# The importance of fieldwork over predictive modeling in quantifying predation events of carnivores marked with GPS technology

> GPS collars tell researchers where a puma stopped, but not whether it made a kill. By visiting almost 1,900 puma clusters on foot in Chile, Colorado and California, CyberTracker-certified field teams showed that the computer models researchers use to save fieldwork can badly miscount kills and miss small prey. The authors conclude that there is still no substitute for reading the signs on the ground.

Summary by Kersey Lawrence, Original Wisdom. Updated 9 November 2026. Page: https://www.originalwisdom.com/tracking-literature/2018-elbroch-lowrey-importance-fieldwork-over-predictive/

## Citation and access

- **Authors:** L. Mark Elbroch, Blake Lowrey and Heiko U. Wittmer
- **Year published:** 2018
- **Journal or source:** Journal of Mammalogy
- **Full citation:** Elbroch, L. M., Lowrey, B. and Wittmer, H. U. 2018. The importance of fieldwork over predictive modeling in quantifying predation events of carnivores marked with GPS technology. Journal of Mammalogy 99(1): 223–232.
- **DOI:** 10.1093/jmammal/gyx176
- **Access status:** Hosted (freely available elsewhere)
- **PDF:** https://www.originalwisdom.com/wp-content/uploads/2026/11/2018_Elbroch_and_Lowrey_etal_the-importance-of-fieldwork-over-predictive-modeling-in-quantifying-predation-events.pdf

## Tags

- **Species:** Mammals > Mountain lion
- **Region:** South America > Chile; North America > United States
- **Tracking:** Beds; Bones; Feeding signs; Hair; Kill sites; Tracking; Tracks and signs identification
- **Behaviour and ecology:** Behaviour; Feeding; Kleptoparasitism; Predation; Predation risk; Scavenging
- **Methods and technology:** Akaike Information Criterion (AIC); Bootstrapping; CyberTracker; CyberTracker Tracker Certification; Geographic Information System (GIS); Global Positioning System (GPS) collars; GPS cluster analysis; Logistic regression; Simulation
- **Data type:** Modelling or theoretical; New field data
- **Observer qualification:** Certified trackers
- **Discipline:** Behaviour and ecology; Wildlife tracking
- **Year published:** 2018

## Detailed keywords

- **Species:** Studied: puma (Puma concolor), called mountain lion in the tag list. Prey: guanaco (Lama guanicoe), huemul deer (Hippocamelus bisulcus), domestic sheep (Ovis aries) and European hare (Lepus europaeus) in Patagonia; elk (Cervus canadensis), mule deer (Odocoileus hemionus) and cattle (Bos primigenius) in Colorado; black-tailed deer (Odocoileus hemionus columbianus) in California; medium-sized prey including American beaver (Castor canadensis) and North American porcupine (Erethizon dorsatum); small prey under 5 kg left out of the models, such as ground squirrels (Spermophilus spp.) and ruffed grouse (Bonasa umbellus). Competitors and sources of disturbance: American black bear (Ursus americanus), Andean condor (Vultur gryphus), people (ranchers and hunters with hounds). Other carnivores mentioned from the literature only: African lion, bobcat, brown bear, jaguar, leopard, Eurasian lynx, wolf, tiger, spotted hyena, African wild dog, cheetah.
- **Location and habitat:** Three study areas. Chilean Patagonia: southern Aysén, Chile; rugged mountains with open Patagonian steppe (53%), high-elevation lenga (Nothofagus pumilio) forest and lower shrublands. Western Colorado, United States: near De Beque; valley rangeland, sagebrush and farmland, pinyon-juniper woodland, Gambel oak forest, and high plateaus of mixed conifer and aspen. Northern California, United States: Mendocino National Forest near Covelo; grassland and chaparral, mixed conifer forest, and true fir forest at high elevations. (The paper labels the study-area coordinates “W” and “S” with latitude and longitude apparently swapped.)
- **Topics:** kill rates, prey selection and diet composition, GPS cluster investigation, predation versus resting sites, handling time at carcasses, prey size, kleptoparasitism by black bears, scavenging condors, fear of people (indirect effects of perceived risk), reduced field effort, model accuracy and precision, prey diversity, field validation, accelerometer and activity data, snow tracking as a check on detectability.
- **Statistics or models:** generalized linear mixed models (logistic, binomial error, logit link, random intercept per puma) fitted with the R package lme4; model selection by Akaike Information Criterion (AIC); optimum probability cutoffs from sensitivity and specificity curves; receiver operating characteristic (ROC) curves; leave-one-out (k-folds) cross-validation; bootstrapped standard errors (10,000 iterations); random subsampling simulations (10,000 per level); Shannon diversity index (H) and evenness (E). Inconsistencies in the paper: (1) the cluster count does not add up: 1,896 visited minus 18 scavenging, 20 not visited within a year and 100 small-prey clusters gives 1,758, not the 1,752 reported; (2) Table 1, California excluding clusters under 12 h, lists 191 clusters but 164 kills plus 24 non-kills is 188; (3) 100 of 973 prey is 10.3%, not the 9.7% reported; (4) the methods say 5%, 15%, 25% and 50% of clusters were randomly omitted, but Figure 2 and the results show 25%, 50%, 75% and 95% of the data sampled (that is, up to 75% removed).
- **Data type:** New field data with simulations: 26 pumas captured (24 used), 1,896 GPS clusters visited. Patagonia 2008 to 2010, Colorado 2011 to 2013, California 2009 to 2013.

## Synopsis

### 1. What they did

Since the early 2000s, researchers have estimated how often large carnivores kill (kill rates) by fitting animals with Global Positioning System (GPS) collars and looking for “GPS clusters”, places where an animal’s locations bunch together. Field teams visit some of these clusters to see whether there was a kill, then build a statistical model to guess which of the unvisited clusters were kills. This saves time and money, but it can go wrong when a carnivore eats prey of different sizes or is pushed off a kill early. The authors tested how reliable this shortcut is.

They studied pumas in three places with different prey: Chilean Patagonia (guanaco, huemul deer and domestic sheep), western Colorado (mule deer and elk) and northern California (black-tailed deer). Pumas were caught mainly with hounds, and also with box traps and foot snares, and fitted with GPS collars that took a location every 30 minutes to 2 hours. The team tried to visit every cluster in the field, searching a 30 m circle around each location for prey remains, and classed each cluster as a kill or not.

Using this nearly complete field record as the “truth”, they then pretended they had done less fieldwork in two ways:

- Selectively: leaving out short clusters (under 12 hours, or under 24 hours), as many studies do.

- Randomly: leaving out a random share of clusters, repeated 10,000 times.

For each version they built a model, predicted the number of kills, and compared it with what was found on the ground. They also compared the variety of prey in the pumas’ diets.

### 2. What they found

Fieldwork

- They captured 26 pumas and used 24. They visited 1,896 clusters; after removing scavenging events, clusters not visited within a year, and kills of prey under 5 kg, 1,752 clusters remained (California 355, Colorado 1,036, Patagonia 361).

- On average they visited 73 clusters per puma, about 38 days after the puma had left.

- 973 prey were found. After small prey were removed, diets were 88% guanaco, 9% sheep and 3% huemul in Patagonia; 80% mule deer and elk and 20% medium-sized prey in Colorado; and 96% black-tailed deer in California.

The models

- In all three areas, the time a puma spent at a cluster was the best and only reliable sign that it was a kill.

- Models built on all the visited clusters came closest to the true number of kills.

- Leaving out short clusters made the models look better by the usual tests, yet they gave wrong answers. In California and Colorado they underestimated kills. In Patagonia, leaving out clusters under 12 hours greatly underestimated kills, and leaving out clusters under 24 hours greatly overestimated them.

- In Patagonia, pumas spent much less time at sheep kills (about 9 hours) than at guanaco kills (about 41 hours), so models missed 72% of sheep kills. Guanaco kills in open grassland were also often abandoned quickly to Andean condors.

- In Colorado and California, black bears often pushed pumas off their kills, shortening the time spent there.

- Randomly leaving out clusters mostly made the estimates less precise. Visiting only 75% of clusters tripled the spread of estimates in California and doubled it in Patagonia.

Diet

Leaving out short clusters made pumas’ diets look less varied and overstated large prey. The effect was biggest in California, where diets included the most small-bodied prey: diversity dropped by 45% (under 12 hours removed) and 57% (under 24 hours removed). In Colorado it dropped by 9% and 16%.

### 3. What they concluded

- Predictive models of GPS clusters should be treated with caution. How long a carnivore stays at a kill depends on prey size, kleptoparasitism (food stealing by competitors), fear of people or other animals, and local ecology, and only fieldwork captures this.

- Skipping short clusters skews the data towards kills and hides small prey and quick kills.

- Until models improve, the authors recommend:

- Sample intensively for set periods (at least 60 days), with cluster rules suited to the species, and visit clusters from all times of day and night.

- Sample different individuals and seasons.

- When models are used, publish the raw data, including how many clusters were and were not visited.

- Also visit a random set of single GPS locations, or use other methods such as snow tracking along random sections of trail, to measure how many kills cluster methods miss.

- For now, cluster models may be best used to decide which clusters to visit in the field.

### 4. How tracking was or was not used

Tracks and signs identification was central to the fieldwork. At each GPS cluster, field teams searched for prey remains and identified prey from hair, skin, rumen contents and bone fragments. They read how the carcass lay and which parts had been eaten, and looked for the presence and position of bite marks to decide whether the puma had killed or scavenged the prey. Clusters without a kill were classed as non-predation events, such as beds. No trailing by people is reported (pumas were treed by hounds for capture). The authors suggest snow tracking along random sections of trail as a way to check how many kills cluster methods miss.

### 5. Tracker or observer certification

Certified. The paper states that all personnel doing field investigations were certified through CyberTracker to ensure reliable and standardized prey identification, citing Evans et al. (2009). The level of certification is not given. The authors also acknowledge that even well-trained observers can miss prey remains.

### 6. Other technologies, skills or innovations

- Argos and Iridium GPS collars (Lotek 7000SAW and Lotek IridiumTrack M), sending locations by satellite.

- Clusters plotted in ArcGIS or Google Earth, then loaded onto handheld GPS units for field searches.

- The authors point to accelerometer or activity data from modern collars as a promising way to tell feeding animals (moving their heads) from resting ones.

### 7. Implications for conservation (from the paper)

- Kill rates and prey selection feed into understanding predator and prey dynamics and wildlife management, so biased estimates can mislead decisions. Inaccurate predictions can also bias kill rates for rare or threatened prey.

- Capturing and handling carnivores carries risk, so researchers should gather as much as possible from every collared animal rather than handle more animals than needed.

- The authors cite a case where a study based on modelled kill rates and a later diet study of the same pumas reached opposite conclusions about kills near housing.

### 8. Implications for tracking (Original Wisdom's observations)

_These are Original Wisdom's observations, not the authors'._

- The study is a strong case for the ground-truthing role of tracks and signs: the GPS shows where a puma stopped, but only people reading hair, bones, bite marks and feeding signs on site can say what happened there.

- Short visits, quick kills and small prey are exactly where models fail, and where careful sign reading at every cluster makes the difference.

- The suggestion to sample random sections of trail by snow tracking points to trailing as a way of measuring what GPS clusters miss.

## Cite this summary

Lawrence, K. 2026. Summary of: The importance of fieldwork over predictive modeling in quantifying predation events of carnivores marked with GPS technology (Elbroch and Lowrey et al., 2018). Original Wisdom, Tracking Literature. https://www.originalwisdom.com/tracking-literature/2018-elbroch-lowrey-importance-fieldwork-over-predictive/