Case-Crossover Study

Case-crossover study is a self-matched, within-individual observational study design in epidemiology and biostatistics formulated by Malcolm Maclure in 1991 to investigate transient, intermittent exposures that act as proximate triggers of acute-onset clinical events. Rather than comparing exposed cases against a separate group of healthy individuals—as in a traditional between-person case-control study—the case-crossover design enrolls only individuals who have experienced the outcome event (cases) and compares each person's exposure status during a brief hazard window (or case window) directly preceding event onset against that same person's exposure status during one or more reference windows (or control windows) sampled from other times when the event did not occur. Because every case serves as their own genetically and demographically matched control, the case-crossover architecture eliminates confounding by all time-invariant between-person characteristics—whether measured or unmeasured—such as germline genetics, biological sex, socioeconomic status, baseline occupational history, and chronic underlying disease severity.

Answering "Why Now?" Versus "Why Me?"

Maclure introduced the case-crossover method to resolve persistent control-selection and confounding by indication challenges when studying triggers of acute myocardial infarction. Whereas a traditional cohort or case-control study addresses the question "Why did this person suffer a myocardial infarction while another person did not?" (which depends heavily on chronic risk factors such as atherosclerosis, hypertension, and lifetime smoking), the case-crossover design addresses the acute kinetic question: "Given that this patient was vulnerable to a myocardial infarction, why did the plaque rupture occur during this specific hour rather than during an earlier hour?"

For a scientific question to be suited to a case-crossover architecture, three structural criteria must be satisfied:

1. Brief, Intermittent Exposure With Rapid Kinetic Induction: The exposure of interest must vary over short intervals within an individual and exhibit a rapid onset and decay of biological hazard (a short induction time and minimal carryover effect), such as a burst of strenuous physical exertion, acute anger, cellular telephone use while driving, an episodic prescription medication dose, or a daily spike in ambient fine particulate matter (PM_2.5), nitrogen dioxide (NO₂), sulfur dioxide (SO₂), or tropospheric ozone (O₃).
2. Abrupt, Precisely Timed Outcome Onset: The clinical event must have a sudden, well-defined onset time rather than an insidious preclinical sojourn time—classic examples include acute myocardial infarction, ischemic or hemorrhagic stroke, ventricular tachyarrhythmia, asthma exacerbation, epileptic seizure, or motor vehicle collision.
3. Stationarity of Baseline Within-Person Exposure Probability: After accounting for design-level time stratification, the probability of exposure across the hazard and reference windows must not exhibit an unadjusted monotonic secular trend.

1:1 Matched Discordant Pair Analysis (Mantel–Haenszel Formula)

In the simplest unidirectional 1:1 case-crossover design, each case i contributes one pre-event hazard window (for example, the 1-hour interval directly preceding myocardial infarction onset) and one prior reference window of equal duration (for example, the corresponding 1-hour clock window 24 hours earlier). Across N enrolled cases, each individual falls into one of four within-person concordance cells:

• a_i (Concordant Exposed): Exposed in both the hazard window and the reference window (X_i,1 = 1, X_i,0 = 1)
• b_i (Discordant Triggered): Exposed in the hazard window but unexposed in the reference window (X_i,1 = 1, X_i,0 = 0)
• c_i (Discordant Reference): Unexposed in the hazard window but exposed in the reference window (X_i,1 = 0, X_i,0 = 1)
• d_i (Concordant Unexposed): Unexposed in both windows (X_i,1 = 0, X_i,0 = 0)

Just as in McNemar's test and 1:1 matched case-control studies, concordant individuals (a and d) contribute zero information to the within-stratum likelihood because their exposure did not vary. Summing the discordant pairs across all N cases (B = ∑ b_i and C = ∑ c_i), the Mantel–Haenszel matched odds ratio (OR_CC) directly estimates the acute incidence rate ratio (IRR):

OR_CC = B / C

With an approximate 95% confidence interval on the natural logarithm scale of ln(OR_CC) ± 1.96 × (1/B + 1/C)^1/2.

Usual Frequency Approach and Conditional Logistic Regression

When each case is queried regarding their usual frequency of exposure over an extended historical baseline (for example, the number of hours spent in vigorous physical exertion over the preceding 12 months), Maclure's usual-frequency method treats each individual i as a separate Mantel–Haenszel stratum comparing the observed exposure in the hazard period against expected person-time exposed versus unexposed across the baseline period. More generally, when M discrete reference windows (m = 1, ..., M) and continuous exposures (such as hourly ambient PM_2.5 in μg/m³) are analyzed alongside time-varying co-exposures (such as ambient temperature and relative humidity), parameters are estimated by maximizing the conditional logistic regression likelihood stratified on individual i:

L(β) = ∏_i=1^N [ exp(β X_i,0) / ∑_m=0^M exp(β X_i,m) ]

where m = 0 denotes the hazard window and m = 1, ..., M denote the matched control windows for subject i.

Unidirectional vs. Bidirectional vs. Time-Stratified Sampling

Because the hazard window occurs just prior to the event—and because an event such as fatal stroke, hospitalization, or behavior modification alters or terminates post-event exposure—early case-crossover studies sampled reference windows only before the event (unidirectional retrospective sampling). However, if the population prevalence of exposure exhibits a secular upward or downward trend over calendar time (for example, a newly launched pharmaceutical drug gaining market share, or seasonal winter heating increasing ambient air pollution), sampling reference windows only prior to the hazard window introduces systematic exposure time-trend bias.

To eliminate time-trend bias, epidemiologists developed three structural sampling refinements:

In environmental public health studies of ambient air quality (where an individual's hospital admission for asthma does not alter citywide outdoor NO₂ or O₃ concentrations after the admission date), time-stratified case-crossover sampling (matching each event day to all other occurrences of the same day of the week within the same calendar month and year, yielding 3 or 4 control days per case) simultaneously controls for seasonal cycles, long-term secular trends, and day-of-the-week behavioral rhythms. When the exposure is an endogenous patient behavior or prescription drug whose use changes after the clinical event, post-event windows cannot be used; investigators instead deploy Sammy Suissa's (1995) case-time-control design, or transition to a cohort-derived self-controlled case series with pre-exposure washout periods.

Although the case-crossover design eliminates time-invariant between-person confounding, it remains vulnerable to within-person time-varying confounding. If two transient exposures co-occur within the same narrow hazard window—such as acute psychological stress, heavy alcohol consumption, and cocaine use, or heat waves and photochemical ozone spikes—unadjusted within-person comparisons cannot separate their individual effects without multivariable conditional logistic modeling of the co-varying trigger. Furthermore, when exposure in the hazard and reference windows is ascertained via retrospective patient interviews rather than objective registry or environmental monitor logs, patients frequently overestimate exposure just prior to a traumatic clinical event, introducing within-person recall bias and differential exposure misclassification.

How does a case-crossover study differ from a self-controlled case series (SCCS)?

Although both are self-matched within-person designs that eliminate time-invariant confounding, a case-crossover study is an outcome-anchored (case-control-derived) design that starts at the moment of the clinical event and looks backward (or within a local time stratum) to compare exposure odds in hazard versus reference windows. In contrast, a self-controlled case series is an exposure-anchored (cohort-derived) design that fixes an observation period, identifies when exposure occurred (such as vaccination), and models the Poisson incidence rate of events inside post-exposure risk windows relative to baseline control person-time.

Why can case-crossover studies not evaluate chronic cumulative exposures?

If an exposure is constant over time within an individual (such as germline BRCA1 genotype, childhood lead exposure, or multi-decade pack-years of smoking), every patient falls into the concordant cells (a or d) where hazard-window exposure equals reference-window exposure. Because only discordant subjects (b and c) contribute to the conditional likelihood, constant exposures yield zero statistical information in a case-crossover analysis, though they can still be evaluated as effect measure modifiers across subgroups.

1. 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. https://doi.org/10.1093/oxfordjournals.aje.a115853
2. Mittleman, M. A., Maclure, M., Tofler, G. H., Sherwood, J. B., Goldberg, R. J., & Muller, J. E. (1993). Triggering of acute myocardial infarction by heavy physical exertion—protection against triggering by regular exertion. New England Journal of Medicine, 329(23), 1677–1683. https://doi.org/10.1056/NEJM199312023292301
3. Suissa, S. (1995). The case-time-control design. Epidemiology, 6(3), 248–253. https://doi.org/10.1097/00001648-199505000-00010
4. Janes, H., Sheppard, L., & Lumley, T. (2005). Case-crossover analyses of air pollution exposure data: referent selection strategies and their implications for bias. Epidemiology, 16(6), 717–726. https://doi.org/10.1097/01.ede.0000181315.18836.9d