Released does not always mean survived
A fish can be returned to the water alive and still die later. Fisheries records may call that fish a live discard or a released animal, while a stock assessment may need an estimate of mortality associated with the capture and release process. These are not competing accounts: the first describes what was seen at the vessel, and the second estimates what happened afterward. Fish discard and release mortality includes immediate death as well as delayed consequences of injury, stress or predation. To read a scientific estimate without changing its meaning, distinguish capture, handling, release condition, subsequent observation and inferred longer-term survival.
The usual numerator is less straightforward than it appears. A study might count fish that died during holding, animals detected by tags as likely dead, or all estimated fishery removals attributable to discarding. Its denominator might be every encountered fish, fish brought aboard, fish released alive, fish tagged successfully, or only fish whose fate could be determined. A mortality rate can be arithmetically correct and still answer the wrong question if the denominator is silently substituted. Readers should always look for the exact population at risk, the time period of follow-up and any excluded individuals.
Release mortality also differs from total fishing mortality. A stock may suffer deaths from retained catch, discards, gear escape and post-release effects. Some assessments incorporate an assumed discard mortality fraction, while others distinguish mortality sources in more detail. The phrase ‘ninety percent release survival’ should not be shortened to ‘ninety percent of caught fish survive’, since retained fish and animals dying before release may not belong to that study denominator. Nor does a survival estimate automatically apply to every gear, season or size class. The scientific job is to describe the conditional quantity the study actually measured.
NOAA Fisheries identifies discard and release mortality as an important research and management issue, including injuries and stress that may produce later death. This means the endpoint matters. A fish active at release may fail to feed, suffer infection or be vulnerable to predation after the observer has left. Conversely a visibly exhausted animal can recover under suitable conditions. A trained observer’s condition score is useful evidence, but it is not the same as measured survival over a defined interval. Translating a condition category as a fate statement removes the very uncertainty the research is designed to investigate.
Follow the fish through a clearly defined fate sequence
A useful fate diagram begins before the release decision. The fish is encountered by gear, captured or retained by the gear, brought aboard or handled in the water, classified by condition and either retained, discarded dead or released alive. The released fish may later survive, die, disappear from observation or be recaptured. Studies vary in which branches they observe. A report limited to live releases has not estimated mortality among all animals that encountered the gear. These distinctions are particularly important when comparing escape devices, hook types or on-deck handling procedures. The reader should reconstruct the fate sequence before using any percentage.
Time zero must be specified. A holding experiment may start counting survival after a fish reaches a recovery tank, while a telemetry study may start at the release point. Fish that die while being transferred can be excluded from one denominator and included in another. Even if both studies report twenty-four-hour mortality, they may not cover the same risk period. A transparent methods section states when observation began, how early deaths were recorded and whether transport, tagging or transfer stress was considered part of the treatment. Without a shared time origin, an apparent comparison of survival rates can mislead.
Some fish cannot be classified because a tag fails, a receiver misses it or an animal leaves the study area. It is tempting to label each disappearance a death, but that confuses observation with biology. The reverse mistake is to count every undetected fish as a survivor. Survival analysis treats incomplete follow-up through explicit censoring or a model of observation, while acknowledging that assumptions may fail when lost individuals differ systematically from those followed. An animal’s last detected position is a measurement, not an automatic death certificate. This simple distinction prevents many false mortality claims.
Units must stay consistent across the chain. A study may report fish counts, individual-level probabilities, mortality over days or mortality per fishing encounter. Fisheries management may later convert these into tonnes of removals using estimated body mass and catch totals. A twenty-percent mortality probability is dimensionless; twenty tonnes of dead discards is a mass estimate. The conversion needs both a meaningful eligible discard quantity and an applicable mortality fraction. A table showing twenty percent cannot, by itself, reveal how many fish or kilograms were lost.
Mechanisms: capture stress, injury and environmental conditions
Capture affects fish through more than one physical route. Exhaustion during struggle, abrasion in nets, hook injuries, pressure changes from depth, thermal exposure, crowding, oxygen limitation and periods out of water can each affect physiological recovery. Which mechanism dominates depends on species, anatomy, capture depth, gear and operating conditions. Some injuries are visible at release; others are not. A fish that swims away demonstrates short-term mobility, not necessarily restored physiology or long-term survival. A careful interpretation describes specific mechanisms without claiming that every captured animal faces the same risk.
Depth-related pressure changes deserve special attention for species susceptible to barotrauma. Gas expansion as pressure decreases can impair buoyancy and produce visible symptoms. The response differs across species and capture profiles, so it is unsafe to generalize from one fishery to another. A descending device may return an affected fish to depth, but whether that lowers long-term mortality requires evidence suitable to the species and conditions. A study of device use must distinguish successfully deploying the tool, restoring depth, short-term observed recovery and longer-term mortality reduction. These endpoints can form a chain without being interchangeable.
Temperature and time are also contextual. A fish brought into air on a hot deck may experience different stress from one released rapidly in cooler conditions. Yet the biological effect of a measured temperature depends on species and exposure duration, not on a universal threshold. Researchers should record water temperature, air exposure, handling sequence, gear soak duration and fish size where feasible. Such variables help explain why a release mortality estimate obtained during one season may not transport to another. When translating results, retain their range of conditions rather than replacing it with a categorical verdict such as ‘safe release’.
Gear design influences which animals are caught and how they are harmed. A selective device may allow some fish to escape before reaching the deck. Those escapees form a different study group from fish landed and then released. Their survival may be affected by contact with the gear, swimming exhaustion or predator exposure, but measurements and assumptions differ. An assessment that combines escape mortality and deck-discard mortality without defining both risks can double-count or omit animals. A strong report clarifies each pathway and keeps the numbers linked to their corresponding mechanisms.
Release condition scores are predictors, not outcomes
Observers often record categories such as vigorous, sluggish, injured or dead at release. These descriptions can support research, but they are not universal biological outcomes. A score may be defined differently by programmes, and trained observers can disagree on borderline cases. A report should provide the scoring rubric, observation timing and any inter-observer calibration. An injury such as bleeding or inversion may be predictive of poor survival in some settings, but the strength of that association must be estimated rather than assumed. The statement ‘poor condition’ is therefore not synonymous with ‘certain mortality’.
Suppose an illustrative sample has fifty fish scored vigorous and fifty scored impaired. Within a monitored follow-up, five vigorous fish and twenty impaired fish die. Conditional observed mortality is 5/50 = 10 percent in the vigorous category and 20/50 = 40 percent in the impaired category. This example suggests the score carries predictive information, but the association is not proof that the score itself caused death. Fish size, capture depth and handling duration may differ between categories. A valid causal claim requires an appropriate design that addresses such confounding.
The overall observed fraction in that example is 25/100 = 25 percent, provided every fish’s fate was adequately observed. If a fishery’s releases are mostly vigorous, its expected fraction under those conditional rates differs from a fishery whose releases are mostly impaired. One cannot transfer a single pooled mortality rate across fleets with different condition distributions without justification. Stratified estimates may be helpful when sample sizes are sufficient. In small studies, however, dividing observations into too many categories creates unstable estimates. The interpretation should balance biological specificity against statistical precision.
Handling improvements are often evaluated through condition scores because long-term monitoring is expensive. An intervention that reduces severe injury is promising, but its effect on delayed mortality should be presented as predicted unless validated by fate data. Changes in score definitions, observer training or composition of sampled fish can create apparent improvements. A well-designed evaluation records intervention exposure, baseline condition, gear, species and environment. It should ask whether the same biological population would have fared differently under the alternative handling procedure, rather than merely comparing unadjusted counts from unrelated trips.
Designing holding, cage and pen studies
One method observes released fish in controlled tanks, cages or sea pens for a specified time. These facilities make direct monitoring easier, yet they can impose additional capture, transport, confinement and water-quality effects. A control group may help identify mortality caused by holding rather than fishing, but controls must be appropriately selected and comparable. A low mortality fraction inside a well-maintained facility does not automatically establish low mortality after release into the wild, where predation, movement and feeding challenges differ. The study’s environment is part of the endpoint definition.
Transport from a fishing vessel to a monitoring pen can exclude fragile individuals that die before arrival or select only fish safe enough to handle. If deaths during transfer are omitted, the resulting survival fraction may look more favourable than the full capture-to-release pathway. Alternatively, poor pen conditions can produce deaths that would not have occurred in the sea. Researchers need to record transfer losses, acclimation procedures, stocking density, oxygen, temperature and disease events. A transparent fate table includes all enrolled animals, not just those present when formal monitoring begins.
Suppose an experiment catches 120 fish, of which ten die before they reach the monitoring facility. The remaining 110 enter holding; twelve die during the following week. Mortality from the original capture count is (10+12)/120 = 18.3 percent if all deaths are attributable within the defined experimental pathway. Mortality among successfully entered holding fish is 12/110 ≈ 10.9 percent. Both arithmetic statements are correct, yet they estimate different quantities. Without a control group one cannot determine how much of the observed mortality was caused by capture versus handling and captivity. State the denominator and attribution caveat.
An ethical and informative study aims to reduce avoidable research harm while still obtaining valid fate information. Welfare practices, permits and animal-care procedures vary by jurisdiction and species. These constraints affect the research design and should not be hidden as unimportant details. If a species is particularly sensitive to confinement, another monitoring method may offer stronger ecological relevance. The general lesson is that an observation method can alter the process it seeks to measure. Good translation preserves the methodological trade-off rather than portraying controlled holding as a direct window into the unobserved ocean.
Acoustic and satellite tags: detection is not survival
Electronic tags can follow released fish without keeping them in cages. Acoustic tags transmit signals detected by receivers when the fish comes within an effective range; satellite-linked tags may supply location or environmental histories under different constraints. Neither records every moment of a fish’s life. Receiver geometry, battery duration, tag shedding, habitat and animal movement influence the chance of detection. When a fish disappears, death is one possible cause among others. A defensible analysis estimates or discusses observation probability and defines which detection patterns are considered diagnostic of mortality.
A tag can change the animal’s behaviour or survival, particularly when it is large relative to body size or implanting it creates additional stress. Tagging impacts should therefore be considered alongside fishing impacts. If the tagged sample excludes the smallest or most injured fish, it may not represent all releases. A study can also lose tags independently of fish fate. A stationary tag detected repeatedly at the seabed might indicate mortality, shedding or predation; ancillary sensors and validation are needed to distinguish them. The strength of the inference depends on the device and study design.
Imagine one hundred tagged fish. After a month, sixty transmit movement patterns interpreted as alive, twenty transmit strong indicators of death and twenty have insufficient follow-up. It is wrong to call survival 60 percent simply because sixty were confirmed alive, or 75 percent because twenty of eighty resolved fates were classified dead. The latter is an estimate conditional on being resolved. If unresolved animals differ systematically in fate, even that conditional fraction can be biased for the full sample. The honest initial report is sixty confirmed alive, twenty classified dead under stated rules and twenty unresolved, followed by model-based estimates with assumptions.
Detection studies can use sentinel tags, range trials, receiver performance checks and alternative movement models to estimate detectability. They can also examine whether fish are likely to leave the array. A tag that disappears near the boundary of a receiver network is a different evidentiary case from a stationary signal repeatedly recorded inside the network. If a study infers mortality from reduced movement, it should test whether resting behaviour, habitat restriction or a predator carrying the tag could produce similar observations. ‘Likely mortality’ should remain distinct from a directly recovered dead animal.
Survival curves, time horizons and censoring
A survival curve describes the fraction of a defined enrolled group expected to remain alive beyond each time, under the analysis’s assumptions. The value at day one is not interchangeable with the value at day thirty. Mortality may be concentrated immediately after release or develop later. A Kaplan–Meier estimator adjusts survival at observed death times by the number still at risk, while handling censored observations under assumptions about their censoring process. It is not merely the number seen alive divided by the initial count. Readers should examine the risk table and follow-up interval before quoting the final percentage.
Consider a deliberately simple example. Ten fish enter monitoring. Two die while all ten are at risk, so the first observed survival step is 8/10 = 0.8. Afterward two surviving fish are lost to follow-up for reasons treated as independent censoring, leaving six at risk. One of these six dies. The next Kaplan–Meier step is 5/6, giving cumulative estimated survival 0.8×5/6 ≈ 0.667 or 66.7 percent. Simply dividing known survivors by the initial ten does not reconstruct that estimate. The validity of 66.7 percent depends on the censoring assumption and accurate event timing.
If loss to follow-up is informative, conventional censoring methods can mislead. For example, badly injured fish might shed tags more frequently, or healthy fish might leave the monitored area faster. Either mechanism connects observation loss to potential fate. Sensitivity analyses can compare plausible interpretations of unresolved animals and indicate whether conclusions remain stable. The most pessimistic scenario treating all unknown fish as dead and the most optimistic treating them as alive are often bounds rather than realistic estimates. Good communication shows where the main estimate lies and what assumptions determine it.
Confidence intervals on survival curves also depend on study size and event distribution. A curve that remains at ninety percent because only ten fish were monitored may carry substantial uncertainty. At later times, few animals may still be under observation, producing wide intervals. A report should not read a long flat curve as proof of durable survival when almost every tag has expired. The number at risk at each time is often as informative as the curve’s line. When translating a figure, preserve the horizon, number at risk, interval type and definition of death events.
Control mortality and attribution of fishing effects
A fish may die during observation for reasons unrelated to capture, including disease, predators or environmental stress. Some experiments use a control group to estimate background death under comparable monitoring conditions. However, subtracting control mortality percentages mechanically is not always statistically valid. Control and captured groups can experience different hazards, and the appropriate correction depends on study design and estimand. A clear report should state whether it estimates all deaths observed after release or mortality attributable to fishing in excess of a defined baseline. These quantities may require different analyses.
An illustrative calculation makes the issue concrete. Suppose 20 of 100 captured fish and 5 of 100 control fish die during the same observation period. Observed risks are 20 percent and 5 percent. Their risk difference is 15 percentage points, a descriptive contrast. It is not a universal release mortality correction that can be applied to every fishery. If controls were healthier or underwent different handling, the fifteen-point difference could reflect bias rather than a pure causal effect. Interpreting it causally requires reasonable exchangeability and comparable follow-up conditions.
Rates can also be expressed as hazard ratios or odds ratios, neither of which is the same as a risk difference. A hazard ratio describes relative instantaneous event rates under its model; it does not say that a given proportion of fish will die. An odds ratio of two is not generally a risk ratio of two when death is common. Translating such results into percentages without a baseline risk creates invented absolute probabilities. The safest public explanation keeps the original measure, then adds a worked absolute-risk illustration only when the required baseline and model assumptions are known.
Experimental randomization can strengthen causal inference, but in practical fishing studies crew, weather, gear and fishing grounds may be clustered. Treating all fish as independent when they came from a few hauls can make uncertainty intervals too narrow. Analyses may include trip or haul effects, or use cluster-resampling. A reader should look for the unit of treatment assignment and the unit of statistical uncertainty. If all treated fish were released from one vessel and all controls from another, vessel differences may explain some of the contrast. Careful writing does not pretend that a complex field design is a simple two-column trial.
Convert release mortality to stock removals without double-counting
Fisheries managers sometimes estimate the portion of discarded catch that ultimately dies, expressed as mass or numbers. A basic accounting example starts with a known quantity of released fish and an applicable mortality fraction. If one thousand fish averaging 0.8 kilograms are released, their total released biomass is 800 kilograms. At an assumed post-release mortality fraction of 0.25, the expected dead-discard biomass is 800×0.25 = 200 kilograms. This calculation assumes the 0.8-kilogram mean is appropriate for the released group and that the quarter mortality rate applies to these fish under their capture conditions.
The same calculation can be performed in numbers: 1,000×0.25 = 250 expected deaths. Multiplying by 0.8 kilograms produces 200 kilograms. Yet using the mean mass of retained fish could bias the estimate if discarded fish are smaller. One should ideally use size-specific weights and mortality rates when data support that detail. A report that says ‘200 dead discards’ without a unit risks confusing fish numbers with kilograms. Some assessments may estimate unobserved discard or escape mortality as a separate component; it should not be added twice to total fishing removals.
Distinguish immediate dead discards from live releases in the bookkeeping. Suppose a hypothetical haul produces 100 fish retained, 20 fish discarded already dead and 80 released alive. If the applicable delayed mortality fraction among live releases is 25 percent, expected additional delayed deaths are 80×0.25 = 20 fish. Total expected deaths associated with retained and discarded outcomes would be 100+20+20 = 140 fish under this simplified count model. Applying 25 percent to all 100 discards and also counting the twenty already dead would double-count some deaths. The model must respect the observed fate categories.
If the fishing mortality estimate is used in stock assessment, uncertainty in discard quantity and mortality fraction should travel into the resulting removals. Assume a mortality fraction plausibly lies between 0.15 and 0.35 for a homogeneous released group of 800 kilograms, with no other error. The corresponding range of expected delayed mortality biomass is 120 to 280 kilograms. This is a simple sensitivity interval, not a statistical confidence interval, because no probability model or parameter dependence was specified. Calling it a 95-percent interval would invent inferential precision.
What meta-analysis can and cannot transfer between fisheries
Researchers may combine discard mortality studies across regions or gears to estimate broader patterns. Such syntheses can identify important risk factors, but the average of heterogeneous studies is not automatically a suitable value for one specific fishery. Species, depth, season, gear, handling time and follow-up method can all alter outcomes. A pooled number should be accompanied by inclusion criteria, weighting, heterogeneity and a description of which settings are represented. If one study uses cage survival and another uses telemetry, their differing observation processes may contribute to between-study variation.
A reader should check whether the synthesis estimates the mean of study effects, a prediction interval for a new setting, or a conditional effect of a predictor. A narrow confidence interval around the pooled mean can coexist with a wide prediction interval when individual fisheries vary substantially. Translating a pooled rate as a guarantee for every vessel is unwarranted. When the new fishery differs markedly from the studied ones, local validation or a conservative sensitivity analysis may be more useful than borrowing the average uncritically.
Selection of studies matters. Published research may focus on commercially important species, places with monitoring resources or gear changes judged likely to improve survival. Failed trials, short follow-up and unobserved mortality can be less visible. A careful review evaluates design quality and missing-data handling rather than simply counting papers. Strong research can still leave unresolved gaps for endangered or poorly studied species. The correct phrase is that available evidence supports a range under specified conditions, not that science has settled a universal release survival percentage.
Not all interventions reduce mortality through the same mechanism. Shortening air exposure, reducing injury, improving gear selectivity or using an appropriate descending device may each act on different stages of the capture pathway. An intervention can decrease the number of fish captured, the number discarded, or the mortality probability among releases. These are separate contributions to total harm. A useful evaluation decomposes them so that one apparent success does not hide an offsetting change elsewhere. A gear with fewer releases might still have more total dead fish if capture or fatal-injury rates increase.
Worked exercise set: denominators and uncertainties
Exercise one: a survey observes 240 fish handled, of which sixty are retained, forty die before release and 140 are released alive. Later evidence supports an estimated twenty percent mortality among live releases. Find expected delayed deaths and the simplified total deaths including retained animals. Answer: delayed deaths = 140×0.20 = 28. Simplified total = 60+40+28 = 128. Do not call 28/240 the mortality rate among released fish; that rate is 28/140 = 20 percent. The expected total fraction is 128/240 ≈ 53.3 percent, under the assumptions that retained fish count as removals and the estimated delayed fraction is applicable.
Exercise two: a study tags eighty live-released fish. Fifty are later detected moving, ten are classified dead under validated signal patterns, and twenty have unknown fates. How many are confirmed survivors, deaths and unresolved? Answer: fifty, ten and twenty. A crude observed death fraction among all tagged fish is 10/80 = 12.5 percent, but this is not a valid estimate of true mortality unless unobserved fates and detection are handled appropriately. Among resolved fish, the fraction labelled dead is 10/60 ≈ 16.7 percent. Both quotients are descriptive and neither automatically equals the stock assessment input.
Exercise three: a model estimates seventy percent survival to day seven and fifty-five percent survival to day thirty. What additional mortality is suggested among those surviving to day seven? Under comparable definitions, conditional survival from day seven to day thirty is 0.55/0.70 ≈ 0.786, so conditional mortality over that later interval is about 21.4 percent. The difference of fifteen percentage points is the additional loss relative to the original cohort, not the conditional risk among day-seven survivors. This example assumes the survival estimates are from the same cohort and coherent analysis, not two unrelated studies.
Exercise four: 500 kilograms of released biomass are assigned a mortality probability of 0.30 with a plausible alternative range 0.20 to 0.45. Compute the point expectation and sensitivity range. Answer: 150 kilograms, with 100 to 225 kilograms under the alternatives. If 50 kilograms of already dead discards are separately documented, add that amount once to obtain 200 kilograms at the point assumption. The range then becomes 150 to 275 kilograms if the dead-discards measurement is treated as exact for teaching purposes. Real analyses should account for measurement uncertainty and possible dependence between quantities.
Common interpretation failures and their repairs
Failure one is turning ‘released alive’ into ‘survived fishing’. The repair is to state that the release condition was observed at a specific time and subsequent survival was not necessarily measured. Failure two is treating lost tags as deaths. The repair is to describe detection coverage, movement outside the array, tag performance and the survival model. Failure three is treating all unknown fates as living. That opposite assumption can be equally misleading. The repair is the same: preserve an unresolved category or apply justified censoring and sensitivity analysis.
Failure four is using a mortality fraction from one fishery as though it were a species constant. The same species can face different injury probabilities in shallow hook-and-line fishing versus deep trawling. The repair is to match estimates to gear, size, depth, temperature, handling and follow-up horizon, then state uncertainty when transportability is weak. Failure five is conflating a predicted mortality model with observed outcomes. The repair names the predictor, validation evidence and prediction interval. Calling a score-based prediction ‘measured deaths’ undermines credibility even if the numerical model is sophisticated.
Failure six is confusing a statistical percentage with a physical quantity. Twenty percent of released animals is a probability or fraction; twenty percent of landed mass is a different quantity. The repair is to label the denominator and units every time. Failure seven is reporting confidence intervals without noting the effective sample size and clustering. The repair is to ask whether hauls, vessels or trips were treated as independent. Failure eight is concluding an intervention works because treatment and control groups differ without examining assignment bias. The repair is to explain the comparison design and remaining confounders.
Finally, avoid making welfare or legal conclusions that a survival study does not test. A fish surviving thirty days after release is not proof of full reproductive recovery, and a mortality estimate alone does not determine whether a fishery complies with conservation law. Regulations, welfare standards and stock-management decisions have their own evidence requirements. A strong explanation shows where the biological evidence is decisive, where it is conditional and where an additional authority must determine the appropriate action. Scientific uncertainty should not be erased for the sake of a simple headline.
Reader questions, transfer and final interpretation
Can a fish swim away and still die? Yes. The question is not whether every released fish dies, but whether apparent immediate recovery reliably predicts survival under the particular conditions. Exhaustion, injury and predation may affect fate after the vessel’s observation ends. The longer the follow-up and the stronger the detection design, the more directly a study can inform delayed outcomes. A report should identify the observation window rather than imply that immediate movement settles the issue for all time.
Why do studies produce very different mortality percentages? They may cover different species, gears, depths, seasons, handling procedures and follow-up durations. Their definitions of the enrolled group and their handling of unknown fates may also differ. Two estimates cannot be ranked as more or less humane without checking what they measure and whether the samples are comparable. A responsible comparison can produce a table of study conditions before it attempts any meta-analysis. Heterogeneity is information about biology and measurement, not merely an inconvenience to average away.
How can release mortality research improve fishery decisions? It can help identify risky capture processes, evaluate changes in gear or handling, prioritize research and improve estimates of fishery removals. It can also show which species and fisheries lack reliable estimates. But useful evidence should retain uncertainty, not become a magic correction factor. A management strategy that uses one point rate should test plausible alternatives, especially where the rate substantially influences stock status or catch advice. In plain language, the goal is better decisions with transparent assumptions.
A final transfer question is useful. A local report says only that ninety fish were released and seven were seen dead on the deck. Can it calculate a robust thirty-day release mortality rate? No. It has documented at least seven observed deaths associated with that handling episode, but it has not followed the fish released alive or defined the population counted in the numerator and denominator for a thirty-day fate estimate. To repair the claim, specify the disposition counts, release condition, follow-up design and uncertainty. The best explanation protects the distinction between a genuine observation and an unsupported extrapolation.
Primary sources and connected guides
For first-party methodological context see NOAA Fisheries’ discard and release mortality science programme and the FAO review of mortality among fish escaping trawl gears. These primary sources support the conceptual distinctions. Every numerical exercise in this article is hypothetical and must not be treated as field data from either source.
For related but different measurements consult eduKateSG’s guides to observer coverage and bycatch expansion, gear selectivity and escape curves, and acoustic telemetry detection limits. Those methods cannot independently substitute for a well-defined post-release survival endpoint.