[R-sig-ME] GLMM question

[Ben Bolker]

Jake Westfall <jake987722 at ...> writes:

> 
> Hi Leena,
 
> Yes, p-values for tests of fixed effects can be found in various
>  ways. See the FAQ here:
 
> http://glmm.wikidot.com/faq
 
> In particular see the sections "What is the best way to test
> hypotheses on effects in GLMMs?" and "Why doesn't lme4 display
> denominator degrees of freedom/p values? What other options do I
> have?"
 
> Note that most of these methods of obtaining p-values will *not*
> also come with an estimated degrees of freedom. My guess is that
> just having the p-value will satisfy the editor and the absence of
> DFs will not be a big deal. But if you decide that you do also want
> the degrees of freedom, you can use the Kenward-Roger procedure
> implemented in the pbkrtest package.

  The situation is a little different for LMMs and GLMMs. You can only
use KRmodcomp() for linear mixed models.  You can use parametric 
bootstrap via PBmodcomp() in pbkrtest, or via bootMer in lme4 proper.

  GLMMs generally assume the data set is large (asymptotic tests),
so they give you p-values (but no dfs).  This is spelled out in
more detail in the FAQ referenced by Jake.  If the editor really
wants dfs for GLMMs, they're going to be disappointed.  
If they want p-values
that take finite-size corrections into account, you will need
parametric bootstrapping or some similar approach.
 
> Jake
> 
> From: leena.hamberg at ...

 [snip]
 
> I would like to ask a question relating to generalized linear mixed
> models. I have used package lme4, function glmer to estimate my
> models (logit link for occurrences, log for counts and identity for
> height models). I presented the results of my models in our
> manuscript (coefficients with SE - significant ones
> highlighted). However, the editor asked me to add p-values, df:s,
> and test statistics to the result section every time I am presenting
> significant or insignificant results. I did as was asked and
> explained that degrees of freedom were not available for these
> models and that when normality was assumed (i.e., in the case of t
> statistics) p-values were not available. However, the editor
> answered as follows:

  You should probably use lmer instead of glmer for your height models.
For completeness you really need to specify the family as well as the
link (I assume binomial for occurrences, Poisson for counts, and
Gaussian for height).  

Out of curiosity: how do you know whether the parameters are
statistically significant or not (at
some unspecified alpha level, probably 0.05) if you don't know the
p-values?
 
> "In my previous e-mail I've requested you to add details of the
> statistical results in your MS (e.g., results of the GLM you've
> done, F-values, Chi2-values, df, P-values, etc.)... ...You did not
> take this comment fully into account and I disagree with your answer
> to this request. On the contrary to what you answered me, R (since
> you used R) provides all the detailed results you are requested to
> provide...
 
> ...Also, even if p-values, df and statistics are tightly
> interrelated, this does to prevent you to give the corresponding
> information in your published work, at least to help potential
> readers to verify what you wanted to say. P-values are always
> available in R - or can easily be found - for Gaussian or not
> normally distributed traits. So, you have to provide all the needed
> information is you MS. For example, every single t-test has to come
> with its df and P-value. If really you are not able to find this in
> R, then you have to use another program..."
 
>  
> So how to proceed? Can df:s and p-values be found in any way 
> using the R? If yes, how this can be done?
> Unfortunately I couldn't solve this problem by myself.
> 
> Here is an example of GLMMs estimated:
> 
> tyvivmaxp10P=glmer(Tkvmaxpit~käsittely+tyvilpm+m3haYHT+saastotKAIK+
> TKvHirvi+(1|ruutu),family=gaussian(link ="identity"), data=pihlajatE10)
> 
> summary(tyvivmaxp10P)
> 
> Linear mixed model fit by REML
> Formula: Tkvmaxpit ~ käsittely + tyvilpm + m3haYHT + 
>   saastotKAIK + TKvHirvi + (1 | ruutu)
>    Data: pihlajatE10
>    AIC  BIC logLik deviance REMLdev
> 994.5 1015 -489.3    999.5   978.5
> Random effects:
> Groups   Name        Variance Std.Dev.
> ruutu    (Intercept)  207.25  14.396
>  Residual             1255.73  35.436
> Number of obs: 100, groups: ruutu, 8
> 
> Fixed effects:
>              Estimate Std. Error t value
> (Intercept)  80.71047   17.52630   4.605
> käsittely2  -45.79590   13.98636  -3.274
> tyvilpm      22.06164    6.23251   3.540
> m3haYHT       0.11672    0.31592   0.369
> saastotKAIK   0.09566    0.11513   0.831
> TKvHirvi1    12.77524    8.55455   1.493

 [snip]