Modeling Headaches that Result from Psychometric and Distributional Properties of Outcome Variables [Online]

Abstract

This talk will focus on the modeling headaches that result when statistical models are applied to different kinds of outcome variables. In particular, I focus on the ways in which the psychometric properties of outcome variables can lead to catastrophic problems when interaction estimates are of interest. I first discuss this problem in the context of an outcome variable that is heteroscedastic. I then discuss results from a suite of simulation studies showing that power and false discovery are common problems when analysis of interactions is undertaken in the presence of non-normal outcome variables (i.e., binary, count, censored, non-interval). Finally, I demonstrate that more conservative approaches (e.g., treating seemingly continuous outcomes as ordinal ones) may both uncover additional insights and protect against the most severe problems in some cases.

Ben Domingue is an assistant professor in the Graduate School of Education at Stanford University. He is interested in how student outcomes are leveraged to inform our understanding of student learning, teacher performance, and the efficacy of other programs. He has a particular interest in the technical issues that make it challenging to draw simple inferences from such student outcomes. While not analyzing item response data, he may be found thinking about the implications for social science of the sudden increase in our capacity to measure human DNA and the promise and pitfalls associated with how this new data may change our understanding of human behavior.

Date: 
Tuesday, April 13, 2021 - 2:00pm
Building: 
Online session
Room: 
Zoom