model output exhibits large standard error

[gackstek]

Hello,
I am doing an analysis of understanding human-wildlife conflict due to depredation of livestock by large predators.
I gathered data by participatory mapping for livestock and conflict areas. I converted my data in detection-non-detection format.

My specific research question is whether there are any spatial covariates that can help me to understand patterns of livestock depredation, while livestock does not use all the habitat equally.

I selected two site covariates (NDVI, distance to PA) that are weakly correlated (R² = 0.17). Further I use one survey covariate (area covered by interviewee). All covariates are z-transformed.

I am using a two-species model to estimate conflict. The precondition is that the site must be used by livestock, which is where it helps to estimate uncertain presence of livestock.

I tested several model configurations and selected the best performing model according to the AIC

Code: Select all

LS_CF.7a <-occMod(model= list(psi~SP+INT+NDVI.z+dist_PA.z,p~area.z),
data=LS.CF.pao,
type="so.2sp.1",
param="psiBA")

While the beta estimators seems to be reasonable, I observe an extremely large standard error:

Code: Select all

coef(Top.LS.CF,"psi")
                         Beta_est           se
A1_psiA                  6.653216 2.407160e+00
A2_psiBA                23.624411 5.084240e+04
A3_psiBa               -99.399756 2.756793e+06
A4_psiA.NDVI.z_psiA     -1.445936 8.183640e-01
A5_psiA.dist_PA.z_psiA   3.364343 1.940491e+00


coef(Top.LS.CF,"p")
                    Beta_est       se
B1_pA[1]            1.193677 0.080211
B2_pB[1]           -1.773291 0.105345
B3_pA[1].area.z_pA  0.542757 0.053497

The fitted values for PsiA (i.e. probability of livestock) do make sense in a way:



But for PsiBA (i.e probability of conflict when livestock is present) all fitted values are 1 with a SE of 0











Am I on the right track and are there any guesses, why I do not get reasonable results?

Thank you for your help and all the best,

Kristina

[darryl]

Hi Kristina
Can you provide some basic data summaries please.

How many sites?

How many surveys?

How detections of sp A? Spread across how many sites?

How detections of sp B? Spread across how many sites?

How many sites where species B was detected, and species A wasn't detected? And similarly, where species A was detected?

Cheers
Darryl

[gackstek]

Hi Darryl,
thank you for your quick reply.

The number of sites is 726.

I conducted interviews with 85 herders. Per site the number of survey varies between 0 - 9 (i.e. for some cells up to 9 Herders could say if there is livestock or conflict)

Species A (Livestock) was detected 739 times, spread across 354 sites.

Species B (Conflict) was detected 360 times, spread across 266 sites.

Since Species A (Livestock) is a precondition of Species B (conflict), Species A was detected in all sites where Species B was detected (i.e. 266)
Consequently there are no sites where Species B is present and Species A not.


I hope this clarifies the structure of my data.

Thank you for your help

Kristina

[jhines]

Hi Kristina,

The estimate of psiBa (presence of conflict, given no livestock presence) = zero is reasonable since it is impossible to have conflict without livestock. When a parameter is zero or one, the standard error is undefined, which usually results in a very large value for the estimated standard error on the beta. So, the PsiBa estimate makes sense for these data.

The estimate of psiBA (presence of conflict, given livestock presence) of 1.0 makes sense if all sites with livestock detected also have conflict detections. If there are sites with livestock detections and no conflict for all surveys, then I expect psiBA to be < 1.0. If all sites with livestock detections have conflict (psiBA=1.0) and no sites without livestock presence have conflict (psiBa=0), then you basically have a single-species model.
I’d be happy to look over the data to verify the results if you’d like.
Jim

[gackstek]

Hi Jim,
you are right. I was expecting this for PsiBa. Conflict without Livestock is not possible by definition.

I am mainly interested in PsiBA. There are sites where I sampled Livestock, but no conflict. In fact, there are 88 site with Livestock and no Conflict.

You can download my Pao object here https://box.hu-berlin.de/f/c9941c9a67ea414c9824/?dl=1. You should be able to reproduce my model with it.

Thank you very much for support.
Best,
Kristina

[gackstek]

Hi Jim,
Are you able to access the data? Do you need more information from me?
Best,
Kristina

[jhines]

Hi Kristina,

Yes, I was able to access the data and I was composing this response when you posted your last message

If psiBA (presence of conflict, given presence of livestock) is what you're interested in, then I think the single-species model will work if you modify the data a little. Since all sites with detections have livestock, you can modify the data such that you only have detections of conflict. This means change all 1's to 0's in the data to indicate that sites/surveys where only livestock were detected have no detection of conflict. You could change all 3's (detection of both livestock and conflict) to 1's, but it is not necessary as Presence only cares if the detection data > 0 for the single-species model. So, you end up with only detection/non-detection of conflict in the data. The results for that model will give you the probability of presence of conflict for sites with livestock.

The 2-species model allows the estimation of the interaction between 2 species, but in this case there is no interaction between livestock and conflict, so the 2-species model is not necessary. You can get estimates of presence of livestock using the unmodified data with a single-species model, and estimates of occupancy and detection of conflict using the data modified as described above. I think the results should be similar to the 2-species results you have, but slightly different due to the uncertainty introduced with the more complicated model.

Jim

[gackstek]

Hi Jim,

Thank you for your prompt reply!

I understand your reasoning for using a single-species model. Since my aim is to estimate conflict across my entire study area, my initial hypothesis was that the presence of livestock would be a key driver.

Am I correct in thinking that a single-species model wouldn’t account for the fact that livestock does not use all habitats equally? Or is this factor implicitly considered in the data?

Looking forward to your thoughts.

Best regards,
Kristina

[jhines]

Yes, but since the estimate of psiBA is 1.0, it means conflict occurs at all sites with livestock. (The single-season model with the modified data also produced an estimate of 1.0 for presence of conflict at sites with livestock.) To get the estimate of conflict over all sites (psiB), you can multiply the estimate of presence of livestock (psiA) by the estimate of presence of conflict, given livestock presence (psiBA) which would simply be psiA.

The estimate of psiBA = 1.0 made me question whether something was wrong, but I think the significant number of NA's in the data makes the intuition less clear (for me). The single-species results (psi=1.0) agree with the 2-species results (psiBA=1.0) so I think that is the best estimate of conflict presence, given the data.