Do glmer and glmmadmb calculate log likelihood on thesame scale ?
Curious. I so boldly stuck my foot in my mouth about using corStruct
= "full" to equate to glmer(), yet the log likelihood remains
unchanged using it or not in another case I ran where I could get the
model to converge. Interestingly, a correlation is printed when
corStruct = "full", but somehow the log likelihood and model df do not
change? I am adding Ben Bolker to this (although he probably would
have seen it soon enough anyway). A) because it may be relevant that
your original model converges in an older version of glmmadmb, but not
the one I am using. B) because I do not understand why the correlation
among random effects does not seem to affect the log likelihood and
model df in glmmadmb but does in glmer.
glmmadmb(formula = y ~ Visit + (Visit | subject), data = epil2,
family = "poisson", corStruct = "full")
gives:
Number of observations: total=236, subject=59
Random effect variance(s):
Group=subject
Variance StdDev Corr
(Intercept) 0.8881 0.9424 (Intr)
Visit 0.5399 0.7348 0.5399
Log-likelihood: -686.899
glmer(y ~ Visit + (Visit|subject), data = epil2, family = "poisson")
gives:
AIC BIC logLik deviance
617.9 635.2 -303.9 607.9
Random effects:
Groups Name Variance Std.Dev. Corr
subject (Intercept) 0.88772 0.94219
Visit 0.53592 0.73206 -0.034
Number of obs: 236, groups: subject, 59
On Thu, Jun 21, 2012 at 7:13 PM, David Duffy <David.Duffy at qimr.edu.au> wrote:
On Thu, 21 Jun 2012, Joshua Wiley wrote:
Just a note that those are not the same models.
Huh.
OR you could change glmmadmb() to fm2 <- glmmadmb(y~Base*trt+Age+Visit+(Visit|subject), ?data=epil2, family="poisson", corStruct = "full")
however, on my machine, there are warnings about the matrix not being positive definite, and the model never converges.
I haven't updated glmmADMB recently (0.6.4 2011-08-30), so it gives the warning, but continues merrily to give: ? ? ? ? ? ?(Intercept) ? Visit (Intercept) ? ? 0.24931 0.24931 Visit ? ? ? ? ? 0.54572 0.54572 and the same likelihood as before Log-likelihood: -655.41 glmer was ?Groups ?Name ? ? ? ?Variance Std.Dev. Corr ?subject (Intercept) 0.24912 ?0.49912 ? ? ? ? Visit ? ? ? 0.53962 ?0.73459 ?0.011 ?'log Lik.' -262.6586 (df=9) Your suggestion Visit + (1 | Visit) + (0 + Visit | subject) gave me ?Groups ?Name ? ? ? ?Variance ? Std.Dev. ?subject Visit ? ? ? 0.77301 ? ?0.87921 ?Visit ? (Intercept) 5.1845e-16 2.2769e-08 'log Lik.' -421.1128 (df=8) I have to admit, the direction of the change in loglikelihood here confuses me a little ;) My lme4 is also a little antiquated: 0.999375-42 (2011-10-02). -- | David Duffy (MBBS PhD) ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ,-_|\ | email: davidD at qimr.edu.au ?ph: INT+61+7+3362-0217 fax: -0101 ?/ ? ? * | Epidemiology Unit, Queensland Institute of Medical Research ? \_,-._/ | 300 Herston Rd, Brisbane, Queensland 4029, Australia ?GPG 4D0B994A v
Joshua Wiley Ph.D. Student, Health Psychology Programmer Analyst II, Statistical Consulting Group University of California, Los Angeles https://joshuawiley.com/