Least likely observations in regression models for categorical outcomes

This article presents a method and program for identifying poorly fitting observations for maximum-likelihood regression models for categorical dependent variables. After estimating a model, the program least likely will list the observations that have the lowest predicted probabilities of observing the value of the outcome category that was actually observed. For example, when run after estimating a binary logistic regression model, least likely will list the observations with a positive outcome that had the lowest predicted probabilities of a positive outcome and the observations with a negative outcome that had the lowest predicted probabilities of a negative outcome. These can be considered the observations in which the outcome is most surprising given the values of the independent variables and the parameter estimates and, like observations with large residuals in ordinary least squares regression, may warrant individual inspection. Use of the program is illustrated with examples using binary and ordered logistic regression.


Issue Date:
2002
Publication Type:
Journal Article
DOI and Other Identifiers:
st0022 (Other)
PURL Identifier:
http://purl.umn.edu/116014
Published in:
Stata Journal, Volume 02, Number 3
Page range:
296-300
Total Pages:
5

Record appears in:



 Record created 2017-04-01, last modified 2017-08-22

Fulltext:
Download fulltext
PDF

Rate this document:

Rate this document:
1
2
3
 
(Not yet reviewed)