Dear Prof Bates and fellow mixed modellers,
I would find it useful to have a brief discussion of the way lmer handles
weights in the present version. This follows from a threat earlier this
month whose subject (Crash fitting lmer in lme4_0.999375-4) is no longer
relevant.
Briefly, I ran some weighted models in lmer that caused R to crash when
using 0.999375-4. There was no crash in 0.999375-5, but the variance
estimates on the random effects were unfeasably high. I get different
estimates in 0.999375-10, but still >10^8. The fixed effects are unchanged
(all with enormous standard errors).
Prof Bates conceded a problem concerning the way sigma is estimated in
weighted lmer. To my knowledge, has not been mentioned in subsequent
threads, so I can't tell if it's been fixed.
It turns out that unweighted models also have certain odd properties with
this dataset. I think this is because I've fit a saturated model in which
the fixed effects are highly correlated. Simpler models give sensible
answers, so I must apologise for fitting stupid models.
However, I'm still unsure whether estimates from weighted models are
trustworthy and would appreciate some advice.
Best wishes, Nick Isaac
PS Further apologies for sending a half-written draft of this message (I'm
still getting used to gmail).
[[alternative HTML version deleted]]