Dear Vickly,
Please keep the mailing list in cc.
The idea is that you need a sufficient number of observations per
parameter. 10 to 20 is often used as a rule of thumb. If you have a lower
number, the model is too complex given the data will probably overfit.
Think about a simple linear model (intercept + 1 parameter for slope).
Although you can technically fit this model when you have 2 or 3
observations, the resulting model is not very useful.
Best regards,
ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek / Research Institute for Nature and
Forest
team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance
Kliniekstraat 25
1070 Anderlecht
Belgium
To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to say
what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey
2016-11-23 10:37 GMT+01:00 Vickly Mobilim <vickmoe7 at gmail.com>:
Hi Thierry,
Thank you for the kind reply! That is very helpful.
May I know more about the calculation? I have never seen it. How do you
use it to know if it is sufficient to build a model?
On Nov 23, 2016 5:27 PM, "Thierry Onkelinx" <thierry.onkelinx at inbo.be>
wrote:
Dear Vickly,
I assume you have measurements on the individual animals and you can
identify the animal during the different exposures. I think you want a
model like this: flash_rate ~ treatment * exposure + temperature + humidity
+ size_ratio + (1|animal_id) This requires -1 + 4 * 3 + 1 + 1 + 1 + 1 = 15
parameters. You have 78 * 3 = 234 observations. That is 234 / 15 = 15.6
observations per parameter, which reasonable to fit the model.
Best regards,
ir. Thierry Onkelinx
Instituut voor natuur- en bosonderzoek / Research Institute for Nature
and Forest
team Biometrie & Kwaliteitszorg / team Biometrics & Quality Assurance
Kliniekstraat 25
1070 Anderlecht
Belgium
To call in the statistician after the experiment is done may be no more
than asking him to perform a post-mortem examination: he may be able to say
what the experiment died of. ~ Sir Ronald Aylmer Fisher
The plural of anecdote is not data. ~ Roger Brinner
The combination of some data and an aching desire for an answer does not
ensure that a reasonable answer can be extracted from a given body of data.
~ John Tukey
2016-11-22 18:01 GMT+01:00 Vickly Mobilim <vickmoe7 at gmail.com>:
Greetings,
I've read several writing of yours about GLMM and I thought it would be
the
best tool to answer my research questions. However, I wasn't sure if I
really need it and my data permit me to use it. That said, I have 78
individuals of firefly divided into four groups (A= 20 indv., B = 20
indv.,
C = 20 indv. and D = 18 indv.). This is due to several limitations that I
can't take more samples of firefly. I will explain the details of the
experiment below.
I'm hoping that you can advise me on this issue, whether you have seen
such
cases of low sample size using GLMM or whether GLMM is not suitable for
my
study.
I expose the fireflies with several intensity of white light according to
their group (Group A = 0.05lux, B = 0.1lux, C = 0.3lux and D = 0.5lux)
then
measure their flash rates and duration before, during and after exposure
to
light (repeated measure design). Temperature, humidity and individual
eye-to-body size ratio were also measured. My main aim was to measure the
impact of several light pollution intensity to their flash rates and
duration and taking temperature, humidity and eye-to-body size into
account.
I realized that calculating changes in their flash rates and duration are
achievable by subtracting post-experiment result with pre-experiment
result
then use unpaired t-test to compare the results. However, my data was not
normal and I used Mann-Whitney U test instead. But this does not take
temperature, humidity and eye-to-body size into account. As I was looking
into the possibility of taking them into account, I found several
modelling
technique that is suitable including GLMM but I am not sure if I can
employ
them because according to a statistician I am in consult with, the sample
size is too small to be developed into a model that it would invite more
problem in analysis.
--
Regards,
Vickly Mobilim
[[alternative HTML version deleted]]