Air Quality Exposure

How to choose a lag structure for air pollution and health outcome analysis

Aerial view of a hazy city skyline at dusk, with smog blurring the buildings and a dim sun low over the horizon.

Lag structure decisions get argued about in journal clubs more than almost anything else in this field, and for good reason. Pick a window that's too short and you miss the effect entirely. Pick one that's too long and you're smoothing real signal into noise, or worse, inviting collinearity that makes your coefficients unstable from one model specification to the next.

There's no universal answer here, but there is a sensible process, and it starts before you touch the health outcome data at all.

Start with the biology, not the dataset

The outcome drives the lag, not the other way around. Respiratory and cardiovascular events triggered by acute exposure, think asthma exacerbations, ED visits for arrhythmia, stroke admissions, tend to cluster in the days immediately following a pollution spike. Lag 0-1 or lag 0-2 is standard starting territory for these, often cross-checked against lag 3-5 to rule out a delayed secondary peak.

Chronic outcomes don't work that way. Birth weight, lung function decline, incident cardiovascular disease over a cohort's follow-up period, these respond to cumulative exposure measured in months or years, not days. A single daily PM2.5 value assigned at the ED visit date is close to meaningless for a question like that. You're looking at trimester averages, annual averages, or a moving window keyed to a biologically plausible induction period from an earlier paper on the same outcome.

Mixing these two logics is the most common lag-structure mistake in cohort work: running an acute-style single-day lag against an outcome that's fundamentally a chronic process, then wondering why the effect estimate is null or unstable.

Distributed lag models and the collinearity problem

Once you've settled on acute vs chronic, the next question is whether to model lag as a single day, an unconstrained set of lag terms, or a distributed lag model (DLM) that constrains the lag-response relationship with a spline or polynomial.

Unconstrained lags, putting lag 0 through lag 7 in the same model as separate terms, sound appealingly flexible until you run them. Daily pollution values are autocorrelated by nature; today's PM2.5 predicts tomorrow's reasonably well in most airsheds. Load eight correlated lag terms into one regression and the individual coefficients start swinging wildly between model runs, even though the cumulative effect across all eight stays stable. That cumulative effect, rather than the single-day estimates, is usually the number worth reporting.

A distributed lag non-linear model (DLNM) handles this more gracefully by constraining the lag-response curve with a spline, which cuts the degrees of freedom you're spending and smooths out the autocorrelation noise. The tradeoff is a choice you now have to justify: how many knots, what lag range, natural spline or polynomial. Reviewers will ask you to defend each one. Pre-registering the lag structure, or at minimum the lag range and functional form, before looking at the outcome data saves a lot of back-and-forth at revision stage.

A few things worth checking before you commit

Run a sensitivity analysis across two or three plausible lag windows and report whether the direction and rough magnitude of the effect holds. A 20% change in lag window that flips your result from protective to harmful belongs in the main text, stated plainly, rather than left for a reader to find buried in a supplement.

Check for seasonal confounding independently of lag choice. A long lag window in a single-pollutant model can quietly absorb a seasonal pattern that has nothing to do with the pollutant itself, particularly for pollutants like ozone that track temperature closely.

And underneath all of this sits a quieter problem: the lag structure only means something if the exposure value assigned to each lag day is actually close to what the participant breathed. A fixed monitor forty kilometers from a participant's home address introduces its own kind of noise into every lag term you specify, acute or chronic, and no amount of spline-fitting fixes that upstream. Building a per-participant exposure series at neighbourhood resolution gives the lag structure something real to work with in the first place.

If your cohort's exposure assignment is still riding on the nearest monitor, it's worth seeing what a daily, neighbourhood-scale retrieval over your study area would change in the lag analysis.

Start a pilot

← Back to the blog