Hi Alan et al.,

Generalized Linear Mixed Models (not to be confused with General Linear Mixed 
Models) are designed for exactly this sort of data.  The Generalized form lets 
you define the distribution to be whatever you want it to be.  With a count 
variable like this, you should start out with either a negbin or poisson 
distribution and a log link, and use AIC (or AICc, depending on your sample 
size) to choose the best-fitting model.  However, in recent analyses I ran 
using count data as dependent variables, I actually found a log distribution 
with either a log or identity (=normal) link to have the best fit.  FYI, if 
you're using a log link and/or distribution and have any '0' values, you will 
need to add 1 to each value prior to running the models to avoid the log(0) 
problem.

SPSS has the capability to run Generalized Linear Mixed Models, as do both SAS 
(Proc GLIMMIX) and R.

Best,
Nicole Michel

*********************************************************
Nicole Michel
PhD Candidate
4060 Stern
Dept. of Ecology and Evolutionary Biology
400 Boggs
Tulane University
New Orleans, LA 70118
Fax: 504-862-8706
*********************************************************

________________________________________
From: Ecological Society of America: grants, jobs, news 
[[email protected]] on behalf of Alan Griffith (agriffit) 
[[email protected]]
Sent: Thursday, July 14, 2011 9:39 AM
To: [email protected]
Subject: [ECOLOG-L] Non-parametric statistics

Hello all,

I have been searching for some advice on appropriate non-parametric statistics 
for the analysis of a dependent variable that fails normality and homogeneity 
assumptions under both sqrt and ln transformations.

First I will describe the dataset.  The data are from a field sample.  I have 4 
years of data from the same set of ecological populations.  The number of 
populations varies year to year.  The number of individuals sampled in a 
population may have varied within and among years.

Here is a description of the model I would like to implement.  Let’s say the 
Dependent Variable is # seeds eaten / plant.  So, I want to implement 
individual plant nested within population (i.e.  a mixed model with population 
identifier as random variable or SUBJECT(PopID)).  YEAR is a categorical 
independent variable, Population Size is one continuous independent variable.  
Total # Seeds produced / plant is another continuous independent variable.  I 
would also like to test interactions.

As I said before, I was not successful in transforming my dependent variable 
using my standard choices (ln and sqrt).  I had found references to using rank 
transformed data in an ANOVA / ANCOVA model, but this was rejected by a 
reviewer.  I am familiar with simple nonparametric tests like Kruskal-Wallis, 
but I do not see how to preserve the complex model with such tests.

My first hope is to find a method, generally accepted by ecologists, that is 
easily implemented in SPSS.  If this is not possible, I can explore more 
complicated analyses with the help of my campus math / stats consultant.

Thanks for you advice.

|   /      \   |  Alan B. Griffith, PhD
\  \  ̗  ̖  /  /   Associate Professor
  \  \( )/  /    Department of Biological Sciences
   \ (   ) /      University of Mary Washington
    /(   )\       (540) 654-1422
  / / ( ) \ \     [email protected]
/  |  ¦¦  |  \
|              |

Reply via email to