[R-sig-ME] warning computing profile confidence intervals with confint()

[Ben Bolker]

Sol Lago <solcita85 at ...> writes:

> Hi everyone,

> I am using a mixed effect model and I would like to 
> provide profile confidence intervals, which were
> suggested in the lme4 documentation as an alternative to
>  traditional pvalues. Below are my model and the
> command I use to generate the confidence intervals:

> m <- lmer(residRT ~ Gram*Number + (0+Gram+Number|Subject) + (1|Subject)  + 
>   (1|Item), data= data)  #model
> confint(m, method="profile", oldNames=FALSE)  #profile CIs
 
> However, when I run confint () I always get this warning: "In
> cov2cor(m) : diag(.) had 0 or NA entries; non-finite result is
> doubtful" (I get it 13 times, one for each parameter in the model).

  This is quite possibly a false positive -- it sounds vaguely
familiar but I don't have a working example (otherwise it would
be listed as an issue on github and I would be trying to fix it)
If you can come up with a small/minimal reproducible example,
or if someone else has one, could you (1) post it here or (2) send
it via e-mail or (3) post it to https://github.com/lme4/lme4/issues
(at the very least, we could use a more informative error message)
 
> My questions are:
 
> (1) How worried should I be? Would you advice not to report the CIs
> given the warning? When I compute normal CIs intead (same command,
> just method="Wald) I get values that very similar to the profile
> confidence intervals. I wonder if I should report normal CIs
> instead: the upshot is that they can be always computed, but I worry
> they are less related to the way the mixed-effects model was
> computed, so it might be conceptually confusing.

  If the Wald and likelihood profiles are similar that's a pretty
good sign.  If the Wald intervals *are* similar, and familiar to
your audience, you might be right that reporting them would be
best (although as far as I know profile CIs are always more accurate
than Wald CIs, so familiarity and computational convenience would
be the only reason to prefer Wald CIs).
 
> (2) If I were to report the profile CIs, which are the standard
> references I should give in the paper? I don't think profile CIs are
> known in my field, so I think I should point people to whichever are
> the standard references for this.

  I don't know about the "standard": most theoretical statistics
textbooks should mention it at least in passing. In ecology you could quote
Bolker 2008 _Ecological Models and Data in R_, or Mangel and Hilborn's
1998 _Ecological Detective_  ...