Research and analysis

COMEAP: shape of the concentration-response curve linking PM2.5 with all-cause mortality

Published 2 October 2025

Applies to England

Summary

This document updates the Committee on the Medical Effects of Air Pollutants (COMEAP)’s views on the shape of the concentration-response function (CRF) that associates mortality with long-term exposure to fine particulate air pollution (PM_2.5). CRFs represent the relationship between a pollutant and an adverse effect on health. They are used to quantify the public health burden of air pollution and to predict the health benefits of reductions in air pollutant concentrations. The CRF linking mortality with PM_2.5 concentrations is important in mortality burden estimates and cost benefit analyses of air pollution policies.

Some studies have suggested that the CRF might not be linear, and there has been particular interest in the possibility that it might be supra-linear – that there might be more adverse health effect per unit change in concentration (for example, per microgram per cubic meter (µg/m³)) at lower levels than at higher exposures. However, as it was not clear to what extent these results may be due to the methods used or characteristics of the populations studied, COMEAP previously considered the evidence to be insufficient to recommend any change from the assumption of a linear^{[footnote 1]} CRF when quantifying the mortality burden attributable to long-term exposure to PM_2.5 or the beneficial mortality impacts of reducing PM_2.5 concentrations (COMEAP, 2022).

In this updated statement, we consider developments in the evidence from studies funded by the Health Effects Institute (HEI) examining associations between mortality and long-term exposures to low levels of PM_2.5. Taking into account the recent analyses and associated limitations, our position remains unchanged: at this time, we do not recommend any change from the assumption of a linear CRF. In our view, the evidence currently available does not support an alternative recommendation. We also continue to recommend that sensitivity analyses are undertaken using the 95% confidence intervals around the CRF.

We continue to recommend efforts to reduce concentrations of PM_2.5, even where exposures are already low and concentration-based air quality standards or targets are met. This is because there remains a lack of evidence of a lower exposure threshold for the adverse health effects of PM_2.5.

We recommend that further research is undertaken to increase understanding of the potential sources of heterogeneity in the size and shape of CRFs reported in epidemiological studies for this, and other, pollutant-outcome pairs. Future research could also explore the influence of using different shapes of CRF in health impact assessments when quantifying mortality attributable to PM_2.5.

Background 

Levels of ambient air pollutants have declined significantly over the last few decades in Europe, North America, and other developed regions. Nonetheless, recent epidemiological studies continue to demonstrate associations between exposure to ambient air pollution and adverse health effects. Notably, 3 large cohort studies funded by the Health Effects Institute (HEI) examining long-term exposures have reported associations with mortality even at low levels of PM_2.5:

• in Canada – ‘Mortality-air pollution associations in low exposure environments – MAPLE’ (Brauer and others, 2022)
• the United States – ‘Assessing adverse health effects of long-term exposure to low levels of ambient air pollution – Medicare’ (Dominici and others, 2022)
• Europe – ‘ELAPSE: an analysis of European cohorts effects of low-level air pollution: a study in Europe’ (Brunekreef and others, 2021)

A proportion of each of the cohorts in these studies experienced low levels of air pollution, defined as levels below the annual average air quality standards that were current at the time in the US (12 µg/m³), Canada (8.8 µg/m³) and Europe (25 µg/m³) (Boogaard and others, 2024).

Some studies have suggested that the CRF might be supra-linear: that is, the effect, per unit change in concentration, at lower levels might be greater than at higher exposures. Of these HEI-funded studies, both MAPLE and ELAPSE reported supra-linear patterns, while the Medicare study reported an almost linear CRF function at exposures below the then US standard.

A supra-linear CRF could have important implications for quantification. If used to estimate the mortality burden associated with exposure to a population exposed to relatively low levels of air pollution, it would likely result in higher estimates than an assumption of a linear relationship extending to low concentrations. A supra-linear relationship would also suggest that greater health benefit might be achieved by reducing the PM_2.5 concentration (for example, a reduction of 1 µg/m³) experienced by a population already exposed to low levels than by the same reduction in the concentration experienced by a population exposed to higher levels of PM_2.5. If this was used in health impact analysis to inform policy development, it might suggest greater benefits of PM_2.5 reductions in areas with lower concentrations than where concentrations were higher, and therefore might support the introduction of policies which would increase inequalities in exposure to PM_2.5 and associated health effects. In practice, in the UK context, the impact of a supra-linear CRF on policy analysis might not be large. Interventions to reduce already low concentrations of air pollutants might be more costly to implement than those addressing higher concentrations, so are likely to be less favoured by cost-benefit analyses. The influence on calculations of population mortality burden, or of the mortality impact of reducing concentrations, will depend upon the proportion of the population exposed to different concentrations.

COMEAP published updated advice on quantifying the mortality effect of long-term exposure to particulate air pollution, based on the association with PM_2.5 in 2022. They noted that there was no consensus on the shape of the CRF at lower levels of PM_2.5. Therefore, they did not consider the evidence sufficient to recommend any change from the current assumption of a linear CRF when quantifying the effects associated with long-term exposure to PM_2.5 (COMEAP, 2022).

Some studies, including ELAPSE (Brunekreef and others, 2021), have examined associations with low exposures to other pollutants and for other outcomes. However, exploration of the shape of the CRF has, to date, largely focused on the relationship between long-term exposure to PM_2.5 and mortality.

Developments in the evidence

Following publications from the three HEI-funded studies – MAPLE, Medicare, and ELAPSE – Chen and others (2023) carried out a harmonised analysis to evaluate the impact of key study design factors (including confounders, the exposure model used, population age, and outcome definition) on PM_2.5 effect estimates. The main analysis used Cox proportional hazards regression or equivalent models to estimate the association of long-term PM_2.5 exposure with all-cause mortality. It was concluded that the magnitude of the association was affected by the confounders adjusted for, exposure model applied, age of the population and, marginally, by outcome definition. ^{[footnote 2]}

The difference in the observed associations across the 3 studies was marginally reduced by applying the harmonized analysis. Some heterogeneity in the associations might be expected, given the diversity of the cohorts. Typically, heterogeneity is likely due to a combination of differences in methodology, concentration ranges and composition of PM_2.5 or other co-pollutants, population characteristics, geographical location and time periods (Boogaard and others, 2024).

Chen and others (2023) also investigated the shapes of the CRFs, and a combined CRF for all of the cohorts was assessed using meta-analysis based on the extended shape-constrained health impact function (eSCHIF) (see Figure 1). Because the Medicare cohort consists of adults aged 65 years or older, these comparative and combined analyses were restricted to subjects aged 65 years or older in the other cohorts, too. The concentration-response curves generally showed an increased risk at higher concentrations, starting with the lowest observed levels within each study population. The Medicare cohort (in red) was an exception: it displayed a slight decrease in risk over its lowest concentrations followed by a sharp increase in risk over median concentrations (7 to 9 µg/m³). The shape of the CRFs differed substantially across cohorts (see Figure 1 and Appendix A). Some were supra-linear, with a flattening out of the tail at higher concentrations. Others were sigmoidal, some were steep at lower and higher concentrations but flatter at median concentrations, and one was nearly linear. The concentrations at which the curves were flattest, or steepest, also differed between cohorts.

The combined CRF (in blue) suggested a monotonically increasing risk starting from the lowest observed exposure level (approximately 4 µg/m³), with a steep slope from 7 to 9 µg/m³. From approximately 9 μg/m³ the slope of curve started to flatten out, and higher than approximately 14 μg/m³ the curve was almost linear. It was highly influenced by the Medicare cohort (in red) at lower concentrations, and dominated by the Swiss cohort (in black) at higher concentrations. As noted by Chen and others (2023), there was wide uncertainty in the combined CRF at lower concentrations (<7 μg/m³) due to the differences between the CRFs in the US Medicare (sub-linear), and the CanCHEC and Norway studies (both supra-linear). Only two of the European cohorts in the ELAPSE study (from Norway and Denmark) contributed to the shape of the combined eSCHIF at concentrations below about 10 μg/m³. Concentrations in the other European cohorts were higher than this.

Figure 1. CRF for exposure and all-cause mortality in population years of age

Graph A: Mean of eSCHIF fits in 8 individual cohorts with counterfactual levels equal to the fifth percentile of the cohort-specific exposure distributions.

Graph B: eSCHIF (blue solid line) and confidence interval (shaded area) combined by meta-analysis. The thickness of the individual lines indicate the weights each cohort contributed to the meta-analysis: Medicare (0.444), Swiss (0.213), CanCHEC (0.084), Norway (0.080), Belgian (0.067), Danish (0.049), Dutch (0.045), Rome (0.017).

Abbreviations: BEL, Belgian (2001 to 2011); CAN, CanCHEC (1991 to 2016); DAN, Danish (2000 to 2015); DUT, Dutch (2008 to 2012); MED, Medicare (2000 to 2016); NOR, Norwegian (2001 to 2016); ROM, Roman (2001 to 2015); SWI, Swiss (2001 to 2014).

Source: copied from Chen and others (2023) Figure 3. Reproduced from Environmental Health Perspectives with permission from the authors

QUARK discussion

At a meeting in October 2023, members of COMEAP’s quantification sub-group (QUARK) heard a presentation on the harmonised analysis (from the programme manager at HEI) and discussed the results of the harmonised analyses and the combined eSCHIF. An updated summary of COMEAP’s views on the shape of the CRF was prepared, based on this discussion. The revised draft was discussed at the next QUARK meeting, held in March 2024, and further edited to reflect the points made during the meeting. The draft updated statement was discussed by COMEAP at its meeting held in April 2024. Revisions made in the light of comments received were discussed by QUARK at meetings held in May and October 2024. Following these discussions, the drafting was refined further, and the statement finalised by correspondence.

Members expressed their concern regarding combining the CRF curves for different cohorts, which differed substantially in shape and data range, into a single curve. It was noted that the large US cohort was highly influential on the shape of the combined curve, particularly at lower concentrations; the shapes of the CRFs for the Canadian and Norwegian cohorts were very different (supra-linear) at these low concentrations. Given that the shape of the CRF for each cohort (including each individual cohort within the ELAPSE study) was so different, the combined eSCHIF was not considered likely to be a good representation of the shape of the CRF in any individual population. Therefore, it was unlikely to be a better representation of the pattern of association within a population of interest than the usual assumption of a log-linear CRF.

We note that Boogaard and others (2024) comment that, given the heterogeneous nature of PM_2.5, there is no reason to believe that a single shape is appropriate for all locations and populations, as spatial differences in components and sources likely play an important role in determining the shape of these associations.

Updated view

Clearly, the shape of the CRF at low concentrations remains a very important issue which requires careful consideration. The studies, and our discussions, are relevant to burden estimates and health impact analyses in the concentration ranges experienced in the UK currently and anticipated in the future. In 2022, the Average Exposure Indicator (AEI) for the UK was 8 μg/m³ (Defra, 2023a) and population-weighted annual mean PM_2.5 concentrations for England, Wales, Scotland and Northern Ireland were 7.8 μg/m³, 6.7 μg/m³, 4.6 μg/m³ and 5.2 μg/m³, respectively (Defra, 2023b).

We acknowledge that some primary studies, as well as reviews and meta-regressions, have suggested that the CRF might be supra-linear. However, there are substantial differences in the shape of the curves across recent studies investigating the shape of the CRF. Our view is that it will be important to understand the possible reasons for these differences before considering using a non-linear CRF in quantification. It is not clear that the shape of the combined eSCHIF is a good representation of the shape of the CRF for any individual cohort or population of interest. We also note that the confidence interval around in the combined eSCHIF is not inconsistent with a linear relationship. Therefore, we would not expect a deviation from the current assumption of a linear association to better describe the CRF.

We note that the uncertainty in the risks estimated by the combined eSCHIF varied at different exposure ranges. In addition, uncertainties in the exposure model are not reflected in the ranges around the eSCHIF, which only reflect the uncertainties in the epidemiological associations. Low concentrations of PM_2.5 are more difficult to measure and model accurately; this introduces another source of uncertainty in the associations reported at the lowest concentrations.

Heterogeneity in the size and shape of the CRFs reported in different studies is likely due to a combination of differences in methodology and other factors such as population characteristics, concentration ranges, geographical location, time-periods, unmeasured confounders, composition of PM_2.5 and other co-pollutants, and the extent of infiltration of outdoor pollution into the indoor environment.

These factors, together with the desire to have a standard approach across endpoints and pollutants, means that we recommend, at this stage, to continue to apply the current assumption of a linear CRF for use in quantification of mortality effects in the UK attributable to long-term average PM_2.5 concentrations. We also recommend continuing to conduct sensitivity analyses using the 95% confidence intervals around the CRF, in order to consider the uncertainty in the estimates.

We regard these HEI-funded studies, and other studies of cohorts which include participants exposed to low concentrations of air pollutants, as valuable new evidence to inform thinking regarding effects in populations exposed to low levels of PM_2.5. We think these studies are less useful when considering quantification of effects in populations exposed to higher concentrations. The larger evidence base from the literature is more relevant to these situations.

Recommendations for quantification and policy

Studies consistently report that there is a relationship between PM_2.5 and mortality even at low concentrations. Despite the recent additional analysis by HEI, there remains no consistent evidence on the shape of the CRF at lower levels of PM_2.5. We do not consider the evidence sufficient, at this time, to recommend any change from the current assumption of a linear CRF. We recommend that a linear CRF is used when quantifying the mortality effects associated with long-term exposure to PM_2.5 in the UK, and that sensitivity analyses are undertaken using the 95% confidence intervals around the CRF.

We also continue to recommend efforts to reduce concentrations of PM_2.5, even where exposures are already low and concentration-based air quality standards or targets are met. This is because there remains a lack of evidence of a lower exposure threshold for the adverse health effects of PM_2.5.

Recommendations for future research

Work is needed to better understand the potential sources of heterogeneity in the size and shape of CRFs reported in epidemiological studies for this (and other) pollutant-outcome pairs. Understanding the reasons for heterogeneity between studies, and within and between regions, would help inform the selection of appropriate CRFs for specific quantification applications.

Future research could include the exploratory use of different shapes of CRF as sensitivity analyses when quantifying mortality attributable to PM_2.5 (estimates of the attributable mortality burden, or health impact analyses of interventions intended to reduce concentrations). Both supra- and sub-linear shapes should be included. As well as the shape of the curves used, the concentrations over which they are steepest, or flatter, will also influence the result of the quantification. This could also be explored in a research context.

Examination of the shape of the CRFs for other pollutant-outcome pairs, including morbidity endpoints, will also be important.

COMEAP

October 2025

References

1. Boogaard and others (2024) ‘Assessing adverse health effects of long-term exposure to low levels of ambient air pollution: the HEI experience and what’s next?’ Environmental Science and Technology
2. Brauer and others (2022) ‘Mortality: air pollution associations in low-exposure environments (MAPLE): phase 2’ Research Report 212. Boston, MA: Health Effects Institute
3. Brunekreef and others (2021) ‘Mortality and morbidity effects of long-term exposure to low-level PM_2.5, BC, NO₂, and O₃: an analysis of European cohorts in the ELAPSE project’ Research Report 208. Boston, MA: Health Effects Institute.
4. Chen and others (2023) ‘Long-term exposure to low-level PM_2.5 and mortality: investigation of heterogeneity by harmonizing analyses in large cohort studies in Canada, United States, and Europe’ Environmental Health Perspectives: volume 131, number 12
5. COMEAP (2022) ‘Particulate air pollution: quantifying effects on mortality’ Appendix B: Summary of COMEAP views on the studies in populations with low-level exposures and the shape of the concentration-response curve
6. Department for Environment, Food and Rural Affairs (Defra) (2023a) ‘Air pollution in the UK 2022’ (Accessed 19 June 2024)
7. Defra (2023b) ‘Population-weighted annual mean PM_2.5 data, Modelled background pollution data’ (accessed 19 June 2024)
8. Dominici and others (2022). ‘Assessing adverse health effects of long-term exposure to low levels of ambient air pollution: Implementation of causal inference methods’ Research Report 211. Boston, MA: Health Effects Institute

COMEAP sub-group on the quantification of air pollution risks in the UK (QUARK)

Chair  

Dr Mike Holland (EMRC and Imperial College London)

Members

Professor Klea Katsouyanni (University of Athens, Greece and Imperial College London)
Professor Duncan Lee (University of Glasgow)
Professor Gavin Shaddick (Cardiff University)
John Stedman (Ricardo Energy and Environment) Co-opted
Professor Francesco Forastiere (Imperial College London)
Dr Suzanne Bartington (University of Birmingham)
Dr Haneen Khreis (MRC Epidemiology Unit)
David Birchby (Logika Group)
Dr Dimitris Evangelopoulos (Imperial College London) Associate Member
Dr James Milner (London School of Hygiene and Tropical Medicine) Associate Member

Secretariat

Dr Christina Mitsakou (UK Health Security Agency)
Alison Gowers (UK Health Security Agency)

COMEAP Chair

Professor Anna Hansell (University of Leicester)

Acknowledgements

We thank Dr Hanna Boogaard, Consulting Principal Scientist at the Health Effects Institute, for giving a presentation of the harmonised analysis, responding to questions of clarification and contributing to the discussion of the QUARK sub-group.

Appendix A

Chen and others (2023) Figure S2

Cohort-specific concentration-response curve for PM_2.5 exposure and all-cause mortality in study populations aged 65 years and older. eSCHIF (blue solid line with 95% confidence interval indicated by the shaded area) fit to the splines (red solid line with 95% confidence interval indicated by the red dotted lines) with counterfactual level equals the 5th percentile of the cohort-specific PM_2.5 exposure distributions. Six spline knot values indicated by green tick marks. Curves truncated at the 5th and 95th percentiles of cohort-specific PM_2.5 exposure distributions (corresponding values specified in Table 1).

Source: reproduced from ‘Environmental Health Perspectives’ with permission from the authors

Model specification in each cohort in the harmonized analysis

Can

Canadian Census Health and Environment Cohort (CHEC) model adjusted for cohort (strata), sex (strata), 5-year age groups (strata), follow-up year (time axis), income, immigration status, visible minority, indigenous identity, neighbourhood composite SES index, and airshed indicator.

Medicare

Model adjusted for sex (strata), 5-year age groups (strata), follow-up year (strata), race (strata), Medicaid eligibility (strata), ZIP code level income, median house value, poverty rate, house-owing rate, high education rate, and Census region indicator.

Belgian

Model adjusted for age at follow-up (time axis), sex (strata), follow-up year (strata), country origin, education, both neighbourhood- and regional-level income, unemployment rate, low education rate, non-western ethnic rate, and regional indicator.

Danish

Model adjusted for age at follow-up (time axis), sex (strata), follow-up year (strata), country origin, income, both neighbourhood- and regional-level income, unemployment rate, low education rate, and regional indicator.

Dutch

Model adjusted for age at follow-up (time axis), sex (strata), follow-up year (strata), country origin, income, both neighbourhood- and regional-level composite SES index, income, unemployment rate, non-western ethnic rate, and regional indicator.

Norwegian

Model adjusted for age at follow-up (time axis), sex (strata), follow-up year (strata), income, both neighbourhood- and regional-level poverty rate, unemployment rate, low education rate, and regional indicator.

Roman

Model adjusted for age at follow-up (time axis), sex (strata), follow-up year (strata), education, neighbourhood-level income, unemployment rate, low education rate, high education rate, and composite SES index.

Swiss

Model adjusted for age at follow-up (time axis), sex (strata), follow-up year (strata), nationality, mother tongue, education, both neighbourhood- and regional-level composite SES index, unemployment rate, low education rate, high education rate, and regional indicators.

1. Or log-linear. Statistical methods used in cohort studies of mortality typically assume a log-linear relationship between exposure and risk. In practice, for a small HR (as found in most air pollution studies) and over a small concentration range (as typically found in a single-site study) there is little difference between a linear and log-linear relationship. This might not be the case when larger concentration differences are being considered. ↩

2. As well as the exposure models being different, it is noted that the sample size in the US Medicare study differed in the harmonised analysis (74.5 million) compared to the original HEI publication (68.5 million) because the covariate adjustment sets were different for these 2 analyses. ↩