Skip to main content
Dryad logo

Monitoring small mammal abundance using NEON data: Are calibrated indices useful?


Parsons, Arielle; Clark, James; Kays, Roland (2022), Monitoring small mammal abundance using NEON data: Are calibrated indices useful?, Dryad, Dataset,


Small mammals are important to the functioning of ecological communities with changes to their abundances used to track impacts of environmental change. While capture-recapture estimates of absolute abundance are preferred, indices of abundance continue to be used in cases of limited sampling, rare species with little data, or unmarked individuals. Improvement to indices can be achieved by calibrating them to absolute abundance but their reliability across years, sites, or species is unclear. To evaluate this, we used the US National Ecological Observatory Network (NEON) capture-recapture data for 63 small mammal species over 46 sites from 2013–2019. We generated 17,155 absolute abundance estimates using capture-recapture analyses and compared these to two standard abundance indices, and three types of calibrated indices. We found that neither raw abundance indices nor index calibrations were reliable approximations of absolute abundance, with raw indices less correlated with absolute abundance than index calibrations (raw indices overall R2 < 0.5, index calibration overall R2 > 0.6). Performance of indices and index calibrations varied by species, with those having higher and less variable capture probabilities performing best. We conclude that indices and index calibration methods should be used with caution with a count of individuals being the best index to use, especially if it can be calibrated with capture probability. None of the indices we tested should be used for comparing different species due to high variation in capture probabilities.  Hierarchical models that allow for sharing of capture probabilities over species or plots (i.e., joint likelihood models) may offer a better solution to mitigate the cost and effort of large-scale small mammal sampling while still providing robust estimates of abundance.


The NEON dataset we used contains capture-recapture data for 63 species over 46 sites from 2013–2019, representing 39% of North American small mammal species. Each site contains 3-8 (mean = 6) replicate trapping arrays of 100 traps set in grids with 10m spacing. Field scientists used baited Sherman live traps for animal capture, checked daily and set within 10 days before or after the new moon. Traps were typically run at monthly intervals for 6 months during the growing season at a subset of sites (core sites), and 3-4 months for the rest of the sites. At each site, half of the trap arrays were run for multiple nights (mean = 3), and the other half were run for a single night. Trapped individuals were tagged with either an individually identifiable ear tag or PIT tag. NEON trapping targets small rodents including cricetids (New World rats and mice, lemmings, voles), dipodids (birch mice, jumping mice), heteromyids (kangaroo rats), small sciurids (flying squirrels), and introduced murids (Old World rats and mice, gerbils). Sampling did not target lagomorphs (rabbits, hares, pikas), mustelids (weasels), large squirrels, or soricids (shrews). Although some of these taxa were incidentally captured, they were removed from this analysis. Some captured individuals (n = 453) were incorrectly identified at the species level in the field, and we used NEON DNA barcoding data to correct those identifications before analysis.

Capture-recapture analysis.—Under the NEON sampling protocol, about half of the trapping grids at each site were sampled multiple nights per trapping session, while other grids were sampled only a single night per month.  Traditional capture-recapture abundance estimates require that multiple nights of sampling are conducted, omitting data from grids that are sampled only a single night.  To accommodate all the data, we used a Bayesian hierarchical approach wherein trapping arrays sampled for multiple nights share a likelihood with arrays that ran only one night, thus providing information on capture probability that exploits the information available both in repeat- and single-night counts.  The approach of Royle et al. (2012) was adapted to estimate abundance at replicate trapping arrays within each geographic site during each month and year of sampling. We use data augmentation to fix the dimensionality of the model (i.e., make an unknown sample size “known”) by choosing an arbitrarily large super population size M of 700 individuals for each species at each trapping array during each closed session, representing a density of roughly 70,000 individuals/km2. The capture histories of all individuals of a given species captured at a geographic site are pooled (i.e., capture histories for each individual over all arrays at the site) and we then add an individual-level covariate describing the array membership, latent for augmented individuals.  Estimates of abundance at each replicate array within the geographic site can be obtained by summing over the individuals associated with each array.  Because sampling at each trapping array took place over a restricted time period (1–3 days), the assumption of constant detection probability p over time is reasonable.  While accounting for individual heterogeneity in capture probabilities has been shown important in small mammal studies, because we are making inference about individuals for which we have no individual detection data (i.e., those captured in arrays running single nights only) we could not adequately model individual heterogeneity in capture probabilities. We ran our model separately for each species within a site/month/year closed session in JAGS (Plummer 2003) via package runjags (Denwood 2016) in Program R (Version 3.5.3; R Core Team 2017).  We skipped species during closed sessions with < 10 individuals captured at a geographic site and those site/month/years with < 3 trapping arrays running, for a total of 9,226 models.  We assessed model fit for each species closed session with posterior predictive checks (PPC) by calculating the sum of squared Pearson residuals. We calculated a Bayesian p-value from posterior simulations and assumed adequate fit if 0.1 < p < 0.9. Most models converged (i.e., Gelman-Rubin statistic < 1.1, confirmed by examining traceplots) after a burn-in of 5,000 iterations and sampling of 300,000 iterations, thinning every 100 samples.  Those that did not converge after that number of iterations were discarded (n = 114). 

Datasets used: 


  • Royle J.A., Converse S.J. 2014. Hierarchical spatial capture–recapture models: modelling population density in stratified populations. Methods in Ecology and Evolution 5:37-43.
  • Royle J.A., Converse S.J., Link W.A. 2012. Data augmentation for hierarchical capture-recapture models. arXiv preprint arXiv:1211.5706.
  • Sollmann R., White A.M., Gardner B., Manley P.N. 2015. Investigating the effects of forest structure on the small mammal community in frequent-fire coniferous forests using capture-recapture models for stratified populations. Mammalian Biology 80:247-254.


National Science Foundation, Award: 582434