When Pixels Bend Dose-Response Curves: Sub-Grid Exposure Spread as a Source of Contradictory Risk Estimates

Environmental health studies rarely know what each person actually breathed. They assign everyone in a grid cell, often a 4 km or 24 km square from an air-quality model or a satellite pixel, the cell's average concentration. Then they fit a flexible dose-response curve. If the true curve is bent (U-shaped, L-shaped, or with a threshold), averaging exposure before applying the curve does not give the average risk. The resulting distortion is not fixed noise. It depends on how much exposure varies inside each cell and on whether that within-cell spread grows with the cell's mean. Two cohorts with the same biology can therefore report different curve shapes simply because they used different grids, or because their cities differ in how sharply pollution falls off around roads.

We propose to test that claim and to build a cheap fix. The fix replaces "the curve evaluated at the cell mean" with "the curve averaged over a predicted within-cell exposure distribution." That distribution comes from satellite-conditioned dense pollution maps, which resolve finer structure than the cell average. The epidemiology side supplies the curve shapes and the inconsistency to explain. The remote-sensing side supplies the mechanism (pixels as spatial averages) and the data needed to estimate within-cell spread.

Core Claim

Write the true individual risk as $g(c)$ for exposure $c$. Suppose a cell has mean $\mu$ and within-cell, population-weighted exposure variance $\tau^2(\mu)$. Here $\tau^2$ is written as a function of $\mu$ because, near emission sources, high-mean cells also tend to be internally heterogeneous. A second-order expansion of the cell-average risk gives

$$\mathbb{E}[g(C)\mid \text{cell}] \approx g(\mu) + \tfrac{1}{2} g''(\mu)\,\tau^2(\mu).$$

Differentiating with respect to $\mu$, the slope an analyst recovers from cell means is

$$\frac{d}{d\mu}\mathbb{E}[g(C)\mid\mu] \approx g'(\mu) + \tfrac{1}{2}\big(g'''(\mu)\,\tau^2(\mu) + g''(\mu)\,\tfrac{d\tau^2}{d\mu}\big).$$

This yields a sharper and less obvious statement than "Jensen's inequality biases non-linear models." If the within-cell spread is constant and $g$ is close to quadratic, the curve shifts up or down but keeps its shape, so the problem is mostly harmless. Shape distortion, including a moved U-shape minimum, a blurred threshold, or a spurious slope, requires one of two things. Either $\tau^2$ changes with $\mu$, or $g$ has kinks or plateaus that the expansion misses.

The relationship between spread and mean differs across cities and grid resolutions. The theory therefore predicts which pairs of studies should disagree and in which direction. The empirical claim is that, on real exposure fields, this mechanism produces curve-shape disagreements between grid resolutions large enough to matter. A second claim is that averaging a spline basis over a predicted within-cell distribution removes most of that disagreement.

Relation to Prior Work

The curve shapes we take as ground-truth templates come from nutritional and clinical epidemiology, where restricted cubic splines (piecewise cubic curves forced to be linear in the tails) have repeatedly found bent responses:

  • a U-shape with a minimum near 4.2 mmol/L for serum potassium and mortality in Fan et al.'s meta-analysis;
  • an L-shape saturating near 9.3 mg/day for vitamin E and cognition in Zhang et al.;
  • a tipping point at 25 g/day of sugar in Yu et al..

These papers motivate the shapes. They involve measured individual exposures, not gridded ones, so we use them only as templates for how bent real curves are, not as air-pollution evidence. The investigator-measured BMI meta-analysis motivates the broader point that the exposure measurement method alone can move effect sizes: 33–44% excess risk with measured BMI versus roughly 19–24% with self-report.

Two papers supply the exposure data:

  • Kleeman et al. produced nested 24 km and 4 km chemical-transport fields for California health studies. They recommended averaging over at least a month, but did not examine how spatial averaging interacts with bent risk curves. Their nested grids give us a natural pair of resolutions over the same population.
  • Gonzalez Avilés et al. built dense ground-level NO₂ maps by fusing coarse Sentinel-5P columns, 10 m Sentinel-2 imagery, and land cover, and evaluated them only against stations. We reuse their approach as the source of within-cell spread, which they never connected to health estimation.

Richardson et al. constrain the design. Their corrected-score estimators in mismex fix confounding plus measurement error, but they assume independent additive error with known variance and a linear exposure effect. Gridded exposure violates all three assumptions: the error is shared within a cell (so-called Berkson error, where truth scatters around an assigned value), it varies across cells, and the effect is non-linear.

Deng et al. observed that ecological risk conclusions on a 3 km grid change with grid choice (the modifiable areal unit problem), but offered no mechanism. Our expansion gives one for the specific case of bent responses.

Research Program

The first component is the estimator. A restricted cubic spline writes $g(c)=\sum_k \beta_k B_k(c)$ for fixed basis functions $B_k$. Under an additive-risk or rare-outcome log-linear approximation, the cell-level model becomes $\sum_k \beta_k\,\mathbb{E}[B_k(C)\mid\text{cell}]$. The model stays linear in $\beta$. Analysts can therefore keep their usual software and simply swap each basis column $B_k(\bar c)$ for its expectation under the predicted within-cell distribution. For logistic or survival outcomes we use Monte Carlo draws from that distribution inside the likelihood.

The second component is the within-cell distribution. For each coarse cell, we take the dense field's values at residential locations, using WorldCover built-up classes as a population proxy. We then widen this spread with the dense model's own prediction error. That error is calibrated by land-cover class on held-out stations, because the dense map is itself smoothed and imperfect.

The third component is analysis. We will estimate $\tau^2(\mu)$ empirically on both California PM₂.₅ and European NO₂. We then check whether the predicted direction of curve distortion, sign of $g''$ times sign of $d\tau^2/d\mu$, matches what simulation shows.

Baselines

The naive reference fits a restricted cubic spline to the cell-mean exposure, as is standard practice. It uses the same synthetic cohorts, knots, and confounders as every other method.

The fine-resolution plug-in assigns each synthetic person the dense-map value at their location and fits the same spline. This is the strongest practical competitor, because it is what a group with access to Gonzalez Avilés-style maps would do. The closest reported number concerns exposure accuracy, not health: their dense autoencoder reached an MAE of 4.98 µg/m³ against 5.50 µg/m³ for point-wise models (Gonzalez Avilés et al.). No reported result was found on Lacuna for dose-response curve recovery from any gridded or dense exposure product. Our method beats this baseline only if it recovers the true curve with lower error and calibrated intervals, given that the dense map itself carries error.

The strongest existing measurement-error method is mismex corrected-score estimation (Richardson et al.). We give it the true error variance and run it with a linear exposure term and with a binned exposure. No reported number was found on Lacuna for its performance under spatially shared or non-linear settings.

The key ablation is the variance-covariate (Taylor) correction. It adds $\tfrac12 \hat g''(\mu)\hat\tau^2$ terms using only the within-cell mean and variance, not the full distribution. This isolates whether the full distribution matters beyond its second moment.

All methods share the same cohorts and spline knots, and receive equal tuning effort on a separate validation split.

Why not just…?

Why not just use the fine-resolution dense map at each address? This is the real threat. If the dense map were nearly exact, plugging it in would solve the problem with no new estimator. It is not exact: its validation MAE of about 5 µg/m³ is a sizable fraction of typical urban NO₂ levels. Its own authors also note that the autoencoder produces smoother fields than UNets, which compresses peaks. Plugging it in therefore just moves the averaging problem to a finer scale. The experiments settle this directly. If the plug-in recovers the curve as well as our method once realistic dense-map error is simulated, we will report that the fine map, and not our estimator, is the fix.

Why not just average over longer time windows? Kleeman et al. show that monthly averages meet performance criteria in time. Our distortion, however, comes from spatial spread within a cell, which persists in annual means: a home beside a freight corridor stays beside it. We will test this by repeating the analysis on daily, monthly, and annual Kleeman fields. The prediction is that the distortion shrinks only to the extent that $\tau^2(\mu)$ shrinks.

Why not just apply mismex or regression calibration? Corrected scores assume independent, additive error with a linear effect. Under shared within-cell error and a U-shaped truth, a linear corrected slope estimates the wrong quantity no matter how well the error is corrected. The baseline run will show how far it falls short.

Why not just add within-cell variance as a covariate? That is our Taylor ablation. If it matches the full-distribution estimator, then the simpler version is the recommendation. We expect it to fail mainly for threshold and L-shaped curves, where the expansion breaks down at the kink.

Proposed Experiments

The main experiment is semi-synthetic. We will place synthetic individuals by built-up land cover across California, using Kleeman's 4 km and 24 km PM₂.₅ fields, and across the European NO₂ station domain of Gonzalez Avilés et al. Each person receives an exposure drawn from the dense field plus calibrated error. Outcomes come from U, L, and threshold curves scaled to match the shapes cited above.

Metrics are:

  • integrated squared error of the recovered curve;
  • error in the location of the minimum or threshold;
  • 95% interval coverage;
  • an inconsistency index, defined as the integrated difference between curves estimated at 4 km and at 24 km over the same people.

The decisive check is whether our estimator shrinks the inconsistency index toward zero while the naive spline does not. We will also check whether the sign of each method's distortion matches the sign predicted by $g''\,d\tau^2/d\mu$.

A second, real-data check targets the one ingredient the method depends on: are the predicted within-cell distributions calibrated? We hold out ground stations and measure how often their observed concentrations fall inside the predicted within-cell quantiles. Poor coverage would invalidate the approach regardless of the simulation results.

No geocoded health cohort appears in the file. A real-outcome replication would therefore require partnering with an existing cohort, and we state this as a limitation rather than a claim.