chisq = 0 and p = 1 in glmer model comparison result

Hi Becky,

The model with the two additional parameters has a lower log likelihood
(-480.3, df=15) than the smaller model (-475.8, df=13). This shouldn?t be
possible, because adding an unconstrained parameter makes the model more
flexible, allowing a closer fit to the data. This anomalous difference in
likelihoods has caused the chi-squared statistic in the LRT to be negative,
giving a p-value of 1, because all of the chi-squared distribution is
positive, i.e. P(chi-squared > test statistic) = 1 when the test statistic
is <= 0. Either one or both of the models haven?t converged, or (less
likely) there is missing data in the added variable so that the data
differs between the two models.

In addition, it?s very unusual to fit random slopes without also including
a fixed effect, as you?ve done for targetWordFactor. The null model allows
the WW-W and NW-W differences in the log odds of being correct to differ
randomly between subjects and between items, but forces the mean
differences (across subjects and items) to be zero. You?d need a good
reason to fit such a model.

Best wishes,
Paul

On 3 Jun 2019, at 14:25, Guillaume Adeux <guillaumesimon.a2 at gmail.com>

wrote:

Hello Becky,

Even though I cannot directly answer your question... a Chisq of 0 with
such a difference in AIC seems indeed suspicious.

To test the effect of your predictor in a GLMM context through LRT tests,
you could (should?) consider using the test_terms function from the

{monet}

package or its little brother function mixed() from the {afex} package.
These packages are specifically designed for such tests.

I hope this helps.

Sincerely,

Guillaume ADEUX



Le lun. 3 juin 2019 ? 15:14, Becky Gilbert <beckyannegilbert at gmail.com>

a

?crit :

Dear list

I have two glmer models, one with a fixed factor (targetWordFactor) and

one

without, and I am comparing them using the anova function to get the LRT
results for the fixed effect of targetWordFactor. The anova results are
showing a chi-square value of 0 and p value of 1. Is this result

possible,

or is it perhaps a sign that I've done something wrong?

Here are the anova results:

anova(accModelNullWord,accModelWord)
#                                  Df    AIC    BIC logLik deviance

Chisq

Chi Df Pr(>Chisq)
# accModelNullWord 13 977.59 1067.0 -475.8   951.59
# accModelWord       15 990.61 1093.8 -480.3   960.61     0      2
1

The targetWordFactor fixed factor has 3 levels (2 contrasts), so the
degrees of freedom in the anova result look correct to me. Here are the
model specifications:

contrasts(pauseDetValidNoFillersExcluded$targetWordFactor)
#        WW NW
# W       0  0
# WW   1  0
# NW    0  1

accModelNullWord <- glmer(correct ~ 1 +
                           (1 + targetWordFactor|subject) +
                           (1 + targetWordFactor|item),
                           data = pauseDetValidNoFillersExcluded,
                           family = binomial(link = "logit"),
                           control = glmerControl(optimizer="bobyqa",
                                                optCtrl =
list(maxfun=2e5)))

accModelWord <- glmer(correct ~ 1 + targetWordFactor +
                       (1 + targetWordFactor|subject) +
                       (1 + targetWordFactor|item),
                       data = pauseDetValidNoFillersExcluded,
                       family = binomial(link = "logit"),
                       control = glmerControl(optimizer="bobyqa",
                                            optCtrl = list(maxfun=2e5)))

Apologies if this question has been asked before - I did search the list
but couldn't find anything.

Many thanks,
Becky

       [[alternative HTML version deleted]]

      [[alternative HTML version deleted]]