2

I have done binary logistic regression for a dichotomous outcome and used 5 predictors (3 continuous and 2 dichotomous); one of the dichotomous predictor gave a big number of OR and 95% CI (108.28, CI= 6.64- 1764.6, $p < .001$). Is such a big number okay to report, or is something wrong? Sample size was 55 cases.

By the way, the continuous independent variables were linearly related to the logit of the dependent variable. I tested this Box-Tidwell procedure by adding the interaction terms between the existing continuous independent variables and their natural log transformations to the logistic regression, and they were nonsignificant.

Glen_b
  • 257,508
  • 32
  • 553
  • 939
Emaddin Kidher
  • 127
  • 1
  • 7
  • 4
    No, this doesn't make sense. You have 55 people, and you're (almost) saying that something is 108 times more likely. You don't have enough data, and have a quasi-continuous separation problem. – Jeremy Miles Jan 06 '14 at 23:42
  • Your confidence interval test is mangled. Please fix it. – Glen_b Jan 07 '14 at 10:08
  • @JeremyMiles: I do not fully agree. As long as the estimate is presented along with a confidence interval for the true OR, then why hide the result? – Michael M Jan 07 '14 at 13:00
  • Thank you all very much for your kind reply. What can I do in such situation? I can get more data. If i report such results, would the examiners reject the conclusion? Is there any alternative test to do or type of results that I can used, some mention log B. What about bootstrap method would that help? thank you again for your contributions – Emaddin Kidher Jan 07 '14 at 13:29
  • See [here](http://stats.stackexchange.com/questions/11109/how-to-deal-with-perfect-separation-in-logistic-regression) for some ways of dealing with separation. – Scortchi - Reinstate Monica Jan 12 '14 at 22:11

0 Answers0