0
$\begingroup$

This question already has an answer here:

I am doing a binary logistic regression analysis. I got one categorical predictor with 7 levels. When I try to do this in Minitab 17 I get an error message: "The model could not be fit. Maximum likelihood estimates of parameters may not exist due to quasi-complete separation of data points." If I exclude 5 of the levels and only test the two I am actually interested in, I do not get this message. But I guess that simply excluding levels does not come without a cost and would need to be corrected for? (Even if these two, control vs special treatment, were the ones I was interested in.)

Does anyone know of a way out of this (other than adding more data since I do not have that option)? In R maybe? Or another test?

$\endgroup$

marked as duplicate by Scortchi Nov 26 '15 at 16:45

This question has been asked before and already has an answer. If those answers do not fully address your question, please ask a new question.

  • 1
    $\begingroup$ The problem is fundamental. Changing to different software can't solve it. As you are aware, the usual underlying problem is that the data sample is too small; in scientific practice one often can't go out to get more data. It would be good to have a different message for you. $\endgroup$ – Nick Cox Nov 26 '15 at 16:46
  • 1
    $\begingroup$ Note that if you mean you've a single predictor with seven levels, & no other predictors, you don't lose information about the two you're interested in by excluding observations at the other levels from the analysis (not that you should need to, but I don't know how Minitab behaves). $\endgroup$ – Scortchi Nov 26 '15 at 16:58
  • 1
    $\begingroup$ I imagine that @Scortchi would agree that even if 5 levels are excluded, they should remain context for the others, e.g. graphically and descriptively. $\endgroup$ – Nick Cox Nov 26 '15 at 17:28
  • $\begingroup$ @NickCox: Oh yes! The only reason for my comment was I don't know if Minitab's giving only that error message & refusing to return a fitted model. What I'd probably suggest, software willing, is maximum-likelihood estimation as usual with profile-likelihood confidence intervals (perhaps a good idea for all parameters if the sample size is small). $\endgroup$ – Scortchi Nov 26 '15 at 17:34
  • $\begingroup$ @Scortchi: You're correct in that Minitab gives that error without returning a fitted model. For one who doesn't really know much statistics, could you please elaborate your suggestion? I would be very grateful. And just to double check: if I understand you and Nick correctly, simply excluding the other five levels is not possible? $\endgroup$ – Sara B. Nov 26 '15 at 19:14