Air Quality Exposure

What buffer radius to use for satellite PM2.5 exposure assignment

Aerial view from above a layer of clouds, with patches of green landscape visible below.

Every exposure assignment protocol eventually hits this question: do you take the satellite PM2.5 value at the pixel containing the geocoded residence, or do you average pixels within some radius around it? The radius should follow from two things: the resolution of your source data, and the geocoding uncertainty in your cohort.

Point estimate or buffer average

A point estimate pulls the single grid cell value at (or nearest to) the residential coordinate. It's simple, it's defensible in a methods section, and it preserves whatever spatial contrast exists in the surface. The problem shows up when your geocoding isn't precise enough to trust a single pixel. If your addresses are matched to a ZIP+4 centroid or a rooftop match with a few hundred meters of slop, a point estimate from a 1 km product is really sampling a somewhat random neighbor of the true pixel, and you've added noise you didn't need to add.

A buffer average smooths over that slop by taking the mean of all pixels within a fixed radius, commonly something in the 1 to 10 km range depending on the cohort. It trades precision for stability: less sensitive to small geocoding errors, but also less sensitive to real block-to-block differences in PM2.5 that might matter for a study of, say, proximity to a highway or an industrial source. If your exposure contrast of interest is at the neighborhood scale, a buffer that's too large will wash it out before it ever reaches your regression.

Sizing the radius to your data's resolution

The radius question only makes sense relative to pixel size. A 3 km buffer around a residence sitting on a 10 km pixel retrieval doesn't do anything. You're still assigning one value to everyone inside that pixel regardless of what radius you write down in the methods. A 3 km buffer on a 1 km product, by contrast, averages across roughly 25 to 30 pixels, which can meaningfully dampen a point source signal that a single-pixel assignment would have caught.

A rule of thumb that holds up across a few different exposure assessment papers: pick a buffer no smaller than your geocoding uncertainty, so you're not assigning spurious precision, and no larger than the spatial scale of the contrast you're studying, so you're not averaging it away. For a cohort with rooftop-level geocoding and a research question about traffic-related PM2.5, that often lands somewhere under 1 km. For a cohort geocoded at the ZIP code level studying regional background exposure, 5 km or more may be appropriate. There isn't a single correct number, which is why most protocols run a sensitivity analysis across two or three radii and report whether the effect estimate holds.

This is also where source resolution changes the calculus more than people expect. Coarse retrievals at ≥250 m or coarser, even after statistical downscaling, still carry some footprint from the native pixel they were derived from. If the underlying product is built to resolve neighborhood-scale contrast in the first place, the buffer-radius decision gets a lot less consequential, because a point estimate and a small buffer average start converging on the same answer. That's the gap Air Quality Exposure was built to close: daily aerosol retrievals downscaled to neighborhood resolution over the exact study area you define, so your exposure assignment isn't fighting the resolution of the input data before it even gets to the epidemiology.

What to put in the methods section

Whatever radius you pick, document the pixel size of the underlying product, the geocoding method and its stated precision, and the radii you tested in sensitivity analysis. Reviewers increasingly expect this, partly because buffer choice has been shown to shift effect estimates in reanalyses, and partly because it's one of the few exposure assignment decisions a reader can actually check against your data description. If you can't state the native pixel resolution, you can't justify the buffer radius, and that's usually the first thing a careful reviewer will ask.

If your cohort's exposure assignment keeps running into this problem because the underlying imagery is coarser than your study area needs, it may be worth seeing what a neighborhood-resolution daily retrieval does to your buffer sensitivity analysis before you lock the protocol.

Start a pilot

← Back to the blog