Love in the Times of Political Polarization. Drivers of Social Distancing and Affective Polarization in Spain Based on a Conjoint Experiment on Partner Choice

Studies on affective polarization have traditionally discussed which causal mechanisms have the greatest influence on the phenomenon. In particular, it has been debated whether partisan identity is more important than ideological and issue positions or the other way around. Without an experimental approach, the evidence is contradictory and sometimes too dependent on the national context. However, in this paper I argue that the conjoint design is the best possible alternative to resolve a far-reaching debate and I use for this purpose an example of a conjoint experiment applied to a representative sample of the Spanish population. In my conjoint experiment, participants were asked to show their preference for a potential boyfriend/girlfriend. Each profile was composed by combining different attributes including: the party the person voted for, their ideology, some social identities (religiosity, economic class and territorial feelings) and their position on different issues. The results of the experiment show that partisan identity is the attribute that most influences the choice of partner, well above ideology. Similarly, I found a clear process of social distancing from radical right voters (Vox). This evidence allows us to reaffirm the protagonism of partisan identities (acting as mega-identities) in the explanation of polarization, also in multiparty systems such as the Spanish one. The paper also discusses why it is preferable to measure the results of the experiment using a Bayesian hierarchical (HB) model. Unlike the aggregate multinomial logit (MNL) model, the HB model allows us to calculate utilities at the individual level and, at the same time, the utilities form a distribution between that is governed by meta-parameters. The analyses carried out, using the 'ChoiceModelR' package, which applies Markov chain Monte Carlo (MCMC) algorithms to estimate a hierarchical multinomial Logit model with a normal heterogeneity distribution, address some of the weaknesses of classical regression and AMCE analyses, especially the effect of preference variability on the validity of the estimators.