Skip to content

How the estimates are made

The dashboard asks one question: what share of cases of a health outcome in a place may be attributable to heat? This page explains, briefly, where that share comes from and how to read it. The detail is in Modeling and Model integration.

What the number is

The attributable share is the fraction of cases that the fitted model links to temperatures above the minimum-risk temperature, for the place and period shown. It describes a population, not individuals, and an association, not proof of cause. The icon array shows the same share as figures out of 100.

The models

The modelling team fits distributed-lag non-linear models (DLNM): each gives an exposure–response curve of risk against daily maximum temperature (Tmax), spread over the time before the outcome.

Outcome Exposure Reference
Low birth weight Mean Tmax of the birth month and the eight months before it; whole pregnancy or one trimester Minimum-risk temperature: 28 °C (Madhya Pradesh), 24.3 °C (Kenya)
Under-five mortality Tmax on the day and the three days before; by age group (neonatal is the main analysis) The area's 1991–2020 mean Tmax

Curves are fitted for a state or a country. Districts, divisions, counties and constituencies apply that curve to their own temperatures; they are not separately fitted models.

Where the temperatures come from

  • Observed: Copernicus ERA5 daily maximum temperature, averaged over each area. Where the model was fitted on a different temperature product, ERA5 is shifted onto the model's scale with a fixed offset per area and month.
  • Forecasted: the Copernicus C3S multi-system seasonal forecast, up to six months ahead. Each system's temperature anomaly is added to the area's ERA5 normal; the figure uses the mean of the systems and the range runs from the coolest to the warmest.
  • Projected: five CMIP6 climate models (about 0.5° grids) under three emission scenarios (SSP1-2.6, SSP3-7.0, SSP5-8.5). A time horizon is the average of its years, compared with the baseline (1981–2010 for low birth weight, 1991–2020 for under-five mortality). Projections are scenario averages, not forecasts of any particular year.

Uncertainty and precision

  • The 95% confidence interval comes from the model's coefficient uncertainty. Shorter projected periods show the spread across climate models instead, which is not a confidence interval.
  • The ratio (odds ratio for low birth weight, risk ratio for under-five) is derived from the share and its interval.
  • The precision badge grades that interval: High (upper/lower ≤ 2.5), Moderate (≤ 5) or Low.
  • A negative share, or one the modelling guides suppress, is shown as a message ("No excess heat risk estimated…") instead of a number, and as "No heat excess" on the map, so it is not read as low risk.
  • Estimates relying on temperatures rarely seen when the model was fitted are flagged.

Limits

  • Area values aggregate coarser climate grids; they do not add spatial detail the data does not have.
  • Missing data are shown as unavailable, never as zero.
  • The estimates support planning. They do not by themselves establish causality, and they are not diagnoses or forecasts for individuals.