www.phidot.org

I am having an odd problem. Using secr ver. 2.9.5, I fit the following model and got the following results:
BC0 <- secr.fit (CHbc, mask = bcmask, model = list (D=~1, g0 ~ 1, sigma ~ 1),
hcov="Sex", verify = FALSE)

Fitted (real) parameters evaluated at base levels of covariates

session = 1, h2 = F
link estimate SE.estimate lcl ucl
D log 1.246207e-03 1.503725e-04 9.845815e-04 1.577353e-03
g0 logit 2.339183e-01 2.715903e-02 1.849200e-01 2.912602e-01
sigma log 1.353240e+03 8.680046e+01 1.193528e+03 1.534325e+03
pmix logit 3.671874e-01 4.260695e-02 2.882925e-01 4.539032e-01

session = 1, h2 = M
link estimate SE.estimate lcl ucl
D log 1.246207e-03 1.503725e-04 9.845815e-04 1.577353e-03
g0 logit 2.339183e-01 2.715903e-02 1.849200e-01 2.912602e-01
sigma log 1.353240e+03 8.680046e+01 1.193528e+03 1.534325e+03
pmix logit 6.328126e-01 4.260695e-02 5.460968e-01 7.117075e-01



Fitting the same model after upgrading to ver 2.10.2, I get different results:

Fitted (real) parameters evaluated at base levels of covariates

session = 1, h2 = F
link estimate SE.estimate lcl ucl
D log 5.620850e-04 6.775755e-05 4.441833e-04 7.112819e-04
g0 logit 1.398790e-01 1.937748e-02 1.060286e-01 1.823323e-01
sigma log 2.672849e+03 1.664143e+02 2.366078e+03 3.019395e+03
pmix logit 3.671893e-01 4.260742e-02 2.882935e-01 4.539060e-01

session = 1, h2 = M
link estimate SE.estimate lcl ucl
D log 5.620850e-04 6.775755e-05 4.441833e-04 7.112819e-04
g0 logit 1.398790e-01 1.937748e-02 1.060286e-01 1.823323e-01
sigma log 2.672849e+03 1.664143e+02 2.366078e+03 3.019395e+03
pmix logit 6.328107e-01 4.260742e-02 5.460940e-01 7.117065e-01


I have not tried to install the previous version to see if I could recreate the old results but I was wondering if anyone had experienced similar problems. The problem does not occur with all my models. I have only identified this issue with 1 model. It is possible that there could be differences in the CH or mask when I ran the models with the old and new versions, but I have double checked and they all seem the same. The only difference I can see is that the 2 versions used different starting values in the MLE, but I would think that would be expected. Thanks for any advice.
Joe

Hi Joe
Are you sure it was 2.9.5? I could understand this if you were using 2.9.4 or earlier as the change in default 'minprob' in 2.9.5 sometimes results in different estimates (or perhaps the copy of 2.9.5 you have was downloaded from the Density website before it was finalised for CRAN).

I will check this if you send me the data offline: inconsistent results scare me, too.
Murray

Email sent. Thanks!

My earlier diagnosis seems to be correct, and using secr 2.10.3 with the old default for minprob reproduces the earlier results:

Code: Select all

secr.fit(..., details = list (minprob = 1e-50))

The default was changed to 1e-200 in 2.9.5 (June 2015) for a good reason - to produce more stable estimates - and you should rely on the more recent results. Most of the time it makes no difference.
Murray

Thanks Murray. I updated secr and re-ran the model but am still getting some weird results. pmix is highly skewed toward males (unusual) and sigma is about twice as high for male bears compared with females (unheard of). I am using hcov=sex in a mixture model on sigma and g0 and I am thinking that it may be a multimodal maximization problem. However, changing the starting values don't seem to help much. I re-ran some models with the hazard rate detection function (1) and things look ok. I know you advise against using it because of the long tails, but my buffer is large (20 km). The data are relatively sparse across this large study area, as you know, which may be a contributing factor. Do you think the hazard rate estimates are okay? Thanks again.
Joe

Hi Joe
You seem to have covered the obvious issues. I wouldn't expect pmix skewed towards males in bears... Perhaps you could post the code you used. I'll see if I can turn up any new angles in the next couple of days.
Murray

Will do. Thanks Murray.

For the record, we found a single extreme female movement in the dataset. The high leverage of this one (legitimate) point probably explained the aberrant results (preference for a long-tailed detection function & odd sex ratio).
Murray