Air Quality Exposure

How to assign address-level exposure in a cohort study

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

Every exposure assignment protocol starts the same way: a list of geocoded addresses and a question nobody wants to answer out loud. Which pollutant surface do we pull from, and how far off is it going to be for the participant who lives fourteen kilometers from the nearest regulatory monitor?

That distance problem is the actual obstacle in most cohort studies. Not the geocoding (that's tedious but solvable), not the statistics (that's downstream). It's the fact that ambient monitoring networks were built to track compliance with air quality standards across a region, not to characterize exposure at ten thousand individual front doors.

Why nearest-monitor assignment falls apart in cohort work

The default approach, assigning each participant the reading from the nearest fixed-site monitor, treats a 20 km catchment as uniform. It isn't. PM2.5 varies with traffic density, local combustion sources, terrain, and meteorology at a scale monitors were never dense enough to resolve. A participant living next to an arterial road and one living in a quiet cul-de-sac three kilometers away can get the same assigned value even though their actual exposure differs meaningfully.

This matters more in rural and peri-urban cohorts, where monitor spacing can run 30-50 km, than in dense urban samples with several stations nearby. If your cohort spans both, the measurement error isn't just large, it's differential: better for urban participants, worse for rural ones, which can bias exposure-response estimates in ways that are hard to diagnose after the fact.

Methods for filling the gaps between monitors

Epidemiologists have a few standard tools for this. Land-use regression models exposure as a function of road density, population, elevation, and land cover around each address, calibrated against monitor data. Kriging and other spatial interpolation methods smooth between known monitor values using distance-weighted or geostatistical functions. Chemical transport models simulate pollutant dispersion from emissions inventories and meteorology.

Each has tradeoffs. LUR needs enough monitors within the study region to calibrate against, which rural and understudied regions often lack. Interpolation methods inherit the spatial resolution of the underlying monitor network, so gaps stay gaps. Transport models are compute-heavy and their output resolution is usually coarser than a neighborhood.

Satellite-derived aerosol optical depth, downscaled and combined with ground monitor data, gets around the monitor-density problem because coverage comes from the imaging swath rather than from wherever a state agency happened to site a station. Pixel resolution and revisit cadence vary by platform, so validate any given surface against your own study's monitors for your own study area before you build exposure estimates on top of it.

Residential history and address changes

Cohort studies rarely have one address per participant. People move, and follow-up exposure has to track where they actually lived during the exposure window relevant to the outcome you're studying, not just their enrollment address. That means geocoding every address in the residential history, time-weighting exposure across address periods, and deciding how to handle gaps (missing move dates, unverifiable prior addresses, PO boxes used in place of a street address).

A workable protocol geocodes the full history first, flags addresses that fail to match at a sufficient precision tier, and only then layers an exposure surface on top. Doing it in the other order tends to hide geocoding failures inside exposure estimates that look plausible but aren't tied to a real location.

None of this is a solved problem with a single right answer. It's a set of tradeoffs between what your monitor network can support, what your cohort's geography demands, and how much residential mobility your follow-up period has to account for. If gaps between monitors are the bottleneck for your protocol, a daily satellite-derived exposure surface built around your study's footprint is worth comparing against whatever interpolation method you're using now.

Run the comparison on a handful of addresses you already have ground-truth monitor data for, and see where the two methods agree and where they don't.

Start a pilot

← Back to the blog