Air Quality Exposure

Berkson vs classical measurement error in air pollution exposure assignment

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

When epidemiologists talk about measurement error in exposure assignment, the conversation around Berkson error and classical error splits air pollution work from most other fields that worry about mismeasurement. Getting the distinction right matters, because it changes what you can say about bias, not just precision, in your regression output.

Classical error: the exposure varies around your measurement

Classical measurement error is the textbook case: your measured value is the true exposure plus independent noise. Personal monitors worn by a subset of your cohort for a week are a good example. The reading varies around each person's true exposure in both directions, and the error bounces independently of the true value. In this setup the standard result holds: classical error on a continuous exposure attenuates your regression coefficient toward the null, and the shrinkage gets worse as measurement noise grows relative to the between-subject variance in true exposure. If you're running a two-stage design where a smaller monitored subsample calibrates a surrogate for the full cohort, you're squarely in classical-error territory, and it's worth running regression calibration or SIMEX rather than reporting the naive coefficient on its own.

Berkson error: everyone near the monitor gets the same number

Berkson error flips the structure around. Instead of your measurement varying around someone's true exposure, you assign the same value to everyone in a spatial unit, a census tract, a nearest-monitor catchment, a model grid cell, and their true individual exposures vary around that assigned number instead. This is the dominant error structure in air pollution cohort work, because nearly every exposure assignment method, nearest-monitor assignment, land-use regression, kriged surfaces, chemical transport model output, works by interpolating or averaging onto a shared surface and then reading off one value per participant address.

The practical consequence is different from classical error, and it's the part reviewers get wrong most often. Berkson error on its own doesn't bias your point estimate toward the null in a simple linear model. What it does is inflate the variance of the true-exposure distribution you're regressing against, which widens your confidence intervals and costs you power, especially with a rare outcome or a cohort on the smaller side. A reviewer who flags "nearest-monitor assignment biases your effect estimate" is usually importing the classical-error result into a setting where it doesn't apply.

Why the two rarely show up alone in practice

Real exposure pipelines mix both. A chemical transport model output carries Berkson structure at the grid-cell level, one value shared across everyone inside a few kilometers, but the model itself has classical-type error against ground truth, because its emissions inventories and meteorology inputs carry uncertainty that doesn't track any individual's true exposure. Land-use regression has Berkson error in how a fitted surface gets assigned to an address, and classical-type error sitting in the fitted coefficients themselves. Exposure-assessment methodologists have pushed for a while now to separate "prediction error," which behaves like Berkson and mostly costs precision, from "input error," which behaves like classical and costs you bias, rather than lump the two together as one problem.

The resolution grid you build exposure on changes which error dominates. Nearest-monitor assignment over a sparse network produces large Berkson error, because the spatial unit sharing one value can span tens of kilometers, with plenty of within-unit heterogeneity in traffic, land cover, and local sources packed inside it. A daily satellite-derived surface downscaled closer to neighbourhood resolution shrinks that shared unit toward the scale participants live and move within, which is a lever you have before you even reach for a statistical correction. That's the thinking behind the downscaled daily retrievals described on the Air Quality Exposure home page: the aim is simply to cut down the size of the unit everyone in a tract gets lumped into.

Identify which error structure you're dealing with before you pick a correction method. Regression calibration and SIMEX are built for classical error. They'll do the wrong thing, or nothing useful, if what you actually have is Berkson error coming off a coarse assignment grid. Start there, and the rest of your sensitivity analysis plan follows from it.

If your cohort's exposure assignment still runs on a sparse monitor network, it's worth checking what a finer assignment grid does to the Berkson component before the design gets locked in.

Start a pilot

← Back to the blog