[R-sig-ME] Which model types accept the correlation-argument from nlme?

Tim Richter-Heitmann trichter at uni-bremen.de
Thu Apr 5 17:45:32 CEST 2018

Dear list,

I am struggling a bit with finding a suitable approach for the following 

I have 60 samples on a grid, which were sampled six times (bimonthly). 
The 60 samples were not resampled each time, instead the sampling 
location was shifted a few decimeters each time. So, i have six 
individual spatial grids.

The outcome variable are counts of microbial marker genes.

My task is to classically find significant differences in abundance for 
bacterial clades between dates.

My test of choice is glht() from the multcomp packages. This requires a 
model-object from which it calculates Tukey-comparisons of means per 
data. So, classically:

mod <- glm.nb(Abundance ~ Date, data=data)

glht(mod, mcp(Date = "Tukey"),  mcp(Date = "Tukey"), vcov=vcovHC).

Now, since i have likely spatially autocorrelated outcome data, and 
possibly random effects of the time, i set up

amod.null <- lme(fixed=Abundance ~ Date, data = data, random = ~1| Date, 
method="ML") # i think i asked a question about this in this list some 
years ago

and use the update function, applying amod.null to c("corExp", 
"corGaus", "corLin", "corRatio", "corSpher"). I check for the best AIC, 
and subject this model to glht().

So far, so good (i hope).

(Sidequestion: Is the comparison of categorical means of autocorrelated 
measurements affected by the same problems as e.g. performing regression 
analysis between two continous variables, of which at least one is 

It seems that lme() is only for normally distributed data, but my 
outcome variable is seemingly best modelled assuming its negative 
binomially distributed (this was tested with GAMs).

So, i am looking for a model type, which is a) acceptable for glht() or 
other multiple comparison tests, b) allows fixed and random effects, c) 
allows correction of spatial autocorrelated outcome variables, d) 
accepts assumptions of negative-binomial / Poisson

count data.  I was made aware that glmmPQL does a lot of the things i 
want, but it doesnt give p-values afaik, and is not usable with glht().

I am now down to chosing between:

- a glm.nb with correction for heteroscedasticity.

- a lme with random effects and corrections for spatial autocorrelations.

What would an experienced modeller do (i am not)?

Thank you!

Dr. Tim Richter-Heitmann

University of Bremen
Microbial Ecophysiology Group (AG Friedrich)
FB02 - Biologie/Chemie
Leobener Straße (NW2 A2130)
D-28359 Bremen
Tel.: 0049(0)421 218-63062
Fax: 0049(0)421 218-63069

	[[alternative HTML version deleted]]

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