The talk presents recent methodological work on comparisons of coefficients across same-sample nested logit or probit models.

In quantitative sociological research, researchers have long compared the coefficient of a predictor variable between models that successively add control or blocks of control variables. In this approach, coefficient change is taken to represent either confounding or mediation, two concepts that are central for testing sociological explanations. Yet whenever researchers compare coefficients between models using logit or probit, the coefficient change does not yield an unbiased estimate of the degree of confounding or mediation.

The talk explains, both intuitively and technically, why this is so, and it demonstrates how a new method solves the problem. The talk also touches on a related issue, namely the problem that arises in comparisons of logit or probit coefficients across layers such as countries, birth cohorts, social class, gender, or race/ethnicity.

The talk shows that this latter comparison problem, having far-reaching consequences for applied sociological research, is as serious a problem as the former, and it briefly discusses how sociologists may deal with the problem.

Lecture on recent developments in the use of nonlinear probability models in sociology