Breeding Status and Detectability

[Bird Counter]

I conducted 5 repeat surveys of breeding birds. I conducted 3 point count surveys on all of the sites, and on a subset of sites I conducted 2 more intensive surveys within the detection area of the point counts. The intensive surveys were used to establish the breeding status and nesting habits for species detected on point counts. Given the intensive nature of the sampling, my sample sizes are modest: 35 sites the first year and 49 the second year.

The season is closed and occupancy is defined as use. That is, the species was encountered at a site during the sampling period at least once. Although surveys were not conducted simultaneously, all surveys occurred during the period of closure. The surveys employed standardized protocols and are independent among methods and replicates.

I established breeding status independent of the point count surveys. I counted a species as breeding if I confirmed the presence of reproductive activity (RA) during nest searches. For species in which I saw no evidence of RA at a site, I classified them as undetermined (because these could be non-breeders, breeders on the periphery of their territories, or breeders that we missed). The overall probability of detecting reproductive activity, p(RA), was 1 – (1-delta)^2 >0.86 for each species in multi-state models.

Using breeding status as a state variable in models of point count detectability (confirmed breeder = 1, undetermined = 0), I found that the probability of detecting the species at a site was very low (p <0.15) when breeding status = 0. Often the probability of detecting confirmed breeders was quite high (p >0.7 or >0.95 over three surveys).

There are two possible interpretations:

1) Confirmed breeders are more detectable than undetermined individuals within the same area; therefore the probability of detecting an individual with undetermined breeding status that is present during the survey is much lower than a breeder.

2) Confirmed breeders are more detectable because they are more likely to use the point count detection area during all three surveys when their nests/fledglings are also within the PC detection area. Undetermined individuals are not present and undetected; instead they are less likely to present in all three surveys. (Note: this could be the case for failed breeders also, but we did not evaluate nest success).


Based on the independently conducted point counts and nest searches, I believe it is the latter. Overall, 89% of all of single observations during nest searches (in which the species was observed in only one of the two surveys) were of lone adults, usually males (63%). It makes sense that males without active nests/fledglings within the detection area are less likely to be present in all of the surveys.

Consequence: there is non-random movement in and out of the detection area based on breeding status during the period of closure. Thus, there is a potential that overall occupancy (use) estimates will be inflated, and uncertainty will be high for one group relative to the other. How non-random it is depends on how many of these undetermined individuals fall into the breeding versus non-breeding categories. This is something I cannot evaluate directly. However, I know the proportion of plots where I detected common species and also evidence of breeding activity, was > 0.50 for all but two species.

Currently, I am taking two approaches to the modeling of ψ-hat. First, I am analyzing the relationship between presence and habitat covariates without accounting for the putative breeding status of the individuals. Second, I am including only known breeders in the models. This is making a difference in the effects of covariates on parameter estimates for some species and not for others. So, there is some evidence that there is bias resulting from using all of the observations versus just those of confirmed breeders. Given the low detectability of the individuals of undetermined breeding status, there is not much I can do with this data separately.

I find this all very interesting, but also a bit vexing. It has added a whole level of complexity to the interpretation of detectability (especially given the +10 number of species I am working with). I should add that this is an a posteriori exploratory analysis. I implemented this study long before I was introduced to occupancy modeling but my study design allows this kind of analytical approach. The impetus to do these analyses came from the low availability of some common species during point counts that I knew to be breeding within the PC detection area.

I am seeking the thoughts of people conducting similar analyses, or who are interested in this issue.

[dhewitt]

This is an interesting study. I wonder if you need to worry that the definition of a "site" (you use the point count area) is actually different between the breeders and undetermined individuals. For example, breeders could have tight territories that they stick to and the undetermined birds just happen to move through -- making them more like transients. Thus, the point count area is really too small to sample the undetermined birds. I'm also curious to know what habitat covariates you're chasing and what kind of precision you are able to get with the undetermined birds, on both Phi and p.

[Bird Counter]

Thanks for your reply. I'll write up a post with the answers to your questions within a day. I conducted the study in post-burn habitat in SW Oregon. I agree with your initial thoughts that the detection area is more of a transit site (e.g., Poulin et al. 2008, Biol. Conserv. 141(4):1129-1137).

[Bird Counter]

Now that I have my data in front of me, I’ll answer your questions.

This is an interesting study. I wonder if you need to worry that the definition of a "site" (you use the point count area) is actually different between the breeders and undetermined individuals. For example, breeders could have tight territories that they stick to and the undetermined birds just happen to move through -- making them more like transients. Thus, the point count area is really too small to sample the undetermined birds.

Yes, this is exactly what I think. Confirmed breeders (that is, those with active nests or young within or near the detection area) are more likely to be detected on multiple surveys than individuals which are outside of their core territories or are non-breeders/failed breeders. Nevertheless, I think an important assumption (often unstated) made about birds sampled with point counts during the breeding season is that individuals present are indeed breeding. If they aren’t, then the environmental data collected from the detection area will not necessarily be indicative of nesting habitat. [This ignores nest success]. From that perspective, the sampled area is not too small. A potential useful application of this kind of information is that when detectability is very low in some areas relative to others, it may be indicative of breeding status (assuming the survey methodology is sound). That was not the goal of this study, but it does provide interesting pilot data that could be used to design a study on a few species of interest (and with better precision).

I'm also curious to know what habitat covariates you're chasing and what kind of precision you are able to get with the undetermined birds, on both Phi and p.

I worked in forest burned with high severity that ranged from 300-1350 m in elevation during the early post-fire period. Given the modest size of the dataset, I am focusing a few habitat covariates (percent ground vegetation, distance from live forest, elevation, snag density and hardwood basal area).

Here are examples real parameter estimates for two species using the most simple multi-state model: ψ,R,p1(.)p2(.)Delta(.). I just downloaded the newest version of Presence yesterday so I ran the models using that.

In the first example, the probability of detecting non-breeders (p1) is very low relative to breeders. The probability of detecting breeders given species presence (p2) is better (but in this case, there is a lot of uncertainty associated with the estimate, which creates other problems). At any rate, there are large difference in the point estimates.

The naïve estimate of ψ1 is 0.63
The naïve estimate of the number of sites with confirmed breeders, ψ2 is 0.48 of the sites where the species was detected.

Individual Site estimates of Psi1:
Site Survey Psi1 Std.err 95% conf. interval
1 002 1 1-1: 1.0000 0.0002 0.0000 - 1.0000
Individual Site estimates of Psi2:
Site Survey Psi2 Std.err 95% conf. interval
1 002 1 1-1: 0.4327 0.1299 0.2128 - 0.6828
Individual Site estimates of P1:
Site Survey P1 Std.err 95% conf. interval
1 002 1 1-1: 0.2306 0.0819 0.1082 - 0.4257
Individual Site estimates of P2:
Site Survey P2 Std.err 95% conf. interval
1 002 1 1-1: 0.7289 0.1118 0.4701 - 0.8906
Individual Site estimates of delta:
Site Survey delta Std.err 95% conf. interval
1 002 1 1-1: 0.6296 0.1426 0.3389 - 0.8493

As a result, this model did not reach convergence and the estimate of ψ1 is meaningless. What is happening is that the non-breeders (more appropriately, undetermined) are assumed to be present but undetected, thus inflating the estimates of ψ1. Am I thinking about this correctly? I understand that it is better to use the term ‘use’ rather than occupied to account for the fact that the species is present at least part of the time in the detected area, and as long as the movement in and out of the detection area is random, this should not bias the results. Unfortunately, I think we often have to assume that movement is random because we don’t have enough information about the movement patterns of the species in question.

The second species has a different issue. The point estimate for putative non-breeders is higher, but the SE is unacceptably high, again creating problems. Note that our ability to detect confirmed breeders was very good.

The naïve estimate of ψ1 is 0.47.
The naïve estimate of the number of sites with breeding, ψ2 is 0.74 of the sites where the species was detected.

Individual Site estimates of Psi1:
Site Survey Psi1 Std.err 95% conf. interval
1 002 1 1-1: 0.5283 0.1379 0.2746 - 0.7682
Individual Site estimates of Psi2:
Site Survey Psi2 Std.err 95% conf. interval
1 002 1 1-1: 0.7253 0.1774 0.3156 - 0.9380
Individual Site estimates of P1:
Site Survey P1 Std.err 95% conf. interval
1 002 1 1-1: 0.3750 0.4053 0.0198 - 0.9468
Individual Site estimates of P2:
Site Survey P2 Std.err 95% conf. interval
1 002 1 1-1: 0.9231 0.0537 0.7313 - 0.9814
Individual Site estimates of delta:
Site Survey delta Std.err 95% conf. interval
1 002 1 1-1: 0.7500 0.0988 0.5164 - 0.8939

I think that the high uncertainty associated with p1 is a result of seeing evidence of breeding in only one of the two surveys at 6 sites.

Using these two species as models, it appears that the higher the discrepancy between p1 and p2 and the fewer the number of sites with breeders, the more influence the non-breeders have.