How researchers decide a heatwave killed people, and why the method is changing

Every summer now produces a version of the same headline: a heatwave hits, and within a few weeks a study says climate change was responsible for some share of the deaths. The number varies, the method is rarely explained, and readers are left to take it on trust.

A guidance document published this month, produced by a large expert panel, sets out what those studies should be doing — and identifies a widely used shortcut it says can be misleading [s1].

Where the guidance came from

Detection and attribution analyses have underpinned conclusions since the mid-1990s that human activities have warmed the atmosphere, ocean and land [s2]. The methods were extended first to extreme events, then to sectoral impacts including economic losses, food security, wildfires, agriculture and health [s2].

Health has lagged. The panel notes that existing health attribution studies remain limited in both the health impacts covered and their geographical scope, so that current understanding is partial — biased towards exposures where climate scientists have already done the detection work, and towards regions with abundant long-term health and weather data [s2]. In practice, most health attribution studies have concentrated on problems where the anthropogenic signal dominates the climate noise and where the health outcome is strongly associated with a climate variable, heat-mortality being the archetype [s2].

The guidance was built in three steps [s2]. An organising team and scientific advisory committee selected 65 scientists through a systematic literature review and snowball sampling, with criteria covering expertise, geography, gender and career stage [s2]. A hybrid workshop convened in London in September 2024, sponsored by the Wellcome Trust and co-convened with the University of Washington, with 35 participants in person and 20 remote [s2]. A core writing team then drafted the guidance, which all authors agreed [s2].

The panel is explicit that the framework is meant to promote transparency about analytic choices rather than impose a rigid approach [s2].

The two families of method

Attribution studies use one of two broad approaches, which the guidance describes as storyline and probabilistic [s2].

The storyline approach breaks an event into causal processes — roughly, thermodynamic and dynamic components — and assesses how much anthropogenic climate change contributed to one or more of them [s2]. A common implementation conditions the analysis on a given atmospheric pressure pattern and then asks what warmer oceans or higher carbon dioxide did within that pattern [s2]. The simplification is useful; the guidance notes it often neglects changes in dynamical processes, which anthropogenic warming can substantially affect [s2].

The probabilistic approach estimates the change in an event's frequency or intensity with and without human influence [s2]. An observed outcome, and the "factual" range from historical simulations including all forcings, is compared with a "counterfactual" range from natural forcings alone [s2]. Counterfactuals are generated from historical-natural simulations — most commonly from the Detection and Attribution Model Intercomparison Project — or from statistical regression-based counterfactuals of observations and reanalyses [s2].

Confidence differs sharply by hazard. The guidance notes substantial and well-supported climate-change influence on heatwaves, backed by strong theoretical understanding of the processes involved, but much less certainty for severe convective storms because of limits on how accurately models represent precipitation [s2].

The method the panel warns about

The clearest practical recommendation concerns the fraction of attributable risk, or FAR.

FAR quantifies the portion of an event's probability attributable to climate change, and studies then apply that fraction to the observed health impact. The guidance identifies three problems with doing so [s2].

First, it does not account for changes in the probability of the health impact itself as the hazard changes magnitude [s2]. Second, it is extremely sensitive to the spatial and temporal scales chosen [s2]. Third, and most fundamentally, it represents the health impact as binary — something that either occurred or did not — when the magnitude of many health impacts increases progressively with the magnitude of the hazard [s2]. Above the optimal temperature for a location, higher temperatures are normally associated with higher mortality [s2].

The guidance gives a clean statement of when FAR is and is not appropriate. If the question concerns a threshold — how climate change changed the likelihood of crossing a temperature that produces 1,000 extra heat deaths and strains public health resources — a FAR analysis is informative [s2]. If the question concerns the mortality of a specific event, it may be unintentionally misleading [s2]. Event attribution studies are accordingly gravitating towards approaches that assess changes in event intensity and the associated changes in the magnitude of impacts [s2].

Underlying all of this is a warning against a tempting piece of arithmetic. Because the relationships between weather variables and impacts are often non-linear, it cannot generally be assumed that the proportion of climate-change-attributable health impacts equals the proportion of an event's likelihood attributable to human influence [s2]. Non-linearity also enters through other drivers, through adaptation such as heatwave early warning systems, and through compounding and cascading risks — a loss of housing or income can set off a chain of further impacts [s2].

What a good study should do

The guidance's recommended steps run from framing to reporting [s1]. Co-develop the research question across disciplines, since it determines every subsequent choice — who is involved, which methods, which data, which causal links, how results are communicated [s2]. Build a transdisciplinary analytic team with foundational grounding in the core disciplines [s1]. Engage stakeholders and decision-makers in defining the study design, including how the exposure event or trend is defined [s1]. Identify, visualise and describe the links in the causal pathway from climate variable to health outcome [s1]. Choose appropriate counterfactual climate data and, where applicable, evaluate the skill of the climate and health models used [s1]. Quantify the attributable change in climate variables, then the attributable health impacts in the context of other determinants of exposure and vulnerability, and report how the recommendations were incorporated into the analytic plan [s1]. An annex provides a checklist [s2].

The step most often skipped is the one about other determinants. The guidance asks whether the status of the health system affected the magnitude and pattern of outcomes, and what underlying vulnerabilities or protections modified the impact [s2]. It also allows that "all else being equal" studies remain legitimate and informative in some settings — notably climate lawsuits, where the relevant question may be how impacts would have differed absent the defendants' emissions [s2].

Why the panel thinks this matters beyond method

The framing throughout is that incomplete understanding and documentation of climate change's health impacts impedes decision-making and legal responses [s2]. Better attribution evidence could feed climate risk assessments, provide leverage for adaptation policy in health and agriculture, inform litigation, and inform the programme of the Loss and Damage Fund established under the UN climate convention at COP27 [s2].

That is a candid statement of purpose, and it is also why methodological discipline matters here more than in most fields. Studies have already quantified climate impacts attributable to the emissions of individual entities such as companies or countries, using reduced-complexity earth system models or statistical approaches apportioning contributions to an entity's share of historical emissions [s2].

The honest limit

None of this makes attribution a measurement. Attribution studies estimate what would have happened in a world that does not exist, using models of that world, and the answer moves with the choice of counterfactual. What the guidance offers is not certainty but auditability: a study that follows it will make its analytic choices visible, so that a reader — or a court — can see what the number depends on.

The panel designed the framework so future iterations can incorporate new scientific developments and diverse knowledge systems [s2]. Its intended audience includes journal editors and research funders as well as researchers [s2], which is the practical mechanism by which guidance of this kind changes what gets published.

Sources

• [s1] The attribution of human health outcomes to climate change: transdisciplinary practical guidance — Climatic Change, published online 23 July 2025. https://doi.org/10.1007/s10584-025-03976-7
• [s2] Same paper, open-access full text — Europe PMC, PMC7618133. https://europepmc.org/article/PMC/PMC7618133