Estimation of Marginal Effects using Margeff

Abstract

This article describes the user-written program margeff, which enables the fast estimation of (average) marginal effects. Besides describing the program, this article offers a new discussion of some problems that are related to computation of marginal effects. I will argue that (1) marginal effects computed at means are not good approximations of average marginal effects, computed as means of marginal effects evaluated at each observations, if some of the parameter estimates are large; (2) both average marginal effects and marginal effects computed at means might produce wrong estimates for dummies that are part of a set of indicator variables indicating different categories of a single underlying variable; and (3) the use of marginal effects computed at means is preferred if some of the regressors are mathematical transformations of other regressors.

A New Piece of the Puzzle: Sexual Orientation, Gender, and Physical Health Status

“…We then conduct pairwise comparisons of the difference in probabilities (the average marginal effects) for Group 1 (e.g., heterosexual men) versus Group 2 (e.g., bisexual men). Using the observed data, instead of the more common approach of setting all model covariates at their means, is beneficial in that it isolates the influence of the gender-by-orientation characteristic (Bartus 2005). Table 1 presents weighted mean and percentage values, stratified by gender and sexual identity, and shows that bisexual adults report the highest rates of poor health.…”

Borrowing in foreign currency: Austrian households as carry traders

“…For ease of interpretation and comparison we follow the same procedure for our dummy and categorical independent variables, as none of these variables describe different categories of a single variable (Bartus, 2005). The sums of the estimated partial derivatives equal zero across the four loan types (No Loan, Other Loan, Housing Loan in euro, Housing Loan in Foreign Currency) of the multinomial logit model because the sums of the probabilities must equal one.…”

Student evaluation based indicators of teaching excellence from a highly selective liberal arts college

“…The more familiar Marginal Effects at Means (MEM) are simply the marginal effects computed at the sample means. Bartus argues that AMEs are more realistic when some of the independent variables are dummy variables (Bartus, 2005). AMEs have been computed while accounting for dummy variables values changing from 0 to 1 and count variables changing by an integer value of 1.…”

“…Employing instructor specific dummies for all instructors is not possible because of near multi-colinearity with the CAUCASIAN variable. Average marginal effects are computed for the Ordered Probit models taking into account that several of the independent variables are integer count variables and dummy variables according to the procedure outlined in Bartus (2005). The Ordered Probit Model was subjected to robustness checks by running subsample regressions of SET scores of individual instructors/courses including only semester and course controls.…”