Extreme event attribution

How can anyone say whether a warming climate caused one particular flood?

A river bursts its banks. Within days someone asks whether climate change caused it, and the reply used to be a well-worn refusal: no single event can be attributed to climate change, only long-term trends. That answer was honest and it was also unsatisfying, because the question people are asking is a reasonable one.

Since roughly 2004 scientists have been answering it anyway — with numbers, sometimes within a week of the event. So what changed? Not the ability to trace causation in a single case. What changed was the question being answered.

Consider a smaller case first. A heavy smoker dies of lung cancer. Did smoking cause that death? Strictly, nobody can say — non-smokers get lung cancer too, and there is no marker on the tumour. Yet we are entirely comfortable saying smoking raised his risk many times over. The causal claim that can be supported is about probability, not about the individual case.

Attribution science makes exactly this move. It does not ask whether warming caused the flood. It asks: how much more likely, or how much more intense, is a flood of this severity in the world we have than in the world we would have had without human influence?

Now, how could you possibly observe the second world? You cannot — and this is where the inference genuinely rests on models, which is worth stating plainly rather than glossing. The standard approach runs a very large ensemble of simulations under today's conditions, and another under counterfactual conditions with human forcings removed and greenhouse gases at pre-industrial levels. Count how often an event of the defined severity occurs in each. The ratio of those two probabilities is the answer.

Notice the phrase "the defined severity". Before anything is computed, the event has to be reduced to a measurable quantity — three-day rainfall over a specified catchment, or the annual maximum daily temperature at a station. That definition is a choice, and different reasonable choices give different answers. It is fixed in advance precisely to keep the result from being tuned.

Then comes the gate that most outsiders miss. The models must first be shown capable of representing this class of event at all — does the ensemble reproduce the observed distribution of such events in the historical period? If it does not, the honest output is that no attribution statement can be made. That gate is why heatwave attributions are typically strong and confident while attributions for droughts, wildfires and tropical cyclone rainfall are more cautious or sometimes absent: heat is well resolved by climate models, and convective rainfall and fire weather are not resolved nearly as well.

What comes out is never a yes or a no. It is a probability ratio with an uncertainty range. The 2003 European heatwave study by Stott, Stone and Allen concluded human influence had very likely at least doubled the risk. The 2021 Pacific Northwest heat dome was assessed as virtually impossible without human-caused warming. And some events come back with no detectable signal at all — a genuine result, not a failure, and one that is published.

One caveat worth carrying: there is a live methodological debate between this probability-based framing and a "storyline" framing that instead asks how a given event unfolded differently because of warmer, moister conditions, without attempting a probability at all. The two answer different questions and both are defensible; treating either as the definitive method is overreach.

Three refinements.

First, "caused" is the wrong verb throughout, and no attribution study uses it. The output is a change in likelihood or intensity. When a headline says climate change caused a flood, it has translated a probability ratio into a causal claim the study did not make.

Second, an inconclusive result is informative. If the probability ratio's uncertainty range spans one, that says the available evidence does not separate the two worlds — possibly because the signal is small, possibly because the models are not up to the event. Reporting it as "climate change did not cause this" inverts the meaning.

Third, models are not the only route. Where long, homogeneous observational records exist, one can fit a statistical distribution of extremes whose parameters shift with global mean temperature, and read the change in return period directly from observations. The strongest studies do both and check the two approaches agree. Where they disagree, the confidence statement is lowered rather than a winner picked.

• Why heat attributions are confident and drought or wildfire attributions are not.
• How rapid attribution within days differs from a peer-reviewed study a year later.
• Whether attribution results carry weight in climate litigation and loss-and-damage claims.

Every term the collection defines is gathered in the glossary.