[R-sig-ME] Calculating SE from GLMM results (using glmer{lme4})

Raldo Kruger raldo.kruger at gmail.com
Sat Sep 26 10:23:29 CEST 2009


Hi Ben,

Thanks for your help. See my responses below (~~~).

Regards,
Raldo



On Thu, Sep 24, 2009 at 3:21 PM, Ben Bolker <bolker at ufl.edu> wrote:
> Raldo Kruger wrote:
>> Dear R users,
>>
>> Please excuse the basic questions, but I’m new to GLMMs and R!
>>
>> I’m analyzing an experiment where Seedling numbers in plots where seed
>> has been sown on restoration sites is the response variable. I’m most
>> interested in determining whether the Nutrients (N) and water
>> absorbing polymer Gel (Ge) additions to the soil substrate contribute
>> positively to the survival of the seedlings, over a 3 year time period
>> (for simplicity I'm just using 3 time periods, each in the same season
>> for the 3
>> successive years).
>> Fixed factors: Nutrients (0 and 1), Gel (0 and 1)
>> Random factors: Site (4 non replicate sites), Year (3 time periods)
>> Response variable: Seedling numbers (counts) / 0.25m2 plot
>> The results are as follows:
>>                Estimate       Std. Error      z value  Pr(>|z|)
>> (Intercept)   4.52982 0.24486 18.5           <2.00E-16        ***
>> N            -0.07922 0.08415 -0.94          0.346489
>> Ge             0.20766        0.08428 2.46            0.013744        *
>> Year          -1.62937        0.04672 -34.88  <2.00E-16       ***
>> N:Ge  -0.44213        0.11898 -3.72           0.000202        ***
>> N:Year        0.11705 0.06322 1.85    0.064125        .
>> Ge:Year       -0.04861        0.0645  -0.75   0.451132
>> N:Ge:Year     0.11458 0.08917 1.28    0.198821
>>
>
>  Some comments:
>
> *  It looks like you fitted year as a fixed effect rather than a random
> effect (probably sensible, since you only have 3 levels / years), and
> incorporated all fixed effect interactions (i.e. N*Ge*Year) ?  However,
> it also looks like you fitted year as a continuous covariate, which
> means that R is trying to fit a linear function of time -- is that
> really what you want?  There's a very large negative year effect -- if
> your year values are coded 1-3, then it suggests you have very few
> seedlings left in year 3?

~~~Yes, you're right - I have used Year as a continuous covariate,
partly because when I first ~~~tried it as a categorical predictor, I
didn't know how to make sense of the results. But ~~~you're right,
Time is not a linear function in this case, since there is a big drop
in ~~~seedling numbers from year 1 to 2, and a much smaller drop from
year 2 to 3. I've ~~~corrected it, and the results are much 'better'
(and i can now make sense of them!). ~~~Either way, there are very few
seedlings left in year 3 (which is what the observed data ~~~shows
too).


>
>  It's also
> worth considering whether you are really getting reliable answers based
> on only 4 sites -- I would also try this with Site as a fixed effect
> and see whether the answers differ considerably.  (I know that,
> philosophically, Site and Year are both random effects, and you may
> run into trouble with reviewers who are used to classical ANOVA

~~~ I've been using glmer, and it doesn't seem to accept a model with
no random factors. Is ~~~that true, and if not, how can i  write the
model without a random factor so that it is ~~~accepted by glmer?

>> 1)    So as I understand (from previous correspondence with R-users) the
>> number of seedlings in the control plots in year 0 is
>>         exp(4.53) = 92.7. Is the standard error calculated with
>> 0.24486 (i.e. 92.7*0.24), or with 92.7*exp(0.24).
>
>  The latter.
>  So for example the approximate confidence intervals would be
> exp(0.453 +/- 2*0.245)
>
>> 2)    And for the N:Ge treatment, the effect is exp(-0.08+0.21-0.44)
>> =0.73 (I.e. a 27% reduction compared to the control), right? So
>>         is the SE for the N:Ge effect calculated as the sum of the
>> SE’s too, i.e. 0.08+0.08+0.12, or is it just 0.12?
>
>  The SE for combined effects is calculated as sqrt(se1^2 + se2^2 + se3^2)
>
>> 3)    Lastly, is it possible to fit two response variables in one GLMM?
>> E.g. seedling numbers and height.
>
>  This would be hard -- you're talking about a multivariate response
> with different measurement scales/error distributions for different
> variables ...
>>
>> Many thanks,
>> Raldo Krüger
>> Msc student
>> University of Cape Town
>>
>>
>
>
> --
> Ben Bolker
> Associate professor, Biology Dep't, Univ. of Florida
> bolker at ufl.edu / www.zoology.ufl.edu/bolker
> GPG key: www.zoology.ufl.edu/bolker/benbolker-publickey.asc
>



-- 
Raldo




More information about the R-sig-mixed-models mailing list