Germany holds general elections on September 24th 2017! This motivated us to prepare a small series of contributions concerning the German electoral systems and the election results. After our last blog contribution on ticket splitting, this contribution by Julian Noseck summarizes what we can learn from predictive models in the German context.

The 2017 German Federal Election: The Wisdom of Predictive Models

As the recent elections in the United States or the United Kingdom revealed, the forecast of electoral results solely based on polls is occasionally error-prone. For Germany, current polls  (state: 11.09.2017) show two clear patterns. First, they see a quite significant lead of the Christian Democrats (CDU/CSU) with their candidate chancellor Angela Merkel (36.5-38.5 %) in comparison to the Social Democrats (SPD) with their front-runner Martin Schulz (21-24 %). Secondly, the polls see four minor parties passing the 5%-threshold and thus gaining parliamentary representation: The Greens (Bündnis 90/Die Grünen), the Free Democrats (FDP), the Left Party (DIE LINKE), and the Alternative for Germany (AfD), each of them receiving between 6.5% and 11%.

In the light of the difficulties concerning poll-based electoral prognoses, the question arises: Does political science provide any other tools to forecast the outcome of the upcoming election in Germany that might be more reliable? This question can definitely be answered in the affirmative: Predictive models serve for this purpose. Forecast models have a long tradition in American electoral research. Generally, predictive models are characterized by going beyond short-term survey data, i.e. the incorporation of longitudinal structural data. This structural data might comprise past election results, economic data, or other factors traditionally seen as crucial determinants of voting behavior by electoral research like long-term partisanship (cf. Campbell et al. 1980). In the following, I present three examples of such models, all applying a different approach and objective with regard to forecasting the election results.

One prominent approach on forecasting the election outcome depicts the so-called “Chancellor Model” that was developed by Helmut Norpoth (Stony Brook University) and Thomas Gschwend (University of Mannheim) (Norpoth/Gschwend 2010). It was used the first time for the 2002 German federal election and successfully predicted the prospective chancellor ever since. The Chancellor Model focuses on forecasting the potential future governing coalitions and relies on three predictors: The popularity of the chancellor candidates Merkel and Schulz (based on polling data), the parties’ long-term support (average vote share in the last three federal elections), and an incumbency factor describing the phenomenon of “government fatigue” (number of terms in office). For the 2017 election, the model clearly predicts a fourth term in office for Merkel, finding potential coalitions of CDU/CSU and FDP or CDU/CSU and the Greens to have an absolute majority of seats in parliament in all likelihood (Norpoth/Gschwend 2017).

In contrast to the Chancellor Model the predictive model set up in the context of the project “zweitstimme.org” was just recently developed by the scholars from the University of Mannheim Thomas Gschwend, Sebastian Sternberg, Marcel Neunhoeffer, Simon Munzert (HU Berlin), and Lukas Stoetzer (University of Zurich). The model combines information from past federal elections since 1949 (“structural component”) with actual data from polls (“polling component”). It differs from the Chancellor Model in several respects. On the one hand, it gives more weight to data from recent polls. On the other hand, it does not include an incumbency factor and the popularity of the chancellor candidates into its calculations (on the different motivations for the two models see this blog contribution  by Thomas Gschwend). The estimation of the model is based on an MCMC-algorithm that simulates the election outcome 100.000 times to receive the probabilities concerning the final results. Figure 1 shows the recent prognosis (state: 10.09.2017) by zweitstimme.org, including the respective confidence intervals and a line for the 5%-threshold. As the figure illustrates, the model currently predicts the CDU/CSU to become the largest party, followed by the SPD, and the other parties all competing for position three.

Figure 1: Election forecast by zweitstimme.org, latest update: 10.09.2017

Source: zweitstimme.org

A third predictive model  for the 2017 German federal election was developed by Mark Kayser (Hertie School of Governance) and Arndt Leininger (University of Mainz) (Kayser/Leininger 2017). Their approach is characterized by analyzing the voting behavior in Landtagswahlen (state elections) to predict the outcome of the federal election, as there is a strong correlation between a party’s election result in a state election and its federal election result in the respective state. As a “structural model” it largely abstains from including polling data and almost completely relies on historical data including economic and political data, election results from state elections since 1961 as well as the respective turnout rates. By strongly focusing on structural factors and by analyzing the sub-national level, the model by Kayser and Leininger differs from the Chancellor Model as well as the model by zweitstimme.org. However, similar to the other forecasts, the structural model clearly predicts the CDU/CSU to take the lead in the upcoming elections. In this video , one of the authors explains the model in greater detail.

In conclusion, the predictions in all models presented above are quite close to the numbers provided by recent polls. Accordingly, it seems like significant changes with regard to these numbers is rather marginal. Most likely, the next government will be a coalition led by the Christian Democrats. It is, however, less foreseeable which governing coalition will finally form (although, some predictive models provide probabilities for – at least mathematically – probable coalitions). The German election and the government formation process will thus remain interesting. Furthermore, political scientists will be interested in comparing which predictive model performs best.

References:

Campbell, Angus/Philip E. Converse/Warren E. Miller/Donald E. Stokes. 1980. The American Voter. 2nd edition. Chicago: University of Chicago Press.

Kayser, Mark A./Arndt Leininger. 2017. “A Länder-based Forecast of the 2017 German Bundestag Election.” PS: Political Science & Politics 50 (3): 689-692.

Norpoth, Helmut/Thomas Gschwend. 2010. “The chancellor model: Forecasting German elections.” International Journal of Forecasting 26 (1): 42-53.

Norpoth, Helmut/Thomas Gschwend. 2017. “Chancellor Model Predicts a Change of the Guards.” PS: Political Science & Politics 50 (3): 686-688.