Air Pollution

A central challenge for both research and policy is that air pollution is not directly observed as a single, person-level exposure. Instead, translating emissions into actionable policy requires understanding four distinct links in a chain: (i) emissions – what sources release into the air; (ii) ambient concentrations – the resulting pollutant levels present outdoors at a given place and time; (iii) personal exposure – what individuals actually inhale, as shaped by time indoors, building quality, and defensive behaviour; and (iv) dose and damage – what the body absorbs and how it affects health and economic outcomes. In this section we focus on the final link – how exposure maps to damage – before turning to defensive behaviour, regulation, and political economy in later sections.

Dose-Response Curve

The final link in the chain above from (iii) exposure to (iv) dose and damage is summarised by a dose–response (or concentration–response) curve: a function mapping a sustained change in PM2.5 exposure into changes in mortality risk and other outcomes. Understanding this relationship matters for academic and policy reasons. It determines where marginal benefits of regulation are highest, how standards should be set, and how sensitive burden and benefit-cost estimates are to modelling assumptions. These questions are especially consequential in LMICs, where exposures are often far outside the range from which most evidence has been drawn. For instance, if the curve is steep at low concentrations, then further reductions in already-cleaner settings can deliver large health gains, supporting stringent standards even below current regulatory limits. If marginal damages remain large at high concentrations, then pollution reductions in the most exposed LMIC settings deliver especially high returns.

An important note here is that most economic studies treat ambient concentration as a proxy for personal exposure, which is often the only feasible approach. However, this assumption is not innocuous and can generate systematic measurement error. In places where people spend most of their time indoors, where building stock is poor, or where outdoor workers face prolonged exposure, the gap between ambient readings and actual inhaled doses can be meaningful, and estimated damages can correspondingly be under or overstated. More recently, researchers are using indoor air pollution monitors (Chowdhury et al. 2026, Garg et al. 2025) and portable pollution monitor backpacks (Berkouwer and Dean 2026) to more accurately measure personal exposure distinct from ambient concentrations.

Epidemiological Evidence from Cohort Studies

The most direct evidence on chronic exposure comes from long-run prospective cohort studies, which track populations over years or decades and relate sustained PM2.5 exposure to all-cause and cause-specific mortality. The landmark American Cancer Society cohort study found robust associations between long-run PM2.5 exposure and cardiopulmonary and lung cancer mortality (Pope et al. 2002). Extended follow-up of the Harvard Six Cities study corroborated these findings and documented continuing risks at concentrations that were, at the time of publication, already below prevailing regulatory standards (Lepeule et al. 2012). A recurring and policy-relevant finding across these cohorts is that the relationship does not exhibit an obvious safe threshold: risk appears to decline monotonically as concentrations fall, with no floor below which further reductions deliver no benefit. The 2021 WHO guidelines, which halved the recommended annual mean PM2.5 standard from 10 to 5 µg/m3, reflect precisely this reading of the evidence – and explicitly note that even 5 µg/m3 should be understood as a policy target rather than a safe limit (World Health Organization 2021, Pai et al. 2022). Yet this evidence base has a critical limitation for applications to LMICs: virtually all of it was generated at annual average concentrations below 20 µg/m3, far below the levels routinely experienced across South Asia and sub-Saharan Africa.

Causal Evidence from Economics

The epidemiological tradition reviewed above rests on statistical control of a limited set of observable confounders and faces challenges of endogeneity and residential sorting. A distinct economics literature has developed quasi-experimental designs, exploiting natural experiments, instrumental variables, and policy discontinuities to isolate causal effects of pollution from confounding factors. These studies typically exploit short-run variation in pollution and relate it to near-term health outcomes. The resulting estimates are more credibly causal than cohort associations, but they answer a somewhat different question: the acute rather than chronic exposure margin. Both bodies of evidence are important for a full accounting of damages.

Benchmarks from higher-income countries

The literature on the causal effects of air pollution on health and human capital has been reviewed comprehensively elsewhere (Graff Zivin and Neidell 2013, Aguilar-Gomez et al. 2022, Brewer et al. 2023); here we summarise the key findings most relevant to LMIC policy and highlight where the evidence base is thin or absent. The most influential early contributions established the approach by examining infant mortality in the US. Chay and Greenstone (2003) exploit geographic variation in particulate pollution shocks induced by the 1981–82 recession and find a 0.35% decline in the infant mortality rate for every 1% reduction in total suspended particulates. They find that these effects are driven almost entirely by deaths occurring within one month of birth, suggesting foetal exposure as the primary pathway. Currie and Neidell (2005) use within-county variation in carbon monoxide across California to document comparable effects on infant health.

The literature has since broadened considerably beyond infant mortality. Schlenker and Walker (2016) exploit idiosyncratic variation in airport runway congestion in California driven by network delays originating on the East Coast as an instrument for daily carbon monoxide exposure. They find significant effects on hospital emergency room admissions and respiratory hospitalisations among nearby residents. A one standard deviation increase in daily pollution levels generates an additional $540,000 in hospitalisation costs for the six million individuals living within 10 km of the studied airports. Di et al. (2017) link Medicare administrative data for over 60 million elderly Americans to satellite and monitor-based pollution estimates, finding a positive association between PM2.5 and mortality with no evidence of a threshold even at concentrations entirely below the then-current US national standard of 12 µg/ µg/m3. Deryugina et al. (2019) use daily changes in local wind direction as an instrument for PM2.5 and find that mortality effects are concentrated in roughly 25% of the elderly population, specifically those with five to ten years of remaining life expectancy. Their machine-learning-based estimates of life-years lost imply that the reduction in US PM2.5 concentrations that occurred from 1999 to 2013 generated mortality reductions worth approximately $24 billion annually.

Two notable papers push the evidence in directions particularly relevant for understanding the full welfare costs of pollution. Deschênes et al. (2017) exploit the NOx Budget Program, a cap-and-trade market that substantially reduced ozone concentrations from 2003. Using a triple-differences design, they estimate both mortality reductions and pharmaceutical expenditure responses. They find that annual reductions in pharmaceutical purchases valued at roughly $800 million and mortality reductions valued at roughly $1.3 billion combined, with defensive investments representing over one-third of total willingness-to-pay for air quality improvements. This result is important beyond the specific setting: it implies that studies measuring only mortality will systematically understate the full welfare cost of pollution by ignoring the costly private actions people take to protect themselves. Bishop et al. (2023) use the expansion of Clean Air Act regulations as a source of quasi-random variation in individuals’ long-run PM2.5 exposure, tracking Medicare beneficiaries aged 65 and above from 2001 to 2013. They find that a 1 µg/m3 increase in decadal PM2.5 raises the probability of a new dementia diagnosis by 2.15 percentage points.

Evidence from low- and middle-income countries

A growing body of work has applied similar designs in LMIC settings, where the policy stakes are highest but the evidence base is considerably thinner. Three features of the literature are worth foregrounding. First, it is geographically concentrated: India, China, and Indonesia account for the large majority of credibly identified studies, while Pakistan, Bangladesh, most of sub-Saharan Africa, and Latin America beyond Mexico and Brazil have limited quasi-experimental estimates. Second, effect sizes in LMICs frequently exceed comparable developed country benchmarks. Third, different identification strategies within the same country sometimes yield strikingly different conclusions, a divergence that itself reveals something important about the dose–response relationship and the limits of policy-based estimates.

Jayachandran (2009) exploits the sharp spatial and temporal patterns of fires across subdistrict-birth month cohorts and finds that the resulting pollution led to approximately 15,600 missing children or roughly 1.2% of affected birth cohorts, with effects twice as large in poorer areas. Arceo et al. (2016) use thermal inversions as an instrument for daily pollution in Mexico City and find effects on infant mortality substantially larger than comparable US estimates, consistent with nonlinear dose–response at higher baseline concentrations and with limited avoidance capacity among lower-income households. Across 30 sub-Saharan African countries, Heft-Neal et al. (2018) exploit natural variation in Saharan dust driven by remote rainfall patterns and estimate that a 10 µg/m3 increase in PM2.5 raises infant mortality by 9% with PM2.5 levels accounting for 22% of all infant mortality.

A natural interpretation of the larger LMIC estimates is that populations sit on a steeper part of a nonlinear dose–response curve at very high baseline concentrations. However, Colmer et al. (2021) exploit thermal inversion variation in Hong Kong (a uniquely high-pollution yet high-income setting) and find large effects of PM on birth weight but no detectable effect on neonatal mortality. Since Hong Kong’s pollution levels are comparable to many developing cities, the authors conclude that mortality damages are high in less-developed countries because of their stage of economic development, not necessarily because they are more polluted.

The Shape of the Dose-Response Curve

Whether the concentration–response (C–R) relationship is convex, concave, or linear has first-order implications for where regulatory effort should be directed and what returns can be expected from pollution reduction across different settings. Hsiang et al. (2019) provide a framework for understanding why the same environmental change can generate very different welfare impacts depending on where a population sits on the damage function.

For mortality and acute health outcomes, recent evidence points towards concavity. Miller et al. (2025) exploit variation in wildfire smoke plume proximity and density across US counties using Medicare data and find that health impacts rise steeply at low exposures and then plateau, flattening around 6 µg/m3 of smoke-induced PM2.5. Several mechanisms can produce this shape. Frailty selection depletes the most susceptible individuals at lower pollution levels, so that as concentrations rise the remaining population is drawn increasingly from those with greater physiological resilience, reducing the marginal mortality response. Key physiological pathways, including airway inflammation and cardio-vascular stress responses, may saturate at moderate exposures, so that additional pollution triggers no further acute biological response beyond what lower doses have already induced. Finally, at high and visible pollution episodes, individuals who can do so reduce outdoor activity, shift travel modes, or wear masks, mechanically flattening the observed ambient-to-outcome relationship even if the underlying biological dose–response remains steep (Hsiang et al. 2019). This behavioural margin is larger where pollution is salient and where households have the resources to act on it.

The mortality C–R function could plausibly be convex at very high concentrations. Compounding damage across organ systems already compromised by sustained exposure, combined with limited healthcare access in LMICs, could generate responses that accelerate rather than plateau in the extreme tail. For instance, Guidetti et al. (2024) find that hospital capacity constraints intersect with air pollution exposure to generate worse health outcomes. Our conclusion is that the C–R relationship at the pollution levels prevalent in South Asia and sub-Saharan Africa remains poorly identified, and burden estimates derived from risk functions calibrated at lower exposures and applied to LMIC populations via extrapolation carry substantial uncertainty (Ebenstein et al. 2017, Jaganathan et al. 2024).

For non-health outcomes such as labour productivity and cognitive performance, the C–R relationship may look quite different. Because these effects are subclinical and difficult to self-attribute, avoidance behaviour may be weaker and the gap between ambient concentration and effective dose could be smaller (Aguilar-Gomez et al. 2022). On the other hand, if the negative effects of air pollution or positive effects from improvements in air quality are more immediately noticeable, avoidance behaviour could be stronger. There is also a much lower likelihood of a frailty selection mechanism of the kind that possibly generates concavity in acute mortality: cognitive and productivity effects operate across healthier populations, with no equivalent pool of highly vulnerable individuals that is depleted at lower exposures.

While there remains uncertainty on C–R shape as it pertains to labour productivity and cognitive performance, the evidence is consistent with a broadly linear or even accelerating relationship across the concentration range. Graff Zivin and Neidell (2012) find a linear ozone–productivity gradient among agricultural workers that is detectable well below the US regulatory standard, with no evidence of a plateau. Adhvaryu et al. (2022) study an Indian garment factory operating at a mean PM2.5 of around 65 µg/m3 and find that a one standard deviation increase in pollution (roughly 45 µg/m3) reduces hourly worker efficiency by around 6%, with effects largest for more complex tasks. Zhang et al. (2018) show that three-year cumulative air pollution exposure has effects on verbal cognition roughly twenty times larger than contemporaneous exposure, consistent with neurological damage that accumulates rather than saturates. La Nauze and Severnini (2025) similarly find that PM2.5 impairs adult cognitive function in a brain-training dataset, with the largest effects concentrated among prime-working-age adults and on novel tasks requiring fluid intelligence rather than practised routines. These findings suggest that to the extent the mortality C–R curve flattens at very high concentrations, the share of total welfare damages attributable to non-health channels may grow substantially with pollution level – with significant consequences for how the returns to regulation are calculated in LMIC settings.

For full reference list see the end of the conclusion chapter.