As social scientists, nearly all of us have a binary dependent variable (with discrete categories 0/1) that is of scientific interest. Most of us also know that appropriate models for these types of outcome variables require use of a nonlinear link function that maps an unbounded set of parameters (e.g., Y*=B0 + B1X1 +B2X2 + … +BkXk) onto a probability space where Y = [0, 1]. Most commonly used link functions result in parameters that are interpretable with respect to the logistic cumulative density function (for logit) or the Gaussian cumulative density function (for probit). Although these coefficients are less directly interpretable than coefficients for continuous outcomes, they are still often used for assessing direction of effects, significance, and even relative magnitude. However, even this limited utility breaks down when scientific interest moves beyond main effects and into exploring interactions in binary models (e.g., effect moderation) or comparing coefficients across binary models (e.g., effect mediation). Rather, with respect to interactions, mediation and cross-model comparisons, the identification assumptions and functional forms prerequisite of logit and probit models make the coefficients they produce little more the nuisance parameters that, if interpreted directly, can lead to incorrect conclusions. This lecture will provide an overview of these issues and suggest several alternative methods for appropriately exploring these questions with binary outcomes, largely focused around examining changes in average marginal effects.

Organisé par la Chaire de recherche en études électorales et la Chaire de recherche du Canada en démocratie électorale.

Contactez semra.sevi@umontreal.ca pour rejoindre l’atelier sur Zoom.

DATE : vendredi 19 juin, 13h-14h