Pricing the Ozone Penalty of Cleaner Air in Cohort Studies

Haze shades the ground. Fine particles scatter and absorb the ultraviolet sunlight that drives ozone formation, so when particle pollution falls, more ultraviolet light reaches the surface and some ozone, a separate harmful gas, rises in its place. Atmospheric chemists call this the "clean air paradox" and have measured it. Absorbing aerosols cut the photolysis rates that start ozone chemistry by roughly 17–22% in China (Absorbing Aerosol Optical Properties and Radiative Effects on Near-Surface Photochemistry in East Asia). Photolysis is the light-driven splitting of molecules such as nitrogen dioxide. On weekends in Beijing, when particle levels are lower, surface ultraviolet light and ozone both rise (Ozone weekend effects in the Beijing–Tianjin–Hebei metropolitan area, China). Health cohort studies, by contrast, treat particles and ozone as two separate exposures. A long-term mortality cohort either ignores ozone (Cardiovascular Mortality, Habitual Exercise, and Particulate Matter 2.5 Exposure: A Longitudinal Cohort Study) or enters it as a second variable to hold fixed.

This proposal argues that holding ozone fixed answers the wrong policy question. If cutting particles causes some extra ozone, then part of ozone lies on the causal path from particles to health: it is a mediator, not merely a confounder. A study that adjusts for it estimates what particle reduction would do if ozone magically stayed put. A clean-air policy is closer to the total effect, which includes the ozone penalty. We propose to estimate that penalty inside a retrospective cohort, meaning one built from records already collected. The key step is a physically computed exposure: the ozone attributable to aerosol shading, obtained by recomputing photolysis with and without aerosols. Raw ozone readings would not do, because they mix this pathway with weather.

Core Claim

The health benefit of reducing fine particulate matter (PM2.5) can be split into a direct part and an indirect part. The indirect part runs through aerosol shading of ultraviolet light, which suppresses ozone. Write $TE$ for the total effect of a PM2.5 reduction on cardiovascular mortality. Write $NDE$ for the natural direct effect, the change if ozone stayed at what it would have been under the original particle level. Write $NIE$ for the natural indirect effect, the change due only to the ozone shift the particle reduction induces. Then

$$TE = NDE + NIE .$$

Our claim is that $NIE$ can be identified with a radiative-chemical mediator, $\Delta O_3^{aer}(s,t)$. This is the surface ozone difference at location $s$ and day $t$ between the real atmosphere and one with aerosol optical effects switched off. It is computed from aerosol optical properties, not read off monitors. Because it is driven by particle optics rather than by the shared sunny, stagnant weather that raises both pollutants, it avoids the main reason naive mediation on observed ozone fails.

The physics also gives a falsifiable prediction. In rural areas where ozone formation is limited by nitrogen oxides, aerosols lower ozone by about 8–10 ppbv. In urban areas limited by volatile organic compounds (VOCs), moderate aerosol loads of 20–45 µg m⁻³ raise ozone by 5–10 ppbv through particle-surface chemistry that produces hydroxyl radicals (The atmospheric oxidizing capacity in China – Part 1: Roles of different photochemical processes). So the sign of $NIE$ should flip between the two kinds of area. A two-pollutant regression cannot express that flip.

We expect the penalty to be modest. Aerosol-driven surface ozone changes during heavy haze are 4.2–10.6 ppb (What have we missed when studying the impact of aerosols on surface ozone via changing photolysis rates?). A finding of a small or null penalty is therefore a legitimate result, and the design is built to report it credibly.

Research Program

The work has three parts: an exposure emulator, a semi-synthetic test bench, and the cohort analysis.

The exposure emulator. The emulator is a fast statistical stand-in for an expensive physical model; it maps aerosol optical properties to the aerosol-attributable ozone change. Its inputs are column aerosol optical depth, the fraction of aerosol extinction that is absorption, boundary-layer height, solar geometry, and a chemical-regime indicator.

It is trained on paired WRF-Chem runs. WRF-Chem is a coupled weather and chemistry model, and the pairing follows the design in What have we missed…: one run computes photolysis without aerosol optics, one with them. The emulator must target the full surface ozone change, not only the change in chemical production. That paper is a constraint on the design. It shows a −44.3 ppb drop in chemical production that is largely offset by +12.0 ppb from extra vertical mixing and +19.6 ppb from reduced deposition to the ground. An emulator built only on photolysis would therefore overstate the penalty roughly threefold.

The photolysis change itself comes from the TUV radiative transfer model (Tropospheric Ultraviolet and Visible), driven by AERONET-derived optical property climatologies as in Absorbing Aerosol Optical Properties…. AERONET is a network of ground sun photometers. That paper supplies the method. It also bounds the precision we can claim, since it notes that brown-carbon and mixing-state optics are crudely represented.

There is a data gap to state plainly. The file's AERONET and WRF-Chem domains are in mainland China, Korea and Japan. The cohort is in Taiwan. Long-term gridded aerosol optical depth and Taiwanese ozone monitor records over 2001–2019 must be obtained from sources outside this file. Transferring the emulator across domains is itself a risk, and it is tested below.

The semi-synthetic test bench. We keep the real exposure fields and real covariates and simulate mortality outcomes from a known structural model. In that model, both PM2.5 and $\Delta O_3^{aer}$ affect the hazard, and regional meteorology confounds both. Because $TE$, $NDE$ and $NIE$ are known by construction, every estimator can be scored on bias and confidence-interval coverage. This is the only setting where ground truth exists, and it is where the main methodological claim is tested.

The cohort analysis. We reuse the 384,128-person MJ Health cohort linked to the Taiwan National Death Registry, with its 1 km² PM2.5 surfaces and time-varying Cox models (Cardiovascular Mortality, Habitual Exercise, and Particulate Matter 2.5 Exposure). We add $\Delta O_3^{aer}$ as a time-varying mediator and estimate $NDE$ and $NIE$ with inverse-probability weighting. Weights are built with the matching and trimming machinery of Matching on the Estimated Propensity Score and Weight Trimming and Propensity Score Weighting.

Confounding by large-scale weather is handled with regional meteorology embeddings in the style of Weather2vec. That work learns spatial summaries of reanalysis weather fields for propensity scores and shows that local-only adjustment badly biases ozone policy estimates. It never touched health outcomes and never modeled light extinction. We use it as a tool, not as the contribution.

Baselines

The naive reference. This is the single-pollutant time-varying Cox model of the MJ cohort paper: PM2.5 only, with local weather adjustment. It estimates something closer to the total effect, but it is confounded by weather that moves both pollutants. That paper reports a 123% higher cardiovascular death risk for inactive, high-PM2.5 participants versus active, low-PM2.5 participants. We reproduce its specification exactly before changing anything.

The strongest competitor. This is the standard two-pollutant time-varying Cox model, with monitor-based ozone added to PM2.5. It is what an epidemiology reviewer would ask for and is the field's default. We found no reported result on Lacuna for a two-pollutant PM2.5–ozone hazard ratio in this cohort. Its comparison numbers will therefore come from our own run on identical person-time, identical covariates, and identical spline degrees of freedom.

Beating it means two things. In the semi-synthetic bench, the proposed method must recover $NIE$ and $TE$ with lower bias and near-nominal coverage where the two-pollutant model, which estimates something close to $NDE$, cannot. In the real cohort, the estimated $NIE$ must have a confidence interval excluding zero in at least one regime, or be tightly bounded near zero.

The Weather2vec-adjusted two-pollutant model. This variant gets the same regional embedding as our method and the same propensity tuning budget. Its purpose is to separate improvements from better confounding control from improvements from the mediator.

The key ablation. We run the full pipeline twice more with different mediators. The first uses observed monitor ozone. The second uses a photolysis-only proxy that leaves out the vertical-mixing and deposition buffering. If observed ozone performs as well as $\Delta O_3^{aer}$ in the bench, the physical mediator is unnecessary.

Why not just…?

Why not just put PM2.5 and ozone together in one Cox model? That model is well defined, but it targets a different quantity: the particle effect with ozone held fixed. If ozone partly lies downstream of particles, that is a direct effect, not the effect of a clean-air policy. The semi-synthetic bench shows this directly by comparing its estimate with the known $TE$ and $NDE$.

This alternative is a real threat in one case. If $\Delta O_3^{aer}$ explains a negligible share of ozone variance in Taiwan, then $NIE$ is near zero and $NDE$ is close to $TE$. The two-pollutant model would then be adequate. We will report that outcome rather than argue around it.

Why not just do statistical mediation on observed ozone? The observed PM2.5–ozone relationship is dominated by shared meteorology and even changes sign with pollution level and altitude. Correlations are negative on polluted days and positive on clean days (Lidar- and UAV-Based Vertical Observation of Spring Ozone and Particulate Matter in Nanjing, China). Mediation on observed ozone therefore mixes the shading pathway with weather. The observed-ozone ablation in the bench quantifies how much bias this introduces.

Why not just adjust for broadband solar radiation? Absorbing aerosols act preferentially in the ultraviolet-A band. Broadband irradiance therefore does not track the photolysis changes that matter (Absorbing Aerosol Optical Properties…). We include broadband radiation as an alternative mediator in the ablation so the bench can settle the question empirically.

Why not just forecast exposures with a time-series foundation model? Such forecasting models address prediction, not the counterfactual "aerosols switched off" ozone this design needs. They also perform worst on reactive ozone in the Air Quality Arena benchmark. That source gives no MASE numbers (mean absolute scaled error, a forecast-accuracy score), so we cite only its qualitative finding.

Proposed Experiments

Emulator validation against data it never saw. The emulator must reproduce the aerosol-driven surface ozone reductions of 4.2–10.6 ppb in Beijing, Tianjin, Shijiazhuang and Zhengzhou (What have we missed…). It must also reproduce the sign of the weekend natural experiment. That experiment shows PM2.5 falling up to 10.7% on weekends, ultraviolet light rising from 12.3 to 12.7 W m⁻², and ozone rising 3.4–9.5% (Ozone weekend effects…).

That paper used carbon monoxide as a stand-in for VOCs. We therefore treat the weekend comparison as a check on sign and order of magnitude only, not as a precise target. If the emulator gets the weekend sign wrong, the pipeline stops there.

Semi-synthetic bias and coverage. On the test bench, we report bias and 95% interval coverage for $TE$, $NDE$ and $NIE$ across all baselines. We vary the true size of the ozone penalty from zero upward. This locates the penalty size below which the cohort cannot detect it.

Negative control. We assign $\Delta O_3^{aer}$ computed for nighttime hours, when no photolysis occurs, as a placebo mediator. It must yield an $NIE$ indistinguishable from zero. A nonzero placebo effect would reveal leftover weather confounding.

Regime test. We stratify $NIE$ by an indicator of whether a location's ozone formation is limited by nitrogen oxides or by VOCs. Physics predicts opposite signs (The atmospheric oxidizing capacity in China – Part 1). Agreement would be strong evidence the mediator captures the intended mechanism.

Exercise subgroup, as a secondary analysis. Outdoor activity raises the amount of ozone a person actually inhales. We ask whether the penalty is concentrated among habitual exercisers. This revisits the null exercise-by-PM2.5 interaction ($p = 0.990$) in the MJ cohort through a pathway that study could not see.