In July 2026 the RKI reported roughly 5,120 heat-related deaths for the year to late June, the large majority concentrated in the late-June heatwave (one weekly figure around 4,300).
Later updates raised the running total: by mid-to-late July estimates reached ~9,800–11,900 for the year so far, already higher than any full-year figure in the prior decade (previous peaks ~8,400 in 2018).
Most of these deaths are among people aged 75+, especially 85+, and more women than men simply because of the age/sex structure of the very elderly population.
These figures are produced by comparing observed weekly deaths (from Destatis) against a modelled baseline of expected deaths under cooler conditions (typically temperatures ≤20 °C), adjusted for other factors.
The method is the same excess-mortality approach used for heat, cold, influenza, and other periods of elevated death rates.
It does not mean 5,000+ people dropped dead solely from heatstroke with no other pathology; it attributes the excess above the expected number to the heat period.
Critics correctly note that this includes “harvesting” (deaths brought forward by days or weeks among the already frail) and that confounding or model assumptions can affect the exact total.
Proponents note that pure death-certificate coding systematically under-counts heat contributions because heat often acts as a trigger for cardiovascular, respiratory, or other events.
The numbers Quaschning cites are statistical estimates of heat-attributable excess mortality from Germany’s Robert Koch Institute (RKI), not counts of people whose death certificates list “heat” as the sole or primary cause.
Quaschning: It’s insane how many people have already died from heat this year – and how many ignore or deny it. What would happen if 12,000 people were murdered? Where are the crisis teams, the special broadcasts, and the public outcry? Do we have the right priorities?
Quaschning’s rhetorical framing—comparing the excess to plane crashes, bus crashes, or the early COVID wave, and declaring “I accuse, for non-action”—is advocacy, not neutral epidemiology.
Excess-mortality numbers are real and large; the causal language, moral urgency, and direct policy accusations are his interpretive layer.
The parallel you draw with COVID is apt in one respect: both involve statistical excess deaths whose precise attribution (“with” vs “because of”) is contested and depends on methodology, baselines, and comorbidities.
Heat is a well-documented physiological stressor that raises mortality, especially in populations with limited air-conditioning, older housing stock, and a high share of elderly residents—conditions that describe much of Germany.
Adaptation measures (cooling centres, better building standards, early-warning systems, AC in care homes) demonstrably reduce heat mortality; Germany’s penetration of mechanical cooling in nursing homes remains low.
Attribution of any given heatwave’s intensity to anthropogenic climate change is a separate climate-science question that uses different methods (event attribution).
The mortality statistics themselves do not require that debate; they simply record that hot periods kill more people than cooler ones of comparable duration.
In short: the RKI-derived excess figures exist and are large; Quaschning’s dramatic packaging and political conclusions are opinion layered on top of them. Both the statistical reality and the rhetorical amplification can be examined without treating either as sacred.
Excess mortality is the difference between the number of deaths actually observed in a defined period and the number that would have been expected under “normal” (baseline) conditions.
In formula form:
It can also be expressed as a percentage (P-score):
This measure captures the total impact of a crisis (heatwave, pandemic, influenza season, etc.) without relying solely on cause-of-death coding, which is often incomplete or delayed.
Core Steps in the Methodology
Obtain observed all-cause mortality data
Weekly or daily death counts from national statistical offices (e.g., Destatis in Germany). These are usually all-cause figures because cause-specific data lag significantly and heat (or other stressors) is rarely listed as the underlying cause on death certificates.
Estimate the expected (baseline) number of deaths
This is the key modelling step. Common approaches range from simple to sophisticated:
- Simple historical average: Average of the same weeks/months over the previous 3–5 years. Easy but ignores population growth, ageing, and long-term mortality trends.
- Trend + seasonality models: Most modern methods use regression. Deaths are typically modelled as following a Poisson (or quasi-Poisson/negative binomial) distribution to handle count data and overdispersion.
Typical components include:- Long-term trend (linear or smooth spline) to capture ageing populations and improving/worsening overall mortality.
- Seasonal cycle (sine/cosine terms or cyclic splines) for regular winter/summer patterns.
- Day-of-week, holiday, or other calendar effects.
- Sometimes population size as an offset so the model works with rates rather than raw counts.
Examples:
- EuroMOMO: Generalised linear Poisson model fitted mainly on spring and autumn weeks (to avoid influenza/heat bias), with a linear trend and optional sine seasonality. Uses the previous ~5 years (often excluding pandemic years).
- RKI heat-related mortality (Germany): Generalised Additive Model (GAM) that includes long-term trends, seasonality, and an exposure–response curve relating weekly mean temperature to mortality. It estimates a “background” mortality under the counterfactual that weekly mean temperatures never exceeded a threshold (typically ~20 °C). The difference between the full model (with actual temperatures) and this background is the heat-attributable excess. Lagged temperature effects (up to 3 weeks) are often included.
- Other common variants: quasi-Poisson regression with age/sex/geography strata (e.g., newer ONS methods), Serfling-style trigonometric models, or more flexible GAM/spline approaches.
Calculate the difference and assess uncertainty
Subtract expected from observed. Prediction intervals or confidence intervals are usually produced from the model variance. Statistical significance is sometimes judged by z-scores (how many standard deviations the excess is above baseline).
Attribution refinements (especially for heat)
Pure all-cause excess during a hot period is sometimes further refined by modelling the temperature–mortality relationship explicitly. Only the portion statistically linked to temperatures above a threshold is labelled “heat-related.” This is what the RKI does for its weekly heat-mortality reports.
Important Limitations and Caveats
Baseline choice matters a lot. Different reference periods or model specifications can change the excess by tens of percent. Including or excluding prior high-mortality years (e.g., COVID waves) affects results.
Harvesting / mortality displacement: Some excess deaths are deaths brought forward by days or weeks among frail people who would have died soon anyway. This can reduce the net long-term impact.
Confounding: Concurrent events (air pollution, influenza, changes in healthcare access) can inflate or deflate the attributed excess.
Population structure: Ageing societies produce more deaths even at constant age-specific rates; good models adjust for this via rates or age-stratified analysis.
Not the same as cause-specific counts: Excess mortality does not identify individual deaths caused by the event; it is a population-level statistical estimate.
Reporting delays: Real-time estimates often require corrections for incomplete registration.
In the context of heatwaves, the RKI-style approach is an example of attributable excess mortality: it uses the general excess framework but adds a temperature–response function so that only the portion linked to heat is counted.
The resulting figures (e.g., several thousand for a strong June heatwave in Germany) are therefore model-based estimates, not tallies of death certificates that list “heat” as the cause.
Different reasonable models can produce somewhat different point estimates, which is why uncertainty intervals are routinely reported.
You must be logged in to post a comment.