Logistic regression: Strange standard errors from glm() in R

Question

To my surprise I found that standard errors and thus Wald confidence intervals became smaller when I removed the intercept from a simple logistic regression model, using glm() and R.

# load an object named "my.df" to the global enviroment
load(url("http://hansekbrand.se/code/test.RData"))
# fit a model with intercept to data
my.fit <- glm(deprived.of.education ~ religion, data = my.df, family = binomial("logit"))
# fit a model without any intercept to data
my.fit.without.intercept <- glm(deprived.of.education ~ 0 + religion, data = my.df, family = binomial("logit"))

# inspect the first fit
summary(my.fit)$coefficients
#                        Estimate Std. Error   z value     Pr(>|z|)
# (Intercept)          -2.8718056 0.03175130 -90.44687 0.000000e+00
# religionChristianity  0.4934891 0.03234887  15.25522 1.519805e-52
# religionHinduism      0.5257316 0.03376535  15.57015 1.161317e-54
# religionIslam         1.5734832 0.03231692  48.68914 0.000000e+00
# religionNonreligious  1.5975456 0.03555164  44.93592 0.000000e+00

# inspect the second fit
summary(my.fit.without.intercept)$coefficients
#                       Estimate  Std. Error    z value Pr(>|z|)
# religionBuddhism     -2.871806 0.031751299  -90.44687        0
# religionChristianity -2.378317 0.006189045 -384.27842        0
# religionHinduism     -2.346074 0.011487113 -204.23530        0
# religionIslam        -1.298322 0.006019850 -215.67354        0
# religionNonreligious -1.274260 0.015992939  -79.67642        0

I understand why the z values are different, because the null hypotheses in the two cases are different. In the first case, with the intercept, the null is "same as the reference category", while without the intercept, the null becomes "zero".
But I do not understand the large difference in standard errors between the two models.

Without the intercept, the standard errors seem to vary with n of each level, i.e. there are many cases of "Christianity" and "Islam", and they have small standard errors, but with the intercept, there is essentially no variation in the standard errors.

Could someone please explain the reason for the differences in the magnitude of the standard errors between the two models?

I would like to calculate probabilities and confidence intervals around them, and I have done so using the estimates from the first model. If I would do that with the estimates from the second model, the confidence intervals would be much smaller, but would they be reliable?

Answer 1

Your coefficients, even when they share common names, are not the same, i.e. their interpretation is different.

In the first model, the effect of religionChristianity is a variation in the outcome wrt the baseline (religionBuddhism), a relative variation. In the second model the effect of religionChristianity is an absolute variation.

The effects are numerically equal, $-2.8718056+0.4934891=-2.378317+5e-07$, but in the first case the effect is a sum of two effects, i.e. you should compare the joint significance of Intercept and religionChristianity in the first model with the significance of religionChristianity in the second one. You should compare a joint confidence interval (first model) with a simple one (second model).

The simple CI for religionChristianity is:

> confint(my.fit.without.intercept)
Waiting for profiling to be done...
                         2.5 %    97.5 %
...
religionChristianity -2.390467 -2.366207

There are several ways to compute joint intervals. Using arm:

> library(arm)
> n.sims <- 1000
> sim.i <- sim(my.fit, n.sims)
> intercept.plus.christianity <- sim.i@coef[,1] + sim.i@coef[,2]
> quantile(intercept.plus.christianity, c(0.025, 0.975))
     2.5%     97.5% 
-2.390826 -2.366828 

Can you see any significant (relevant) difference?