1

I had a very large value for Exp(B) in SPSS binary logistic regression. What is wrong and what should I do?

The coefficients are shown on the picture

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
Elena Vassiljeva
  • 13
  • 1
  • 3
  • Which value of $B$ is "very large"? The largest one I can see is $0.703$ for "D2(1)", because it is more than twice its standard error, but there is nothing unusual or alarming about that. – whuber Apr 22 '15 at 16:29
  • 4
    Try search for "complete separation." It seems some of the predictors appear to be a constant or some zero cells were created. – Penguin_Knight Apr 22 '15 at 16:33
  • 1
    @Penguin_Knight I've created the hauck-donner-effect tag for this express purpose! – Sycorax Apr 22 '15 at 17:10

1 Answers1

2

This usually indicates a problem with Quasi-Complete (categorical variables) or Complete Separation (continuous variables).

Quasi-complete: Where you have one category of a class variable x that has only one type of outcome (y=0 or y=11)

Complete Separation: some breaking point in a continuous variable where target has only one type of outcome on either side. For example, if $x>5$ then y=1 in all cases and if $x\leq 5$ then y=0 in all cases.

ShainaR
  • 173
  • 6
  • Complete separation is only when you have y=0 for x=c for some value of c (and 0 and 1 might be flipped) but just having a point where either above or below you only see a single type of response doesn't result in complete separation. – Dason Oct 03 '16 at 17:22
  • Right. I guess my answer was not exactly clear on that. updated. – ShainaR Oct 05 '16 at 17:46