Contextual Influence or Individual Attributes? Multilevel Modelling Applied to the Analysis of Electoral Outcomes

Abstract

Traditional regression models typically rest on two key assumptions: 1) that the response variable is numerical and continuous (commonly assumed to follow a normal distribution), and 2) that the observations are independent of one another. However, in many social science research contexts, these assumptions often do not hold. Frequently, the response variable is qualitative—such as binary, nominal, or ordinal—and the independence of observations is often violated, as outcomes from one observation may be correlated with others. Generalized mixed-effects models, also known as hierarchical generalized models, address these limitations. These models are particularly well-suited for cases where the characteristics and behaviors of an observation are thought to be influenced by the cluster or group to which it belongs. This allows researchers to account for both individual-level and contextual-level data within a single model, as well as the interaction between these levels. This paper aims to promote the use of hierarchical modeling as a valuable tool for analyzing data that is structured in levels or clusters, particularly when the response variable is not continuous or normally distributed. To illustrate the technique, we present supporting statistical theory and apply it to analyze results of Brazilian national elections.