Modeling Categorical Outcomes: Advanced Methods for Models with Binary, Ordinal, and Nominal Outcomes


Nonlinear models are notoriously difficult to interpret. This workshop explores advanced methods of interpretation that exploit the full potential of models for binary, nominal and ordinal outcomes. Many applications of these models include complications that prevent interpretations using regression coefficients — methods from the linear regression model can no longer be used. Examples include:

  • Nonlinear relationships such as when the probability of employment increases with age until middle age before the probability begins to decrease.
  • Interactions that allow the effect of a variable to differ by the level of another variable. For example, the effect of education on employment might differ by gender.
  • Predictions and marginal effects can be used to compare effects across groups, including extensions of the Blinder–Oaxaca decomposition.
  • Often we want to compare the effects from different models, such as determining if the negative effect of gender on tenure decreases after controlling for productivity. Or, we want to test whether gender has a larger effect on the probability of tenure than on the probability of promotion to full professor.
  • Generalized marginal effect let you examine the effect of changing multiple variables or making proportional changes, such increasing income by 5% rather than $1000.
  • In nonlinear models, the effect of a variable differs among individuals. The distribution of effects lets you answer questions such as which individuals will receive the greatest benefit of an intervention to encourage students to remain in school?
  • Ordinal models make strong assumptions about the dependent variable that are often violated. Specification tests, graphical methods comparing predictions, and statistical tests comparing compare marginal effects can guide your model selection.

Prerequisites: Participants must be thoroughly familiar with linear regression.

Software: These methods are illustrated using Stata, but can be applied in R and other languages. Attendees who use R, SPSS, or SAS can use lab to explore how to use each method in other software. While the instructors are not experts in other software, we can help you translate methods from lecture into other programs. You should have the software you use installed on your laptop. You will be given a temporary license for Stata that you can use during the workshop.

For further information see the workshop webpage or contact the instructor at

Fee: Members = $1700; Non-members = $3200

Tags: probit, logit, logistic regression, categorical outcome

Course Sections