Air Quality Exposure

How to validate satellite-derived exposure estimates against personal monitors

A small weather monitoring station standing alone in an open grassland under a clear sky.

A cohort study that assigns exposure from the nearest fixed monitor, often 15 or 30 km from a participant's address, has always had a measurement error problem baked into the design. Satellite-derived surfaces get adopted because they promise something finer than that single-point assignment. But adopting one without checking it against what people actually breathe just trades one unverified assumption for another. The validation step answers a specific question: for a given address and study day, does the satellite estimate agree with what a personal monitor measured on that same person, that same day.

Building the validation subset

Most cohorts can't put a personal monitor on every participant for the full follow-up period, so validation works off a subset, usually somewhere between 30 and 100 people, wearing PM2.5 monitors for a run of consecutive days, sometimes repeated across seasons to capture both winter inversion days and summer wildfire smoke. Nephelometer-based units like a pDR give you continuous readings; filter-based gravimetric sampling gives you an integrated 24-hour mass concentration that's slower but still the reference standard in a lot of panel studies.

The part that trips people up is timing. A personal monitor logs continuously, but a satellite-derived estimate is only as good as the overpass or retrieval window it's built from, so the comparison day has to be defined the same way for both streams, usually as a calendar day average rather than trying to match an instantaneous pass time to a monitor reading. You'll also need to decide which location anchors the comparison: the grid cell containing the home address, the work address, or some blend based on the participant's daily diary. Pick one before you run the analysis, not after you see which one correlates better.

It's worth stating plainly: a personal monitor captures total exposure, indoor and outdoor, weighted by where the person spent their day. A satellite-derived surface captures ambient outdoor concentration at a point on the ground. These are not the same quantity, and a validation exercise that ignores that will overstate how far off the satellite estimate is.

Comparing the two measurement streams

Don't reach for ordinary least squares here. OLS assumes one measure is error-free and the other carries all the noise, and neither a personal monitor nor a satellite retrieval qualifies as a gold standard with zero error. Deming regression or reduced major axis regression treats both sides as measured with error, which is the honest starting point. From there, a mixed-effects model with participant as a random intercept handles the repeated-measures structure, since each person contributes several comparison days that aren't independent of each other.

Bland-Altman plots, adapted for repeated measures, show you where the agreement breaks down: often at the high end of the concentration range, during smoke events, or on days with a lot of time spent indoors near a combustion source like a gas stove or a wood fire. An intraclass correlation coefficient gives you one number for how much of the variance is shared across the two methods, but don't stop there. Plot the residuals against concentration, against season, against urban versus rural grid cells. The pattern in the residuals tells you more than the summary statistic does.

Reading the error structure before you trust it

Once you've got agreement metrics, the next question is whether the mismatch behaves like classical error or Berkson error, because the two have opposite effects on a health-effect estimate downstream. Classical error, where the satellite value varies randomly around the true personal exposure, tends to attenuate your effect estimate toward the null. Berkson error, more typical of a spatially smoothed surface assigned to everyone in a grid cell, widens your confidence intervals without necessarily biasing the point estimate. A downscaled, neighbourhood-resolution product behaves differently here than a coarse regional one, which is part of why a daily retrieval built at a finer spatial grain for the study footprint is worth testing against the panel before you commit a full cohort's exposure assignment to it.

Once the validation subset gives you a calibration equation, bias term, or correction factor, run it both ways: apply it to the full cohort's satellite-derived exposures, then rerun the health model with and without the correction to see if your effect estimate moves. A correction that leaves the effect estimate roughly where it started gives you a defensible exposure surface; a correction that shifts it substantially is worth knowing now, before a reviewer asks for it.

If you're scoping a validation study and want to pull a daily surface over your exact cohort footprint to test it against a personal-monitor panel, that's the offer Air Quality Exposure is built around.

Start a pilot

← Back to the blog