Preliminary questions on N-mixture models

224 views
Skip to first unread message

david rosenbaum

unread,
Feb 15, 2021, 12:12:03 PM2/15/21
to unmarked
Hi all,

I'm a MS student interested in using N-mixture modelling to estimate abundance from a count dataset I've collected. I've only just started exploring them and have some beginner conceptual questions.

First, do these models assume that surveyed sites are occupied? My survey methods have a high detection probability as estimated from static occupancy modeling ( ~0.7) but I have several sites with zero detections across all their respective repeated surveys. However, when examining output from N-mixture models I fit using pcount(), I still get low abundance estimates from sites with zero detections. Due to the sites with all zeros, would fitting models using a ZIP mixture be reasonable? 

Additionally, the detection probabilities given by N-mixture models are much lower than those given by occupancy models. Do detection probabilities for N-mixture denote detectability of the species, as in occupancy models, or do they describe detection probability of detecting an individual?

Finally, I have ran into what I'm learning is a common issue: the "good fit, bad estimate" problem. NB mixture models I fit have the lowest AICc and have a more reasonable goodness of fit compared to ZIP and P models (c-hat is around 1 for my NB global model but >2 for both ZIP and P). However, all NB estimates are sensitive to K, unlike ZIP and P models. Given this situation, what following steps would be recommended? I have thought of selecting P or ZIP models using QAICc to account for overdispersion these provide, while retaining the finite estimates. Would I be on the right track there?

Thank you for any help or information,

David Rosenbaum


Philip Shirk

unread,
Feb 15, 2021, 2:03:43 PM2/15/21
to unma...@googlegroups.com
A few quick thoughts: 
  1. The "true" abundance is modeled with a Poisson (or Negative Binomial, or something similar) distribution. Those distributions do allow for 0's, so the models aren't exactly assuming that the sites are occupied. What they're giving you is an estimate of the abundance/density at the sites. That estimate comes from the mean of a Poisson (or other) distribution, and when you have observations at some sites, the mean of the distribution will never be exactly 0. It might be very low, but not 0. 
  2. The detection probabilities are different. In occupancy models the detection probability is the probability of detecting a species if it is present. In N-mixture models, the detection probability is the probability of detecting any given individual that is present. 
  3. This one's a bit tougher. N-mixture models can be quite sensitive to some of the modeling assumptions, and when you're working with real-world data it can be difficult to know when a model is leading you astray. I would be inclined to trust the NB a little more than the P model because the NB can account for overdispersion in counts. K is really a nuisance parameter. Theoretically it should be infiniti, but we set it lower for computational reasons. One approach is to just keep making K larger until estimates stabilize. Another approach that you could take would be to fit the models using the UBMS package, which allows you to easily add random effects to the model (I think - I haven't actually used UBMS myself). Adding random effects will help to account for overdispersion beyond what the NB can account for (e.g. overdispersion in detection probability). Fitting models in a Bayesian framework with the UBMS package also has the benefit of assessing the models in more of a simulation framework (i.e. comparing the data that you observed to data simulated from the model fit - ideally they will look very similar). You can do the same thing in a frequentist framework, too, it's just less commonly done.
Philip Shirk
Postdoctoral Scholar
Cooperative Institute of Marine Resources Studies
Oregon State University


--
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unmarked+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/unmarked/66114411-ccde-4beb-8fe1-6a2c53fa4274n%40googlegroups.com.

david rosenbaum

unread,
Feb 16, 2021, 10:50:08 AM2/16/21
to unma...@googlegroups.com
Hi Philip,

Thank you for the insight! I have tried adjusting K (even up to 1,000,000!) for my NB models but the estimates never stabilize and become quite unreasonable when K is this high. I will do some research on UBMS.

David Rosenbaum

You received this message because you are subscribed to a topic in the Google Groups "unmarked" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/unmarked/xQ_TErL5vUU/unsubscribe.
To unsubscribe from this group and all its topics, send an email to unmarked+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/unmarked/CANG%3DJfEXf2shuefr1nq2YL1fd73Ag4_0WbC89UQbjwW0X9XdYw%40mail.gmail.com.

Philip Shirk

unread,
Feb 17, 2021, 12:24:25 AM2/17/21
to unma...@googlegroups.com
That's not a good sign. Your data might not follow the model assumptions closely enough to get a good model fit. Dennis et al. 2015 talks about situations where the choice of K may cause issues.

Philip Shirk


Reply all
Reply to author
Forward
0 new messages