Splitting Air-Pollution Exposure by Its Cause to Test Blood-Pressure Effects

Large cohort studies relate year-to-year changes in a person's fine-particle exposure (PM2.5, particles small enough to reach deep into the lungs) to later health. The exposure change comes from two sources. One is emissions: factories, traffic and fuel policy. The other is weather: wind, rain, and the height of the air layer near the ground that holds pollution in, called the boundary layer. Atmospheric scientists have long separated the two, a step they call "deweathering." Cohort epidemiologists have not.

This matters because weather also changes blood pressure directly. Cold days, for example, raise it. An estimate that mixes both sources of pollution change can therefore absorb a temperature effect and credit it to particles. We propose to split each participant's exposure history into an emission-driven part and a weather-driven part. We would then estimate the effect of each part on continuous blood-pressure trajectories across repeated checkups, and use agreement or disagreement between the two estimates as a built-in test for confounding.

Core Claim

If particles themselves raise blood pressure, then one microgram per cubic metre should have the same effect whether a factory or a stagnant boundary layer put it there. The two components can be written as

$$\Delta PM_{it} = \Delta E_{it} + \Delta W_{it},$$

where $\Delta PM_{it}$ is the change in person $i$'s average exposure between visit $t-1$ and visit $t$. The term $\Delta E_{it}$ is the emission-driven part: the model-predicted concentration under typical weather. The term $\Delta W_{it}$ is the remainder, driven by that interval's actual weather.

We would regress within-person blood-pressure velocity on both components separately, with flexible adjustment for temperature and humidity. This gives two coefficients, $\beta_E$ and $\beta_W$. Under the hypothesis that particles cause the effect and the model is correctly specified, $\beta_E = \beta_W$. A gap between them is informative in a specific direction. If $\beta_W$ is inflated, temperature or seasonal effects are leaking into the estimate. If $\beta_E$ is inflated, secular trends that move with regional emission cuts are the likely cause; examples are changing medication use, diet, or urbanization.

The sharpest version of this goes further. It uses only the weather-driven variation that comes from boundary-layer height, wind and rain, after temperature and humidity are adjusted for. These factors move PM2.5 strongly but are much less plausibly direct causes of blood pressure. That variation can serve as a near-random "natural experiment" instrument.

To our knowledge, nothing in the literature we reviewed treats deweathering as a tool for causal identification in cohort studies. In atmospheric science, deweathering is used to remove weather as a nuisance. In cohort studies, exposure is used raw. We use the decomposition to test and sharpen the exposure–biomarker estimate.

Why this is the right cut, and what it is not

The obvious combination of these two literatures would simply deweather the exposure and rerun the cohort models. We think that is wrong for health questions. People breathe the air that is actually there, including pollution trapped by a low boundary layer. Discarding the weather-driven part throws away real exposure and makes the emission-only estimate the most exposed to trend confounding. Our design keeps both parts and compares them.

The motivating gap comes from three Taiwanese cohort studies. Dynamic Changes in Long-Term Exposure to Ambient Particulate Matter and Incidence of Hypertension in Adults followed 134,978 adults with serial checkups. It related the change in satellite-derived PM2.5 between visits to incident hypertension and reported a hazard ratio of 0.84 (95% CI 0.82–0.86) per 5 µg/m³ reduction. The same MJ cohort underlies Effects of air pollution and habitual exercise on the risk of death and Habitual exercise is associated with reduced risk of diabetes regardless of air pollution. All three use raw satellite estimates and threshold diagnoses as outcomes.

On the atmospheric side, two papers show that this raw exposure mixes weather and emission change. Separating emission and meteorological contributions to long-term PM2.5 trends over eastern China attributes multi-year PM2.5 trends to both sources. The 10-year trend of PM2.5 components in Hong Kong documents rain washout cycles that hide emission trends. Alexis Lau co-authored both the Hong Kong atmospheric work and the Taiwanese cohorts. The two methods have simply never been joined.

The deweathering method comes from Transferable Models to Understand the Impact of Lockdown Measures on Local Air Quality. That paper fits generalized additive models (regressions built from smooth curves of each weather variable) on multi-year data, then freezes the weather curves. We reuse that recipe. The same paper also constrains us: it notes that such models ignore secondary chemistry and are usually fitted to daily rather than hourly series. We therefore restrict the main analysis to PM2.5 averaged over the interval between visits, where daily resolution is adequate. Ozone is left as an explicitly harder case.

The biomarker side draws on work showing that direction and speed of change carry information a single reading misses. hsCRP trajectories and incident metabolic syndrome found that a "Decreasing" inflammation trajectory still carried an adjusted hazard ratio of 1.408. Albumin changes in acute myocardial infarction found that the shape of serial albumin measurements predicted mortality better than the admission value. Situational spikes such as white-coat blood pressure (Physical activity and high blood pressure among Chinese children) are exactly the noise that a trajectory model smooths out. For this reason we model continuous systolic and diastolic pressure across visits rather than a binary hypertension label.

Research Program

Exposure decomposition. For each 1 km grid cell or district, we fit a transferable additive model of daily PM2.5 on meteorology from a reanalysis product. The weather inputs are boundary-layer height, wind speed and direction, precipitation, temperature and humidity, plus day-of-week and a slowly varying trend term. The emission-driven series is the model's prediction when that cell's weather is repeatedly redrawn from its own multi-year climatology, following the resampling logic in the eastern-China paper. The weather-driven series is the observed series minus the emission-driven series. We then average both components over each participant's inter-visit interval at their residential address.

Validation of the decomposition. Before any health analysis, we check that the decomposition recovers known emission shocks. We would refit on the Swiss, Beijing and Wuhan monitoring data used in the lockdown paper and require that our emission component reproduce its deweathered NO2 changes: −29.4% for Eastern Switzerland, −28.1% for Beijing and −52.8% for Wuhan. We also require that the Swiss PM10 rise of +9.8% during a Sahara dust episode be assigned to the non-emission part.

Trajectory model. We fit a mixed-effects model of blood pressure over irregularly spaced visits. It includes a person-specific intercept and slope, and treats the per-year change between consecutive visits as the outcome. The main terms are $\Delta E$ and $\Delta W$. Adjustments include smooth, lagged temperature and humidity terms, season of visit, calendar year, age, body mass index, antihypertensive use, and the lifestyle variables already collected by the MJ cohort. Irregular spacing is handled by expressing change as a rate per year and averaging exposure over the actual interval.

We deliberately do not start with the global–local projection method of Globally Aligned Longitudinal Enrichments. Its gains were shown on multi-modal imaging data in ADNI, and a rate-based mixed model is the more interpretable first step for a single biomarker.

Boundary-layer instrument. We regress $\Delta W$ on boundary-layer height, wind and precipitation anomalies, conditional on temperature, humidity and place-by-season effects. The fitted value then serves as an instrument in a two-stage estimate of the blood-pressure effect.

Negative control. The weather-driven component measured in the interval after a visit should not predict blood-pressure change before that visit. A non-zero association here signals residual confounding by persistent place-and-season patterns.

Secondary outcome. Fasting glucose trajectories, also measured at MJ checkups, serve as a second outcome.

Baselines

The naive reference point is the existing approach reproduced on the same records. This means a time-varying Cox model of incident hypertension on raw between-visit PM2.5 change, as in Dynamic Changes in Long-Term Exposure to Ambient Particulate Matter. We also refit its continuous analogue: the same mixed model with raw $\Delta PM$ in place of the two components. The reported benchmark is a hazard ratio of 0.84 (95% CI 0.82–0.86) per 5 µg/m³ reduction. For the continuous blood-pressure-velocity analogue, no reported result was found on Lacuna. Our reproduction of the published Cox estimate is a requirement before anything else is interpreted.

The strongest competitor is raw PM2.5 plus flexible meteorological adjustment. Concretely, this is the same mixed model with smooth, lagged spline terms for temperature, humidity, wind and precipitation. It is what a careful epidemiologist would do without our decomposition. For a fair comparison it receives the same records, the same weather inputs, and the same number of spline degrees of freedom as our first-stage model, with smoothing chosen by the same cross-validation. On real data, "beating" this baseline cannot mean a smaller error, because the true effect is unknown. It is defined on the semi-synthetic benchmark described below. There we count a win only if our component or instrument estimates show lower bias and correct confidence-interval coverage in at least one realistic confounding scenario, while matching the baseline where no confounding is simulated.

The obvious-integration baseline uses only the deweathered, emission-driven exposure. It isolates what is gained by keeping the weather-driven part. The key ablation replaces the two-component model with a single component. A second ablation drops the restriction of the instrument to boundary-layer height, wind and rain, and lets it use all weather variables including temperature. If the temperature-inclusive instrument gives estimates far from the restricted one, that confirms temperature's direct effect is the contaminant we claimed.

Why not just…?

Why not just add weather covariates to the Cox or mixed model? This is the most serious alternative, and it may be enough. If flexible temperature and humidity adjustment already removes the confounding, the raw-PM estimate will match both component estimates on real data, and $\beta_E$ will equal $\beta_W$. Our decomposition then becomes a diagnostic that certifies the simpler model rather than a better estimator. We would report that plainly.

The reason we doubt it is enough is that covariate adjustment cannot say which variation in PM2.5 drives the estimate. Nor can it detect confounding by emission-linked secular trends, which no weather term addresses. The semi-synthetic experiment settles the question directly. We simulate blood pressure with a known PM effect plus a lagged, nonlinear direct temperature effect and a trend confounder correlated with regional emission cuts, then check which estimator recovers the planted effect.

Why not just deweather the exposure and rerun the existing models? Health effects follow the air people actually breathe. Weather-trapped pollution is real exposure, not noise. The lockdown paper's Sahara-dust example cuts both ways: it shows raw series can mislead about emissions, but the dust was still inhaled. The emission-only baseline above quantifies what is lost. We expect it to have the widest confidence intervals and the greatest vulnerability to secular-trend confounding.

Why not just keep the binary incident-hypertension outcome? Threshold diagnoses discard within-range change and are pushed around by single-visit spikes. The hsCRP and photoplethysmography-age results show that trajectory information changes risk estimates substantially. AI-derived PPG age reports a hazard ratio of 2.37 for a single elevated reading, rising to 5.53 when the elevation persists across visits. We run both outcomes. The continuous one is justified only if it gives tighter intervals for the same component comparison.

Why not just use a chemical transport model for the split? A full weather-and-chemistry simulation such as WRF-Chem, used in Ozone pollution over China and India, would handle chemistry better. It depends on emission inventories that are uncertain at 1 km over fifteen years, and it costs far more to run. If such simulations exist for Taiwan, we would use them as a second decomposition. Agreement between the two methods on $\beta_E$ and $\beta_W$ would strengthen the conclusion, and disagreement would mark the result as decomposition-dependent.

Proposed Experiments

The first experiment is the decomposition sanity check on the lockdown-paper monitoring networks. It passes if the emission component comes within a few percentage points of the published deweathered NO2 changes and assigns the Swiss dust episode to the weather part.

The second experiment is the semi-synthetic benchmark. We take the real exposure and weather histories and real visit schedules and simulate blood pressure under several scenarios: no confounding, a direct temperature effect, emission-trend confounding, and both together. We score every baseline and our estimators on bias and on coverage of 95% intervals. This is where "beating the strongest baseline" is defined.

The third experiment is the real-data analysis on the MJ cohort. Access is restricted; the file documents the cohort's use in three studies but gives no access route. We report $\beta_E$, $\beta_W$, the instrument estimate, the negative-control estimate, and the test of $\beta_E = \beta_W$, for systolic pressure and for fasting glucose.

The decisive outcome has three possible readings:

  • If all three estimates agree and the negative control is null, the published PM–hypertension link gains an unusually strong causal footing.
  • If only $\beta_W$ is inflated, temperature leakage is present in the existing literature.
  • If only $\beta_E$ is inflated, emission-linked secular trends are, and part of the reported 0.84 hazard ratio reflects those trends rather than air quality.

If MJ access fails, the same pipeline can run on the UK Biobank records described in the PPG paper, which include blood-pressure-related outcomes. Continuous repeated blood pressure is available there only for a subset, and exposure would need a separate satellite product not described in the file.