Dear Thierry,
many thanks for your answer. I checked the output of my models again, and
the random term when time was both fixed and random, indeed was always
almost zero.
I think, i should clarify my sampling design briefly.
The plot was subdivided in 30 subplots.
A subplot was subdivided into 12 sampling locations on a regular grid.
For each time point, a unique pair of 2 neighboring sampling locations
were sampled.
Meaning, the x,y-coordinates are different for each sampling date,
together they form a perfect grid with 360
points.
I can see using locationIDs, but technically they are not from the same
exact location for each date;
which is why i liked the 'correlation' argument in the lme models, in
which i could use x,y coordinates.
Is there a way to incorporate this into the glmer.nb model you have
proposed?
Thank you very much!
Tim
On 24.11.2015 10:30, Thierry Onkelinx wrote:
Dear Tim,
Have a look at the INLA package (www.r-inla.org <http://www.r-inla.org>).
It allows you to model spatially correlated random effects, temporally
correlated random effects, use a negative binomial distribution and specify
linear combination (needed for the posthoc tests). Downside: it's not for
the faint of heart.
Having time as factor both in the fixed and random part is useless. See
http://rpubs.com/INBOstats/both_fixed_random
Assuming that you revisited the same locations, then a reasonable simple
model would be:
fit <- lme4::glmer.nb(abundance ~ time + (1|locationID))
pro:
- negative binomial
- repeated visits to the locations acknowledged
- post hoc test of time via glht
contra:
- compound symmetry correlation for location instead of spatial
correlation
- no temporal correlation
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
2015-11-23 22:39 GMT+01:00 <trichter at uni-bremen.de <mailto:
trichter at uni-bremen.de>>:
Dear list,
i am very new to mixed models. My data encompasses species
composition matrices from six different time points with spatial
correlation structure. For each species, i want to know if there
is a pure effect by time, f.e. if abundance changes can be purely
explained by time alone.
I used to glht() with time being a simple factor (so not
accounting for the repetitive nature of my data), but this seems
inapprobiate/wrong. So, i am actually trying to do:
fit <- lme(fixed=abundance ~ time, random=~1|time, data,
correlation=corxxx(form=~x.pos + y.pos))
with time being a factor with 6 levels (a side question would be,
if it would be better to use "time" as.time?)
Because my data is actually negative binomially distributed, i was
advised to use glmmPQL, but this gives me only intercepts, no
significancies or ways to compare models by log likelihood or AIC.
The basic question is, if that syntax is correct? Because i have
seen many examples looking at interactions, but never anything
where the only fixed predictor is also random. I do get an output,
which i can interpret and which resembles what i can actually see
from boxplots.
The overarching question is, if there are post-hoc tests for
repeated measurements of spatially autocorrelated, non-normally
distributed data.
Thank you, Tim