0

Is the only reason why fitted probabilities of 0 or 1 occur is that some of your predicting variables(x) are perfect linear combinations of the target(y) variable? Is there any other reason?

Sycorax
  • 76,417
  • 20
  • 189
  • 313
italo
  • 121
  • 6
  • The original question was closed as unclear but I don't think the most recent edit has helped clarify things - "why fitted probabilities occur" should probably say "why fitted probabilities of 0 or 1 occur"? – Silverfish Jan 26 '18 at 10:33
  • I have updated the question. Thank you @Silverfish! – italo Jan 26 '18 at 14:32

1 Answers1

0

For the reasons, in addition to the situation that the predictor x is a perfect linear combination of the response y, quasi-complete separation will also induce the same error.

Theoretically, this occurs when the maximum likelihood estimate of your regression coefficient does not exist. Simply put, for example, imagine we have an 2*2 contingency table formed by the "problematic" predictor x and the response y, it is possible that there is an 0 in the table, which obstructs the calculation of MLE. This situation (quasi-complete separation) induce the "fitted probabilities numerically 0 or 1 occurred" error. Hope it helps.

son520804
  • 451
  • 2
  • 9
  • 1
    You're welcome. Btw, if you are confused about the solution about this problem, you could also consult this post and a nicely formatted solution is written by @Scortchi. https://stats.stackexchange.com/questions/11109/how-to-deal-with-perfect-separation-in-logistic-regression – son520804 Jan 23 '18 at 01:13