Home
Teaching
Research
SPost Commands
Workflow of Data AnalysisContact and vita
Links
Recommendations
FTP downloads

Group comparisons in regression models
Working PaperI am pleased to note that Alexandra Seremina has translated into Romanian the article "Group Comparisons in Logit and Probit Using Predicted Probabilities". It is available courtesy of azoft. I greatly appreciate her work. The following is the latest draft of the 2005 paper that I have sometimes cited in substantive papers. It contains my latest work on how predicted probabilities can be used to compare groups in regression models for binary (and ordinal and nominal) outcomes. J. Scott Long. 2009. "Group Comparisons in Logit and Probit Using Predicted Probabilities." Working paper draft 20090625. Click here to download. The draft paper uses black and white figures since that is what is most common in journals. But, for several graphs, color versions are much easier to understand. Here are color versions of some of the figures: Figure 4, Figure 5, Figure 6, Figure 7, Figure 8, Figure 9, Figure 10, and very cool 3D color Figure 11. Software for group comparisonsThe analyses require software that: 1) estimates binary logit or binary probit; 2) computes out of sample predictions of probabilities (i.e., predictions at specified values of the RHS variables, but values that might not occur in the observed data); 3) tests of differences between predictions. Most standard packages can produces 1 and 2. The SPost package by Long and Freese can compute tests of differences and makes it relatively simple to produce plots such as those shown in the paper above. However, doing this can be very tedious. Accordingly, I have written a wrapper to the SPost program prvalue that simplifies group comparisons. This package is called GC (for group comparisons) and consists of several utility programs. To install this Stata package and documentation, you must be in Stata 9 or later and connected to the internet. Enter the command: findit scottlong and follow instructions to download the package. The example programs that begin with gcpaper*.do are used for the results in this paper. After installation, type help gc for further details. jslgcadodocumentation.pdf gives an overview of the commands. Warning: This set of programs makes it much simpler to do these analyses, but it is not fully tested. Let me know if you have problems. FYI, the 3D figure was created using Matlab. 
© 2014 J. Scott Long 