Methods for studying wild animals’ causes of death

This is the second post in a series. Read part one → Part three →

Key takeaways

• Cause of death is an important area of wild animal welfare research.
• Accurate cause-specific mortality rates are difficult to obtain.
• Multistate modeling can provide insight where information about cause-specific mortality rates is incomplete.
• Experimental studies and improved monitoring technology can improve our ability to diagnose cause of death, estimate cause-specific mortality rates, and understand compensatory mortality.
• Future research should identify and focus on the causes of death that involve the most suffering.

In a world with perfect monitoring of wild populations, biologists could be immediately alerted to the death of an animal, show up on the scene, and perform a necropsy to determine the animal’s precise cause of death. They would also know everything about the age and health history of this animal, having monitored them from birth. Unfortunately, it is not practical to monitor most wild populations at nearly this level of detail. Our understanding of how and why wild animals die must be pieced together from statistical models and sparse empirical data. In this post, I will introduce some of the tools and approaches used to study wild animals’ causes of death.

Death assemblages

The most straightforward way to study wild animals’ causes of death is to scour a habitat for corpses. The resulting collection is sometimes called a “death assemblage.” When the body of a recently deceased animal is recovered, indicators such as injuries, signs of disease, and stomach contents may make it possible to determine what killed them (Behrensmeyer and Miller 2012).

Determining cause of death from serendipitously recovered corpses depends on finding animals soon after their deaths. Death assemblage surveys are therefore most useful for animals whose remains are more likely to be discovered, like coastal marine mammals. Their large size means they take a while to decay, and their bodies often wash up on beaches where they are highly visible and quickly reported to authorities or researchers (Tinker et al. 2016). Recovery of well-preserved corpses is much less common for animals who are small or live in lush environments like forests, which typically provide more concealment and support more rapid decomposition. 

The probability of recovering a corpse can also be biased by whatever caused the animal’s death (Gerber et al. 2004). For example, the bodies of animals who were killed by predators may be less likely to be recovered because they are almost entirely consumed on the spot or are carried away from the site (e.g. hyenas stashing bones of prey, Kuhn 2005). On the other hand, animals who die from starvation, disease, old age, or something else that leads to their gradual weakening may attempt to conceal themselves in their final hours. While their bodies are likely to remain intact for longer, they may go unnoticed by researchers.

Certain causes of death are also inherently more difficult to diagnose, so researchers need to be wary of simply ignoring data in an “unknown” category that may be disproportionately filled with instances of hard-to-distinguish causes of death. For example, in a survey of bones found in a hyena den, presumably belonging to their past prey, Kuhn (2005) identified roughly a quarter of the specimens to species level. Without further information, any quantitative claims about the composition of the local ecosystem based on these data would have to be taken with serious caution because the remaining 75% of “unknown” specimens might be drawn unequally from underlying populations. For example, animals who are morphologically similar to other species, or animals who are not preferred prey for hyenas, might be underrepresented in resulting estimates. 

Bias is a pervasive problem in cause of death surveys. A naive survey of rabbit corpses in a particular forest might conclude that most recent deaths were attributable to some rapidly debilitating disease, while in fact the majority have been carried off by birds of prey without a trace. To estimate the true rates of alternative causes of death from corpse recovery data, researchers need to account for the fact that the forensic clues to certain causes of death may be more or less detectable than others, depending on a host of circumstances.

Accounting for incomplete information

The inevitable incompleteness of corpse recovery surveys can be partially offset by explicitly modeling sources of uncertainty along the chain of events from death by a specific cause, to corpse recovery, to cause identification. This approach is sometimes called multistate modeling (Hougaard 1999; Gauthier and Lebreton 2009). These models make a distinction between the objective chain of events that occurred to an individual animal and the observable evidence of those events (Koons et al. 2014). For example, researchers generally do not directly observe rodents being preyed on by raptors. A fraction of predation events leave behind bits of fur or the partially eaten body of the rodent. Researchers then have a particular probability of encountering these bits of evidence and deducing the correct cause of death.

Box 1

To illustrate what a multistate model looks like in practice, imagine that a flock of birds begins to roost every day in a tree outside where you live. The foliage is dense, so you cannot see all of the roosting birds at any one time through your window. One of the birds (“Bird X”) has distinctive markings, with which you can usually distinguish them from the others. One day, you can’t see Bird X in the tree. Your first worry is that Bird X might have died, but then you remember that they might just be out of view behind a branch. A week passes and Bird X still hasn’t been seen in the tree. Is it more likely that you simply have not caught sight of Bird X for seven days in a row, or that Bird X died sometime in the week?

This problem can be solved by estimating the exact proportion of birds in the tree who can be seen on any given day (“detection probability”), the probability of recognizing the identity of Bird X if you set eyes on them (“identification probability”), and the average daily survival probability for each bird in this population. These can then be incorporated into a model where Bird X is considered to be in one of two states: alive or dead. Using the model, we can determine whether our series of observations of the tree outside is better explained by Bird X being alive or by being dead. Marescot et al. (2015) applied a similar sort of model to observations of tagged black-tailed deer to estimate their age-specific survival rates and the impact of predation on their population growth.

While multistate modeling is frequently used to estimate survival rates in wild animal populations, similarly nuanced studies of cause of death are rare. This may reflect both a general disinterest in cause-of-death research and a greater confidence in the raw data that does exist. Studying human-caused deaths or species for which there is a good infrastructure for monitoring removes some uncertainty and lessens the need for multistate models. These statistical methods are most important to apply when studying populations where the vast majority of individuals’ deaths go undocumented, leaving room for large biases in the data. For example, Gerber et al. (2004) investigated cause-specific mortality of sea otters. To do so, they took advantage of an extensive network of marine mammal research groups to recover and analyze carcasses that washed up on the beach. Using a demographic model, they estimated that their sample represented just under half of all mortality during the study period. To sample half of all deceased individuals in a wild population of marine mammals is a remarkable achievement, but it highlights how much potential there is for unequal cause-specific recovery rates to produce biased results, especially in less intensively monitored populations. In marine mammals, for example, different predator species may hunt at different distances from shore, or near rocky cliffs as opposed to smooth beaches. Either of those factors could affect the probability of a carcass washing up onshore and being recovered by researchers (Joly et al. 2009).

The biggest weakness of purely observational studies of wild animal mortality is that it is difficult to know what data are missing. The simplest way of overcoming this is to tag a set number of animals at the beginning of a study and only count the remains of tagged individuals. This sort of study design is still subject to biased recovery rates, but the magnitude of bias is limited because corpses of untagged individuals are excluded. Schaub and Pradel (2004) used such an approach to estimate the proportion of white storks killed in airborne collisions with power lines in Switzerland (Figure 1a). They found that simply counting corpses overestimated the proportion of juvenile white storks killed by power lines by about 25%. It makes sense that the bodies of birds killed by power line collisions would be more conspicuous than birds who die otherwise, because the ground around power lines tends to be more exposed to allow for maintenance access. Power lines may also serve as natural concentrating points for associated deaths, whereas deaths by other causes may be uniformly distributed across the landscape.

Limiting analyses to tagged animals is effective at preventing overestimation of particular causes of death, but it can also leave researchers with a lot of missing data from individuals whose bodies were not recovered or who lost their tags. Steps can be taken to improve the recovery rate of tagged animals, such as attaching radio trackers to animals alongside conventional tags (Tavecchia et al. 2011; Figure 1b). Records of when a tagged animal was last seen alive can also complement corpse recovery data, because they allow for independent estimation of survival rates and rates of tag loss. Without this information, corpse recovery data alone cannot distinguish between an animal who is still alive and one who died but was never recovered. Fortunately, technical improvements in the precision and durability of radio trackers are making it practical to do this remotely. Even with relatively basic GPS tags, statistical models can distinguish tag loss or technical failure from actual death of the tagged animal with almost perfect accuracy (Sergio et al. 2018). Electronic tags capable of recording animals’ fine-scale movements or vital signs can also provide evidence related to an animal’s cause of death, such as how sudden it was or what posture the animal adopted in their final moments (Cooke 2008; Brown et al. 2013). Advances in the field of animal remote sensing will be helpful for learning about cause-specific mortality at a level of detail that is currently impossible for researchers to observe.

Figure 1

Multistate models treat wild animals’ fates — observed or unobserved — as the product of a series of conditional probabilities reflecting the combination of events over their life. A) A basic model that considers whether an animal is alive or dead, what caused their death, and whether their corpse was recovered (Schaud and Pradel 2004). B) A more complex model that estimates the probabilities of a bird dying from poisoning or electrocution, and ultimately the probability of a given bird being sighted again alive or dead (Tavecchia et al. 2011). In both of these models, the state transition probabilities are estimated based on the history of encounters between researchers and each tagged animal over the course of a study. For example, if researchers observe an individual after not sighting them for multiple field seasons, then that sighting might both increase the estimated survival rate and decrease the estimated detection rate.

Experimental study of cause-specific mortality

Rather than simply observing the fates of wild animals, experimental approaches directly exclude or introduce specific causes of death. The most common form of this approach is to exclude predators with fencing or other deterrents. For example, Smith et al. (2012) excluded all medium-sized mammalian predators from some nesting grounds of the gopher tortoise, but not others, and found that tortoises born in the predator exclusion zones were almost twice as likely to still be alive one year later. Similarly, Conner et al. (2011) found that exclusion fencing reduced the number of cotton rats killed by mammals. However, increased predation by birds and snakes largely made up for the decreased predation by mammals, especially after a forest fire destroyed much of the undergrowth that could otherwise conceal the rats. 

Although most experimental research on cause-specific mortality to date has focused on hunting or predation, a similar approach could be used to study the effectiveness of interventions against other common causes of death, including curable diseases or hunger (c.f. Rickett et al. 2013). However, interventions like supplemental feeding, disease prevention and predator exclusion may result in compensatory mortality, where individuals spared from one cause of death are at elevated risk of dying soon after because of another cause. The most effective way of learning about compensatory mortality risk is to use an experimental rather than observational study design, in conjunction with the same best practices for observational studies as described previously (Bergman et al. 2015). 

Outlook for welfare-focused cause of death research

From the perspective of wild animal welfare, understanding cause-specific mortality is morally urgent. A few shifts in the current approach to cause-of-death research will help us determine which issues to prioritize to help animals. Most importantly, much more research should be focused on the most common deaths in the wild: deaths of juvenile animals and members of species with large global populations. Because they are so numerous, these individuals’ experiences of dying are especially important to understand if we value all sentient animals’ welfare equally. These animals tend to be more difficult to monitor due to their smaller body size, but studying their mortality is becoming ever easier with new technology. At the current rate of progress, organisms weighing around one gram could be tracked via GPS within a decade or less (Kays et al. 2015). 

One major source of uncertainty in wild animal mortality research is the diagnosis of an animal’s cause of death. This is especially true for low-tech studies, where corpses are encountered by chance, rather than sought at a particular location based on radio tracking and a mortality signal. As a result, corpses may be significantly decomposed or scavenged. Even under ideal circumstances, it can be difficult to determine the order in which apparent injuries occurred. In lieu of perfect monitoring, researchers could perform detailed analysis on a subset of the recovered bodies to evaluate the reliability of their field diagnoses of the rest, as Joly et al. (2009) did with sea otter carcasses. Studies could also be dedicated to intensive monitoring of a relatively small number of animals so that when they die the cause can be known with high confidence and researchers can quickly find their corpses to document field-identifiable characteristics and how these change over the coming days or weeks, as well as estimate recovery rates. The field of taphonomy has long filled a similar supportive role for paleobiology (Behrensmeyer and Miller 2012). Upfront investment in this sort of meta-research could enable subsequent studies of wild animal mortality to accomplish more with less (c.f. Loannidis 2018).

Welfare-focused researchers should aim to identify and focus on the causes of death that entail much more suffering than others. This could involve a reframing of cause-specific mortality research where rather than asking how animals of a given species are likely to die, researchers ask who the victims of a particular cause of death are likely to be. Rather than tracking marmots and determining what proportion were preyed on by cougars, for instance, researchers could monitor cougars using cameras or fecal analysis to learn who their prey were (Popanom et al. 2011). This sort of cause-centric approach might also be appropriate in contexts where a single ultimate cause can lead to several different proximate causes of death which vary in their severity. For example, researchers could survey a forest in the aftermath of a fire and try to determine what proportion of fatalities occurred by burning, asphyxiation, subsequent starvation, or other proximate causes related to the fire (Gutiérrez and Javier de Miguel, 2020). Simply verifying that a certain cause of death occurred at least once in a given population may be useful to inform future work. If so, low-cost cause-specific mortality studies could use a form of rarefaction analysis to judge how comprehensively the causes of death present in a population have been sampled, with a focus on enumerating the causes rather than estimating their relative frequencies. 

Efforts to improve the health and lifespans of our fellow humans are supported by publicly available data on the most common causes of death. This is part of how governments decide which public health interventions and fields of medical research to allocate funding to. A comparable wealth of data could help us identify which wild animals need the most urgent aid, and what form that aid should take. In the next part of this series, I will give an overview of some of the data that does exist on the causes of wild animal mortality.

References

1. Behrensmeyer, Anna K., and Joshua H. Miller. 2012. ‘Building Links Between Ecology and Paleontology Using Taphonomic Studies of Recent Vertebrate Communities’. In Paleontology in Ecology and Conservation, edited by Julien Louys, 69–91. Springer Earth System Sciences. Berlin, Heidelberg: Springer. https://doi.org/10.1007/978-3-642-25038-5_5.

2. Bergman, Eric J., Paul F. Doherty, Gary C. White, and A. Andrew Holland. 2015. ‘Density Dependence in Mule Deer: A Review of Evidence’. Wildlife Biology 21 (1): 18–29. https://doi.org/10.2981/wlb.00012.

3. Brown, Danielle D., Roland Kays, Martin Wikelski, Rory Wilson, and A. Peter Klimley. 2013. ‘Observing the Unwatchable through Acceleration Logging of Animal Behavior’. Animal Biotelemetry 1 (1): 20. https://doi.org/10.1186/2050-3385-1-20.

4. Conner, L. Mike, Steven B. Castleberry, and Anna M. Derrick. 2011. ‘Effects of Mesopredators and Prescribed Fire on Hispid Cotton Rat Survival and Cause-Specific Mortality’. The Journal of Wildlife Management 75 (4): 938–44. https://doi.org/10.1002/jwmg.110.

5. Cooke, Steven J. 2008. ‘Biotelemetry and Biologging in Endangered Species Research and Animal Conservation: Relevance to Regional, National, and IUCN Red List Threat Assessments’. Endangered Species Research 4 (1–2): 165–85. https://doi.org/10.3354/esr00063.

6. Gauthier, Gilles, and Jean-Dominique Lebreton. 2008. ‘Analysis of Band-Recovery Data in a Multistate Capture-Recapture Framework’. Canadian Journal of Statistics 36 (1): 59–73. https://doi.org/10.1002/cjs.5550360107.

7. Gerber, Leah R., M. Tim Tinker, Daniel F. Doak, James A. Estes, and David A. Jessup. 2004. ‘Mortality Sensitivity in Life-Stage Simulation Analysis: A Case Study of Southern Sea Otters’. Ecological Applications 14 (5): 1554–65. https://doi.org/10.1890/03-5006.

8. Gutiérrez, J. and de Miguel, F. J. 2020. Challenges posed by fires to wild animals and how to help: A literature review, Oakland: Animal Ethics, retrieved from https://www.animalethics.org/fire-wild-animals-help.

9. Hougaard, Philip. 1999. ‘Multi-State Models: A Review’. Lifetime Data Analysis 5 (3): 239–64. https://doi.org/10.1023/A:1009672031531.

10. Ioannidis, John P. A. 2018. ‘Meta-Research: Why Research on Research Matters’. PLOS Biology 16 (3): e2005468. https://doi.org/10.1371/journal.pbio.2005468.

11. Joly, Damien O., Dennis M. Heisey, Michael D. Samuel, Christine A. Ribic, Nancy J. Thomas, Scott D. Wright, and Irene E. Wright. 2009. ‘Estimating Cause-Specific Mortality Rates Using Recovered Carcasses’. Journal of Wildlife Diseases 45 (1): 122–27. https://doi.org/10.7589/0090-3558-45.1.122.

12. Kays, R., M. C. Crofoot, W. Jetz, and M. Wikelski. 2015. ‘Terrestrial Animal Tracking as an Eye on Life and Planet’. Science 348 (6240): aaa2478–aaa2478. https://doi.org/10.1126/science.aaa2478.

13. Koons, David N., Marlène Gamelon, Jean-Michel Gaillard, Lise M. Aubry, Robert F. Rockwell, François Klein, Rémi Choquet, and Olivier Gimenez. 2014. ‘Methods for Studying Cause-Specific Senescence in the Wild’. Methods in Ecology and Evolution 5 (9): 924–33. https://doi.org/10.1111/2041-210X.12239.

14. Kuhn, Brian. 2005. ‘The Faunal Assemblages and Taphonomic Signatures of Five Striped Hyaena (Hyaena Hyaena Syriaca) Dens in the Desert of Eastern Jordan’. Levant 37 (1): 221–34. https://doi.org/10.1179/lev.2005.37.1.221.

15. Marescot, Lucile, Tavis D. Forrester, David S. Casady, and Heiko U. Wittmer. 2015. ‘Using Multistate Capture–Mark–Recapture Models to Quantify Effects of Predation on Age-Specific Survival and Population Growth in Black-Tailed Deer’. Population Ecology 57 (1): 185–97. https://doi.org/10.1007/s10144-014-0456-z.

16. Pompanon, Francois, Bruce E. Deagle, William O. C. Symondson, David S. Brown, Simon N. Jarman, and Pierre Taberlet. 2012. ‘Who Is Eating What: Diet Assessment Using Next Generation Sequencing’. Molecular Ecology 21 (8): 1931–50. https://doi.org/10.1111/j.1365-294X.2011.05403.x.

17. Rickett, Jennifer, Cody J. Dey, Jillian Stothart, Constance M. O’Connor, James S. Quinn, and Weihong Ji. 2013. ‘The Influence of Supplemental Feeding on Survival, Dispersal and Competition in Translocated Brown Teal, or Pateke (Anas Chlorotis)’. Emu - Austral Ornithology 113 (1): 62–68. https://doi.org/10.1071/MU12053.

18. Schaub, Michael, and Roger Pradel. 2004. ‘Assessing the Relative Importance of Different Sources of Mortality from Recoveries of Marked Animals’. Ecology 85 (4): 930–38. https://doi.org/10.1890/03-0012.

19. Sergio, Fabrizio, Alessandro Tanferna, Julio Blas, Guillermo Blanco, and Fernando Hiraldo. 2019. ‘Reliable Methods for Identifying Animal Deaths in GPS- and Satellite-Tracking Data: Review, Testing, and Calibration’. Journal of Applied Ecology 56 (3): 562–72. https://doi.org/10.1111/1365-2664.13294.

20. Smith, Lora L., David A. Steen, L. Mike Conner, and Jessica C. Rutledge. 2013. ‘Effects of Predator Exclusion on Nest and Hatchling Survival in the Gopher Tortoise’. The Journal of Wildlife Management 77 (2): 352–58. https://doi.org/10.1002/jwmg.449.

21. Tavecchia, Giacomo, Jaume Adrover, Antoni Muñoz Navarro, and Roger Pradel. 2012. ‘Modelling Mortality Causes in Longitudinal Data in the Presence of Tag Loss: Application to Raptor Poisoning and Electrocution’. Journal of Applied Ecology 49 (1): 297–305. https://doi.org/10.1111/j.1365-2664.2011.02074.x.

22. Tinker, M. Tim, Brian B. Hatfield, Michael D. Harris, and Jack A. Ames. 2016. ‘Dramatic Increase in Sea Otter Mortality from White Sharks in California’. Marine Mammal Science 32 (1): 309–26. https://doi.org/10.1111/mms.12261.