Discrepancy between MARK and PRESENCE

[geoffwah]

Dear All

A problem that has been driving me mad all morning. I am running a single-season occupancy analysis, with both site and sampling covariates. I've run the models in MARK (v. 5.1, build 2600), and then used PRESENCE (v. 2.0, build 060818.1130) to run Mackenzie and Bailey GOF tests on my subglobal models (ie. I don't have a single global model, and hence am running GOF's on all subglobal models as per Burnham and Anderson, p306).

Weird thing is, one of my subglobal models has a different -2LogLike value in PRESENCE to that given in MARK. In PRESENCE it is 211.8943, and in MARK it is 218.45282. This model has two site covs with an additive structure (as do all the other subglobal models), and hence I compared -2LogLike's of models with only one of the site covs, but the same in all other aspects (to try and nail down which cov might have been responsible). Ie. rather than Psi(cov1+cov2), I tried Psi(cov1) and Psi(cov2) in each. PRESENCE and MARK give precisely the same -2LogLike's for these models. Thus, the discrepancy only occurs for this one additive model - all the rest are fine!

I would greatly appreciate any thoughts on the source of this problem.

Much thanks

Geoff

[darryl]

Geoff,
As PRESENCE gives a lower -2log-like I'd believe it over MARK Just kidding.

Does either PRESENCE or MARK give you any warnings? Have you used the same covariate values in both and turned off MARKs standardization option? Have you tried using the estimates from PRESENCE as initial values in MARK and vice-versa? Do the estimates provided by either look 'funky'?

Darryl

[geoffwah]

Hi Darryl

Much thanks for getting back to me so quickly.

No warnings on either output, and checked the design matrices several times to make sure they are correct. Also, used the same design matrices for the single cov models (ie, just retrieved the full model and deleted the relevant column), both of which work no probs. Have also started from scratch, redoing the data files for both, and making sure to create them from precisely the same spreadsheet. To no avial. And just tried your suggestion of running the models with the starting values from each package, but both produce precisely the same results as they did without specifying starting values. Now sitting amongst a pile of my own hair!

Parameter values make sense in both, but there are some substantial discrepancies (see below). The primary one is the acov1000 site covariate, which is an autocovariate term. Am giving these models a whirl using the psuedolikelihood approach to autologistic modelling (without detection uncertainty in the calculation of the autocovariate). Might this be the source of the problem? But it doesn't manifest in any other models including the autocovariate term....

Would much appreciate any further thoughts!

Cheers

Geoff

MARK

LOGIT Link Function Parameters of {Psi(Area+Acov1000),p(Effort+Date+Night)}
95% Confidence Interval
Parameter Beta Standard Error Lower Upper
------------------------- -------------- -------------- -------------- --------------
p(int): 2.3518825 1.1148366 0.1668027
p(effort): 0.0117052 0.0081571 -0.0042827
p(date): -0.0220435 0.0072817 -0.0363155
p(Night): 0.6520550 0.3922528 -0.1167606
Psi(int): -3.0708387 0.6197452 -4.2855393
Psi(area): 0.7294124E-04 0.4685233E-04
Psi(acov1000): 0.9944856 0.2502594

PRESENCE

B1 :detection p1 2.778188 (0.486099)
B2 :detection p1Effort 0.005324 (0.006094)
B3 :detection p1Date -0.024049 (0.002723)
B4 :detection p1Night 0.604680 (0.376088)A1 ccupancy psi -6.283744 (0.899702)
A2 ccupancy psiarea 0.000912 (0.000368)
A3 ccupancy psiacov1000 1.791563 (0.429354)

[geoffwah]

Ohh yeah, and MARK's standardisation option is swtiched off...

[cooch]

Ohh yeah, and MARK's standardisation option is swtiched off...

Be careful - very - with twitching standardization on and off. The way MARK handles standardization has changed in the past year, and *may* be different to how PRESENCE handles things.

I'd suggest a thorough read of the individual covariates chapter in the 'MARK book', where the standardization of covariates issue is treated in some detail.

[geoffwah]

Thanks Evan

Haven't been standardising throughout, and my understanding is that this should give me identical results to PRESENCE. This appears to be the case, as results are identical for all except the one problematic model.

Cheers

Geoff