The paper presents a method and R package for estimating density and capture related parameters using simulations and inverse prediction. This approach allows analysis of spatial capture-recapture data that are not readily fitted using ML or MCMC methods. I read the paper and worked through the examples provided in the Appendix. Here are my few, mostly minor comments and suggestions.

The title and part of the introduction, rather broadly, promise an approach for dealing with “awkward” spatial capture-recapture data. Current functionality of the package caters to situations that involve single-catch traps and trap interference, possibly in the context of spatially varying density. Perhaps the title should reflect this narrower scope? Alternatively, could a solution for one other type of detection-level non-independence be presented/exemplified, for example one of the mechanisms mentioned in the Discussion (post-collection subsampling or group-living)?

Line 34: I suggest specifying that you mean a physical trap/restraining device. Although, I suppose it could also apply to other detectors, for example single-use hair snags.

L 77: The description in the text, aided by Figure 2 and the examples in the Appendix, is a concise depiction of the approach. A diagram/flow chart (perhaps integrating the panels of figure 2) could give readers a more intuitive understanding of the approach.

L 81: Although proxies are explained later, I suggest adding a descriptive sentence about proxies here to keep the reader from wondering. Maybe something along the lines “Proxies are measures related to the parameter of interest; for example, the number of unique individuals detected can serve as a proxy for density.”

L 97: You acknowledge that the choice of proxy is “something of an art”. Nonetheless, could you provide a bit more guidance here? Other than an approximately linear relationship (after link application), what should readers look for in their search for proxies? This is something we are also struggling with as we explore proxies for use goodness of fit testing in spatial capture-recapture analysis, and any insights would be appreciated.

Line 137: Consider using a different letter for the number of detections of an individual; n is already used for the number of distinct individuals detected.

L 178: For people running ipsecr.fit on a computing grid, parallelizing simulations should allow for speedy execution. Some of the more complex models will run slowly on personal computers with a limited number of cores. To achieve a reduction in computation time for those users, would it be feasible to allow a different (lower) number of simulations for iterations of simulation and inverse prediction until the hyperbox formed by the set parameters includes the inferred parameters? Or is this already implemented via the tuning parameters?

L 222: Without simulations demonstrating that the method “works well with some data sets”, we must take your word for it (or test it ourselves). I do not think an extensive simulation study exploring a wide range of scenarios is needed but a small demonstration with known “true” parameters would help. As the single-catch problem with varying density has been pointed out in the Abstract, perhaps it can serve for such a simulation study, comparing bias and error in key parameters for models that do and do not ignore the limits on individuals per trap and traps per individual during an occasion.

Table 1: A more detailed description of some arguments may help here or in the R help file. Also, from the description of the approach it was not apparent to me that multiple hyperboxes can be constructed. Could this be made clearer and explained in the text?

L 248: The non-target data (NTFeb1996.txt and NTFeb1997) were not provided with the submission, at least I did not find them. To run this part of the code, I generated random interference data using the numbers provided in Table 2. Will the actual data be provided with the paper or package?

Table 3: Out of curiosity, could the overall proportion of disturbances be replaced by multiple proxies that inform about the spatio-temporal configuration of disturbance (e.g., proportion of traps disturbed and proportion of occasions with disturbances), to advantage? Presumably both proxies are impacted by the spatial/temporal configuration of interference, and could perhaps be even more influential when density is variable?

Richard Bischof
Norwegian University of Life Sciences

I am satisfied with the explanations provided and revisions made by the author. Congratulations on a nice paper!