By construction, the AME of an independent variable in a logistic regression model depends not only on its coefficient, but also on the predicted values of the dependent variable. The formula for the AME of the -th independent variable is
As mentioned in the previous post, one of the claims made by Mood (2010) is that coefficients of nested models are not comparable, because tend to increase when predictor variables are added to a model, even if these predictor variables are uncorrelated. This post examines how big a problem this is and whether it can be avoided by the use of average marginal effects. This investigation is based on a few simulation studies and an empirical example.
One the most cited article in social sciences Carina Mood (2010) points a couple of problems in the interpretation of results from logistic regression. These problems have led many authors to move away from interpreting logistic regression coefficients, but base their conclusions on average marginal effects (AMEs) instead. Some authors even abandon logistic regression as a method to analyse binary dependent variables altogether and use linear probability models instead (e.g. Maneuvrier-Hervieu et al. 2025). In this and the following blog posts I’ll take a look at Mood’s claims and ask how serious the problems uncovered by Mood are and how helpful her suggested solutions are.
Our paper on "After secularisation? A comparative analysis of religious cleavages in Western Europe" is now published as open access by the European Journal of Political Research (Elff et al. 2026)
The recently published version 0.99.31.6 of the memisc package also contains an
%$$% operator that simplifies routine data preparation steps that hitherto would
involve calls to the function within(). It is analogous to the operator %$%,
which is provided by the “magrittr” package, but is also defined by this
package.
These operators are illustrated by the following code examples.
Version 0.99.31.6 of package “memisc” was recently (3rd March 2023) published on
CRAN. One of the
new features of this version is the %if% operator which allows
to assign values to subsets of observations. To see how it
works, consider the following example:
I implemented this feature on suggestion from a colleague who missed such a feature for data preparation.
The package 'RKernel' allows you to run R in Jupyter notebooks etc. I have been working on the package on and off since summer 2021. I started because I was dissatisfied with the default rich output that the already existing IRkernel produces. Also, the help system seemed to be broken with version <= 1.3 of 'IRkernel' when used with a more recent version of R (>=4.3?).
My new book "Data Management in R: A Guide for Social Scientists" is now in print and is announced by SAGE to appear on 26 December 2020, right on time for Boxing Day!
Supporting material can be found on the page "Data Management in R: A Guide for Social Scientists".