st: GLLAMM for an ecological dataset with issues

[Benita Kaytor <Benita.Kaytor@unbc.ca>]

Hello Statalist users,



I have a dataset comprised of data from two studies, both of which estimate the percent cover of vegetation species in permanent plots over time.  My objective is to determine if plant groups responded differently under the influence of two disturbance factors which could potentially alter growing conditions.  Below are the main issues complicating this dataset:

-Both studies have a nested design structure, but one has two extra levels compared to the other.  I addressed this by collapsing the data of one study so they both have the following top-down hierarchy: site, plot, vegetation group (percent cover data are for vegetation groups).

-Not all sites were measured every year, so the number of repeated observations ranges from 3 to 5.

-Percent cover data have a non-normal distribution with most observations being zero, tailing off to the right with a few observations being >100 percent (due to pooling of individual species data into groups).

-I have both fixed and random effects that I want to include as predictor variables.



In order to relax the requirements of simpler models and use the data in their rawest form, I am attempting to use a GLLAMM.  I have finally been successful getting a simple version of the GLLAMM code to work, but there are a few questions I have not yet been able to resolve reading the GLLAMM manual and other literature:

-It took 45 minutes to complete a test run with only 1 iteration, and my computer nearly crashed before finishing.  Approximately how long will it take to run with an appropriate number of iterations (I have 1428 observations), and am I in danger of damaging my computer to work the CPU at 99%?

-I'm not sure how to reflect the mixed effects in my coding. So, a) below is my first code that worked with one iteration, but I think I did not specify the hierarchical levels right:

a)

generate int const = 1
eq cons: const
eq rpt: mpbyr

gllamm pcover mpbyr map caribou, i(veggroup2 plotid2 site) nrf(2,1,1) eq(cons rpt cons cons) iterate(1) adapt trace



-In b), I show my improved code which I haven't run yet because of how long it will take; I think I have the hierarchy right with the random intercept equations specified at the right levels, but I'm not sure if I've incorporated the mixed effects properly.  From the way I understand it, veggroup and caribou should be fixed, and MAP, MATemp, and mpbyr should be random, but I have not distinguished them in the coding.  I'm not sure if veggroup should it be treated as a predictor or not as it is the unit upon which the observations are made.  I'm also not sure if the "adapt" option is appropriate.

b)

eq cons: const
eq rpt: mpbyr
eq bec1: map
eq bec2: MATemp
eq veg: veggroup
eq car: caribou

gllamm pcover mpbyr map MATemp caribou, i(plotid2 site) nrf(2,5) eq(cons veg cons rpt bec1 bec2 car) adapt trace



Hopefully I haven't intimated any potential helpers with all my questions. I have a limited understanding of stats in general, so my vocab is somewhat limited. I  look forward to any thoughts/advice/recommendations, and will probably need help interpreting the results, too...



Thanks for your consideration on this!

Benita Kaytor



MSc Candidate

University of Northern British Columbia

Prince George, British Columbia

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/