Risk-Adjusted Case-Crossover Design

The standard case-crossover design relies solely on the self-matching to control confounding, which eliminates only time-invariant individual characteristics. The risk-adjusted version additionally models time-varying covariates — such as concurrent drug use, temperature, or season — as explicit terms in the conditional logistic regression, removing residual confounding that the self-match alone cannot handle.

The time-stratified referent selection strategy is generally recommended because it avoids secular-trend bias by selecting control days from the same calendar stratum (e.g., same month and year) as the case day. Bidirectional symmetric designs are an acceptable alternative. Unidirectional strategies (only earlier or only later control periods) should be avoided unless you can demonstrate that exposure is trend-free over the study period.

Conditional logistic regression conditions on the matched set formed by each individual's case and control periods, which correctly accounts for the within-person pairing. Standard logistic regression would ignore the matching structure, produce invalid standard errors, and fail to cancel the time-invariant confounders that the self-matched design is designed to eliminate.

Yes — the risk-adjusted case-crossover design is a mainstay of pharmacoepidemiology with administrative claims data. Each patient who experiences an acute outcome serves as their own control by comparing drug exposure in the hazard window to control windows earlier in their claims history. Time-varying covariates such as concurrent drug fills, calendar season, and hospitalisation history can be extracted from the same database and included as adjusters.

Sample size depends on the rarity of the outcome, the expected odds ratio, and the within-person exposure variability. Because each case contributes its own controls, the effective sample size is the number of cases — not the total population. Simulations or published formulas for matched case-control studies can be adapted; a minimum of several dozen cases is typically needed to detect moderate effect sizes, and hundreds of cases are preferred for stable adjusted estimates.