On 30 Jul 2018, at 13:38, Phillip Alday <phillip.alday at mpi.nl> wrote:
In a previous message, one of the warnings was '2=3=4'. Assuming that
there's nothing weird about any internal sorting, that would mean these
models:
~ d + p + s + t + (1 | random) + d:t + p:s + p:t + s:t
~ d + p + s + t + (1 | random) + d:t + p:s + p:t + s:t + p:s:t
~ d + p + s + t + (1 | random) + d:p + p:s + p:t + s:t + p:s:t
They are similar but not identical in formula form: the first one is
missing the three-way interaction, while the the last two differ in the
two-way interaction involving d (d:t vs d:p). Are the models rank
deficient? i.e. are there combinations of factors that don't exist such
that these model terms get dropped? Try looking at these models and
seeing if there if a term is missing:
summary(get.models(models,subset=delta<5)$`9168`)
Or maybe see if the effective terms in each model are equivalent:
mod3 <- get.models(models,subset=delta<5)$`9168`
mod4 <- get.models(models,subset=delta<5)$`9120`
mean(sort(names(fixef(mod3))) == sort(names(fixef(mod4))))
If that last line return 1, then the models have identical fixed
effects, which combined with their identical random effects, you indeed
make them identical.
And this is a rather weird error -- I'm also grasping at straws here.
Phillip