Matched Case-Crossover Design — Time-Matched Self-Controlled Epidemiological Study

The original case-crossover design can be biased when exposure trends occur over calendar time (e.g., pollution rising in winter) or vary by day of week, because control windows at different calendar positions will systematically differ from the hazard window. Explicit matching of control windows to the hazard window on day of week and calendar period removes these sources of confounding, producing an unbiased comparison under the assumption that the matching variables fully capture the temporal structure of confounding.

Bidirectional designs — placing control windows both before and after the event — are generally preferred because they are symmetric around the event and therefore not susceptible to secular exposure trends. Unidirectional designs using only prior windows are appropriate only when the event alters subsequent exposure (e.g., hospitalisation changes ambient exposure), making post-event windows uninformative. Otherwise, unidirectional designs should be avoided because they are sensitive to any monotone trend in exposure.

The window length should match the clinically plausible induction period — the interval within which exposure is believed to trigger the outcome. For pollution-triggered cardiovascular events, this is typically 0–24 hours; for physical exertion, it may be as short as one hour. Control windows must be exactly the same length. Using a window that is too long dilutes the exposure contrast; using one that is too short may miss the biologically relevant time lag.

Only partially. Time-invariant confounders are eliminated by the within-person comparison. Time-varying confounders that are explicitly matched (day of week, season) are controlled by the matching. Time-varying confounders that are not matched and not constant over the study period (e.g., concurrent medication use, disease progression) remain potential sources of bias and should be measured and included as covariates in the conditional logistic regression.

There is no universal threshold, but power in the matched case-crossover design is driven by the number of discordant matched sets — that is, case-control window pairs where the exposure status differs between the hazard and control windows. If exposure is rare and nearly constant within most individuals, even large case samples yield few discordant sets and the study will be underpowered. A formal power calculation based on the expected discordance fraction is strongly recommended before data collection or record linkage.