Statistical Methods II: Generalized Linear Models

Catalog description

This course, the second of the statistics sequence for social work doctoral students, will focus on generalized linear models. Building upon the work done in your prior class, we will first learn how generalized linear models differ from linear models. We will briefly review models for binary (dichotomous) outcomes before examining models for outcomes that are ordinal, nominal (three or more categories), and counts. Then we will examine the idea of two-part models, where you can model outcomes through two (or more) equations, and explore some of its instantiations. You will then be introduced to two key concepts – first, fixed and random effects, which are important to understand if you are considering taking a next course on longitudinal data analysis (aka multilevel models). Second, I will give you a brief introduction to multilevel models, which will lay the conceptual groundwork for such a next course on these models. With your prior class, this class, and a future longitudinal/multilevel class, you will have completed the bulk of the slopes-and-intercept approach to data analysis. A foundation in this family of statistical analyses is necessary to your learning variance-covariance matrix-based approaches (e.g., path models, latent class/growth models, or structural equation models) that you may learn in future classes. All of the approaches that you have learned in this, and in your prior class, are associational, not causal. Causal knowledge, however, is the goal of science. Hence, we will end this course with an introduction to the orthodox theory of causality in research, called the potential outcomes framework. In this section, we will address two important approaches to enhancing causal inference - using existing variables ("observables") or leveraging time or something extraneous ("unobservables").