Why is AIC from multinom and corresponding glm different? I'm reading the example of the book: "S-PLUS (and R) Manual to Accompany Agresti’s Categorical Data Analysis (2002) 2nd edition " (page 55) and when I try to reproduce the example I get this:
snoring<-c(0,2,4,5) 

logit.irls<-glm(cbind(yes=c(24,35,21,30), no=c(1355,603,192,224))~snoring, 
family=binomial(link=logit)) 

summary(logit.irls)

Null deviance: 65.9045  on 3  degrees of freedom
Residual deviance:  2.8089  on 2  degrees of freedom
**AIC: 27.061**

Ok, I get a Akaike information criterion of 27.061.
But when I use the multinorm function from the "nnet" library I get this:
modLogit = multinom(cbind(c(24,35,21,30),c(1355,603,192,224))~snoring)

summary(modLogit)

Coefficients:
  (Intercept)    snoring
2    3.866248 -0.3973366

Std. Errors:
  (Intercept)    snoring
2   0.1662144 0.05001066

Residual Deviance: 837.7316 
**AIC: 841.7316** 

Why the AIC (and the Residual Deviance) are too big?
 A: It's because the minus-log-likelihood function is really only defined up to an additive constant and the absolute values of AIC and residual deviance depend on this constant. Hence, for different implementations you can, in general, expect different choices of constant. Hopefully, the difference in AIC (or the difference in the residual deviance) between two models is the same whether you use multinom or glm.
N.B. In multinom the residual deviance and AIC are computed based on the minus-log-likelihood of independent observations of 0-1-variables, thus 
$$l = - \sum_{i} y_i \log \hat{\pi}_i + (1-y_i) \log(1-\hat{\pi}_i)$$
and AIC$= 2l + 2p$. This is the case even when the data are given in tabulated form. For the AIC computation in glm the minus-log-likelihood is based on independent observations of binomial variables, thus 
$$l = - \sum_{k} n_{k1} \log(\hat{\pi}_k) + (n_k - n_{k1}) \log(1-\hat{\pi}_k) + \log {n_k \choose n_{1k}}.$$
When data are in tabulated form as in the question, the difference between the two reported AIC's is precisely twice the sum of the log of the binomial coefficients. 
