2025/11/19 by Phil Swatton · 1 voice
Mathematics · Social Sciences · #Benford’s Law and Fraud Detection #Computational and Text Analysis Methods #Electoral Systems and Political Participation
paper · doi:10.31235/osf.io/m9ft3_v1
openalex publication_date 2025/11/19 · openalex created_date 2025/11/20 · openalex updated_date 2026/07/14
Empirical political science generates knowledge of the world through abductive reasoning, or inference to the best explanation. However, to date, no attempt has been made to compare the explanatory quality of the sociological modelling tradition, which treats vote choice as downstream from the socio-economic positions of voters, and the spatial modelling tradition, which treats vote choice as downstream from the ideology of voters. I argue that the paradigm of predictive inference, or machine learning, offers three contributions towards comparing these traditions within the abductive framework of political science. This depends on a change in emphasis, from seeking to produce maximally predictive models according a given metric to treating the predictive power of well-chosen variables as a theoretically interesting quantity. Taking the 2019 UK general election and 2019 EU parliament election as my primary and secondary case studies respectively, I use supervised machine learning to compare models using demographic variables and ideological variables to predict vote choice. Using British Election Study data I find that models using demographic variables are not much more accurate than assuming all voters vote for the largest party. I find that models using ideological variables perform better, but even these models struggle to predict third party vote choice. Using European Election Study data from 15 countries, I confirm that this result generalises, albeit there is too much uncertainty to confirm it in any particular country. I perform some post-hoc analysis using mutual information and PCA to give further support to my evidence. To my knowledge, this is the first time mutual information has been used to study vote choice. I conclude by arguing that in line with the three contributions my evidence shows that the spatial modelling tradition represents a better explanation in the abductive reasoning framework, and highlight several promising directions for future research.