1

I analyzed four binary variables with a binary logistic regression.

I used for the analysis of all four binary y-variables the same set of 9 x-variables (a-i).

My problem deals with one of this four regressions. The results of the binary logistic regression shows that the x-variable "c" takes surprising high values regarding the regression coefficient B, Wald and Exp(B) (see table 1) indicating that this variable problematic.

That is why I excluded this variable "c" from this one of four regressions.

I am now in a difficult situation. I already wrote the scientific publication about this analysis, and I have to send it within a few days to the Journal. When I did this analysis, I asked my supervisor about this one variable c and he said it would be ok to leave this one variable out, whereas I did not do it in the other three regressions. But he is really no expert in statistics.

I am trying to include a few sentences in the paper to explain why I did it. If the Reviewers are not fine with it, it would be o.k.. I would be fine with calculating it again and I think the paper all in all is great, and I would be surprised if I would get a rejection.

But the explanation should be plausible with respect to the special characteristics of this variable c within the regression results. The problem is that I cannot calculate until submitting the paper.

Could you give me some advice? What do you think about the regression results regarding this variable "c"?

(I know, I break the rule of thumb of 10 events per item in logistic regressions; that´s a limitation I already mentioned in this paper)

Table1 regression results

enter image description here

Table2 correlation matrix enter image description here

Ole
  • 81
  • 6
  • 1
    See this [question and answer](http://stats.stackexchange.com/questions/11109/how-to-deal-with-perfect-separation-in-logistic-regression). – mdewey Oct 04 '16 at 10:35
  • Ok, thanks. Do you think the coefficients of variable "c" are inflated? And it is maybe an argument to say, the coefficients are inflated and therefore the variable was left out? – Ole Oct 04 '16 at 11:09
  • I think you have separation as discussed in that link. Leaving the variable out means leaving out your best predictor. Is that what you really want to do? – mdewey Oct 04 '16 at 12:14
  • its kinda complicated. In fact im using a zero-one-inflated beta regression for fractional data with masses at zero and one [0,1]. a higher value of the variable of c is both related to a higher possibility of having a value of (0,1) and not 0. And on the other hand a higher probability of having a value of (0,1) than 1. Therefore the interpretability is not he best. If i would use a one-part model, there would be no effect of this variable. Furthermore, other variables have a much stronger effect when I include variable c. So I am quite mistrustful about this variable – Ole Oct 04 '16 at 13:21
  • In your original question you clearly stated you were using a binary logistic regression and your output said the same. Now you are telling us you are using a zero-one inflated beta regression. I am lost for words. – mdewey Oct 04 '16 at 19:59

0 Answers0