Case-Crossover Design — Self-Matched Epidemiological Study of Transient Exposures
- Introduced by
- Malcolm Maclure
- Year
- 1991
- Type
- Observational epidemiological study design
- Family
- Clinical / epidemiology
What does it do?
The case-crossover design is an observational epidemiological method that estimates whether a transient exposure triggers an acute event by comparing each case's exposure during a brief hazard window immediately before the event to their own exposure during earlier control periods.
Because each person serves as their own control, all stable personal characteristics are automatically adjusted for, making the design especially powerful for studying intermittent exposures and sudden-onset outcomes such as myocardial infarction, stroke, or injury.
The intuition behind the method
Imagine you want to know whether heavy physical exertion triggers heart attacks. Instead of comparing people who had heart attacks with different people who did not, the case-crossover design asks: was this person more physically active in the hour before their heart attack than they typically were during similar hours on prior days or weeks? Each patient answers this question using their own history as the reference. This within-person comparison removes confounding by everything that is constant about that individual — their genetics, chronic conditions, and lifestyle — focusing the analysis purely on the transient exposure of interest.
History
The case-crossover design was introduced by Malcolm Maclure in a 1991 paper in the American Journal of Epidemiology as a method specifically tailored to studying whether brief exposures trigger acute-onset events. The design borrows the logic of the crossover clinical trial but applies it observationally: each case provides their own control data from a prior reference period. Murray Mittleman and colleagues refined control-period selection strategies in 1995, and subsequent methodological work by Navidi (1998) and Lumley and Levy (2000) addressed trend bias and the bidirectional versus unidirectional debate. The design has since been widely adopted in cardiovascular epidemiology, environmental health research, and pharmacoepidemiology.
When to use it
Use the case-crossover design when your research question is whether a transient, intermittent exposure acutely triggers a sudden-onset event (e.g., does air pollution spike the risk of asthma attack in the next few hours?). It is ideal when stable confounders are a major concern, because within-person matching eliminates them automatically. The design requires that the exposure varies within the same person over short time intervals and that the induction time between exposure and event is short and well understood. Do NOT use it when the exposure is stable or changes only slowly over a lifetime (e.g., smoking status, genetic factors) — within-person variation will be near zero, making the design uninformative. Avoid it also when the event is rare but the outcome has a long latency, or when recall bias of the exposure cannot be adequately controlled.
Limitations and considerations
- Can only study transient exposures that vary within the same person over the relevant timescale — stable exposures cannot be assessed.
- Recall bias is a serious concern when exposure ascertainment relies on participant memory, particularly when cases may unconsciously over-report pre-event exposure.
- Trend bias (secular changes in exposure over time) can affect unidirectional designs; bidirectional designs may introduce exposure dependency bias if outcomes alter subsequent exposure.
- The induction period — the lag between exposure and event — must be specified a priori; misspecification of the hazard window can substantially distort estimates.
- Cannot estimate absolute event rates or population-attributable risks; only relative measures (odds ratios) are produced.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Complete control for all fixed individual-level confounders (genetics, stable lifestyle, socioeconomic status) by design.
- Efficient — only cases are needed; no separate control group must be recruited.
- Particularly suited to studying transient environmental or behavioural triggers of acute cardiovascular, respiratory, or injury events.
- Matches naturally to administrative data sources (hospital admissions, environmental monitoring data) allowing large-scale analyses.
- Reduces selection bias that arises when choosing an appropriate external control group.
Applications
- Estimating the short-term triggering risk of myocardial infarction associated with strenuous physical exertion, anger, or sexual activity.
- Assessing whether short-term spikes in ambient air pollution (PM2.5, ozone) increase the hourly risk of asthma emergency visits or cardiovascular hospitalisation.
- Investigating whether alcohol consumption in the hours preceding a motor vehicle crash acts as a trigger.
- Pharmacoepidemiological studies of drug-event associations where the drug is taken intermittently and the adverse event onset is acute.
- Studying whether occupational exposures on specific shifts or tasks trigger acute musculoskeletal or cardiovascular events.
- Environmental epidemiology linking weather extremes (heat waves, cold snaps) to acute outcomes using person-level matched time windows.
How it works
How it works
All individuals who experience the acute outcome of interest (e.g., myocardial infarction, fall, traffic crash) during the study period are enrolled as cases. Only people who have actually experienced the event contribute data; there is no separate group of non-cases to recruit, making the design efficient in settings where case registries exist.
A hazard window is defined as the brief interval immediately preceding the event — for example, the one or two hours before symptom onset. Exposure status during this window is the key variable: was the triggering exposure (e.g., strenuous exertion, air pollution spike, alcohol intake) present or absent?
Control periods are time windows from the same individual's recent past when the event did not occur. Common strategies are bidirectional (periods both before and after the event) and unidirectional (periods only before the event). Control periods are typically matched on time of day and day of week to avoid confounding by circadian or weekly rhythms.
Exposure during both the hazard period and all control periods is ascertained — ideally from objective records (electronic health data, environmental monitors, activity trackers) or structured interview recall. Exposure must be transient and capable of varying within the same person over time; stable, time-invariant exposures cannot be studied with this design.
Because each person contributes one hazard period and one or more control periods, the data are analysed using conditional logistic regression matched on subject identity. The resulting odds ratio estimates the relative risk of the event following the exposure compared with non-exposure periods for the same individual.
Formula
OR = exp(beta), from conditional logistic regression matched on individual
Common pitfalls
- Misspecifying the hazard window length — if the window is too wide it dilutes the transient effect; too narrow and true exposure is missed. The choice must be biologically justified, not data-driven.
- Ignoring trend bias in unidirectional designs — if a person's baseline exposure is trending upward over time, control periods from the past will systematically show lower exposure than the hazard period, inflating the odds ratio.
- Using the design for slowly changing exposures — when exposure barely varies within a person (e.g., body mass index, smoking intensity over a few weeks), conditional logistic regression will be based on nearly no within-person contrast, producing unstable or uninterpretable estimates.
- Overlooking exposure dependency — in bidirectional designs, the event itself may change subsequent exposure (e.g., a heart attack leads to reduced exertion), biasing control-period exposure estimates.
Frequently asked questions
How does the case-crossover design differ from a standard case-control study?
In a traditional case-control study, cases (people who had the event) are compared with separately recruited controls (people who did not). Confounding by stable characteristics — age, sex, lifestyle — must be handled by matching or adjustment. In the case-crossover design, each case serves as their own control by contributing exposure data from their own earlier time periods. This eliminates all fixed personal confounders automatically but limits the design to transient exposures.
What is the difference between the unidirectional and bidirectional case-crossover design?
In the unidirectional design, control periods are drawn only from time before the event, avoiding any exposure changes the event itself might cause. However, if exposure is trending upward over time, this introduces trend bias. The bidirectional design takes control periods both before and after the event, which cancels time trends but may introduce exposure-dependency bias if the event alters subsequent exposure. The choice depends on whether exposure trends or event-driven exposure changes are the greater concern.
What statistical method is used to analyse a case-crossover study?
Conditional logistic regression matched on individual identity is the standard approach. Each stratum (one hazard period and one or more control periods from the same person) is analysed jointly, and the model estimates an odds ratio for the transient exposure effect net of all within-person stable characteristics.
Can I study more than one transient exposure simultaneously?
Yes. Multiple transient exposures can be included in the conditional logistic regression model simultaneously, provided each exposure varies sufficiently within individuals across the hazard and control periods. Interaction terms between exposures (e.g., exertion combined with anger) can also be examined, as demonstrated in early cardiovascular trigger studies.
Is the case-crossover design appropriate for environmental time-series data at the population level?
The classic case-crossover design uses individual-level matching. For population-level environmental data (daily pollution counts versus daily hospital admissions), time-stratified case-crossover analysis can be used; it is closely related to, and often mathematically equivalent to, Poisson-based time-series regression. The individual-level design is preferred when person-level exposure data are available, as it provides direct within-person causal inference.
Related methods
Sources
Maclure, M. (1991). The case-crossover design: A method for studying transient effects on the risk of acute events. American Journal of Epidemiology, 133(2), 144–153.
Open DOI referenceCopied.Mittleman, M. A., Maclure, M., & Robins, J. M. (1995). Control sampling strategies for case-crossover studies: An assessment of relative efficiency. American Journal of Epidemiology, 142(1), 91–98.
Open DOI referenceCopied.