Written and maintained by CASRAI Editorial Board
Last updated
Case-crossover design and the self-controlled case series (SCCS) are both within-person observational designs: each subject serves as their own comparison. Instead of comparing a group of exposed people to a separate group of unexposed people, both designs compare the same person’s exposure or event pattern across different time windows. That single structural choice does something a matched case-control design or a multivariable-adjusted cohort study cannot do as cleanly: it removes confounding by any characteristic that does not change within the comparison window — genotype, chronic health status, personality, most of socioeconomic status, baseline behavior — algebraically, by design, rather than by measuring the confounder and adjusting for it statistically. If a confounder is time-invariant over the window being compared, it is identical in both the “case” observation and the “control” observation for that person, so it cancels out of the within-person contrast.
The two designs solve different versions of the “each subject as their own control” problem. Case-crossover is built for a transient exposure that precedes a single (or first) acute event. SCCS is built for recurrent events with a definable, time-limited post-exposure risk period. Picking the wrong one for your data structure produces a design that technically runs but answers a subtly different question than the one you meant to ask.
Why self-matching removes time-invariant confounding
In a standard confounding problem, a third variable is associated with both the exposure and the outcome, distorting the exposure-outcome association unless it is measured and controlled for — by restriction, matching, stratification, or statistical adjustment. Every one of those strategies depends on the confounder being correctly identified and correctly measured; an unmeasured or mismeasured confounder leaves residual bias.
Self-controlled designs sidestep that dependency for a specific class of confounders. When the comparison is the same person at two different times, any characteristic that is constant across those two times — sex, genetic background, most chronic conditions, baseline personality traits, and (over a short enough window) socioeconomic circumstances — is identical in both observations. It does not need to be measured, named, or adjusted for, because it appears on both sides of the within-person comparison and cancels out mathematically, the same logic that makes crossover trial designs and matched-pair analysis (McNemar-type designs) efficient for paired binary outcomes, and that makes a fixed-effects specification remove unit-level confounding in panel data — see fixed vs. random effects for the general-purpose version of the same idea.
What self-matching does not remove: confounding by anything that changes within the comparison window. Time trends, seasonality, day-of-week patterns, and aging are the standard threats, and each design has a specific defense against them (below). This is the tradeoff to hold onto: self-controlled designs trade the need to measure time-invariant confounders for a new, narrower burden — correctly modeling or matching out time-varying factors instead.
Case-crossover: transient exposures, a single acute event
Case-crossover design compares a person’s exposure status in a brief hazard window immediately before their event to their exposure status in one or more referent (control) windows when the event did not occur — commonly the same time of day on a day or days without the event. It was introduced by Malcolm Maclure in 1991 to study transient triggers (such as brief bouts of unaccustomed physical exertion) of acute-onset events like myocardial infarction, and the same logic has since been applied to exposures like cell phone use before a traffic collision, analgesic use before a fall, and allergen exposure before an asthma attack.
Case-crossover fits when three conditions hold: the exposure is transient or intermittent (its presence or absence can meaningfully differ between the hazard window and a nearby referent window), the outcome is an acute, discrete event with a fairly precise onset time, and the exposure’s effect, if any, is short-lived (the hazard window is short enough that the exposure plausibly triggers the event rather than merely preceding it incidentally).
The design’s central methodological problem is referent window selection, because a poorly chosen referent window reintroduces exactly the time-varying confounding that self-matching is supposed to avoid:
- Time-trend bias. If the exposure’s prevalence is itself rising or falling over the study period (a seasonal medication, a habit that is becoming more common), a referent window that sits at a different point in that trend than the hazard window will make the exposure look associated with the event even with no causal effect.
- Bidirectional (symmetric) referent selection. Using referent windows both before and after the hazard window (e.g., one week before and one week after) for control-window sampling largely cancels out a linear time trend, and is the most common current practice.
- The case-time-control extension. When exposure prevalence is changing sharply and symmetric sampling is not enough, adding a separate group of non-cases (a case-time-control design) to estimate and subtract the background time trend directly is the standard fix.
Analytically, a 1:1 case-crossover reduces to a matched-pair comparison: only pairs where exposure status differs between the hazard and referent window (discordant pairs) are informative, exactly as in a matched case-control analysis or a McNemar-type paired design. Multiple referent windows per case are handled with conditional logistic regression, conditioning on the individual.
Self-controlled case series (SCCS): recurrent events, a defined risk window
SCCS was developed by C. Paddy Farrington in the mid-1990s specifically to evaluate vaccine safety, where the practical and ethical difficulty of assembling a comparably-vaccinated-and-unvaccinated cohort made a conventional cohort or case-control design costly, and where the events of interest (e.g., seizures, Guillain-Barré syndrome, intussusception) are individually rare but can recur. SCCS uses only people who had at least one event — there is no separate unexposed comparison group to recruit at all — and asks whether the rate of events for that person was higher during a pre-specified post-exposure risk window than during the rest of their observed time (the “control” or baseline period).
SCCS fits when the event can recur (or, at minimum, when you’re willing to model it that way even for a single occurrence), the exposure has a discrete, identifiable time of administration (a vaccine dose, a drug start date), and there is a defensible, pre-specified risk window during which a causal effect, if real, would plausibly manifest (e.g., “1–14 days post-dose” for a specific adverse event, set from prior biological or pharmacological knowledge, not chosen after looking at the data).
The design rests on two assumptions that do not arise in case-crossover, and both deserve a direct check before trusting the result:
- Event occurrence must not alter the future probability or timing of exposure. If having the event makes a subsequent exposure (e.g., a later vaccine dose) less likely — because a clinician or parent delays or withholds it — the independence assumption is violated and the standard SCCS estimator is biased. Extensions exist for this “event-dependent exposure” case.
- The event must not end observation (censoring by outcome). A case-fatal event truncates the person’s remaining control time asymmetrically relative to when in the risk/control period it occurred, which biases the risk-window vs. control-window comparison unless corrected for.
Analytically, SCCS is fit as a conditional Poisson (or equivalent) regression on each person’s own event and person-time data, conditioning on the individual and, typically, on age group, to estimate an incidence rate ratio (IRR) comparing the risk window to the person’s own control time.
Choosing between them
| Dimension | Case-crossover | Self-controlled case series |
|---|---|---|
| Event structure | Single (or first) acute event per person | Recurrent events per person |
| Exposure structure | Transient, intermittent | Discrete onset with a defined post-exposure risk window |
| Comparison | Hazard window vs. referent window(s) for the same person | Risk window vs. the person’s own remaining observed time |
| Main estimator | Matched odds ratio (conditional logistic regression) | Incidence rate ratio (conditional Poisson regression) |
| Key bias to defend against | Time-trend bias in exposure prevalence | Event-dependent exposure; outcome-related censoring |
| Typical use cases | Physical exertion and cardiac events, mobile phone use and collisions, medication use and falls | Vaccine safety surveillance, drug-induced acute events with recurrence (e.g., seizures) |
Worked example: two simulated datasets
The numbers in this section come from a seeded Python simulation written for this page (seed 20260829), not from a real study or real patient data. They illustrate how each design’s estimator behaves on a dataset with a known, built-in true effect — they are not a claim about any real exposure or outcome.
Case-crossover scenario. 500 simulated cases were each assigned an exposure status (present/absent) independently for a 1-hour hazard window and a matched 1-hour referent window one week earlier, using a control-window exposure prevalence of 5% and a built-in true matched odds ratio of 3.0. The simulation produced 411 concordant-unexposed pairs, 0 concordant-exposed pairs, 58 pairs exposed only in the hazard window, and 31 pairs exposed only in the referent window. The matched OR from the discordant pairs (58/31) is 1.87, 95% CI 1.21–2.89 — a single simulated sample of this size lands within sampling error of, but not exactly on, the built-in true value of 3.0, which is itself a useful illustration of how much a 500-pair case-crossover study can and cannot pin down.
SCCS scenario. 300 simulated people were each followed for 365 days, with one exposure event at a random day and a 14-day post-exposure risk window, a baseline event rate of 0.01/person-day, and a built-in true incidence rate ratio of 2.5 during the risk window. The simulation produced 95 events over 4,200 risk-window person-days (rate 0.0226/day) and 1,078 events over 105,300 control-period person-days (rate 0.0102/day), for an IRR of 2.21, 95% CI 1.79–2.73 — again consistent with, though not identical to, the built-in true value of 2.5.
The contrast worth noticing: the case-crossover analysis above uses only two binary observations per person (exposed/unexposed in each window), while the SCCS analysis uses continuous person-time and a recurring event count per person. That difference in what each design asks of the data is the real dividing line between them — more than the specific exposures or outcomes in either example.
Where these designs stop being the right tool
- Chronic or cumulative exposures without a discrete onset (long-term diet, lifetime occupational exposure) don’t fit either design — there is no meaningful “hazard window” or “risk window” to define, because the exposure doesn’t turn on and off. A cohort study or standard case-control study is the better fit.
- Neither design estimates absolute risk. Both produce relative measures (odds ratios, incidence rate ratios) conditional on the person having had at least one event; neither yields a population-level incidence or absolute risk difference on its own.
- Case-crossover’s referent-window choice is a real analytic decision, not a default. Getting it wrong (ignoring a time trend, choosing an implausible hazard-window length) changes the answer, not just its precision.
- SCCS needs a defensible risk window set in advance. A risk window chosen by scanning the data for where the event rate happens to be highest is a data-dredging problem dressed up as a design feature.
- Neither design is the same as an N-of-1 trial. An N-of-1 trial is an experiment — treatments are randomized and (usually) blinded within a single patient. Case-crossover and SCCS are observational: nothing is assigned, and the “self-controlled” label describes the comparison structure, not an intervention.
Frequently asked questions
Is case-crossover the same thing as a crossover trial?
No. A crossover trial is an experimental design: the same participant receives two or more treatments in sequence, usually with randomized order, to compare their effects. Case-crossover is observational: no treatment is assigned by the investigator, and the “exposure” is whatever naturally occurred (or didn’t) in the hazard window before an event that already happened.
Can case-crossover or SCCS be used for a chronic, always-present exposure?
Not usefully. Both designs rely on exposure status differing across the comparison window for at least some subjects. An exposure that is constantly present (or constantly absent) for the whole observation period generates no discordant information for case-crossover, and no contrast between risk and control periods for SCCS.
Why would SCCS be preferred over a cohort study for vaccine safety?
SCCS needs only people who experienced the adverse event of interest — it does not require assembling and following a separate unvaccinated (or differently-timed) comparison cohort, which is often impractical for a nationally recommended vaccine and, for some questions, ethically fraught. Because every time-invariant characteristic of each case is automatically controlled for by the within-person comparison, SCCS also needs less covariate data than a comparable cohort analysis would.
What’s the minimum sample size for either design?
There’s no fixed threshold; it depends on the discordant-pair rate (case-crossover) or the event count within the risk window (SCCS), the same way any matched or Poisson-based analysis depends on the number of informative events rather than the raw headcount. A study with many cases but a rare or weakly time-varying exposure can be substantially underpowered despite a large N.
