Clownfish Shrink to Survive Heatwaves, Actually Shorter, Not Just Thinner

A Fish That Got Shorter

Melissa Versteeg swam back to the same anemone she had visited a month earlier, held a pair of callipers against the same fish, read off a number, and found it smaller than the number in her notebook.

Her first reaction was not excitement. She felt the drop you feel when you think you have made a mistake. She told reporters afterwards that she had "a bit of a panic," got other people to measure the same fish independently, and kept re-measuring until she believed it [15]. The fish really had lost length. So had another. So had ninety-eight more.

Bony vertebrates do not do this. Growth in fish is indeterminate. It carries on through adulthood rather than stopping at a fixed adult size, and biologists have always treated that carrying-on as one-directional. A fish can slow down. A fish can stall. In a famine a fish gets thin, loses condition and burns through its own muscle, while the skeleton underneath it stays exactly the length it always was. The skeleton does not move.

In Kimbe Bay, Papua New Guinea, between February and August 2023, the skeletons moved.

One hundred fish out of 134 got shorter. Not thinner. Shorter.

The setting matters. 2023 opened the fourth global coral bleaching event. Water around the study anemones ran hot, up to 4 °C above the long-term average for that patch of reef [1]. Temperature loggers sat right at the anemones, recording every two minutes, which is unusually fine-grained for a reef. Most reef thermal data comes from satellites averaging over a kilometre or more. Here the temperature record belongs to the animals being measured.

And here is the part the club kept circling. Fish that shrank were likelier to be alive at the end. A single shrinking event carried a survival probability up to 78 percent higher than a fish that never shrank, and every fish that shrank more than once made it through [1]. Then came the detail that turned a physiology result into something else. Among breeding pairs, the ones where both partners shrank survived better than the pairs where only one shrank, or neither.

Growth Was Supposed to Have One Direction

Be precise about the surprise here. "Animal gets smaller" is not by itself news. Fish populations in warm water have been shrinking for decades. The pattern has a name, the temperature-size rule, and a large and quarrelsome literature attached to it [17, 18]. But that literature is about averages. It compares a cohort raised at 26 °C with a cohort raised at 30 °C, or a population in 1980 with the same population in 2020, and finds the warmer one matures earlier and tops out smaller. Nothing in it requires any individual animal to run backwards. Selective mortality of large fish produces the same population-level curve. So does faster maturation. So does slower growth. The individual can be perfectly monotonic while the average slides downward.

What Versteeg and colleagues have differs in kind. They measured the same named animal again and again, with its own temperature logger, and showed that this specific fish, the one in that anemone, was 68 millimetres in March and something under 67 in April.

Precedents exist. They are wonderful and there are not many of them.

In January 2000, Martin Wikelski and Corinna Thom published two pages in Nature reporting that Galápagos marine iguanas shrink during El Niño events [2]. Their animals lost up to 20 percent of body length, nearly seven centimetres on a big male, and grew it back afterwards when the algae returned. The iguanas that shrank most survived longest, and people remember where they were when they read it. Bone was being resorbed and rebuilt, in an adult, because the food ran out.

The other case is smaller and stranger. Common shrews go into winter with a skull measurably shorter than the one they had in summer, the braincase losing something like 15 percent of its height, and then partly regrow it in spring [3]. Zoologists call this Dehnel's phenomenon, after the Polish worker who first described it, and it happens to individuals, verified by X-raying the same shrew repeatedly. A mammal reversibly shrinking its own skull sounds like a transcription error. The shrews do it anyway.

So the category exists. What did not exist before this paper was a case in a ray-finned fish, in the wild, tracked individually, with a survival benefit attached and a social partner in the picture. Add the scale of the thing. A clown anemonefish shrinking by one percent loses well under a millimetre. You could see the marine iguana result from across a lava field. This one needed somebody to visit the same anemone six times with a calliper and refuse to blink.

Callipers, Underwater, Three Times Each

Here is how you do it. Worth knowing, because the method is the whole argument.

You pick a bay. Kimbe Bay, 5.16° south, 150.50° east. You find anemones with clownfish in them. You map them. Each fish gets identified by its own bar pattern, because the white bands on a clownfish are as individual as a fingerprint and do not change over a season.

Then you go back. Every four to six days, for 180 days. Six full measurement rounds across five lunar months.

You measure total length with callipers, underwater, to the nearest half millimetre. You do it three times on the same fish. If any one of your three readings sits more than a millimetre off the other two, you throw all three away and start that fish again [1].

Sit with that last rule. The typical shrink being reported runs one to two percent of total length. The average dominant female measures 67.50 millimetres. One and a half percent of that is a hair over one millimetre. The re-measurement threshold and the effect size are almost the same number.

No gotcha here. The protocol looks the way it does for exactly that reason. Three readings averaged, with a rejection rule, drives the standard error on the mean of the three well below the standard error of any single reading, and the signal is defined by a consistent direction across 640 observations rather than by any one fish on any one day. Still, Figure 1 puts the two scales side by side. Judge for yourself.

The Pair

Start with how these animals live. The arrangement is one of the odder ones in the sea, and without it the tandem result means nothing.

An anemone holds a group. At the top, a single large female. Below her, a single breeding male. Below him, a queue of non-breeders who do not breed at all and are, functionally, waiting. Everyone is ranked by size, and the ranking is not approximate. Peter Buston showed twenty years ago that each fish in the queue sits at a fairly precise size ratio to the fish above it, close to 0.80, and that the subordinate holds the ratio by modifying its own growth to stay small enough [4, 6].

Why bother? Because the fish above you will evict you if you get too close in size, and eviction from an anemone on a coral reef is a death sentence [5]. Nowhere to go. The predators outside are immediate and numerous. So the subordinate grows slowly on purpose. It stays visibly smaller and keeps its place in the anemone. Buston and Cant later called this a regulated size hierarchy, a queue held together by a credible threat and a voluntary restraint [6, 7].

The sex system sits on top. Clown anemonefish are protandrous hermaphrodites, which means they hatch male and the largest one becomes female. If the female dies, the breeding male changes sex and takes her place, and the next fish in the queue moves up. The pair at the top is monogamous and typically stays together for years [8]. Some pairs have held their anemone longer than the researchers have been diving on it.

Which is the background against which the tandem finding lands, and it lands beautifully.

If shrinking is a stress response, you would expect it to be individual. Fish gets hot. Fish shrinks. But the paper found that when both members of a breeding pair shrank, survival probabilities ran higher than when only one of them did (P = 0.020), and separately that a dominant female was less likely to shrink when her partner was already close to her in size [1]. Read those two together and a picture assembles itself. Shrinking endangers a subordinate when the dominant does not shrink too, because the size gap closes and the eviction threat sharpens. A dominant shrinking alone runs the mirror risk. The safe move is to do it together.

We are wary of over-reading a P-value of 0.020 in 67 pairs, and we say so properly in §10. But the shape of the finding is the interesting part. A physiological response whose payoff depends on what your partner is doing at the same moment is a different sort of object from a physiological response. Something has to be negotiated, or at least coordinated. Nobody knows how.

The Oxygen Arithmetic

From here to the end of §9, everything is ours. The paper measured fish. We did arithmetic about fish. The arithmetic is deliberately simple, so a reader can check it.

Two quantities move when seawater warms. They move in opposite directions.

First, how much oxygen the water can hold. Gas solubility falls as temperature rises. We used the Garcia and Gordon refit of the Benson and Krause tables, the standard equation for oxygen in seawater [15, 16], at salinity 35 and one atmosphere. Our pre-heatwave reference is 28.5 °C. Saturated seawater there holds 195.31 µmol of oxygen per kilogram, or 6.393 mg per litre. At 32.5 °C it holds 183.53 µmol per kilogram, or 6.008 mg per litre. A fall of 6.03 percent.

Second, how much oxygen the fish wants. Metabolic rate in an ectotherm rises with temperature. We used \(Q_{10} = 2.0\), so demand doubles per 10 °C. A conventional value for teleost resting metabolism [12]. Over the same 4 °C, demand rises 31.95 percent.

Divide one by the other. Hold body size fixed. Supply per unit demand falls to 0.7121 of its starting value. A squeeze of 28.79 percent.

Write it out:

\(R\) is the supply-to-demand ratio, normalised to 1.0 at a 67.50 mm fish in 28.5 °C water. \(L\) is total length. \(d_g\) is the exponent by which gill surface area scales with body mass, and that input will cause us trouble in §6 and §10. The exponent 0.75 is the standard metabolic scaling exponent. Mass goes as \(L^3\) under constant shape, which is why both exponents get multiplied by three.

Figure 2 plots those three quantities against temperature. No adjectives required.

What 1.4 Percent Buys You

Now put the fish back in.

The size term in that equation carries the exponent \(3(d_g - 0.75)\). If gill area scales with mass more slowly than metabolism does, so \(d_g < 0.75\), the exponent goes negative and a shorter fish gets a better ratio. Daniel Pauly has argued this for over four decades under the name gill-oxygen limitation: gills are two-dimensional surfaces trying to feed a three-dimensional body, the mismatch worsens as the animal grows, and warming brings the ceiling down [9]. Versteeg and colleagues point to that mechanism in their discussion, carefully, as a hypothesis rather than a finding [1].

We set \(d_g = 0.70\) as our baseline. A choice, and a generous one. It sits at the gill-limitation-friendly end of the published range. Shrinking helps at all only under that assumption. Then we asked a question with a clean answer. How much warming does a real shrink cancel?

A 1.4 percent reduction in total length changes body mass by 4.14 percent. The supply-to-demand ratio improves by 0.212 percent. Convert that gain back into degrees, on the same solubility and \(Q_{10}\) curves. You get 0.0248 °C.

Twenty-five thousandths of a degree. Against an anomaly of four degrees.

We checked this four ways because none of us believed it. The arithmetic holds. Push it further. Cancel the full 4 °C in this accounting. A 67.50 mm fish would have to reach 7.02 mm. A reduction of 89.60 percent. Not a shrinking fish. A different animal, or a larva.

So the ratio argument, taken on its own terms and given its best exponent, fails by two orders of magnitude to explain why shrinking would help.

But a second way of keeping the books gives a very different number. Instead of asking about supply per unit demand, ask about total demand. A smaller fish needs less oxygen and less food in absolute terms, because its metabolic rate scales as \(M^{0.75}\), and that reduction is not a ratio effect at all. A 1.4 percent shrink cuts absolute metabolic demand by 3.12 percent. Expressed as degrees of warming cancelled, 0.4577 °C, eighteen times the ratio answer. It claws back 11.44 percent of the heatwave's energetic cost instead of 0.62 percent.

Neither number is large. One of them at least reaches the order of magnitude that matters to an animal running a deficit with eight more weeks to last.

Bench Notes from the Table

Started on the whiteboard with surface area to volume. Smaller thing, more surface per unit of interior, easier to get oxygen in. Everybody meets this idea in a biology lesson and it feels obviously right.

Obviously right, and also the wrong quantity. A fish does not take up oxygen through its skin. It takes it up through gills. Gill surface area is not proportional to body surface area. The whole question is what exponent it follows. Our surface-to-volume table stays in the output file, because honest bookkeeping keeps the dead ends. At the rank-1 mean the body's SA/V is 0.35672 per millimetre. A 1.4 percent shrink lifts that by 1.42 percent. Then we deleted the conclusion we had written under it.

0.0248 °C.

Three of us re-derived it separately. One in Python. One in a spreadsheet. One on paper, with a table of logarithms. He wanted to see every step. One more in the interactive model after we built it. Same answer four times. Not a bug. What the exponents say.

A genuinely bad twenty minutes followed. We had come in expecting to write that the physics checks out. It did not. The temptation to reach for a different exponent was strong. Openly discussed, and voted down.

Why divide? The fish does not care about a ratio. It cares whether it can pay its bills this week.

That reframing is the whole of §6's second half. Worth the afternoon. Absolute demand falls as \(M^{0.75}\). The shrink is worth 0.4577 °C in that currency. Suddenly the magnitudes are at least in conversation. Note carefully that this rescues nothing about gill limitation. A different story happens to use the same shrink.

Ran \(d_g\) from 0.60 to 0.90. We wanted to know how sensitive we were. The answer: totally. At 0.75 the size term vanishes from the ratio entirely. Above 0.75, shrinking makes the ratio worse, and the same arithmetic that gave us +0.0248 °C at \(d_g = 0.70\) gives −0.0248 °C at \(d_g = 0.80\).

So we cannot tell you the sign of our own result. Not without a number nobody has measured for this species. We put it in Figure 3B rather than in a footnote.

Straw poll on the question "is the gill-limitation explanation the main thing going on here?" Two yes, nine no, four abstained on the grounds that the club has no business voting on fish physiology. The abstentions are correct and we recorded them anyway.

The Ledger

Every number in this article, in one place, with its source. The tag column says who is responsible for it. "Measured" means Versteeg and colleagues measured it. "Ours" means it came out of our script. "Assumed" means we picked it and you are entitled to disagree.

Line I4 decides whether §6 has a positive sign or a negative one. Lines R1 to R10 all inherit it. Line R11 inherits nothing from it. The survival model is a separate exercise with separate faults.

Survival Odds, Rebuilt from Scratch

Start with the Cox model. The paper's 78 percent comes out of one. It handles time-to-event data and copes gracefully with fish still alive when the study stopped. We do not have their data and could not fit that model if we did. So we built a cruder thing and asked what it could see.

We simulated 3,000 breeding pairs, 6,000 fish. Each one shrank or did not, at the rank-specific rates the paper reports. 71 percent of rank 1, 79 percent of rank 2. Shrinkers got a magnitude from a lognormal distribution averaging 1.4 percent of total length. Survival came out of a logistic model carrying an intercept, a slope on shrink magnitude and a bonus for sitting in a pair where both partners shrank. We then solved for the magnitude slope by bisection, stopping at an odds ratio of exactly 1.78, the paper's figure. Outcomes were drawn once. Then we fitted the model back with iteratively reweighted least squares, in numpy, no statistics package.

The fit recovers an odds ratio of 1.7310, with a 95 percent confidence interval running 1.5121 to 1.9816. That sits 2.75 percent below the value we planted, ordinary sampling noise on 6,000 draws, and exactly the gap you should expect between a truth and a single estimate of it.

The dose-response is in Figure 4. A fish that does not shrink has a fitted survival probability of 71.59 percent. Shrink 1.4 percent and that becomes 76.14 percent. Do it in tandem with your partner and it becomes 83.52 percent.

Now the tandem test, the one we actually cared about. Restricting to shrinkers only, so the comparison is tandem against solo with magnitude held in the model, the tandem coefficient is +0.47676 with a standard error of 0.08359. An odds ratio of 1.6109. A Wald \(z\) of 5.70. And \(P = 1.2 \times 10^{-8}\). Raw survival rates now, no model at all. Non-shrinkers 71.79 percent. Solo shrinkers 75.77 percent. Tandem shrinkers 83.42 percent. Tandem beats solo by 7.65 percentage points.

The interactive bench runs the same experiment in your browser. Its own generator, seeded the same way. On its first run at the default settings it returns an odds ratio of 1.6088, with a 95 percent interval of 1.4037 to 1.8438, against an expected value of 1.7666 computed in closed form from the same coefficients. Press the run button a few times and watch it wander. The wandering is the honest width of a single experiment, and it is why nobody should read an odds ratio to four decimal places, ourselves included.

Read that \(P\)-value correctly. We built the tandem effect into the generator. We then detected the tandem effect. The \(P\)-value states something about our sample size and nothing else. Small, because 6,000 is a large number. What the exercise shows is that the effect the paper reports at P = 0.020 in 67 pairs has a shape a logistic model can recover cleanly when given enough fish. None of that makes the effect real.

One last thing came out of this exercise and it changed how we read the paper. We asked what a study of 134 fish could have detected. Observed mortality of 8 percent, and a plain survived-or-not endpoint. We ran 4,000 such studies. The 95 percent confidence interval excluded 1 in 15.28 percent of them. A study that size, with that few deaths and a binary outcome, would miss a true odds ratio of 1.78 roughly six times out of seven.

Not a criticism of the design. The opposite. It explains why the design looks the way it does: repeated measures, every four to six days, per-fish temperature, and time-to-event models that use every day a fish stayed alive rather than a single yes-or-no at the end. The information had to come from somewhere.

The Best Case Against It

Suppose you are the unconvinced referee. Your job is to describe a world in which every measurement in the paper is correct and the headline conclusion is still wrong. Here is the strongest version we can build, and we think it is genuinely strong.

Objection one: the mechanism is missing, and the authors say so. The paper contains no measurement of how a fish shortens. No bone histology, no vertebral counts, no imaging, no resorption markers. The authors state it plainly: their study "includes no information on the mechanism of shrinking," and testing the physiological hypotheses will require laboratory work crossing heat stress with food availability [1]. Without a mechanism, "shrinking" names a pattern in a column of numbers. The cause could be vertebral compression, cartilage loss at the joints, resorption of bone, a change in posture under stress, or some combination behaving differently from any of those. Every downstream story, including the entire physiological argument in our §5 and §6, rests on an unexamined step.

Objection two: the gill-oxygen limitation theory is contested and may simply be wrong. Our whole ratio calculation assumes \(d_g < 0.75\). That assumption sits inside an argument which has run for years and shows no sign of resolving. Lefevre, McKenzie and Nilsson have argued at length that gill-limitation models are not built on valid physiological principles, that gills are plastic structures which acclimate to demand rather than sitting as a fixed ceiling, and that projections built on them are unreliable [10, 11]. Scheuffele and colleagues went after the central prediction empirically and found that gill surface area scales roughly in proportion to metabolic rate, exactly the condition under which body size drops out of the ratio [13]. Then Lonthair and colleagues ran the long experiment, linking growth, gill area and metabolism in brook trout across a temperature gradient, and concluded that smaller body size under warming is not caused by gill-oxygen limitation [14]. Pauly has replied to the critics in detail and the theory is not dead [9]. But an honest reading puts the mechanism our §6 leans on on one side of a live dispute, with the newer experiments sitting on the other.

Objection three: shrinking may be a symptom that correlates with surviving rather than a cause of it. The study watched wild animals in the wild, and nobody assigned any fish to shrink. Consider a fish in good condition with fat reserves, a well-placed anemone and a cooperative partner. Such a fish may be able to afford a controlled reduction in body length, and may also be likelier to survive a heatwave, with no causal arrow running between the two. Or the arrow points backwards: a fish that was going to die may stop shrinking in its final weeks, because whatever process drives shrinking costs energy the dying fish no longer has. Both stories produce a positive association between shrinking and survival. The Cox models cannot separate them, and nothing in the design can.

Objection four: the effect sizes sit uncomfortably close to the measurement floor. A one percent shrink on a typical breeding male is roughly half a millimetre, which is one calliper tick. Averaging three readings does real work against random error, and the direction of the effect across 640 observations is the actual evidence. But systematic error does not average away. If handling stress, or the posture a fish adopts when a diver approaches for the fourth month running, produces a small consistent bias in the same direction, no amount of replication removes it. We are not asserting that this happened, only that the study cannot rule it out, and that the honest place to settle it is a laboratory with imaging rather than a reef with callipers.

Objection five, which is ours and applies to us. Our own model has a sign depending on a number nobody has measured for this species. Move \(d_g\) from 0.70 to 0.80, a shift well inside the published spread, and our 0.0248 °C becomes −0.0248 °C, meaning shrinking now hurts. We cannot resolve that, and no amount of Monte Carlo will resolve it, because the uncertainty lives in the choice of model rather than in the noise around it.

Now the answers, because a steelman left standing is just a hedge.

Objection one is correct and is the most important thing in this article, though it describes a research programme rather than a refutation. The pattern is real whatever causes it, and the pattern is the finding. Objection two weakens our §6 considerably and weakens the paper's discussion a little, but it barely touches the observation, because the paper does not need the gill-limitation story to be true for the fish to have got shorter. Objection three is the serious one and will be answered only by a controlled experiment, which is what the authors themselves say is needed. Objection four is answered partly by the protocol and partly by time: if the effect is an artefact of divers and callipers, another team in another bay will fail to find it. Objection five is ours to carry, and it is why Figure 3 has two panels instead of one.

What remains is an observation that looks solid, a survival association that looks real without being causally established, and a mechanism that is wide open.

What We Think It Is

Here is where the club ended up, after six sessions and one bad afternoon.

The oxygen story, in the narrow form where a smaller body gets a better ratio of gill supply to metabolic demand, does not carry the weight people want to put on it. Our arithmetic says a real shrink is worth about two hundredths of a degree in that currency, and that the sign of even that depends on an exponent nobody has pinned down. If you came here expecting the physics to explain the biology, it does not.

The energy story does better. A fish 4 percent lighter needs about 3 percent less oxygen and about 3 percent less food, every day, for as long as the heat lasts. In a five-month event, with the water running hot enough to bleach the coral around you and the plankton you eat thinning out, 3 percent a day compounds into something that might genuinely decide whether you are alive in August. Not an exciting mechanism. Just arithmetic. Arithmetic is often what survival turns out to be.

The social story is the one we find hardest to put down.

Here is a pair of fish that have shared one anemone for years, held in a hierarchy maintained by the credible threat that the larger one will throw the smaller one out. Both get hot. Both run an energy deficit. And the thing that predicts whether they are both alive at the end is whether they got smaller together, keeping the gap between them intact while the absolute numbers slid downward. Neither gives up a place. Neither closes the distance. They both step back.

We do not know how that is coordinated. Nobody does. It could be nothing more than two animals in the same anemone meeting the same temperature and the same food shortage and responding the same way, with no coordination at all, which would leave the survival correlation intact and would make a considerably duller explanation. We hope somebody checks.

What we can say is what we found when we sat down with the numbers and tried to make them behave. We expected a clean physiological explanation and got one that fails by a factor of a hundred. We expected the social result to be a nice detail, and it turned out to be the part that survived scrutiny best. A fish can be a millimetre shorter in April than it was in March. Somebody had to swim out there with a calliper, six times, and be willing to write the smaller number down.

Everything from §5 onward is the club's own simplified model, not the paper's analysis. Versteeg and colleagues did the diving and the measuring, and built the survival models. We did arithmetic a school student can check, and we have shown you which input breaks it.

References