Dirichlet-Swing: understanding spatio-temporal aspects of political elections in heterogeneous societies through agent-based simulation

Abstract

Many countries have a system of electing members to their governing bodies through district-based elections. In each district, the party with maximum votes wins the corresponding “seat” in the governing body. However, the final seat distribution is strongly dependent on the geographical distribution of voters of different parties, and the party with most (or least) voters may not win the most (or least) number of seats if their voters are non-homogeneously distributed over the districts. This is further complicated in heterogeneous societies, where political preference of voters depends on their social identities, which is also related to their districts of residence. Projections of outcomes by sample surveys tend to fail in such situations. The aim of this paper is to explore how electoral outcomes are influenced by the geographical distribution of voters and community-centric voting preferences. We consider agent-based modeling of voters along with their locations, community memberships and voting preference. Our models represent the relations between these factors with their uncertainties through conditional probability distributions involving latent variables with Dirichlet Processes. Our models also represent spatio-temporal factors in elections – how geographical proximity between districts influence the voting preferences, and swing of votes across successive elections. We propose two novel models for vote swing between successive elections based on Dirichlet Processes, which is far more powerful than the existing models of Uniform Swing and Proportional Swing. For any choice of parameters, our models can be used to simulate a full election by Monte Carlo Sampling, and such simulations provide us a range of possible outcomes. We can also simulate surveys and study how their projections can deviate from the actual results. We discuss inference approaches to estimate the parameters to fit the model to actual district-based elections held in India.