The Gold Standard of Ecology | Known-Fate Models, Collar Failure and Informative Censoring

1. Define the state and the monitoring interval

Back to contents

For every animal and interval, specify whether it was known alive at the start, known alive at the end, known dead during the interval, not yet entered or no longer observable. Weekly, monthly and seasonal intervals answer different questions and can handle event timing differently. Do not let software defaults choose the biological time scale.

Known fate does not mean continuous omniscience. It means the analysis treats fate as observed for the intervals included. The field protocol must justify that treatment through search frequency, mortality sensors, recovery work and rules for ambiguous detections.

2. See the likelihood behind the table

Back to contents

In a simple interval, animals alive at the start either survive with probability S or die with probability 1 minus S. Staggered entry allows newly marked animals to join later risk sets. An individual contributes only while under observation. Covariates can model variation by season, age, sex, treatment or environment.

The simplicity is valuable, but it can hide assumptions. Individuals must be correctly identified, survival events correctly classified, monitoring sufficiently reliable and censoring conditionally independent of survival. Dependence among animals or packs and uncertain death times may require richer models or robust uncertainty estimates.

3. Distinguish death, failure, tag loss and emigration

Back to contents

A silent collar may mean the animal died where signals cannot be received, left the search area, shed the collar, experienced antenna damage or exhausted its battery. These outcomes imply different biological histories. Coding every silence as alive, dead or censored can each be wrong.

Create explicit fate categories at data entry, retain last reliable detection and document search effort. Bench-test tags, deploy double-marking where justified, use drop-off information and search beyond the core area. Fate assignment should be a reproducible evidence rule, not a convenient final spreadsheet edit.

4. Understand non-informative censoring

Back to contents

Censoring is non-informative when, conditional on modelled information, the time an animal becomes unobservable does not carry additional information about its survival. Administrative study end is often close to this ideal. A collar that fails on a preset battery schedule may be manageable if battery life is independent of animal risk and correctly recorded.

Informative censoring occurs when loss is related to fate: trauma damages the collar and kills the animal; poaching removes transmitters; animals in high-risk habitat have poor reception; weak individuals lose collars after weight change. Standard known-fate estimates can then be biased because the missing histories differ systematically from retained ones.

5. Design telemetry reliability before release

Back to contents

Record manufacturer, firmware, battery specification, duty cycle, attachment, mortality-switch logic and expected end of service. Bench-test a sample under field-like conditions and keep deployment dates. Establish a search escalation protocol for missed fixes, including aerial, ground and neighbouring-network checks where feasible.

Design can separate processes. Paired tags or visible marks can estimate tag retention; stationary reference tags reveal receiver outages; scheduled drop-offs distinguish device recovery from animal recovery. These additions cost resources, so simulate which failure rates materially change the survival question before deployment.

6. Worked example: monthly survival with disappearing collars

Back to contents

In an invented three-month study, 20 animals enter alive. During month one, one confirmed death occurs, giving an interval estimate of 19 divided by 20, or 0.95. Two collars then go silent and are censored. In month two, two of the remaining 18 observable animals die, giving 16 divided by 18, about 0.889. No deaths occur in month three. The product is about 0.844.

Now suppose later searches show both silent collars were damaged during fatal snaring events. Censoring removed deaths, so 0.844 was too high. If both batteries instead failed at their independently logged service limit while the animals were later seen alive, censoring is more defensible. The arithmetic did not change; the observation mechanism did.

7. Model failure or bound its consequences

Back to contents

When tag failure information exists, combine the survival process with a failure-time model, use multistate or hidden-state approaches, or conduct inverse-probability weighting under justified assumptions. Competing risks may separate natural death, harvest and tag-related loss. Not every dataset identifies all these processes.

When identification is weak, use sensitivity analysis. Recalculate survival if a stated fraction of unknown fates were deaths, survivors or emigrants. Plot estimates across plausible failure hazards and search effectiveness. Bounds may be more honest than a precise joint model supported by too little information.

8. Changed-context transfer check: acoustic fish tags

Back to contents

Move from GPS collars on mammals to acoustic tags in fish. A missed detection may result from receiver downtime, environmental noise, migration through an unsampled corridor, tag expulsion or death outside the array. Dense detections do not guarantee known fate once the fish leaves coverage.

The same logic transfers: distinguish biological survival from device function and observation. Reference tags, receiver diagnostics, gate arrays and external networks become the reliability tools. Censoring at array exit is defensible only for an estimand and model that acknowledge movement and detection.

9. Know the limits and common misconceptions

Back to contents

Known-fate models do not automatically estimate population survival. Collared animals may differ from unmarked animals, capture and attachment can affect survival, and study-area departure can masquerade as loss. More frequent GPS fixes improve movement information but may shorten battery life. A mortality sensor indicates immobility, not cause of death.

Censoring is not a neutral deletion command. It asserts that the unobserved remainder can be represented by observed risk sets after conditioning. Model selection cannot test that assumption directly when fate is absent. Field evidence and sensitivity analysis carry much of the burden.

10. Audit animals, devices, searches and decisions

Back to contents

Publish inclusion rules, entry dates, interval definitions, last contacts, confirmed deaths, uncertain fates, censoring dates, search escalation, tag recovery and device diagnostics. Retain the uncoded evidence behind each final history. Report exposure time and numbers at risk by interval, not only model coefficients.

Provide code, candidate models, covariate definitions and goodness-of-fit or predictive checks. Compare results under alternative censoring dates and plausible reclassification. If collar manufacturer or batch predicts disappearance, investigate it directly. A reproducible survival estimate begins with a reproducible fate table.

11. Frequently asked questions

Back to contents

Is a missing signal a censored animal? Only after a documented rule and search process. Can battery failure be non-informative? Yes, if timing is independent of survival after conditioning and is accurately known. Should lost animals be treated as alive? No automatic rule is safe. Does known fate mean detection probability is one? The model commonly assumes fate is known for included intervals; field design must support that.

Can sensitivity analysis fix bias? It cannot reveal truth, but it shows which conclusions depend on unverified fate assumptions. Are collars harmless? Not necessarily; attachment effects should be monitored and reported.

12. The gold standard models the missing mechanism

Back to contents

Survival inference is strongest when animal fate, device function and observation are separate parts of the study. Known-fate models provide a transparent framework for staggered monitoring, but their clarity disappears when silent collars are coded without evidence.

The gold standard designs reliability checks before release, documents every fate decision and tests how informative loss would change conclusions. When death and device failure cannot be distinguished, the result should retain that uncertainty. Honest survival estimates are built as much from searches and hardware records as from the final likelihood.