Nearest fixed monitor exposure misclassification: what it does to a cohort study

Most cohort protocols still assign exposure the same way: geocode the residential address, find the closest EPA AQS or state monitor, pull its daily PM2.5 value, done. It's cheap, it's defensible in a methods section, and reviewers rarely push back on it. It's also wrong in a specific, measurable direction for a large share of your cohort, and the direction of that error usually isn't random.
Why the nearest monitor isn't the same as the participant's air
A fixed-site monitor reads one point. PM2.5 at that point correlates with PM2.5 a kilometer away mostly because of regional transport, not because local sources are the same. A participant living near a highway on-ramp, a rail yard, or a wood-burning block has a different local contribution layered on top of that regional background, and the nearest monitor, if it's 8, 15, or 40 km away, never sees that local layer at all. You get the regional signal plus noise that looks like signal.
This is the part that trips people up in review: the error isn't uniform across the cohort. Dense urban areas tend to have more monitors, so urban participants get assigned values from a site that's genuinely closer to their actual exposure. Rural and peri-urban participants, and often lower-income urban blocks with fewer regulatory sites nearby, get assigned a monitor that's farther away and less representative. That's monitor sparsity doing its damage quietly: it doesn't just add noise, it adds noise that correlates with the same socioeconomic and geographic variables you're trying to adjust for or stratify on.
What that does to your effect estimate
Classical measurement error, the kind where the assigned value varies randomly around the true exposure, independent of the true value, tends to attenuate your effect estimate toward the null. That's the pattern everyone cites when they say "our estimate is probably conservative." But nearest-monitor assignment doesn't behave that way in sparse areas. When the gap between monitor and residence is large and systematically worse for some strata, you're closer to Berkson error mixed with a correlated bias, and Berkson error doesn't just add noise to your variance, it can shift the point estimate itself, not just shrink it. Papers that treat monitor-distance error as purely classical and report it as "probably biasing us toward the null" are often being too generous to their own design.
The practical symptom: effect estimates that shift noticeably when you restrict the analysis to participants within 10 km of a monitor versus the full cohort. If your sensitivity analysis shows that split, you already have the misclassification. The question is whether you've quantified it or just footnoted it.
Monitor sparsity by cohort geography
Run the numbers before you write the exposure assignment section. Pull your cohort's residential coordinates, calculate distance to nearest regulatory monitor for every participant, and histogram it by stratum (urban/rural, income quartile, region). If a meaningful fraction of participants sit beyond 15-20 km from any monitor, that's not a footnote, it's a design constraint that belongs in your limitations section with a number attached, not a sentence. Reviewers at exposure-focused journals increasingly ask for this distance distribution as a supplementary table.
What a downscaled surface changes
Satellite aerosol retrievals downscaled to neighbourhood resolution don't replace the ground network, they fill the space between monitor sites with a continuous estimate instead of a step function that jumps at the midpoint between two fixed stations. For a cohort spread across a county or a multi-county catchment, that turns a monitor-distance problem into something closer to a resolution problem, which is a different, more tractable kind of error to characterize. Air Quality Exposure builds that daily surface over the exact study area you define, so the assignment step stops depending on which monitor happened to be closest.
If your protocol still assigns exposure by nearest monitor, it's worth running the distance histogram before your next grant renewal, and worth seeing what a finer surface would do to the same cohort.