelihood to enter cabinet and not to explain which coalition forms from a finite set of potential coalitions. Hence, we deliberately decided not to use a discrete choice modelling in this study, a methodological approach that has been applied more frequently in recent studies of government formation.

2. Including a measurement that assesses which party is the median party together with a second measurement that calculates its distance from the median would introduce multicollinearity, because the latter proxy is a perfect predictor of the former. We chose to use the later measurement in our analysis only.

3. Our data is derived from ParlGov (Döring and Manow 2010), a data infrastructure on parties, elections and cabinets, that collects electoral results and cabinet compositions and systematically links this information to existing data on party positions. The documentation of ParlGov includes a list of sources used to code cabinet parties and electoral results. ParlGov covers only those Central-East European countries that are also members of the European Union. The particular data set we use (including seats change) was created from ParlGov by a software script and we added information about party positions from the Comparative Manifesto Project (Budge et al. 2001; Klingemann et al. 2006). Replication data for this study are available at http://dvn.iq.harvard.edu/dvn/dv/johan-hellstrom.

4. For our example, it may affect a party’s probability to get into government at time, t, if a party is in government at time t – 1. In other words, the dependent variable is time-dependent and first-order auto/serial-correlation may be present.

5. An alternative would be to use conditional (or fixed) effects logit to account for unobserved heterogeneity. However, fixed effects estimates can be biased for short panels with a lagged dependent variable such as our measure of incumbency. In addition, fixed effects models rely solely on within-group variation and any time-invariant variables (e.g. country dummies) cannot be estimated (cf. Beck 2001). Nonetheless, we also estimated the same models using conditional (or fixed) effect estimators and the results gave nearly analogous results compared to those reported here.

6. Rather than producing a single estimation for each missing observation, we set up Amelia to generate multiple datasets (in our case, five), each with unique values for the missing observations. The variance in the imputed values across these five datasets reflects the uncertainty about the observation’s true value. Our estimations are run with each of the five datasets, and the results are combined by using a procedure designed to reflect the appropriate uncertainty levels for each of the missing values. In addition, we also estimated the same models presented in this paper on the original data and obtained similar, but not identical, results (not shown here).

7. We replicated the approach used by Mattila and Raunio (2004) with our dataset and measured gains and losses in the parliament by using two indicator (dummy) variables for gains and losses. The indicator for loss of seats was not significant, but this was attributed to the fact that its effect was too similar to the reference group (neither gain nor loss of seats). If changing the reference group to those parties that gained seats, it was also significant and thus shows that losing seats somewhat lowers the average probability of getting into government.

8. However, if we use party expert data in order to measure party positions (cf. models 6 and 8 in Table A2 in the appendix), we find a contradictory result. These results indicate that the distance from the median party on the left/right dimension matters for a party’s chances of getting into government, but less so than in Western Europe. Whether this result is a genuine effect of politics in the CEE region or if it occurs because of shortcomings in the expert data (i.e. lower coverage and insufficient time variation) is difficult to say, and we leave further investigation of this issue for future research.

9. The plots in Figure 1 are based on model 7 and 8 in Table 3, setting all other covariates to their mean values. The probabilities reported in the text for the impact of being the largest party are based on calculations using the average predicted probabilities.