[R] problem with logLik and offsets

[Jarrod Hadfield]

Hi,

I'm trying to compare models, one of which has all parameters fixed 
using offsets.  The log-likelihoods seem reasonble in all cases except 
the model in which there are no free parameters (model3 in the toy 
example below).  Any help would be appreciated.

Cheers,

Jarrod

x<-rnorm(100)
y<-rnorm(100, 1+x)

model1<-lm(y~x)
logLik(model1)
sum(dnorm(y, predict(model1), summary(model1)$sigma,log=TRUE))

# no offset - in agreement

model2<-lm(y~offset(rep(1,100))+x-1)
logLik(model2)
sum(dnorm(y, predict(model2),summary(model2)$sigma,log=TRUE))

# offset and free parameters - in agreement

model3<-lm(y~offset(rep(1,100))+offset(x)-1)
logLik(model3)
sum(dnorm(y, predict(model3),summary(model3)$sigma,log=TRUE))

# offset only - discrepancy

sum(predict(model3)-c(1+x))

# yet predict is correct