Multi-national case-crossover builder

Before you use this design

A case-crossover that breaks one of these four assumptions still returns an odds ratio with a confidence interval. Nothing warns you. Read the plain line first; the paragraph under it is the version for a methods section or a reviewer.

1. 1. The exposure has to be transient

   Something people start and stop — a short course, an as-needed drug, a dose change. Not a drug they simply stay on.

   Why it matters, and what to do

   Why · Only cases whose exposure status differs between the hazard window and at least one control window contribute to the conditional likelihood. A case exposed throughout, or unexposed throughout, contributes nothing; if most of the cohort is a continuous user, the estimate rests on an unrepresentative discordant minority.

   What to do · Report the number and proportion of exposure-discordant cases. If it is small, say so, and consider a cohort or SCCS design instead. [Maclure 1991]

2. 2. The outcome has to be abrupt in onset

   A single identifiable moment — an MI, a fall, a bleed — not the slow onset of a chronic condition.

   Why it matters, and what to do

   Why · Maclure derived the design for a brief exposure producing a transient change in the risk of a rare, acute-onset disease. Rarity is not, however, what licenses the odds ratio to be read as a rate ratio: the referent windows sample the case's own person-time, and under that sampling the exposure odds ratio estimates the incidence rate ratio directly. What abruptness buys is a usable index date — gradual onset makes it arbitrary, which mis-places every window relative to the true event time, and an event whose onset precedes its record by an unknown lag pushes the true aetiological moment out of the hazard window altogether.

   What to do · State how the index date is defined and how well it approximates onset (admission date is a proxy, not onset). [Maclure 1991]

3. 3. Exposure prevalence must not trend over time within a person

   If people were simply becoming more (or less) likely to be on the drug as time went on, the design finds an effect that is not there.

   Why it matters, and what to do

   Why · The hazard window is always more recent than the control windows, so any monotone within-person trend in exposure probability — a drug being adopted, a dose being escalated as prognosis worsens, a treatment being stopped as a patient becomes frail — is confounded with the effect of interest. This is exposure-time-trend bias, and it does not shrink with sample size.

   What to do · Choose a variant that addresses it, and be clear which trend it addresses. Time-stratified referent selection (Janes 2005) removes bias from calendar-time trends in exposure prevalence. A case-time-control (Suissa 1995) or case-case-time-control (Wang 2011; Huang/Yang/Tsai BMJ 2023) subtracts the portion of the trend shared with a comparison series. None of them removes a trend specific to the individual — a dose escalated as that person deteriorates — so if that is the mechanism you fear, say so and argue it separately. Report the uncorrected and corrected estimates side by side. [Suissa 1995; Janes 2005; Wang 2011]

4. 4. The windows must not overlap, and control time must be eligible person-time

   Hazard and control windows have to be separate stretches of time in which the person was alive, enrolled, and could have been dispensed the drug.

   Why it matters, and what to do

   Why · Overlapping windows double-count person-time and break the conditioning; control windows falling before database enrolment, before the drug was marketed, or before the person entered the age band produce a structural zero that is read as 'unexposed'. Immeasurable time (in-hospital days with no outpatient dispensing record) has the same effect.

   What to do · Require continuous enrolment covering the earliest control window; report how in-hospital time was handled. The builder checks window overlap for you below. [Maclure 1991; Janes 2005]