[R-sig-ME] Fwd: lme4 Question

[Ben Bolker]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

[cc'ing to r-sig-mixed-models.  Sorry about previous accidental e-mail.]

- -------- Forwarded Message --------
Subject: lme4 Question
Resent-Date: Tue, 02 Dec 2014 19:09:03 -0500
Date: Tue, 2 Dec 2014 17:08:54 -0700
To: bolker at mcmaster.ca

[snip]

I have ran a bunch of analyses a while back in version 1.0-6, and have
since updated to version 1.1-7. I have gone back to the initial analyses
since I updated the package, and the results have changed. Everything is
mostly the same except for the standard errors and p-values. What was
significant is not anymore in several cases, even though the code and data
are the exact same. See output below:

*Output (lme4_1.0-6) *

[image: Inline image 1]

*Output (lme4_1.1-7) Same output in versions 1.1-6 / 1.1-5*
[image: Inline image 2]

BMB> Summary from image:

> standard errors of first two estimates change a fair amount; 
> Z-statistics shrink from {2.6,-2.1,2.9,-1.2,0.9 } to 
> {1.3,-1.8,2.5,-1.2,0.88}

[for future reference, it works better on the mailing list if you cut
and paste text]


This statement is the only real change associated with SE that I found
between the versions that I have been using:

*From the news for version 1.1-4*: Standard errors of fixed effects
are now computed from the approximate Hessian by default (see the
use.hessian argument in vcov.merMod); this gives better (correct)
answers when the estimates of the random- and fixed-effect parameters
are correlated (Github #47)

I dug into the subject and it appears that this is the cause of my
inflated SE and p-values. I tried: summary(a1fit1, use.hessian=FALSE),
and as you suggested in the pdf manual, this returned my results from
before.

So, this is suggesting that my previous results are incorrect and that
I should use the estimates from the newer version (which means a full
start over)? Or is there justification for using my previous results
and using the hessian = FALSE command? I.e., how do I know if my fixed
and random effect parameters are correlated?

BMB> Sadly, I believe that the new results are correct.  I can't
see any particular justification for using the old results.  However,
I would certainly recommend that you try computing confidence intervals
by profile likelihood -- that will be more accurate than the Wald
confidence intervals given by summary() in any case (and independent
of the decision about how to compute the variance-covariance matrix
of the parameters).

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)

iQEcBAEBAgAGBQJUfm2RAAoJEOCV5YRblxUHhHcH/iCSlX1LJ2TdQkMjO2z8ztBX
1Yo3fnza4O3X+11PSWtLGb6aXCwvjC1HImbXgNmWnKli3lCyNYn4HTzvTwqss7U2
M807hc2RJy+0LADITHi5KQaGkMcma8AZFzx02L2JkAc20BURI39G/NODKX+kTsHk
AR5lunwt9JY9TwbSv3qVO/HccK7NgdTdjHV3Yqyp0NZQR8XhxOE7yoKAsJRTGvMg
yChjXClidKW3Bno/Hc6mIhCamWB05y9P2llS3VHcooIXxS5f9SM6K/Sj5Ilw6JF6
Fps+aCjMdqCbVi8/Y+DmOB30Mp+WQrirnoykxiUVJkzgVAr8PSUyeg5oIJOTs38=
=5JoD
-----END PGP SIGNATURE-----