We have comprehensive coverage on the topic of perfect separation in logistic regression. When it happens in R we usually see two warnings:

Warning messages:
1: glm.fit: algorithm did not converge
2: glm.fit: fitted probabilities numerically 0 or 1 occurred

Is it possible that the algorithm did converge and fitted probabilities numerically 0 or 1 occurred nonetheless?

  • 3
    $\begingroup$ The problem seems related to the "vanishing gradient" issue in deep learning (i.e. for logistic near 0 or 1, gradient is near 0). The answer may depend on details of the numerical implementation, but the question seems more general. So I think it is on topic. $\endgroup$ – GeoMatt22 May 2 '17 at 20:31

R is giving you two different warnings because these really are two distinct issues.

Very loosely, the algorithm that fits a logistic regression model (typically some version of Newton-Raphson) looks around for the coefficient estimates that will maximize the log likelihood. It will estimate the model at a given point in the parameter space, see which direction is 'uphill', and then move some distance in that direction. The potential problem with this is that when perfect separation exists, the maximum of the log likelihood is where the slope is infinite. Because a search algorithm has to be designed to stop at some point, it doesn't converge.

On the other hand, no matter where it stops, whether it converged or not, it is (theoretically) possible to calculate the model's predicted values for the data. However, because computers use finite precision arithmetic, when they perform the calculations, they eventually need to round off or drop extremely low decimal values. Thus, if the arithmetically correct value is sufficiently close to 0 or 1, when it is rounded, it can end up being 0 or 1 exactly. Values can have that property and be within the normal range of the data due to an extremely large (in absolute value) slope estimate due to complete separation, or they can just be so far out on X that even a small slope will lead to the same phenomenon.

# I'll use this function to convert log odds to probabilities
lo2p = function(lo){ exp(lo) / (1+exp(lo)) }    
set.seed(163)                             # this makes the example exactly reproducible
x  = c(-500, runif(100, min=-3, max=3), 500)  # the x-values; 2 are extreme
lo = 0 + 1*x
p  = lo2p(lo)
y  = rbinom(102, size=1, prob=p)

m  = glm(y~x, family=binomial)
# Warning message:
# glm.fit: fitted probabilities numerically 0 or 1 occurred 
# ... 
# Coefficients:
#             Estimate Std. Error z value Pr(>|z|)    
# (Intercept)   0.3532     0.3304   1.069    0.285    
# x             1.3686     0.2372   5.770 7.95e-09 ***
# ...
#     Null deviance: 140.420  on 101  degrees of freedom
# Residual deviance:  63.017  on 100  degrees of freedom
# AIC: 67.017
# Number of Fisher Scoring iterations: 9

Here we see that we got the second warning, but the algorithm converged. The betas are reasonably close to the true values, the standard errors aren't huge, and the number of Fisher scoring iterations is moderate. Nonetheless, the extreme x-values yield predicted log odds that are perfectly calculable, but when converted into probabilities, become essentially 0 and 1.

predict(m, type="link")[c(1, 102)]      # these are the predicted log odds
#         1       102 
# -683.9379  684.6444 
predict(m, type="response")[c(1, 102)]  # these are the predicted probabilities
#            1          102 
# 2.220446e-16 1.000000e+00 
  • 1
    $\begingroup$ thanks for your answer. it is very clear. I suspect they are two separate issues, but always see they come together (when perfect separation occur.) Today, I saw only 2nd warning, without first one, my first reaction is "does my algorithm converge"? That is why I am asking this question... $\endgroup$ – hxd1011 May 3 '17 at 1:03
  • $\begingroup$ why we define lo2p function? why not directly use plogis? is that because plogis has protections on numerical overflow? $\endgroup$ – hxd1011 May 3 '17 at 13:29
  • $\begingroup$ @whuber commented my another question on log underflow, is that related? Thanks!! $\endgroup$ – hxd1011 May 3 '17 at 13:38
  • 2
    $\begingroup$ @hxd1011, I mostly do it to make the code more immediately intuitive for people who don't use R. $\endgroup$ – gung May 3 '17 at 14:47
  • $\begingroup$ thanks, you are really thoughtful, I think that is why you got many upvotes, $\endgroup$ – hxd1011 May 3 '17 at 14:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.