Currently I am analyzing a dataset using logistic regression, I ran it in R using the glm function to run a multivariate logistic regression with 12 predictors. Some of these are quite collinear as they resulted from feature engineering. I was thus not really surprised that my algorithm did not converge and turned out NA for some of the engineered features. I proceeded to run it in two parts and report the results in that way.

However, my collaborator wanted to check the results to make sure no error occurred. She prefers using JMP and thus ran it in that software with the built-in analysis available. JMP reported estimates for all parameters without complaining about convergence.

My question is thus, how do I deal with this? Are the parameters estimates obtained sound or might this indicate hidden convergence issues?

  • $\begingroup$ Rows of NA's in the table of coefficient estimates often means exact linear dependence between features. If that's true in your case, why would you want to keep all engineered features? $\endgroup$
    – dipetkov
    Sep 19, 2022 at 18:07
  • $\begingroup$ The reason is that the engineered features are considered to be clinically relevant. However, the raw features might be the real reason they are so and are seldom compared. My idea was to input them in the same logistic regression to see what would come out. Normally I would accept that to much colinearity occured and fit two models instead. However, this would render direct comparison impossible if I'm not mistaken (models will be non-nested), and additionally I'm reluctant to do so since both statistical software do not agree on the convergence issue. $\endgroup$ Sep 20, 2022 at 8:46
  • $\begingroup$ I actually don't understand what you mean by fitting two models. And, if the transformed features are considered clinically relevant, why not keep only those? If the NA's are indeed due to exact linear dependencies, then this you don't lose information by dropping inputs which are exact linear combinations of the remaining inputs. $\endgroup$
    – dipetkov
    Sep 21, 2022 at 10:10

1 Answer 1


Software implementations of the same model often use different estimation methods, so the fact glm and JMP (which I am not familiar with) produce different results is no surprise. Further, even if they both used the same estimation algorithm they may do so with different default settings (e.g., convergence criteria, starting values), which may also lead to different results. So the fact that one implementation of your logistic regression model did not converge is not surprising.

If I were in your position, I would compare how glm and JMP estimate your model parameters. However, if you are not interested, I would suggest simply using results obtained from JMP (i.e., the software where your model converged) since convergence issues should never be ignored.

Finally, you could try using point estimates obtained from JMP as starting values in glm. It would be interesting to see if this would help with convergence.

  • $\begingroup$ Thanks for your reaction. I found out which algorithm is employed by which method. If I'm informed correctly, glm uses fisher scoring, whereas JMP employs newton's method. Furthermore I found some indication that Fisher tends to converge better than Newton, so I'm actually quite confused. I would compare both, however, since R omits 4 of my variables and then re-estimates all the others, it leads to wildly different p-values and parameter estimates. I could just use the JMP output but would this be wise? $\endgroup$ Sep 20, 2022 at 8:50
  • $\begingroup$ Considering the fact that a) Fisher should converge better and b) criteria for convergence might differ. Couldn't it then be the case that JMP "hides" the non-convergence issue in a sense by being less strict? $\endgroup$ Sep 20, 2022 at 8:51
  • 1
    $\begingroup$ Nevermind, I found an error in my code. The non-convergence originated from this error. Nonetheless, I wish to thank you for your thoughtful answer, it let me think on convergence issues in general which might come in handy in the future. $\endgroup$ Sep 20, 2022 at 9:17
  • $\begingroup$ Great, i’m glad things worked out! $\endgroup$ Sep 20, 2022 at 11:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.