R help - Feb 2008 - Generalized nonlinear mixed model function?

I am wondering if there is an R function that could estimate a generalized 
nonlinear mixed model.
>From my reading it seems that nlme from the nlme package can fit nonlinear 
mixed models, while lmer from the lme4 package can fit generalized linear 
mixed models. 

One alternative I?ve found is gnlmix from the repeated package, although 
this only allows for a single random effect.

Is there anything else out there that I have missed?

Thanks,
Phil

Phillip van Mantgem
USGS Western Ecological Research Center
Sequoia and Kings Canyon Field Station
47050 Generals Highway #4
Three Rivers, CA  93271-9651     USA
-----------------------------------------------------------

The motivating problem is estimating average trends in forest mortality 
rates with unequally spaced census intervals. The census interval has an 
exponential effect on survival (i.e., p = annual survival, year.interval = 
census interval, and p^year.interval; annual morality = 1 - p), leading me 
to use a nonlinear model. Our data are composed of counts of live and dead
trees, so I?ll need to use a binomial or poisson model.

Our data look like the following...

# plot identifier (random effect)
# eventually I?ll need to add another random term for species identity
plot <- rep(c("A", "B", "C"), each = 3)

# census identifier
census <- rep(1:3, 3)

# year of census
year <- c(1960, 1989, 2004, 1960, 1989, 2004, 1955, 1989, 2004)

# interval between census years
year.interval <- c(NA, 29, 15, NA, 29, 15, NA, 34, 15)

# count of live trees
n.live <- c(1509, 1249, 1106, 1986, 1616, 1383, 3529, 2831, 2511) 

# count of dead trees
n.dead <- c(NA, 260, 143, NA, 370, 233, NA, 698, 320) 

forest.mort <- data.frame(plot, census, year, year.interval, n.live, 
n.dead)
	[[alternative HTML version deleted]]

(Sorry for the re-posting. My first attempt failed.)

I am wondering if there is an R function that could estimate a generalized
nonlinear mixed model.

From my reading it seems that nlme from the nlme package can fit nonlinear
mixed models, while lmer from the lme4 package can fit generalized linear
mixed models.

One alternative I?ve found is gnlmix from the repeated package, although
this only allows for a single random effect.

Is there anything else out there that I have missed?

Thanks,
Phil

Phillip van Mantgem
USGS Western Ecological Research Center
Sequoia and Kings Canyon Field Station
47050 Generals Highway #4
Three Rivers, CA  93271-9651     USA
-----------------------------------------------------------

The motivating problem is estimating average trends in forest mortality
rates with unequally spaced census intervals. The census interval has an
exponential effect on survival (i.e., p = annual survival, year.interval census
interval, and p^year.interval; annual morality = 1 - p), leading me
to use a nonlinear model. Our data are composed of counts of live and dead
trees, so I?ll need to use a binomial or poisson model.

Our data look like the following...

# plot identifier (random effect)
# eventually I?ll need to add another random term for species identity
plot <- rep(c("A", "B", "C"), each = 3)

# census identifier
census <- rep(1:3, 3)

# year of census
year <- c(1960, 1989, 2004, 1960, 1989, 2004, 1955, 1989, 2004)

# interval between census years
year.interval <- c(NA, 29, 15, NA, 29, 15, NA, 34, 15)

# count of live trees
n.live <- c(1509, 1249, 1106, 1986, 1616, 1383, 3529, 2831, 2511)

# count of dead trees
n.dead <- c(NA, 260, 143, NA, 370, 233, NA, 698, 320)

forest.mort <- data.frame(plot, census, year, year.interval, n.live,
n.dead)

Phillip J van Mantgem <pvanmantgem <at> usgs.gov> writes:
> 
> I am wondering if there is an R function that could estimate a generalized 
> nonlinear mixed model.
> 
  Short answer: no (not really/not that I'm aware of).

  Long answer: AD Model Builder?  WinBUGS?  shortcuts like
nlme applied to the proportion dying,
with appropriate scaling of the variance (weights argument)?
   
   I'm afraid that a nonlinear generalized mixed model with
two random effects, although fairly straightforward to define,
is pretty darned difficult to solve reliably ...

  cheers
    Ben Bolker

Hi Philip,

your data are event times because you're monitoring the same trees in 
each plot over time, the event being death of a tree.

Therefore methods from survival analysis are more appropriate.
Start out having a look at the package survival, possibly considering 
a Cox model with adjustment for clusters or with strata (the function 
coxph()) or a parametric model using the function survreg().


Christian