[R-sig-ME] Log likelihood of a glmer() binomial model .
Rolf Turner
r@turner @end|ng |rom @uck|@nd@@c@nz
Mon Apr 22 09:49:09 CEST 2019
On 22/04/19 4:27 AM, Juho Kristian Ruohonen wrote:
<SNIP>
> As alluded to by Ben, predict() can certainly provide fitted
> probabilities for the validation set with the random effects taken into
> account. This is achieved by the re.form = NULL argument. However -- and
> I'll be happy to be corrected on this -- the problem is that dbinom()
> will calculate a (log)likelihood of the observed responses assuming a
> regular binomial PMF, which does not apply in the case of a
> mixed-effects model. Thus the result will not equal the loglikelihood
> that is maximized in the fitting process, i.e. will not equal
> logLik(validationFit) unless it's a standard logistic GLM.
OK. I have finally understood the stupidity that I was perpetrating.
It can expressed succinctly as thinking that E(f(P)) = f(E(P)). Which
is true only if f() is an affine function. I used to go off my trolley
complaining when undergraduate students committed such atrocities. :-(
(In the foregoing f() may be taken to be the binomial probability
function with some fixed "x" value, and P to be the success probability.
So even if predict.merMod() delivered E(P) as I previously thought it
did, my so called cross-validated log likelihood would still be incorrect.)
Sorry for all the noise.
cheers,
Rolf
--
Honorary Research Fellow
Department of Statistics
University of Auckland
Phone: +64-9-373-7599 ext. 88276
More information about the R-sig-mixed-models
mailing list