Binary logistic regression warning message model summary help!!

[flo891245]

Hi, I have to run a binary logistic regression, to test the effect of perceived benefits (interval, continuous) on personal info disclosure (binary, 0;1). I also include control variables in the model, of which one is age. This control variable is coded from 1-8, since it was measured on an ordinal scale (1 = < 18 years, 2 = 18 - 24, etc.). However, the below the model summary table it is stated that: "Estimation terminated at iteration number 20 because maximum iterations has been reached. Final solution cannot be found". When I check the coefficient estimates of age, they are all extremely large.

Therefore, I did a median split for age and the problem specified above is gone. However, I am not sure whether I need to do this median split (or something else) or if the above stated warning message does not pose a threat to interpret the results. The thing is that I do not need to interpret the effect of Age, but rather the effect of perceived benefits on the dependent variable.

 

[hlsmith]

It sounds like the model is failing to converge. Multiple things could be going on. The most common is complete separation, meaning that in one of the groups all members are in one of the outcome groups. You can check for this.

Work arounds sometimes can be, changing the number of iterations the program should try. This doesn't fix the underlying issue but may allow the model to eventually settle. I think running exact or penalized models could help too.

 

[fed2]

I think that for logit models this is usually caused by a 'plane of complete seperation', ie responses are completely divided by the predictors. That would explain why the recode fixed it.

penalized models

probably the way to go. 'firths method'?

 

[hlsmith]

Yeah, with small samples or sparsity - Firth's is a simple and not too esoteric option.