Hi folks,
I am using logistic regression to model habitat selection of woodland
caribou. I have a series of vegetation variables (e.g., pine, spruce
etc.) and a variable meant to capture the spatial autocorrelation of
successive animal relocations. Animal relocations (1s) and random
relocations (0s) serve as the dichotomous dependent variable.
When I run the vegetation variables as a group the results are intuitive:
positive coefficients for vegetation types that caribou typically select
and negative coefficients for vegetation types that caribou typically
avoid. However, when I add the autocorrelation variable and rerun the
model the coefficients increase in magnitude and most variables become
positive. Therefore, if an animal was avoiding a particular vegetation
type with the addition of the autocorrelation term the animal now selects
that type!
This phenomenon is consistent across all of the animals I tested. The
autocorrelation variable is always positive and is strongly significant.
Vegetation variables that switch signs are statistically significant in
both the veg. and veg + autoc. models.
I have looked for the usual suspects including suppressor variables,
collinearity, outliers, etc. I can find no mention of such an
occurrence in the literature. Has anyone had similar experiences with
either linear or nonlinear regression? I am at the end of my rope. All
suggestions and advice are appreciated.
Thanks.
Chris Johnson
University of Northern BC
[log in to unmask]
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|