4
$\begingroup$

I'm modeling habitat suitability for a large, mobile animal using occurrence data (presence only - I have no true absence data in this case) collected from camera traps (stations with automatic motion/heat sensing cameras). At each camera trap location I have 0 to several occurrences recorded.

Typically, when such data is modeled, they are modeled as count data, and detection probability is included in order to mitigate for covariates that affect detection, which otherwise may be identified as covariates that affect suitability. For example, an area where detection is less likely due to, say, dense vegetation, may be mistaken as less suitable due to that vegetation structure when in fact that may not be the case (or even the opposite may be true).

My question is, if the count variable were to instead be used as a detection/no-detection variable (so, a binary) would the probability of detection have less of an effect on the model (*when modeling relative suitability, not probability of occurrence)? i.e. if you didn't account detection probability, would using a binary response variable give you a less biased estimate of habitat suitability than if you were to use count data? And if so, why? It doesn't seem like probability of detection would change, or would be more homogeneous across the study area with this change, but it seems like the effects would be smaller in magnitude by not reinforcing differences in detection probability with multiple occurrences.

Also, I understand that it's not desirable to toss out all the other occurrences assuming you can calculate detection probability. I'm asking this in an attempt to better understand the possible differences in outcome between these two modeling approaches when using such data, as I've seen both methods in the literature, but have not seen this question specifically addressed. If you could reference any publications that may shed light on this, I'd be very appreciative.

$\endgroup$
  • $\begingroup$ Do you in any way have the option of estimating the detection probability over a season, e.g. from being able to identify the individuals in the pictures? Do you have data for several years or only a single year? $\endgroup$ – fileunderwater Apr 6 '16 at 12:44
  • $\begingroup$ yes, that actually has already been done. the data i'm talking about was used to model occupancy with a capture/recapture model. capture probability was 0.05. $\endgroup$ – CSB Apr 6 '16 at 15:12
2
$\begingroup$

Disclaimed: Not really an answer, but too long for a comment


If I understand you correctly, this must depend entirely on the accumulated probablity of detection over the entire detection period (the time the cameras were out). If the accumulated probability of detection across all habitat types is high (approaching 1, given that the animal is present), then the detection probability shouldn't be a problem for the binary occurrence variable. An indication of this would be if you have sites (cameras) with either zero or a very high number of occurrences, which should indicate that the binary occurrence is rather robust. On the other hand, if sites have 0-4 occurrences in total it will be extremely difficult to separate non-occurrence from random factors or detectability. I assume that the detection probability across habitats is unknown though (as well as the accumulated probability of detection), but you might be able to make an educated guess. Also, by using a binary response, you will completely exclude the possibility to estimate e.g. an effect of habitat quality (for instance a difference in abundance between habitats), since counts of 1 or 13 will both be represented by 1 in the binary detection variable. In the Q, you talk about "modeling relative suitability" which indicates that you are indeed interested in something similar to this (relative habitat quality). In that case, I think that a binary response will be too coarse to provide useful information.

One possible toy scenario could be that you have two habitat types A and B (10 camera locations/sites of each) with a binary detection variable as:

A: 1 1 1 0 1 0 1 1 0 1 (species found in 7 sites)
B: 0 1 0 0 1 0 1 1 0 0 (species found in 4 sites)

In that case, the species can obviously exist in both habitat types (as you have categorized them), so the higher amount if zeros in B can represent (among other things) 1) lower occupancy of this habitat (in a metapop perspective), 2) that the quality of B is heterogeneous which is not captured by your habitat categorization (so parts of B is just as suitable as A), 3) lower population abundance in B compared to A (but the same detectability), or 4) lower detectability in B compared to A (but the same abundance). I don't really see how you can separate between these cases with only the binary variable, and I don't believe that using a binary detection variable will solve any specific problems compared to using the count data.

This is all very hand-wavy, but can hopefully give you some useful ideas/perspectives.

$\endgroup$
  • $\begingroup$ thanks for the thorough reply. i was thinking along the same lines as your first point - that cumulative detection probability over the detection period might be high enough to not cause any major issues using binary data. detection period was 20 days, using 300+ traps, and generated 100+ occurrence locations (~180 detections). counts range from 0-5. the only specific problem using binary data would solve is the problem of time, and user expertise. the question was raised as to whether a more user friendly method (e.g. maxent) could be used with this data to make meaningful inferences. $\endgroup$ – CSB Apr 6 '16 at 15:08
0
$\begingroup$

To be able to separate abundance from probability of detection, you need to have marked animal. If this is the case, you can then build a history of capture for each animal and model the probability of detection.

Have a look at this article : http://www.sciencedirect.com/science/article/pii/S0006320712005071. For some recommendations on the use of spatially explicit model when using camera traps to estimate population parameters (abundance...).

To answer your question, theoretically if there is independence between the detection at one site, using binary data will underestimate the probability of detection but if your detection are dependent let say animal moving in a group so each times you detect an animal you have a strong chance of detecting more than one then using binary data would make sense.

Hope it helps !

Cheers

$\endgroup$
  • $\begingroup$ Nico - I understand the relationship between detection probability and abundance, mark-recapture, and why it needs to be accounted for. My question however, is if you didn't account detection probability, would using a binary response variable give you a less biased estimate of habitat suitability than if you were to use count data? And if so, why? $\endgroup$ – CSB Apr 4 '16 at 18:12
  • $\begingroup$ It all depend if your detection are dependent or independent ? What is the animal you are studying ? $\endgroup$ – Nico Coallier Apr 4 '16 at 18:15
  • $\begingroup$ Assume they are independent, and it's a large felid. $\endgroup$ – CSB Apr 4 '16 at 18:16
  • $\begingroup$ and in your model you ''y'' variable is the occurence of the felid you are studying ? And your question is : Is there more probability of detecting an felid in a specific habitat ? $\endgroup$ – Nico Coallier Apr 4 '16 at 18:19
  • $\begingroup$ not really, i guess you could frame it as a question of "probability of detection". but let's frame it as a very basic habitat modeling question: given this set of occurrence locations, what is suitable habitat for this species? i know probability of detection can bias that estimate. my question is whether the bias would decrease if using binary data. $\endgroup$ – CSB Apr 4 '16 at 18:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.