Notice: This website is an unofficial Microsoft Knowledge Base (hereinafter KB) archive and is intended to provide a reliable access to deleted content from Microsoft KB. All KB articles are owned by Microsoft Corporation. Read full disclaimer for more details.

QA: Why are the coefficient estimates produced by rxLogit sometimes different than those produced by R’s glm() or SAS’s PROC LOGISTIC?


View products that this article applies to.

If rxLogit() converges within the maximum number of allowed iterations, and if the same is true of the other program(s), the most likely reason for a difference in estimated coefficients is that different “contrasts” are being used. Another possible reason is that if the model is singular (rank deficient) and variables are dropped to remove the singularities, the values of the remaining coefficients will depend upon which variables are dropped, and this may differ among the programs (this can be viewed as a type of contrast). 

If any of the programs fail to converge, then their estimates are unreliable and cannot be compared. Use of different initial values may lead to convergence. For instance, set the initialValues argument equal to 0 to uses all 0’s as starting values. Or set a vector with the correct number of elements. 

rxLogit displays the condition number of the final variance-covariance matrix as an indicator of the numerical reliability of the results. It also will terminate estimation if the model is too ill-conditioned to yield reliable estimates.

↑ Back to the top


Keywords: kb

↑ Back to the top

Article Info
Article ID : 3104288
Revision : 1
Created on : 1/7/2017
Published on : 10/29/2015
Exists online : False
Views : 55