Seven in ten adults in England saw a heat-health alert this summer. Four in ten of those took no action at all, and the commonest reason was not confusion but irrelevance. A single conversion rate will not locate that loss; a staged one will.
Western Europe has just had its warmest summer on record. England issued heat-health alerts through it, and roughly seven in ten adults encountered one. Of the people who saw an alert, forty-one per cent took no action to protect themselves.
The figures come from a UNU-INWEH policy brief by Khosravi, Assan and Madani, with a supporting survey analysis published in Energy Research & Social Science. Alongside the headline they report that eighteen per cent of people in England hold a high perception of heat risk, that around thirty per cent say they know little or nothing about protective actions, and that seventy-six per cent of the heat-risk practitioners interviewed described informing the public as their dominant communication objective.
Put the first and last of those together and the shape of the problem appears. The practitioners' objective was achieved. Information was delivered to seven people in ten. The objective was the wrong one.
The easy conclusion is that this is a communications failure, which invites a communications response: clearer wording, more channels, better graphics. The survey's own breakdown of why people did not act argues against that reading. Among those who gave reasons, nineteen per cent cited a low perception of personal risk, nine per cent did not recognise the alert, seven per cent were unsure what action to take, and five per cent found the alert difficult to understand.
Only the last two of those are information problems, and together they are the smaller half. The largest single category is people who received the alert, recognised it as an alert, understood what it said, and concluded that it was not about them.
That is a relevance failure, and it is a different engineering problem. A clearer alert does not fix it. A more frequent alert may make it worse.
The reason this distinction is tractable rather than merely interesting is that there is a well-developed model of it. Lindell and Perry's Protective Action Decision Model holds that a warning does not cause a protective action. It enters a decision process alongside environmental and social cues, and that process has stages the warning does not control: exposure, attention and comprehension first, then perceptions of the threat, of the available protective actions, and of the stakeholders issuing the advice, then a search for and assessment of options, and only then implementation. Situational facilitators and impediments sit across the whole sequence, and they are frequently decisive.
The practical consequence for measurement is that the loss between warning and action is not one loss. It is a product of retentions at six or seven junctions, each with its own cause and its own remedy. A single conversion rate, people who acted divided by people reached, is the product of all of them. It will tell you that something is wrong and nothing about where.
Heidenreich and Thieken's application of the model to heat shows how much, and how little, the chain explains. In a survey of 1,417 residents of three medium-sized German cities, the model's constructs accounted for about sixteen per cent of the variance in protection motivation, nineteen per cent in protective response and twenty-three per cent in emotion-focused coping. Temperature, risk communication, gender, age and home ownership were significant contextual predictors; perceived personal vulnerability, response efficacy, response costs, preparedness and perceived external responsibility were significant psychological ones.
Two things follow. The first is a caution: a carefully specified behavioural model, applied with a decent sample, explains under a quarter of the variance in what people actually do. Anyone promising a predictive preparedness index should be asked what they expect to achieve that this did not.
The second is the usable finding. The same study embedded an experiment, and respondents who received a warning that included action recommendations felt significantly better informed and reported more intended specific protective behaviours than those who received a warning without them. The cheapest available improvement is not a clearer hazard statement. It is a named action attached to it.
The predictor in the German data most likely to be skipped is perceived external responsibility. People who believe the authorities are responsible for managing the risk take fewer protective actions themselves.
For a civil protection agency this is genuinely uncomfortable, because it means visible institutional competence can substitute for household preparedness rather than complementing it. A city with excellent cooling centres, a reliable ambulance service and a well-run alert system may record lower private preparedness than a city with none of those things, and on a naive index it would score worse.
That has a direct implication for how Mission Adaptation work is evaluated. A household preparedness score, measured alone, is not a measure of resilience. It is a measure of resilience minus whatever the state is already doing, and the two need to be read together or not at all. We have made a version of this argument before about what a number can bear; this is the sharpest instance of it we have come across.
The new Horizon Europe project AMPLIFY started on 1 October 2026 and runs to 30 September 2029, with an EU contribution of just under €4 million, coordinated by University College Cork under topic HORIZON-CL3-2025-01-DRS-01. Its approach is to build four transdisciplinary Creative Hubs that combine scientific knowledge with storytelling, theatre, music and digital arts, with a transferable hub methodology, a learning platform, an AI-supported reflection tool and an open-access toolkit as outputs.
To an infrastructure engineer that reads like a soft project. Read against the England data it is aimed precisely at the stage where the largest loss occurs. Storytelling is a technology for making a general statement feel personally applicable, and personal applicability is the junction at which nineteen per cent of the English sample stopped. Whether it works is an open question; whether it is pointed at the right junction is not.
The risk is in how it will be judged. Projects of this kind are assessed on outputs produced and people reached, and reach is the cheapest and least informative quantity in the whole chain. If AMPLIFY concludes in 2029 having produced an excellent toolkit and a reach figure, the thing worth knowing will not have been measured.
The obvious metric is a warning-to-action conversion rate: people who took the recommended action divided by people reached. It is a useful headline and a poor instrument. The better version is a cascade with a measured retention at each junction, collected on one base, in one instrument.
Using the England figures to illustrate the shape, not to supply the answer:
Reached. About seventy per cent of adults encountered an alert, so roughly thirty per cent are lost before the chain starts. This is the only stage currently measured well anywhere.
Recognised. Of those reached, nine per cent did not identify what they saw as an alert. Reach and recognition are not the same quantity and are routinely reported as one.
Understood. Five per cent found the alert hard to understand, while around thirty per cent of all respondents reported little or no knowledge of protective actions. Comprehension of the message and knowledge of the response are separate things, and only the first is usually tested.
Judged relevant. Nineteen per cent cited low perception of personal risk. The largest single loss in the chain, and the one for which almost no intervention is routinely evaluated.
Found a feasible action. Seven per cent were unsure what to do, which in Lindell and Perry's terms is a failure of protective action perception rather than of threat perception, and has a different remedy.
Acted. Forty-one per cent of those who saw an alert did nothing.
These percentages rest on different denominators, some on all adults and some on those who saw an alert, so they cannot be multiplied into a cascade. That is not a flaw in the study. It is the finding. Nobody is currently collecting the stages on a common base, which means the sequence cannot be decomposed even where the individual numbers exist. Fixing that requires no new technology and no new theory, only a single instrument asking the same respondents about every junction.
Survey within seventy-two hours, not at the end of the season. Recall of whether a message was seen, recognised and believed decays fast, and the stages collapse into each other in memory.
Recruit for the non-actors. Door to door in a sampled set of streets, not through the alert channel. The thirty per cent who were never reached are the group the design most needs and the sampling frame least provides.
Run the same instrument across three to five communities exposed to the same event. Comparable exposure with different institutional and cultural channels is the only way to separate the effect of the message from the effect of the place.
Ask what the person did instead. Someone who ignored official advice and phoned a neighbour has not failed to act. They have used a channel the evaluation did not count, and our interview guide sets out how to surface that without leading them to it.
For consortia writing into Mission Adaptation and Cluster 3, the practical conclusion is to stop promising composite readiness indices and commit instead to measuring one stage-specific loss properly. For civil protection authorities, it is to put a verb in the alert. For funders, it is to notice that reach is the number most often delivered and the one that answers the least.
If this reads close to a decision you are weighing, we would like to hear about it. A short note is enough to start.
The Lab is an independent research team, not an advocacy organisation. We do not campaign, endorse products, or accept commissions where a finding is expected in advance. Where a partner's own theory of change is what needs testing, that is a legitimate engagement, and the Lab tests it honestly.