Why does the log-likelihood ratio test need a larger maxfun?

Tue, Apr 19, 2016 2:08 PM

Hi Zhaohong,

The following warning indicates that your results cannot be trusted:

 > convergence code 1 from bobyqa: bobyqa -- maximum number of function
 > evaluations exceeded

This means that the results are most likely not the MLEs. You should 
rerun the model with higher values of maxfun. For example:

control = g/lmerControl(optCtrl = list(maxfun = 1e6))

Why you do not see the warning when fitting seems weird.

Hope that helps,
Henrik


Am 19.04.2016 um 17:48 schrieb Zhaohong:

Dear All,

I am running a log-likelihood ratio test on two mixed models (differing
only in one variable) to test the significance of that variable. The two
mixed models (I set the maxfun=500000 for both) are able to converge with
no warnings, but the anova(model1, model2) gives warnings as follows:
        Df    AIC    BIC  logLik deviance  Chisq Chi Df Pr(>Chisq)
..1    92 1450.7 2029.6 -633.33   1266.7
object 93 1452.5 2037.7 -633.22   1266.5 0.2098      1     0.6469
Warning messages:
1: In commonArgs(par, fn, control, environment()) :
   maxfun < 10 * length(par)^2 is not recommended.
2: In optwrap(optimizer, devfun, x at theta, lower = x at lower, calc.derivs =
TRUE) :
   convergence code 1 from bobyqa: bobyqa -- maximum number of function
evaluations exceeded
3: In commonArgs(par, fn, control, environment()) :
   maxfun < 10 * length(par)^2 is not recommended.
4: In optwrap(optimizer, devfun, x at theta, lower = x at lower, calc.derivs =
TRUE) :
   convergence code 1 from bobyqa: bobyqa -- maximum number of function
evaluations exceeded

The Pr(>Chisq) given from this test is 0.6469, but the Pr(>|t|) from the
lmerTest is 0.00160 ** , which seems more likely the case, because the
t-value from the lmer model summary is 3.445, with a total Number of obs:
3996.

So I thought maybe I need to increase the maxfun for the anova() test. I
rerun the test with the command anova(model1,
model2, ,control=lmerControl(optCtrl=list(maxfun=5000000))), and got the
same results:
        Df    AIC    BIC  logLik deviance  Chisq Chi Df Pr(>Chisq)
..1    92 1450.7 2029.6 -633.33   1266.7
object 93 1452.5 2037.7 -633.22   1266.5 0.2098      1     0.6469
Warning messages:
1: In commonArgs(par, fn, control, environment()) :
   maxfun < 10 * length(par)^2 is not recommended.
2: In optwrap(optimizer, devfun, x at theta, lower = x at lower, calc.derivs =
TRUE) :
   convergence code 1 from bobyqa: bobyqa -- maximum number of function
evaluations exceeded
3: In commonArgs(par, fn, control, environment()) :
   maxfun < 10 * length(par)^2 is not recommended.
4: In optwrap(optimizer, devfun, x at theta, lower = x at lower, calc.derivs =
TRUE) :
   convergence code 1 from bobyqa: bobyqa -- maximum number of function
evaluations exceeded

I am wondering what I should do in this situation then?

Thanks a lot!

	[[alternative HTML version deleted]]

Why does the log-likelihood ratio test need a larger maxfun?

Thread (2 messages)