Truly descriptive research with models

• In Event: Causal Inference for Sociology

Abstract

Descriptive research---inquiry that summarizes the world as it exists---is one of the most foundational social science tasks. Descriptive research questions can be powerfully simple. For example, what proportion of the U.S. labor force is employed? Yet many descriptions in sociological research are produced within the context of multiple regression models. We review the theoretical motivations for these ostensibly descriptive models, highlighting the motivations that appeal implicitly to causal goals and those that are truly descriptive. We propose that the mingling of descriptive and causal goals has come at a cost: many truly descriptive research questions remain unanswered in sociology. We highlight how many sociological theories are truly descriptive, and we show an approach to descriptive research centered on predicted outcomes ŷ instead of regression coefficients β̂. We show how this approach enables the use of machine learning estimators for descriptive questions, and we discuss the limited interpretations one can make when carrying out descriptive analysis. We illustrate our approach with two use cases. Broadly, our framework is a call for truly descriptive research with models.

Authors

• Ian Lundberg, University of California, Los Angeles

• Kristin Liao, University of California-Los Angeles