[R-sig-ME] Cumulative link mixed model appropriate in a 2x4 design?
Rune Haubo
rhbc at imm.dtu.dk
Thu Sep 20 12:44:40 CEST 2012
On 20 September 2012 12:22, Klemens Weigl <klemens.weigl at gmail.com> wrote:
> Dear David and Jarrod,
>
> so far I tried a lot :-).
>
> 1st solution:
> To run a linear mixed model with 'the 2 groups/treatments' as fixed effect
> and 'the 4 time points' as random effect. It showed a significant group
> effect. Then it is interesting on which time point. My colleagues just want
> 4 t-tests for independent measures at each time point. If run them it
> showed - Bonferroni-corrected - only one significant time point.
>
> 1) Is this approach okay?
>
>
> 2nd solution:
> I also wanted to apply the cumlutative link mixed model with L.a..
> --> the response was 'size' (continuous)
> --> fixed effect 'treatment' (2 groups)
> --> random effect 'weeks' (4 time points).
> The problem: I always got:
> ----------------------------------------
> FEHLER:
> response needs to be a factor
> ----------------------------------------
>
> I used the following R-code:
> ------------------------------------------
>> library(ordinal)
> ...
>> fm1 <- clmm2(size ~ treatment, random=weeks, data=data)
> and then:
>> fm1 <- clmm2(size ~ treatment, random=weeks, data=data, Hess=TRUE,
> nAGQ=10)
> ------------------------------------------
>
> In the wine data example of the clmm2-tutorial the bottles of wine first
> were rated from 0 to 100 and then 'transformed' into the rating of 1=
> "least bitter", 5= "most bitter".
>
> 2) Do I also have to transform the continuous 'size' response into
> something like a 1 to 5 rating, before I can apply the clmm2-model?
Yes, essentially you would have to do that in order to apply clmm or
clmm2. I would however not use a mixed effects model for these data at
all. If you had more than one observation per mice, that could be
relevant, but not here. That means I would also treat the time
variable as fixed. I would also start out with a linear model / 2-way
ANOVA. A cumulative link model on a coarsened version of 'size' could
possibly be relevant if you are concerned about the normality
assumption (but a transformation of 'size' might just be better) or if
you are just dying to get predictions in terms of probabilities. I
would be more concerned about correlation between time points, in
which case gls from the nlme package could add an AR1 structure to the
model, but it doesn't sound as if you have a dataset large enough to
identify a correlation structure and it might not be important.
>
> First I thought there - compared to t-tests for independent samples where
> one also can easily apply the Mann-Whitney-U-test if the data are
> non-normal... - might also be no problem just to apply clmm2 and it should
> at least show me some results.
> But it showed: response needs to be a factor.
As it says: the response needs to be a factor for a clm or a clmm to be fitted.
Hope this helps,
Rune
>
> 3) Is the problem just 2) or lies the problem somewhere else?
>
> I appreciate your answers.
>
> Kind regards,
> Klemens
>
>
>
>
>
>
> 2012/9/13 Jarrod Hadfield <j.hadfield at ed.ac.uk>
>
>> Hi,
>>
>> With normal response, you are right in thinking that you don't need a
>> cumulative link mixed model. A linear mixed model (with group as a random
>> term?) should suffice.
>>
>> Cheers,
>>
>> Jarrod
>>
>>
>>
>>
>>
>>
>> Quoting Klemens Weigl <klemens.weigl at gmail.com> on Wed, 12 Sep 2012
>> 15:58:44 +0200:
>>
>> Dear R-sig-mixed-model-group!
>>>
>>> Basically I've got a fairly simple dataset with a 2x4 design (two
>>> independent variables = i.v.) and a continuous response variable (only one
>>> dependent variable).
>>> 1st i.v.: two different treatments
>>> 2nd i.v.: 4 time points: after 2, 4, 6 and 8 weeks --> at each time
>>> point: mice with tumor cells are killed and the tumor growth was analyzed.
>>> Therefore no repeated measures. Every mouse can be just in one of the two
>>> treatment groups in just one of the 4 time points.
>>> The data are normally distributed, but with unequal and small 'n' in each
>>> group (ranging from 8 to 14 mice per group).
>>>
>>> Objective: to test wether or not one treatment is better than the other
>>> treatment over the 4 time points all together?
>>>
>>> Someone was suggesting "cumulative link mixed model with Laplace
>>> approximation" for this task.
>>>
>>> Well I am wondering if the clmm with Laplace approximation is appropriate
>>> for this task, because the response variable is "continuous" and not
>>> ordinal (as written in the clmm2_tutorial) Am I loosing much power if I
>>> apply it?
>>>
>>> I'd be interested if someone might have some arguments for or against the
>>> application of clmm with L.a. in that design-setting - or a better
>>> solution?
>>>
>>> Kind regards,
>>> Klemens
>>>
>>> [[alternative HTML version deleted]]
>>>
>>> ______________________________**_________________
>>> R-sig-mixed-models at r-project.**org <R-sig-mixed-models at r-project.org>mailing list
>>> https://stat.ethz.ch/mailman/**listinfo/r-sig-mixed-models<https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models>
>>>
>>>
>>>
>>
>> --
>> The University of Edinburgh is a charitable body, registered in
>> Scotland, with registration number SC005336.
>>
>>
>>
>
>
> --
> ______________________
>
> M.Sc. Klemens Weigl
> Applied Statistics, Biostatistics
> +43 680 215 67 58
> klemens.weigl at gmail.com
> www.psychologische-statistik.at
> ______________________
>
> [[alternative HTML version deleted]]
>
> _______________________________________________
> R-sig-mixed-models at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
--
Rune H B Christensen, PhD
DTU Informatics, Section for Statistics
Technical University of Denmark, Build. 305, Room 122,
DK-2800 Kgs. Lyngby, Denmark
Phone: (+45) 45 25 33 63
Mobile: (+45) 30 26 45 54
More information about the R-sig-mixed-models
mailing list