Categorical data tested for several factors: which test should I use?

#1
Hi guys,

I'm currently doing a research project as part of my internship on mangrove damage along a river as a consequence of a major flood event. Shoreline was analysed for 1 meter sections. For each section my data comprises of the categorical categories mortality (yes/no) and certain other categories such as damage (yes/no).
Now i want to test several factors (location from rivermouth, innerbend/outerbend, species type, etc) in order to see what their relative contributions are. As a section might be affected by several factors in different ways, while checking for one factor i want to 'eliminate the effects of another factor'. What test should i be using? I read that loglinear regression can be applied to categorical data, but can this be used for several factors at a time?

Any help is appreciated!

Robert
 

bugman

Super Moderator
#3
As I see it, you have a couple of options.

Log-linear models are poisson regression models and so deal specifically with count data, so if you are counting the number of damaged or dead mangroves in a given section, then this might be the way to go. If I recall though, log-linear models do not distinguih response and predictor varaibles.

logistic regression models might also be suitable for your data if you are concerned primarily with the binary outcome of a trees condition (dead / alive).

Another consideration is that you have two responses - dead / alive and damaged / undamaged you could merge these so that you have four conditions for the varaible "tree condition" and the counts of each within each cell for each in which case, log-linear models would be the preffered approach.

However, if you were to assign a tree one of: dead, alive, damaged (basically combining your two repsonses) you would be able to fit a multinomial regression.


Anyway, each aprroach has its strengths and weaknesses, so it would pay to read up on them and decide where to go based on your objectives and how you collected your data.
 

noetsi

Fortran must die
#5
If you have a few categorical levels of the response variable (the dependent variable) multinomial logistic regression may work. If you have a few levels of the response variable that are ordered you could try ordered logistic regression.