Ecological fallacy vs. individual exposure misclassification

A reviewer flags your exposure assignment and writes "possible ecological fallacy" in the margin. Another reviewer, on the same manuscript, writes "exposure misclassification likely attenuates your estimate." Are they talking about the same problem? No. They're pointing at two separate failure modes that happen to show up in the same paragraph of every air pollution cohort paper's limitations section, and conflating them muddies both your methods write-up and your sensitivity analyses.
Ecological fallacy: the wrong level of analysis
Ecological fallacy happens when you draw conclusions about individuals from data that was only ever measured at the group level. The classic setup: county-level PM2.5 averages matched to county-level mortality rates, with an inference drawn about what pollution does to a person. The bias lives in the mismatch between the unit you measured and the unit you're making claims about. A county with high average exposure and high average mortality tells you nothing about whether the specific people who died were the specific people with high exposure. Durkheim's suicide data and the ecological correlations literature from the 1950s built entire methodological cautions around exactly this trap.
In cohort studies this shows up less often in its pure form, since most cohorts assign exposure at something finer than a county. But it creeps back in whenever you fall back on county- or ZIP-level regulatory monitor averages because that's what's available, and then interpret the result as if it describes individual biological response. If your exposure variable is a group average and your outcome is individual, you've got at least a partial ecological problem baked into the design, independent of how good your statistical model is.
Exposure misclassification: the wrong value for the right person
Individual-level exposure misclassification is a different animal. Here you've correctly assigned exposure at the individual level, say, by geocoding each participant's residence and pulling the nearest monitor reading or a modeled surface value for that point. The problem is that the value assigned is wrong, or at least noisy, relative to what the person actually breathed. A participant living 18 km from the nearest regulatory monitor gets that monitor's reading as a proxy, and on days with local smoke, a nearby construction site, or a temperature inversion trapping traffic pollutants in a valley, the monitor and the participant's actual air can diverge substantially. The unit of analysis is correct. The number assigned to that unit isn't.
Ecological bias doesn't have a predictable direction: it depends on the within-group variance structure, and it can inflate or deflate an effect estimate depending on how exposure and the outcome's individual-level confounders are distributed within your groups. Classical random exposure misclassification, by contrast, tends to bias effect estimates toward the null when it's independent of outcome and symmetric. Differential misclassification, where measurement error correlates with outcome or with participant characteristics like urban/rural residence or socioeconomic status, breaks that assumption and can go either direction. The two biases also call for different corrections on paper: a Monte Carlo sensitivity analysis built for exposure misclassification does nothing to fix an ecological mismatch, and a finer spatial unit does nothing to fix noisy individual assignment.
What actually reduces each one
Fixing ecological fallacy means moving the exposure assignment down to the individual, or at least to a much smaller spatial unit than county or ZIP, so that within-group exposure variance stops swamping between-group variance. Fixing exposure misclassification means closing the distance between the participant's actual residence and the point where exposure was measured or modeled, which is a spatial resolution problem more than a unit-of-analysis problem.
A downscaled daily exposure surface built around your actual study geography addresses the second kind directly, assigning a value tied to each participant's geocoded address rather than a shared regional average. That's the specific gap Air Quality Exposure was built to close, for cohorts that need per-participant daily values rather than the nearest fixed monitor's reading standing in for an entire ZIP code.
If your cohort's exposure assignment still runs through a regional average or a far-off monitor, it's worth seeing what a daily, participant-level surface over your exact study area would look like instead.