Spatial-Mark Resight - SEs of 0 & Inability to model sex

[Paolo Strampelli]

Hi secr group! I hope this finds all well.

I am writing as I have ran into two issues. I apologise in advance if it’s something obvious – it’s my first time using mark-resight models! I have looked through this forum as well as others, but to no avail.

I am implementing sightings-only SMR models in secr 4.1, to estimate density of a lion population from camera trap data. I applied standard SECR models to this data with success, by identifying animals through individual markings. However, given that about 25% of the detections could not be IDd, I decided to try and apply mark-resight models.

As there was no way to know for sure whether unresolved detections were of marked but unidentifiable or of unmarked individuals, all unresolved detections were initially classified as marked but unidentifiable. This was to avoid misidentifying a marked individual as an unmarked individual and introducing a positive bias (as per Rich et al. 2019’s paper, also on lions). This left me with a Tm file (and no Tu or Tn file).

First Issue - SEs of 0

My first issue is that, while the estimated densities are realistic, my SEs and CIs are of 0 – as so:

Fitted (real) parameters evaluated at base levels of covariates

       link     estimate  SE.estimate          lcl          ucl

D       log 2.534081e-04 0.000000e+00 2.534081e-04 2.534081e-04

g0    logit 5.953364e-03 1.258175e-03 3.932561e-03 9.003200e-03

sigma   log 4.793594e+03 5.193539e+02 3.878897e+03 5.923991e+03

 

Has anyone had a similar issue before?

Interestingly, when I read the Tm file as a Tu file instead (so assuming all unidentified were actually unmarked), I obtained very similar density estimates as when it was read in as the Tm file, and realistic SEs and CIs (so ‘fixing’ the problem, if you will). However, for the reasons explained above I do not think this is what I should be doing for this specific dataset.

Similarly, if I read the Tm file + a Tu file with a few randomly-placed 0s – even with only one  single zero, actually – it behaves fine. If, I do the same but the Tu is all 0s, on the other hand, I go back to getting SEs of 0.

Second issue - Inability to model effect of sex

In addition, none of the SMR work when trying to model sex as a covariate (giving NAs for both D and SE). I saw that the manual mentions that finite mixtures are only “partially implemented” – does anyone know whether this is likely to be the reason, or if it is indeed possible to model the effect of sex on sigma and g0, and my issue stems from something else?  

Any help would be massively appreciated – I’ve been trying to figure this out for a while and there’s nothing more I can think of trying!

Thanks so much in advance for anyone who could provide any input. I’ve attached the input files, as I’ve seen others do so on this forum (except the 'dummy' Tu file mentioned above, as it was simply the same as Tm but with all 0s interspersed with a few random 1s). Please do let me know if there’s anything which is unclear or which I did not explain well.

Thanks and all the best,

Paolo

[Ben Augustine]

I'm sure Murray can help with fitting issues, but I had some general comments.

What is the goal of including the unidentified photos--to deal with unmarked individuals or to try to increase precision by using more data? The main idea of mark resight is that one subset of individuals is marked and another is unmarked. The unmarked individuals must be considered because they have a capture probability of 0 if you require their individual identity and thus introduce extreme individual heterogeneity in capture probability and negative bias in N and D. If you cannot tell if any observations are in fact unmarked, a spatial mark resight will not solve this problem of individual heterogeneity in detection. If all unknown ID observations are classified as "marked but not identifiable", the density of the unmarked portion of the population will be estimated very close to 0. This may be related to the estimation problems you are seeing, but probably doesn't explain them all.Further, you should not pool observations of potentially unmarked individuals with marked individuals in the "marked but not identifiable" class. The correct class is "unknown marked status", though I don't believe secr handles that class. But without any unmarked observations, reclassifying these observations will not solve the problem.

Regarding precision: if you must classify all unknown ID photos as "marked but not identifiable", all data are of the same class of individual--marked. Therefore, a more appropriate model would be one with only 1 class of individual--a "random thinning" model. This has been developed, but not published to my knowledge. The reason it has not been published is that the unknown ID observations only increase precision in sampling scenarios with almost no identified photos.

Given these points, I believe you should stick with regular SCR and accept you might underestimate density if there are indeed some unmarked individuals in the population.

Ben

--
You received this message because you are subscribed to the Google Groups "secr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to secrgroup+...@googlegroups.com.
To view this discussion on the web, visit https://groups.google.com/d/msgid/secrgroup/c079f0ab-e320-4090-ad05-8acb62dba132%40googlegroups.com.

[Paolo Strampelli]

Dear Ben,

Thank you very much for your prompt and well-explained reply. Apologies in the delay in mine, had to carry out some unexpected travels.

The goal of including the unidentified photos was to try and increase precision by using more data, yes. I figured that it would lead to similar (or slightly higher) estimates than the SECR, and with better precision.

Everything also you say makes sense, thank you. And yes, I was already leaning towards regular SECR, given the issues and that I do not seem able to test the effect of sex on sigma and g0, which I know considerably includes the fit of the model (although I am yet to figure out whether this has not been implemented yet or if it's because of a failure on my part).

Either way, thank you very much for that explanation, it was really interesting and helped clear things up!

All the best,
Paolo

To unsubscribe from this group and stop receiving emails from it, send an email to secr...@googlegroups.com.

[Jamie Bolam - 王英龍]

Dear all,

I am having a similar issue to the one described in this thread, and would deeply appreciate everyone's opinions to sense-check what my best course of action is with my dataset. 

For some context, our study species (bears!) has individually unique chest markings, but if they aren't visible in the camera trap frame, it can be pretty difficult to identify individuals (except from external features like body size, snare injuries, etc). 

Our dataset is still quite young and small, with maybe ~12 individuals recorded on our ID station camera traps (yet to process and add the most recent batch of data), most of which were unfortunately only captured on camera once. ~2 individuals were recaptured on separate occasions. Next year we will make a more clustered design to maximise the chances of resighting individuals multiple times, but at least have useful information on their ranging behaviour based on the spread of individuals.

I have run some SECR models which worked, but given our so far small dataset with limited recaptures, we naturally have pretty huge error margins. I can only run null models so far, given the limited data. We also have quite a few records of individuals that were too far from the camera (35-45% of our records ; 7+ instances), or angled so we couldn't see their chest, so I want to try SMR to see if we can include these in our data. Yesterday I prepared my SMR models based on the secrbook guide.

Since I am dealing with a naturally marked population, and we don't know how many individuals we have,  I should use a sightings-only SMR model (markocc = all 0) and include 

details = list(knownmarks = FALSE)

in my model formula.

I was debating how to classify these unidentified individuals as marked (but not identified) or unmarked. Technically, each individual IS marked naturally; if we get a good camera shot, we can identify them, but otherwise it may be difficult in some situations. But as Ben explains above, SMR requires some individuals which are unmarked for the model to work properly. 

Similar to Paolo above, if I consider these unidentified individuals as marked (Tm) and have my Tu records blank, I get a seemingly reasonable estimate, but my SE is 0 for my density estimate and so I have no proper confidence intervals, which is not ideal. 
# link     estimate  SE.estimate          lcl          ucl
# D       log 4.726727e-04 0.000000e+00 4.726727e-04 4.726727e-04
# g0    logit 1.177079e-02 1.733693e-02 6.412413e-04 1.810687e-01
# sigma   log 9.824394e+03 2.446715e+04 6.255360e+02 1.542976e+05
# pID   logit 6.611378e-01 1.158410e-01 4.145798e-01 8.431430e-01

This model also allows me to include details = list(knownmarks = FALSE). 

If I include details = list(knownmarks = FALSE) in my models where I include the unidentified individuals in Tu instead of Tm, the model fails and I get NAs for all model parameters. But, it works if I omit knownmarks = FALSE. Is this because of my small dataset? 

The mask of this data used a buffer of 12200, which was the suggest.fit result from my SECR models.

Another option I am considering is changing how we classify what counts as Tu (unmarked) vs Tm (marked but unidentified individuals). But to be honest, I am pretty stumped at figuring out a way to distinguish which scenarios a photo should be classified as unmarked vs marked but unidentified. Any suggestions are welcome.

- e.g. if we are confident that despite not seeing the chest markings fully, it is a new and unique individual (e.g. a cute baby bear we just caught on camera a few weeks ago), we can class it as a unique individual (e.g. marked but not identified)? But then we might as well class it as a marked and identified individual and include it in our capture history, right?

- If we really have no way of telling which individual it is, class it as unmarked - but then what if it is an individual we HAVE marked, but we just can't tell?

Either way, I know my dataset is pretty limited, the main goal is seeing how we can get the best estimate and make best use of our data given our limited resources. Next year we hope to have a stronger dataset.

Any advice would be deeply appreciated!

I've attached my input data here.

Cheers!

Jamie