I've been doing some work on regression, and a paper in particular caught my attention where they utilised two categorical variables in a logistic regression. From my understanding, if you're using two categorical variables, does that not just come down to conditional probabilities? Out of interest, I wrote some code, and was surprised to see that the method returned a p value (reproducible code below). I'm struggling to understand what the null and alternative hypotheses are in this context, so that I can interpret what the p value is telling me. Any assistance in dissecting the resulting model summary would be greatly appreciated.

Example interpretation:

P(Dep = Class1 | Pred = C) = 20/120 = 1/6

Is the model summary telling me this?


foo <- data.frame(Pred   = c(rep("A",80),rep("B",20),
                  Dep    = c(rep("Class1",120),
fit      <- glm(Dep ~ Pred, family=binomial(link='logit'), data = foo)


glm(formula = Dep ~ Pred, family = binomial(link = "logit"), 
    data = foo)

Deviance Residuals: 
     Min        1Q    Median        3Q       Max  
-1.17741  -0.00003  -0.00003   0.00003   1.17741  

              Estimate Std. Error z value Pr(>|z|)
(Intercept) -2.157e+01  3.268e+03  -0.007    0.995
PredB       -7.168e-11  7.308e+03   0.000    1.000
PredC        2.157e+01  3.268e+03   0.007    0.995
PredD        4.313e+01  4.992e+03   0.009    0.993

(Dispersion parameter for binomial family taken to be 1)

    Null deviance: 269.205  on 199  degrees of freedom
Residual deviance:  55.452  on 196  degrees of freedom
AIC: 63.452

Number of Fisher Scoring iterations: 20
  • 1
    $\begingroup$ Your formula looks like it is backwards. Shouldn't it be Dep ~ Pred? $\endgroup$
    – Wayne
    Dec 13, 2016 at 13:20
  • $\begingroup$ Good Point, I'll adjust that. $\endgroup$ Dec 13, 2016 at 13:25
  • $\begingroup$ If you reverse the formula the resulting model should give P(Dep = Class2 | Pred = C) = 1 and P(Dep = Class2 | Pred = D) = 1. Look at a cross table of Pred and Dep and you'll notice that Class1 is not observed when Pred is either C or D $\endgroup$
    – Niek
    Dec 13, 2016 at 13:33
  • $\begingroup$ This should give you the conditional probabilities that you're looking for: 1/(1+exp(fit$coefficients[1])); 1/(1+exp(fit$coefficients[2])); 1/(1+exp(fit$coefficients[3])); 1/(1+exp(fit$coefficients[4])) $\endgroup$
    – Niek
    Dec 13, 2016 at 13:39
  • $\begingroup$ Ok, sorry to be a pain, but I'd changed the numbers slightly so that the example makes a bit more sense. So in the new example, I'd expect that P(DepCClass1 | Pred = C) = 20/120 = 1/6, however when applying your formulae, I see 1. $\endgroup$ Dec 13, 2016 at 13:53

1 Answer 1


Your intuition is right.

The logistic regression model gives identical inference to a Pearson chi-sq test of "independence" for categorical data (in the long run). In both cases, the null hypothesis is that the conditional probabilities are equal to the marginal probabilities. You can show with some algebra that this necessarily implies that the odds ratio is 1.

The minute differences in the actual logistic regression and Pearson test statistics owes to how they're computed. The $p$values you get in R from calling summary.glm come from a Wald test whereas the Pearson test is a related Score test.

The problem with your example is that you have a singular logistic model. One advantage of the score test is that it can provide test statistics for models like this one which "explode". For a more sane example, consider the following:

foo2 <- as.data.frame(sapply(foo, sample)) ## permute to avoid singularity
fit <- glm(Dep ~ Pred, data=foo2, family=binomial)
waldtest(fit, test='Chisq')

Gives us:

> waldtest(fit, test='Chisq')
Wald test

Model 1: Dep ~ Pred
Model 2: Dep ~ 1
  Res.Df Df  Chisq Pr(>Chisq)
1    196                     
2    199 -3 1.7774     0.6199
> chisq.test(table(foo2))

    Pearson's Chi-squared test

data:  table(foo2)
X-squared = 1.7882, df = 3, p-value = 0.6175
  • $\begingroup$ +1. In terms of "exploding", it is possible with Bayesian logistic regression to keep the coefficients from exploding by imposing reasonable priors. Not sure that this is actually useful, but it might be. $\endgroup$
    – Wayne
    Dec 13, 2016 at 15:01
  • $\begingroup$ @Wayne it's a good suggestion, I don't know how anything but an informative prior could prevent separation in logistic regression. Maybe we ought to rethink how "bad" informative priors are. When I've encountered singular logistic regression in practice, though, I simply comment on those findings and present the results from a score test, also inverting to obtain a CI which includes $\infty$. $\endgroup$
    – AdamO
    Dec 13, 2016 at 17:49

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.