Air Quality Exposure

Handling residential mobility when assigning exposure in a cohort

Aerial view of a suburban neighborhood with curving streets and many individual houses.

Every longitudinal cohort has movers. Some relocate once between baseline and the first follow-up visit. Others cross three or four addresses over a ten-year study. If your exposure assignment protocol only captures the enrollment address, you're building an exposure history for a cohort that doesn't exist, the one that stayed put.

Address change is person-time, and the exposure file needs to treat it that way, with a start date, an end date, and a location for every residence a participant held during follow-up.

Build the address history before you touch exposure data

Start with the address-period table, not the pollutant surface. For each participant, you need a start date, an end date, and a geocoded location for every residence during follow-up. Most cohorts pull this from a mix of sources: enrollment forms, annual contact updates, forwarding address matches, sometimes a commercial locator service for participants who go quiet between visits. Reconcile these into non-overlapping intervals before you assign a single exposure value to anyone.

Gaps matter. A participant with a six-month unaccounted stretch between address two and address three isn't a data-entry problem to paper over with linear interpolation. Flag it, document your assumption (last known address carried forward is common, and defensible if you say so), and carry that flag into the analysis dataset so reviewers can see where the uncertainty lives.

Short-duration addresses deserve a rule, not a judgment call made address by address. Some teams drop residences under 30 days as transient; others fold them into the adjacent period. Either is fine. Pick one before you start assigning exposure and apply it to the whole cohort.

Weight exposure by person-time, not by address count

Once the address-period table is built, exposure assignment is really a weighted-average problem. A participant who spent 18 months at address A and 6 months at address B before a follow-up visit gets a time-weighted exposure across that interval, not an unweighted mean of the two addresses and not whichever address happens to be on file the day you pull the data.

This is where the resolution of your exposure surface starts to matter more than it did at baseline. If a participant moves four blocks within the same metro area, a county-level or even a 10 km monitor-interpolated surface won't register the move at all; the assigned exposure barely changes even though the actual air at the new address might be meaningfully different, closer to a highway, farther from a park, upwind of a different source mix. A surface that resolves at neighbourhood scale is what makes the address-period weighting exercise worth doing in the first place. Without it, you're weighting precisely by time against an exposure estimate that was never precise by space.

Check for mover bias before you report a null

Movers and stayers are rarely the same kind of participant. People who relocate during a multi-year cohort tend to differ systematically, younger, renting versus owning, responding to a life event like a job change or a new baby, sometimes moving because the old neighbourhood had problems they wanted out of. If your exposure-outcome association looks different in a sensitivity analysis restricted to stayers, that's not a nuisance finding to footnote. It's telling you something about who your residential mobility correction is actually correcting for.

A standard check: run the primary model on the full cohort with time-weighted exposure, then rerun it restricted to participants with zero address changes during follow-up. If the effect estimate shifts meaningfully, report both and say why you think the difference exists, selection into movership, exposure misclassification among movers, or genuine effect modification by time-since-move.

Document the address-period construction method in your methods section with the same rigor you'd apply to the pollutant model itself. Reviewers increasingly ask, and a cohort with a buried or inconsistent mobility-handling rule is a harder paper to defend than one that states it plainly in a sentence or two.

If the bottleneck right now is the exposure surface itself rather than the address logic, Air Quality Exposure builds daily, neighbourhood-resolution estimates over the exact footprint your cohort moves through, so the address-period weighting you've already worked out has something precise to weight against. Get in touch if you want to talk through your study area.

Start a pilot

← Back to the blog