2
$\begingroup$

I'm running a conditional logistic regression to assess habitat selection in coyotes. The conditional logistic regression essentially compares the environmental covariates of used locations (1's) to semirandomly generated "available" locations (0's) which are matched in strata to determine if certain covariates are more closely associated with use. These strata account for the fact that each available location is associated with a used location.

I'm particularly interested in whether the selection for areas of high-human density changes based on other variables like the amount of natural habitat in a 1 km radius. I'd also like to include sex in the model, but because of the conditional nature of the model sex never varies between used and available locations in the same strata and therefore the effect can't be estimated.

Does it makes sense in this case to include sex in an interaction but not as a main effect?

Improve this question
$\endgroup$
3
  • $\begingroup$ Can you clarify the data structure? It sounds like one record per coyote for each location. There is a record of 1 if the coyote picked that location and a record of 0 for each other location not picked. Is that true? What do you specify as the cluster indicator? $\endgroup$
    – AdamO
    Commented Oct 14, 2022 at 15:34
  • $\begingroup$ @AdamO Each used location is recorded by a coyote's GPS collar and is unique to the coyote. Each of these locations is matched with 10 available locations. Available locations are generated using randoms draws from the distribution of step lengths (distance between successive used locations) and turning angles (the direction of a used location relative to the previous location) of the sample. A used location's (i) and it's matched available locations originate from the previous location (i-1). Each strata contains 11 locations (one 1 and ten 0) and their covariates. $\endgroup$
    – albondiga
    Commented Oct 14, 2022 at 17:05
  • $\begingroup$ It sounds actually like the very type of data analysis for a geometric series using a complementary log log regression. $\endgroup$
    – AdamO
    Commented Oct 15, 2022 at 0:25

1 Answer 1

Reset to default
3
$\begingroup$

This is a limitation of conditional logistic regression and other fixed effects models for analyzing paired/stratified data: you cannot estimate the effect of between-stratum variables (i.e., variables that do not vary within strata) on the outcome. An alternative is to use generalized linear mixed models, i.e., models with random effects for stratum membership. You can include stratum-level predictors in the model and examine their relationship with the outcome while accounting for stratum membership using a random effect. In R this can be accomplished using lme4::glmer().

Improve this answer
$\endgroup$
1
  • $\begingroup$ Great, thank you! $\endgroup$
    – albondiga
    Commented Oct 13, 2022 at 20:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.