[R-sig-ME] GLMMpql and GEE question

[Hannah L. Linder]

Hi,

I am an M.S. student at the University of Washington School of Aquatic and
Fishery Sciences.My thesis involves the comparison of many models that you
could use to analyze monitoring data. A big part of this comparison is
looking at models with and without autocorrelation (my data is a univariate
time series). I was hoping to compare a GLS, GLM, and GLM with
autocorrelation for a non-normal data set using their RMSE values. I was
originally intending to use a GLM-GEE, because I have seen them used in the
literature within my field, but I noticed the glmmPQL function allows for
different corARMA correlation structure and the gee only allow for an ar-1
correlation structure. So now, I believe that I would rather use the
glmmPQL for the purpose of comparing a model that allows for
autocorrelation but is normally distributed (GLS), one that is non-normal
with no autocorrelation (GLM), and one that is non-normal with
autocorrelation. I am wondering if there is a big difference between the
glmmPQL model and a glm-gee? I know the gee is a marginal model, and a glmm
models random effects, but in the case of a univariate time series (which
is essentially a single group) I am not sure how this would make a big
difference.


If anyone has any time to provide suggestions on better understanding the
difference between these two models, or if it is appropriate to use a glmm
rather than a gee in this case, I would greatly appreciate it.

Thank-you very much,
Hannah Linder

[[alternative HTML version deleted]]

[Thierry Onkelinx]

Dear Hanna,

GLMM doesn't make sense if you have only one level. It requires
theoretically at least two, in practice even more to get reliable
fits.

I would have a look at the INLA package (www.r-inla.org). It allows
both several non-gaussian distributions and correlated random effects.
In case of a univariate timeseries, then time itself is the random
effect.

library(INLA)
inla(Count ~ f(Time, model = "ar1"), family = "poisson") # AR1
inla(Count ~ f(Time, model = "rw1"), family = "poisson") # random walk order 1
inla(Count ~ f(Time, model = "rw2"), family = "poisson") # random walk order 2

see http://www.r-inla.org/models/latent-models

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

Post by Hannah L. Linder
Hi,
I am an M.S. student at the University of Washington School of Aquatic and
Fishery Sciences.My thesis involves the comparison of many models that you
could use to analyze monitoring data. A big part of this comparison is
looking at models with and without autocorrelation (my data is a univariate
time series). I was hoping to compare a GLS, GLM, and GLM with
autocorrelation for a non-normal data set using their RMSE values. I was
originally intending to use a GLM-GEE, because I have seen them used in the
literature within my field, but I noticed the glmmPQL function allows for
different corARMA correlation structure and the gee only allow for an ar-1
correlation structure. So now, I believe that I would rather use the
glmmPQL for the purpose of comparing a model that allows for
autocorrelation but is normally distributed (GLS), one that is non-normal
with no autocorrelation (GLM), and one that is non-normal with
autocorrelation. I am wondering if there is a big difference between the
glmmPQL model and a glm-gee? I know the gee is a marginal model, and a glmm
models random effects, but in the case of a univariate time series (which
is essentially a single group) I am not sure how this would make a big
difference.
If anyone has any time to provide suggestions on better understanding the
difference between these two models, or if it is appropriate to use a glmm
rather than a gee in this case, I would greatly appreciate it.
Thank-you very much,
Hannah Linder
[[alternative HTML version deleted]]
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models