another problem with offsets/downdated VtV

[bbolker]

I think we've seen similar issues before, but this one is cleaner/more reasonable.

Test data set is here

A Poisson model with an offset. If we leave the scale of the elapsed time in hours, we fail with a "Downdated VtV error" problem. If we scale the offset to days, the fit succeeds. Is there some sort of solution based on clamping the link function ... ?

h3 <- readRDS("misc/issues/offset_test.rds")
try(glmer(y~ns(x,2)+offset(log(elapsed))+(1|f),h3,family=poisson)) ## fails
glmer(y~ns(x,2)+offset(log(elapsed*24))+(1|f),h3,family=poisson)