Skip to content

generalized linear mixed models: large differences when using glmmPQL or lmer with laplace approximation

3 messages · Greg Snow, Ben Bolker

Original
Ben Bolker
#
Martijn Vandegehuchte wrote:
Well, Laplace should be better anyway.  (If the difference were in
the other direction -- non-significant with Laplace and significant with
glmmPQL -- I would still tell you to use Laplace.)

  To speak to Greg Snow's comment ("don't worry about p-values, just
look at predictions") -- this is really tough.  I still don't know
what to do about the compromise between how statistics should be done
and how journal editors seem to insist it should be done ...

  cheers
   Ben
Greg Snow
# 
You make reference to my comment below, but I think you overstate my position a bit (the words in quotes are not a direct quote of what I said).

The original poster mentioned that 2 different methods gave 2 different models, one possibility is that one method gave a wrong model (biased in a non-good way), another possibility is that the predictor variables are correlated enough that there are multiple good models.  I merely pointed out that comparing the predicted values to the original values would be one way to possibly distinguish between the 2 cases.

Focusing too much on the predicted values can lead to overfitting, so we should not depend only on that.  P-values are useful in some cases, so I would not say "don't worry about the p-values" as a general statement.

The issue of editors wanting p-values even when they answer the wrong question is part of the result of statisticians doing to good a job of training other researchers.  Now it is our responsibility to continue to train them as to when to use certain tools.

--
Gregory (Greg) L. Snow Ph.D.
Statistical Data Center
Intermountain Healthcare
greg.snow at imail.org
801.408.8111
1 day later
Ben Bolker
# 
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Greg Snow wrote:
(the words in quotes are not a direct quote of what I said).

  Fair enough.  Sorry about that.
one possibility is that one method gave a wrong model (biased in a
non-good way),

another possibility is that the predictor variables are correlated
enough that

there are multiple good models.  I merely pointed out that comparing the

predicted values to the original values would be one way to possibly
distinguish between the 2 cases.

  Looking at the parameters, they seemed to be pretty similar to me,
although of course the details of the data (range of predictor
variables) matters too.
so we should not depend only on that.  P-values are useful in some cases,

so I would not say "don't worry about the p-values" as a general statement.

  Point taken.
is part of the result of statisticians doing to good a job of training
other researchers.

Now it is our responsibility to continue to train them as to when to use
certain tools.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFI7Sh9c5UpGjwzenMRAsXeAKCMiqRLxikc8u8zxnXPK702D0XoNgCgjm2q
y/NYurfZ/jK/AOK7aGkdDkc=
=f++H
-----END PGP SIGNATURE-----