Air Quality Exposure

The modifiable areal unit problem in air pollution exposure assessment

A hazy city skyline seen from a distance in soft morning light, buildings fading into the haze toward the horizon.

Every cohort study that assigns air pollution exposure by geographic unit, zip code, census tract, county, is making a boundary choice it usually doesn't examine. The modifiable areal unit problem (MAUP) is the name for what happens when that choice changes your results because the zones changed, even though the underlying pollution didn't.

MAUP has two parts that get conflated in practice. The scale effect is what happens when you aggregate the same pollution surface into bigger or smaller units. A pollutant-health association estimated at the county level often looks different than the same relationship estimated at the tract level, even though nothing about exposure or disease changed between the two runs. The zonation effect is subtler: redraw the boundaries at the same scale, shift a tract line by two blocks, and the estimate moves again. Both effects are artifacts of the unit, not the exposure, and neither one averages out just because you run the model more times.

Where this shows up in a real exposure assignment

Most cohort studies don't deliberately choose an areal unit for pollution exposure. They inherit it from whatever administrative geography the address data already comes tagged with: census tract for socioeconomic covariates, zip code for enrollment records, county for regulatory monitor siting. Exposure then gets assigned by matching each participant's residential unit to whatever pollution value is available for that unit, often a single value per tract pulled from the nearest regulatory monitor or an areal interpolation between monitors.

That's where ecological bias creeps into a nominally individual-level study. The tract is the unit of record, so two participants in the same tract get the same exposure value regardless of whether one lives a block from a highway on-ramp and the other half a mile into a quiet residential grid. The within-tract variance in actual pollutant concentration collapses to a single number, and whatever health effect depends on that variance gets attenuated or distorted depending on how the tract happens to be drawn in that vintage of the boundary file.

Regulatory monitors compound the problem. There are so few of them relative to the heterogeneity of ambient pollution that a single monitor can sit fifteen or twenty kilometers from a given participant's address and still count as the "nearest" one available. Assigning that monitor's value to every participant in its catchment is itself a crude areal unit, just one shaped like a Voronoi cell instead of a census polygon, and it repeats the same problem even though the monitor's own readings are perfectly accurate.

Why finer resolution shrinks the problem without making it disappear

MAUP doesn't go away because a finer grid cell replaces a tract. If you assign exposure from a surface built at, say, 250 meter resolution instead of a tract polygon, you're still aggregating, just at a smaller areal unit. What changes is how much heterogeneity gets averaged out before it reaches your model. A grid cell a few hundred meters on a side captures the difference between the highway-adjacent block and the quiet grid in a way a multi-square-kilometer tract cannot, even though both are, strictly, areal units subject to the same mathematical problem.

For a cohort with geocoded residential addresses, that distinction matters more than another round of post-hoc sensitivity analysis. You can run your model at tract, zip, and county scale and report that the association holds across aggregation levels, but if every one of those levels sits well above the scale at which pollution actually varies block to block, you've only shown the result is insensitive to a set of units that were all too coarse to begin with. The more useful check is whether the exposure surface resolves variation at something close to the scale a participant's residence pattern would experience, which usually means pushing resolution down from administrative-geography scale toward neighborhood scale, then assigning exposure per participant rather than per zone.

Air Quality Exposure builds a daily exposure surface at that finer scale for a study area you define, so exposure gets tied to the participant's address rather than inherited from whichever polygon happens to contain it. If MAUP has been turning up as unexplained variance in your sensitivity checks, it's worth seeing what per-participant assignment at neighborhood resolution does to the estimate.

Start a pilot

← Back to the blog