multi-state covariate

[lucrezia]

Hello,

I´m running a single-season multi-state model (psi1=all individuals, psi2=juveniles). I have 144 traps which have been sampled (2 sampling occasions each time) in three trapping sessions (June, July and September).
I have used the trapping session as a categorical variable and calculated a spatial auto-covariate as in Klute et al.2002 (Autologistic regression modeling of american Woodcock habitat use with spatially dependent data. - In: J. M. Scott, et al. (eds), Predicting species occurrences: issues of accuracy and scale. Island Press, pp. 335-343.) per each trapping site.

Depending on whether i include the trapping sessions alone (September is the reference session), or also the spatial auto-covariate (atc), to model psi2, I get very different beta estimates (and s.e.).

Model 1
psi1 (.), psi2(atc+session), p1=p2(atc),dlta(.)

betas
psi1 1,38304 (0,244379)
psi2 -1,322193 (0,840892)
psi2 atc 6,264628 (1,598814)
psi2 June -1,157692 (0,700668)
psi2 July -1,266022 (0,720406)
p11 -1,379998 (0,27194)
p11atc 3,732918 (0,542212)


Model 2
psi1 (.), psi2 (session), p1=p2(atc), dlta(.)

betas
psi2 3,73243 (5,776224)
psi2 June -3,211845 (5,584732)
psi2 July -2,86538 (5,541285)
all other being equal (except of course for psi2atc which is not modelled here)

Based on dAIC and AICwgt, model 2 has basically no support at all (dAIC=31,61, AICwgt=0). Calculating the weighted model average (and weighted unconditional s.e., very large) for each trapping session over all models I get final psi2 values which are much lower than what I would have expected from the data set (i.e. it was a rodent peak year, lots of juveniles..). I was wondering whether this difference in beta estimates seems possible or else it suggests that something is wrong in the analyses..

Also, in model 2, psi2 for September is so high that I assumed the large s.e. associated with it to be due to boundary effect. However, this does not seem to explain the large s.e. for June and July.. ?

Thanks in advance for any hint and hope that the post is not repetitive for the forum,
Lucrezia

[darryl]

Hi Lucrezia,
So presumably traps where only surveyed in 1 of the 3 trapping sessions? This may result in some spatial and temporal confounding.

If zero is not within the range of your atc covariate then that could cause your beta estimates to bounce around somewhat between models, especially the intercept term, so you might want to standardize atc by subtracting by some value that's close to the observed mean. However, I would probably expect the session effect to still be similar unless session and atc are correlated to some degree.

The SE's from model 2 do look suspiciously large, did you get any warnings in the output? You might also want to try different starting values, you may have struck a local maximum instead of the global maximum. Trying fitting a psi2(.) model, and take those as the initial values for the psi2(session) model. You'll have to insert a couple of 0's in the right spot for the session effects.

Cheers
Darryl