Distinguishing Occasional Abstention from Routine Indifference in Models of Vote Choice

Abstract

R esearchers commonly employ multinomial logit (MNL) models to explain individual-level vote choice while treating "abstention" as the baseline category. Though many view abstainers as a homogeneous group, we argue that these respondents emerge from two distinct sources. Some nonvoters are likely to be "occasional voters" who abstained from a given election owing to temporary factors, such as a distaste for all candidates running in a particular election, poor weather conditions, or other temporary circumstances. On the other hand, many nonvoters are unlikely to vote regardless of the current political climate. This latter population of "routine nonvoters" is consistently disengaged from the political process in a way that is distinct from that of occasional voters. Including both sets of nonvoters within an MNL model can lead to faulty inferences. As a solution, we propose a baseline-inflated MNL estimator that models heterogeneous populations of nonvoters probabilistically, thus accounting for the presence of routine nonvoters within models of vote choice. We demonstrate the utility of this model using replications of existing political behavior research.

Modeling-Related Processes With an Excess of Zeros

“…This example highlights both the issues related to ignoring an excess of zeros and the benefits in addressing them. If the two steps were ignored, the nuanced effects of these covariates would be lost and the estimates of the effects would be biased, because a standard model with no inflation would lead to a correlation between the error terms and the explanatory variables (Bagozzi and Marchetti 2014; Dunne and Tian n.d.).…”

The baseline-inflated multinomial logit model for international relations research

“…In such instances, one’s MNL estimates will be biased downward, as a significant portion of status quo cases will be impervious to the effects of covariates on transition probabilities from status quo (or peace) to other choice categories. Furthermore, when covariates have contrasting effects on the probabilities of both inflation and active multinomial choice, MNL coefficients may become further biased in indeterminate directions (Bagozzi and Marchetti, 2013).…”

“…reference) category or as one of the primary choice outcomes, although for ease of exposition I focus on the baseline inflation case here. To assess my claims, I explore a recently developed BIMNL model (Bagozzi and Marchetti, 2013), which—in a comparable fashion to zero-inflated count models—accounts for inflated cases via the estimation of a “first stage” logit equation. In essence, the logit equation to the BIMNL model estimates each observation’s latent propensity for inflation, and then conditions the BIMNL model’s subsequent MNL-stage coefficient estimates upon these estimated inflation probabilities.…”

“…To address these concerns, the present study considers the applicability of a newly developed baseline-inflated multinomial logit (BIMNL) model (Bagozzi and Marchetti, 2013) to outcomes of discrete unordered choice in international relations research. Building upon extant zero-inflated models for discrete binary and ordered outcomes (Beger et al, 2011; Bagozzi et al, 2015; Harris and Zhao, 2007), the BIMNL model accounts for baseline category inflation by simultaneously estimating (a) an observation’s likelihood of being baseline inflated (via a binary logit equation) and (b) an observation’s probability of selecting each outcome of nominal choice—conditional on not being baseline inflated—via a multinomial logit (MNL) equation.…”

“…Building upon extant zero-inflated models for discrete binary and ordered outcomes (Beger et al, 2011; Bagozzi et al, 2015; Harris and Zhao, 2007), the BIMNL model accounts for baseline category inflation by simultaneously estimating (a) an observation’s likelihood of being baseline inflated (via a binary logit equation) and (b) an observation’s probability of selecting each outcome of nominal choice—conditional on not being baseline inflated—via a multinomial logit (MNL) equation. By estimating these two processes within a single system, the BIMNL model is able to recover unbiased covariate estimates for one’s polytomous dependent variable of interest (Bagozzi and Marchetti, 2013). I accordingly evaluate the effectiveness of the BIMNL for international relations research via replications of two existing MNL analyses.…”