Separating Dose from Avoidance in Air-Pollution Emergency Visits

When air pollution spikes, emergency departments should fill up. Studies of acute pollution often find a nearly straight-line relationship between particulate concentration and emergency demand. Yet at extreme levels, hospital visit counts are widely thought to flatten. The usual reading is that biology saturates. We propose a different explanation that can be tested: during a visible pollution crisis, and especially once an official "stay indoors" alert is issued, people with mild symptoms stop coming to hospital. Hospital counts then mix two signals. One is the true effect of the dose on the body. The other is a behavioral filter that switches on sharply. If that filter is ignored, health-burden estimates for exactly the worst days come out biased downward, and the patients who do arrive look sicker than usual.

The key observation is that biology does not know where an air-quality alert boundary sits, but behavior does. Official alerts flip on when a monitor-based index crosses a fixed category cutoff. The pollutant dose a neighborhood actually receives varies smoothly across that cutoff, and it can be estimated more accurately than the alert-triggering monitor by fusing satellite, weather and sensor data. We therefore treat the alert cutoff as a natural experiment. Any jump in visits at the cutoff that the smooth dose curve cannot explain is avoidance. The jump should appear among walk-in, low-acuity patients and not among ambulance arrivals. Two of the input directions play distinct roles. Multi-source environmental fusion supplies the smooth dose. Healthcare-avoidance research supplies the expected signature of the filter. The tipping-point literature enters as a secondary, falsifiable hypothesis about whether the filter switches abruptly and leaves a delayed rebound.

Core Claim

Hospital emergency visits on high-pollution days can be written as true morbidity multiplied by a care-seeking probability. That probability is not the fixed demographic constant that surveillance corrections assume. In the reconstruction of true from observed incidence summarized by Massad et al., the observed count is divided by a constant $P(\text{Care} \mid \text{Symptomatic})$. We instead write, for acuity level $a$, neighborhood $i$ and hour $t$,

$$\lambda^{\text{obs}}_{a}(i,t) = \lambda^{\text{true}}_{a}\big(d_i(t), Z_{it}\big)\cdot p_a\big(A_{t}, d_i(t)\big),$$

Here $d_i(t)$ is the fused ground-level dose, $Z_{it}$ are ordinary confounders (season, weekday, temperature), $A_t$ indicates whether a public alert is in force, and $p_a$ is the probability that someone at acuity $a$ who has symptoms comes to hospital. Identification rests on two assumptions, and both are stated so they can be attacked. First, $\lambda^{\text{true}}_a$ is continuous in dose at the alert cutoff. Second, the highest-acuity stratum, ambulance arrivals with high triage category, has $p_a$ close to one and roughly unaffected by alerts. Under these assumptions, the discontinuity in the low-to-high acuity ratio at the cutoff estimates the alert-induced drop in $p_a$. Dividing it out yields an avoidance-corrected concentration–response curve.

What would be new is the decomposition itself. The file contains no paper that separates dose effect from care-seeking effect in pollution-health data. The closest work, the Shenzhen ambulance case-crossover study Associations Between Hourly Ambient Particulate Matter Air Pollution and Ambulance Emergency Calls, fits a distributed-lag non-linear model to ambulance calls only, using city-wide monitor averages. It reports a near-linear response with no threshold. It never looks at the walk-in channel, where avoidance would show up.

Why the pieces fit

The healthcare-avoidance literature tells us what signature to look for, although its evidence comes from COVID, not pollution. Vollmer et al. found that London ED attendance fell 35% against an ARIMA (autoregressive time-series) forecast. The fall was driven by walk-ins while ambulance arrivals held up, and it did not vary with distance from hospital. Akhtar et al. found in Qatar that stroke mimics fell from 87 to 34 per month and mild strokes from 77.8% to 65.7% of admissions, while mean severity scores rose. Meyerowitz-Katz et al. found that utilization dropped before mandates took effect. These papers motivate the hypothesis. They also constrain it. Akhtar et al. could not separate true pathology from self-selection, and our design has to do better than that.

The fusion literature supplies the method for dose. Ahmad and Lin built a two-stage gradient-boosted model in which an inner model predicts the pollutant mixing height, the depth of air over which surface emissions spread. Feeding that prediction to the outer model raised cross-validated $R^2$ for ground NO₂ from 0.73 to 0.93. Li et al. used geographically weighted regression, a regression whose coefficients vary by location, on MODIS aerosol data. AirPhyNet fills gaps with a physics-guided network built on advection and diffusion. All three stop at concentration maps. We reuse them for a different reason than accuracy. A neighborhood-level dose that differs from the alert-triggering monitor reading separates the dose a person receives from the message they hear. Without that separation, dose and alert would be almost collinear, and the decomposition would rest only on functional form.

The tipping-point literature enters as a hypothesis, not as the core model. Li, Ye, Qian and Huang show that when a system is pushed past a tipping point at rate $V$, collapse does not happen at once. It is delayed by a window that scales as $(\lambda V)^{-1/3}$, where $\lambda$ sets how sharply the system curves near the tipping point. The secondary mental-health referral study documents a delayed surge of urgent referrals after an initial avoidance drop. Together these suggest that avoided mild visits may resurface days later as urgent visits. The delay might also shrink when the pollution ramp-up is faster. Li et al. analyze a deterministic, single-variable system, and a population aggregate need not obey the scaling law. We treat this as an exploratory test.

Research Program

The work has three stages. In the first, we build an hourly, neighborhood-scale PM2.5 and NO₂ dose surface for the study city. It follows the nested mixing-height design of Ahmad and Lin. Night and cloud gaps are filled with an AirPhyNet-style transport model run from the released kethmih/AirPhyNet code. We also reconstruct the exact monitor-based index that triggered each public alert.

In the second stage, we assemble emergency encounters that carry arrival mode and triage acuity. The natural target is a Chinese city with frequent alert days, so that hospital ED records can be paired with an ambulance dispatch registry of the Shenzhen type. The file does not identify a dataset that already links ED arrival mode to pollution episodes. Securing one is the main practical risk, and the plan depends on it. We then fit the decomposition above as a Poisson model. Dose enters through a smooth spline that is continuous across the cutoff. The alert indicator and its interaction with acuity and arrival mode enter as jump terms. The fit uses a regression-discontinuity bandwidth chosen by cross-validation.

In the third stage, we divide out the estimated $p_a$ to produce corrected curves. We then test the delayed-rebound and rate-dependence hypotheses on post-episode urgent visits.

Baselines

The naive reference point is the uncorrected count model: a time-stratified Poisson regression of all ED visits on monitor-average PM2.5 with the usual calendar and weather controls. If avoidance exists, this baseline will show the flattening at high doses that we are trying to explain. It is run on identical data and controls.

The strongest existing method is the distributed-lag non-linear model, which lets the effect of pollution spread over several lagged hours, as applied in the Shenzhen study. That study reports a 1.8% increase in all-cause ambulance calls per 24.0 μg/m³ IQR rise in PM2.5 within 0–4 hours. We run it twice, on ambulance calls and on all ED visits, with the same lag structure, spline degrees of freedom and tuning budget as our model. For the full decomposition there is no reported number to beat: no reported result found on Lacuna for avoidance-corrected concentration–response estimation. Our method counts as better if it does three things. It must predict held-out high-acuity, ambulance-arrival counts on alert days at least as well as the DLNM. It must produce a corrected all-visit curve that agrees with the ambulance-only curve in the high-acuity stratum. And it must recover a known avoidance effect in the positive control described below, which the DLNM by construction cannot attribute.

For exposure, the strongest baseline is the nested mixing-height model. Its reported cross-validated $R^2$ of 0.93 and RMSE of 4.19 μg/m³ for NO₂ come from China-wide GEMS data and are a reference, not a target on our city. We re-fit it on our stations with the same folds.

Two ablations isolate the key ingredients. The first replaces the fused dose with the alert-triggering monitor value itself. This tests whether the dose–message separation that fusion provides changes the avoidance estimate or merely its precision. The second drops acuity and arrival-mode stratification, which tests whether the anchoring assumption carries the identification.

Why not just…?

Why not just fit a flexible spline or DLNM to total ED visits? A spline can reproduce the flattening, but it cannot say whether the flattening is biological or behavioral. Both stories produce the same total-count curve. Only the cutoff discontinuity and the acuity mix can tell them apart. The naive and DLNM baselines are kept in the study precisely so that we can show they fit equally well while implying different burdens.

Why not just analyze ambulance calls, which avoidance barely touches? This is a real threat. Vollmer et al. suggest that ambulance arrivals are largely spared by avoidance, so an ambulance-only DLNM, as in the Shenzhen study, may already give an unbiased curve for severe morbidity. Our proposal adds value only for the low-acuity burden, which is the largest share by volume, and for showing where ambulance-only estimates stop being representative. If corrected all-visit and ambulance-only curves coincide across all acuities, we will report that avoidance is negligible under pollution alerts. In that case, ambulance-only analysis is vindicated.

Why not just run a regression discontinuity at the alert cutoff with no structural model? This is the honest core of the method, and we adopt it rather than compete with it. The multiplicative model adds only the acuity-anchored rescaling needed to turn a jump into a corrected curve. If the plain discontinuity on walk-ins is null, nothing else in the proposal survives.

Why not just blame logistics such as transit disruption or poor visibility? Smog rarely closes roads. Vollmer et al. also found no distance gradient in the COVID drop. We test this directly by stratifying the discontinuity by distance band and transport-access indicators. A behavioral filter predicts flat jumps across distance bands, and a logistical barrier predicts larger jumps far from hospital.

Proposed Experiments

The first check is a positive control. We apply the decomposition to London lockdown attendance with the structure of the ICHNT data described in Vollmer et al., treating the lockdown announcement as the "alert." The method should attribute most of the 35% drop to walk-in, low-acuity care-seeking and leave ambulance, high-acuity counts nearly unchanged. Failure here invalidates the estimator before any pollution claim is made.

The second check uses negative controls. Encounters that cannot be deferred, such as cardiac arrest arrivals by ambulance, should show no jump at the cutoff. Placebo cutoffs placed away from the real category boundaries should also show none.

The decisive pollution experiment estimates the jump in low-acuity walk-in visits at the alert cutoff, conditional on the fused dose. A significant negative jump confined to that stratum, flat across distance bands and stronger under fused dose than under the monitor-only ablation, supports the claim. A null jump means apparent saturation is not avoidance. That result would also be worth reporting, because it would strengthen the no-threshold reading of the Shenzhen study.

Finally, as an exploratory test, we look for an excess of urgent visits in the days after alerts are lifted, analogous to the referral acceleration in the mental-health study. We then ask whether the lag to that excess shrinks with faster pollution ramp-up as $V^{-1/3}$. We expect this scaling only if a single dominant threshold mechanism governs care-seeking, and we would report a mismatch as evidence against carrying the tipping-point formalism over from ecology.