Health impact

What the /health page calculates

For an area you choose — a local authority, or a shape you draw on the map — Argus estimates the health benefit that would follow if long-term average air pollution there fell to the WHO 2021 guideline levels, and holds that comparison over twenty years.

The output is a comparison between two futures for the same population: one where exposure stays at today’s modelled level, and one where it meets the guideline. Everything reported — deaths prevented, cases of illness avoided, quality-adjusted life years, monetised value — is the difference between those two.

The model

The engine is asclepius, an implementation of the Air Quality Lifecourse Assessment Tool (AQ-LAT) framework: a Markov cohort model that ages a population through annual cycles, applying disease incidence and mortality rates that depend on the pollution the cohort is exposed to.

Reference: Hall, J., Zhong, J., Jowett, S., et al. (2024). Regional impact assessment of air quality improvement: the air quality lifecourse assessment tool (AQ-LAT) for the West Midlands combined authority (WMCA) area. Environmental Pollution 356: 123871. https://doi.org/10.1016/j.envpol.2024.123871

An earlier version of the toolkit workbook (v3, March 2023) is deposited under CC-BY 4.0 as Air Quality Life Assessment Tool (Hall, Jowett, Zhong, Mazzeo, Thomas & Bartington, 2023; https://doi.org/10.25500/edata.bham.00000935). The parameter set Argus reproduces is v4.1 (September 2023), distributed separately by the University of Birmingham’s WM-Air project (https://wm-air.org.uk/project/health/). The two differ in the PM₂.₅ mortality coefficient, so figures reproduced from the deposited v3 workbook will not match ours.

Concentration-response functions

The link between exposure and health outcome is a set of relative risks per 10 µg/m³ of long-term exposure. Argus carries these. Where an outcome has coefficients for both pollutants, the toolkit’s anti-double-counting rule applies only the larger-magnitude of the two, so the pair is not additive:

References:

• Committee on the Medical Effects of Air Pollutants (2023). Summary of COMEAP recommendations for the quantification of health effects associated with air pollutants, Table 1. UK Health Security Agency, updated 8 September 2023. https://www.gov.uk/government/publications/air-pollutants-quantification-of-associated-health-effects (Open Government Licence). The UK’s expert committee. The PM₂.₅ mortality coefficient itself comes from COMEAP (2022), Statement on quantifying mortality associated with long-term exposure to PM2.5, which recommends RR 1.08 (95% CI 1.06, 1.09) per 10 µg/m³.
• Chen, J. & Hoek, G. (2020). Long-term exposure to PM and all-cause and cause-specific mortality: A systematic review and meta-analysis. Environment International 143: 105974. https://doi.org/10.1016/j.envint.2020.105974 — the meta-analysis behind the 1.08 mortality coefficient.
• Cesaroni, G., Forastiere, F., Stafoggia, M., et al. (2014). Long term exposure to ambient air pollution and incidence of acute coronary events: prospective cohort study and meta-analysis in 11 European cohorts from the ESCAPE Project. BMJ 348: f7412. https://doi.org/10.1136/bmj.f7412
• Scheers, H., Jacobs, L., Casas, L., Nemery, B. & Nawrot, T.S. (2015). Long-term exposure to particulate matter air pollution is a risk factor for stroke. Stroke 46(11): 3058–3066. https://doi.org/10.1161/STROKEAHA.115.009913
• Hamra, G.B., Guha, N., Cohen, A., et al. (2014). Outdoor particulate matter exposure and lung cancer: a systematic review and meta-analysis. Environmental Health Perspectives 122(9): 906–911; and Hamra, G.B., Laden, F., Cohen, A.J., et al. (2015). Lung cancer and exposure to nitrogen dioxide and traffic: a systematic review and meta-analysis. Environmental Health Perspectives 123(11): 1107–1112.
• Jacquemin, B., Siroux, V., Sanchez, M., et al. (2015). Ambient air pollution and adult asthma incidence in six European cohorts (ESCAPE). Environmental Health Perspectives 123(6): 613–621. https://doi.org/10.1289/ehp.1408206
• Khreis, H., Kelly, C., Tate, J., Parslow, R., Lucas, K. & Nieuwenhuijsen, M. (2017). Exposure to traffic-related air pollution and risk of development of childhood asthma: a systematic review and meta-analysis. Environment International 100: 1–31. https://doi.org/10.1016/j.envint.2016.11.012

Where the coefficients come from is now documented. The primary sources above are stated in Hall et al. 2024’s supplementary appendices (Appendix E lists each function; Appendix F gives the derivations), which we obtained and cross-checked in August 2026. Two findings from that cross-check are worth knowing before quoting a number. First, for coronary heart disease the toolkit’s appendix documents the coefficient as 1.28 per 10 µg/m³ — Cesaroni’s hazard ratio of 1.13 per 5 µg/m³, correctly compounded — but the toolkit implements 1.26, consistent with the “13% per 5” result having been scaled linearly rather than compounded; the value 1.26 appears nowhere in the cited paper. Argus reproduces the implemented 1.26, which slightly understates the CHD burden relative to the toolkit’s own cited source. Second, the appendix lists stroke as “1.09” where both the toolkit and its cited source (Scheers) agree on 1.13 — there the appendix, not the toolkit, is in error. Both findings are being put to the toolkit’s authors. Note also that asthma and lung cancer sit outside COMEAP’s endorsed set — COMEAP makes no quantification recommendation for either outcome — so those functions rest on the cited meta-analyses rather than UK committee advice, and Cesaroni’s CHD estimate carries a confidence interval that includes no effect (0.98–1.30).

All functions are applied log-linearly with no lower threshold. For PM₂.₅ and mortality this follows COMEAP’s current position: its 2 October 2025 statement finds “a lack of evidence of a lower exposure threshold for the adverse health effects of PM2.5”, does not recommend any change from the assumption of a linear concentration-response function, and continues to recommend reducing PM₂.₅ “even where exposures are already low” (https://www.gov.uk/government/publications/comeap-shape-of-the-concentration-response-curve-linking-pm25-with-all-cause-mortality). COMEAP has issued no equivalent statement on the other outcomes here; the same functional form is applied to them by the toolkit’s convention, not by COMEAP recommendation.

Population and health data

The cohort is assembled from Office for National Statistics and NHS sources, at the vintages the model reports with every result:

Disease and mortality rates are available at local-authority level. When you select a smaller area — a drawn shape, a group of neighbourhoods — the model applies that authority’s rates to the people who live inside your selection. Variation below local-authority level therefore comes from age structure and exposure, not from local differences in underlying health. This is the caveat shown with every result, and it is the main reason to treat small-area figures as comparative rather than absolute.

Exposure

Exposure is the modelled annual-mean background concentration at the population-weighted centroid of each area — the point that represents where people in that area actually live, rather than its geographic middle. For a drawn selection, each neighbourhood (LSOA) is sampled at its own centroid and the results are population-weighted. Neighbourhoods where the baseline model has no coverage are excluded from the calculation and reported as skipped alongside the result; the population figure counts only the neighbourhoods included.

Concentrations come from the Pan baseline model. Its accuracy — including errors beside busy roads that are much larger than the headline average — applies to these figures too.

Population-weighted centroids are published by ONS (https://geoportal.statistics.gov.uk/), used under the Open Government Licence. Contains OS data © Crown copyright and database right 2021.

Time, money and discounting

Results run over a 20-year horizon. Mortality benefits are phased in rather than applied immediately, following the workbook’s own five-step ramp (cumulative 0.3, 0.475, 0.65, 0.825, 0.825): roughly a third of the effect in the first year, with the full effect from year six. COMEAP’s recommended cessation lag — 30% in year 1, 50% across years 2 to 5, and the remaining 20% across years 6 to 20 — is available as a sensitivity.

Costs and quality-adjusted life years are both discounted at 3.5%, following the toolkit. That matches the HM Treasury Green Book’s standard appraisal rate for costs; the Green Book itself recommends a lower 1.5% rate for health effects such as QALYs (https://www.gov.uk/government/publications/the-green-book-appraisal-and-evaluation-in-central-government), so discounting both at 3.5% is the toolkit’s choice, reproduced here — it makes the reported QALY figure smaller, not larger.

A QALY is valued at £20,000 — the toolkit’s default, at the lower (conservative) end of the £20,000–£30,000 per-QALY range conventionally used in NICE appraisal. Cost savings cover NHS primary, secondary and prescribing costs, social care, and lost productivity.

The monetised total follows the toolkit’s own convention: it values the undiscounted QALY gain at £20,000 and adds the discounted cost savings. That asymmetry is the published workbook’s, reproduced so our results match it; the QALY count reported alongside is the discounted figure. The “show the working” panel on the /health page displays both QALY figures so the difference is visible rather than buried.

Verification

Reproduction of the published toolkit is tested continuously:

• Cell-level equivalence. The engine reproduces the AQ-LAT v4.1 workbook’s own cached run to floating-point tolerance — the cohort and mortality path end to end from a ward name and an exposure alone, and each disease outcome’s attribution arithmetic cell for cell; asthma, coronary heart disease and lung cancer case counts are also re-derived end to end the same way.
• Published-result equivalence. It reproduces Hall et al. 2024’s West Midlands regional totals across 192 wards, to within the tolerance the published figures support. The paper’s per-authority breakdown reproduces only if one authority’s mortality multiplier is applied to every ward — evidence about the published procedure, which we reproduce faithfully rather than treat as validated.
• End-to-end reproduction in Argus. In July 2026 every numeric field the API serves was independently re-derived and matched to floating-point precision, and the geometry that decides which neighbourhoods a drawn shape selects was checked against an independent geometric oracle (shapely ground truth) on a committed suite of 45 adversarial shapes and around 2,000 points, replayed continuously through the shipped selection code.

Reproducing a toolkit faithfully is a statement about arithmetic, not an endorsement of every parameter in it. That is what independent review is for; a review plan is in place and reviewers are being approached.

Last reviewed 2026-08-10

Background concentration estimates

Monitors are sparse. To say anything about a place without one — which is most places — Argus uses a modelled background surface called Pan.

What it estimates

Pan estimates the long-term annual mean background concentration of NO₂ and PM₂.₅ at a point, covering the UK and Ireland. “Background” means the general level of pollution in an area, not the sharp local peak beside a specific busy road. “Annual mean” means it says nothing about today, this week, or a particular hour.

It is available through the public API at /api/public/baseline, and it supplies the exposure figures the health-impact model uses.

Method and inputs

Pan is a layered lookup surface: it combines published modelled concentration layers with its own land-use regression adjustment, using the finest available at each point.

Each response reports which layer supplied the value, so you can tell a 20 m London estimate from a 1 km national one. Note that the 200 m near-road adjustment exists for NO₂ only: PM₂.₅ outside London is served at the 1 km PCM resolution.

Accuracy

Pan’s accuracy is measured, not asserted. It is validated against Argus’s own holdings of reference-network annual means for 2024 (AURN, LAQN, AQE, SAQN, WAQN and Northern Ireland monitors, each requiring at least 75% data capture for the year — as held by Argus, so partly provisional pending ratification, see monitoring data):

That average hides the important pattern. Errors are strongly site-dependent, and are largest exactly where pollution is highest:

A background surface is not trying to reproduce a kerbside canyon, so large near-road errors are expected rather than a defect — but they are not in one reliable direction. In this validation Pan reads low on average at urban-traffic monitors (mean bias −7 µg/m³) and high on average at roadside and kerbside monitors (+17 µg/m³). Treat any figure derived from Pan at a busy-road location as indicative only, and read health-impact figures for road-adjacent populations as uncertain rather than conservative.

For PM₂.₅ the same pattern is far weaker: 0.5 µg/m³ at rural background sites rising to 2.7 µg/m³ at roadside, on a much lower absolute scale.

Uncertainty intervals

A prediction interval is published for PM₂.₅ and withheld for NO₂.

The interval method (split-conformal prediction, calibrated on the same monitor comparison) widens the band according to the predicted concentration. For PM₂.₅ that behaves acceptably: per-quartile coverage runs 86–100% against a 90% target. For NO₂ it does not — because Pan’s NO₂ errors depend on what kind of place you are asking about, not on how high the prediction is. At locations the model predicts to be clean the band under-covers badly (77% against the 90% target), which is precisely the failure the diagnostic exposes: some points predicted clean are in reality polluted. Publishing the band would overstate confidence exactly where it is least warranted. The mean absolute error is published instead, and the interval is withheld until a site-type-aware method is available.

The calibration figures above come from the committed calibration artifact; the headline mean absolute errors and the calibration cohort are also reported by /api/public/baseline/about.

Not included

Pan is a static baseline. It carries no hour-to-hour, day-to-day or seasonal variation, and it does not respond to weather.

The hourly map surface

The concentration overlay on the homepage map and on /concentration is a separate modelled product: the Pan annual baseline scaled hour by hour using live measurements from the monitoring networks. Its own methods page is still to be written; until then, treat the overlay as indicative context rather than a citable estimate.

Last reviewed 2026-07-31

Monitoring data

Every reading Argus displays comes from an existing monitoring network. Argus operates no instruments of its own; it collects, stores and presents what the networks publish.

Networks

The first six are reference networks: instruments meeting the regulatory standard for compliance monitoring. Breathe London is a low-cost sensor network — valuable for spatial density, but different in character. LMAM sites are also reference-grade instruments, run by councils and published through Defra.

Only the six reference networks feed the comparisons used to validate the baseline model; both Breathe London and the locally-managed council networks are excluded from that validation.

Data are retrieved through Defra and network publication feeds, and for bulk historical periods through the openair-format archive files the networks publish per site and year.

Reference: Carslaw, D. C. & Ropkins, K. (2012). openair — an R package for air quality data analysis. Environmental Modelling & Software 27–28: 52–61. https://doi.org/10.1016/j.envsoft.2011.09.008 — the package and data conventions this archive format comes from.

Cadence and freshness

Most networks are polled every 15 minutes. Welsh and Northern Irish data are refreshed roughly daily upstream, so Argus polls those every four hours — more frequent checking would return the same data.

A gap in a chart usually means the network published nothing for that period. Instruments go offline for calibration, servicing and faults, and some feeds drop out intermittently.

Ratification

Readings arrive provisional and are ratified later — typically some months afterwards — when the operator has completed calibration checks and quality assurance. Ratification can change or remove values. Recent data should be read as indicative; older data is more likely to be settled. Argus does not currently re-pull ratified values automatically, so figures covering recent months may differ from the same period viewed later on the network’s own site.

Capture rates

Where Argus computes an average over a period, it applies a minimum data capture requirement so that a monitor which reported for a handful of hours does not appear alongside one that reported continuously. The annual means used to validate the baseline model require at least 75% capture. Thresholds vary by view — the homepage’s “most improved” comparison, for example, requires roughly 500 readings across its 30-day windows, a proxy for about 80% capture.

Units and averaging

Concentrations are in micrograms per cubic metre (µg/m³) — except carbon monoxide, reported in milligrams per cubic metre (mg/m³) — at hourly resolution, aggregated to daily, weekly or annual means as the view requires. Where Argus compares against a guideline, it compares like with like: a 24-hour guideline against daily means, an annual guideline against annual means. See WHO guidelines.

Last reviewed 2026-07-31

WHO guidelines and how Argus compares against them

Argus compares measured and modelled concentrations against the World Health Organization’s 2021 global air quality guidelines. These are health-based recommendations, not UK statutory limits — the legal limits under the Air Quality Standards Regulations 2010 are higher for several pollutants. Comparing against the guideline answers “how healthy is this air?”, not “is this lawful?”.

Reference: World Health Organization (2021). WHO global air quality guidelines: particulate matter (PM₂.₅ and PM₁₀), ozone, nitrogen dioxide, sulfur dioxide and carbon monoxide. https://iris.who.int/handle/10665/345329

Guideline values used

Comparisons are like-for-like: a daily mean is compared against a 24-hour guideline, an annual mean against an annual guideline. Argus does not compare an hourly reading against a 24-hour guideline.

The WHO also defines a peak-season ozone guideline (60 µg/m³, averaged over the six months with the highest ozone), which Argus does not use.

Classification bands

Where Argus labels a value rather than just plotting it, it uses the ratio of the value to the relevant guideline:

These bands are Argus’s own presentational convention, chosen so that a reader can see at a glance whether a place is close to the guideline or a long way from it. The WHO does not define them, and they should not be read as WHO-endorsed categories. The underlying ratio is always available in the data.

Interim targets

The WHO also publishes interim targets — staged milestones for places a long way above the guideline. Argus does not currently display them. The guideline is the destination; the interim targets are a path to it, and they may be a more useful comparison for some areas.

In the health model

The health-impact model uses the 2021 annual guidelines for PM₂.₅ (5 µg/m³) and NO₂ (10 µg/m³) as the target scenario: its figures answer “what would be gained if this area’s long-term average met the guideline?”.

Last reviewed 2026-07-31

Summary of key assumptions

Every headline parameter on this page, in one place.

Licensing and attribution

• Monitoring data: Crown copyright, used under the Open Government Licence; London Air Quality Network data © Imperial College London.
• Breathe London data: used under the network’s published terms.
• ONS geography and population data: Office for National Statistics licensed under the Open Government Licence. Contains OS data © Crown copyright and database right 2021.
• NHS England data (QOF disease prevalence; lung cancer incidence): Open Government Licence.
• AQ-LAT toolkit (v3 deposit) and Hall et al. 2024: CC-BY 4.0.
• COMEAP publications: Open Government Licence.
• LAEI 2022: Greater London Authority, Open Government Licence v3.0.
• Modelled-baseline inputs (via Pan): Defra Pollution Climate Mapping, the National Atmospheric Emissions Inventory, and DfT traffic count data — Open Government Licence v3.0; ONS Census 2021 and NRS Census 2022 — Open Government Licence v3.0; Microsoft Global ML Building Footprints — CDLA-Permissive-2.0; OpenStreetMap contributors — Open Database Licence.
• Generated using Copernicus Atmosphere Monitoring Service information (2021, 2023). Neither the European Commission nor ECMWF is responsible for any use of that information.

Last reviewed 2026-08-10