Self-Controlled Case Series
- Introduced by
- C. Paddy Farrington
- Year
- 1995
- Type
- Within-person case-only design for transient exposures and acute outcomes
- Family
- Self-controlled causal designs / pharmacoepidemiology
What does it do?
The self-controlled case series, or SCCS, is a case-only study design for estimating the association between a transient exposure and an acute event by comparing each individual's event rate during exposed time windows with their rate during unexposed time windows.
Developed by Paddy Farrington in 1995 for vaccine safety evaluation, it uses data only on people who experienced the outcome, and because each person serves as their own control, it automatically eliminates all fixed within-person confounders — genetics, sex, chronic conditions, socioeconomic position — without ever measuring them. A conditional Poisson likelihood removes the individual-level baseline rate and yields a relative incidence comparing risk to control periods. Whitaker, Farrington, Spiessens and Musonda's 2006 Statistics in Medicine tutorial is the standard practical guide to fitting and interpreting the model.
The intuition behind the method
Suppose you want to know whether a vaccine briefly raises the risk of a rare seizure. The honest worry with a conventional cohort or case-control study is confounding: vaccinated and unvaccinated children differ in countless ways, healthy or not, that you can never fully measure. The SCCS sidesteps this with a clever twist — it only studies children who actually had the seizure, and then asks, within each child's own timeline, were the seizures more concentrated in the short window just after vaccination than in the rest of the follow-up? Since you are comparing each child to themselves, anything about that child that does not change over time — their genes, their family, their underlying frailty — cancels out automatically; it simply cannot explain why their seizures cluster in one part of their own timeline rather than another. All you need is the timing of the exposure and the timing of the events for the cases. The method does have to take care of things that do change with time, especially age, since both vaccination schedules and seizure risk move with a child's age, but those are modeled explicitly.
History
The self-controlled case series was introduced by C. Paddy Farrington in a 1995 Biometrics paper on relative incidence estimation from case series for vaccine safety, motivated by the difficulty of finding unconfounded comparison groups for vaccinated children. The design extended the case-only logic of within-person comparison into a likelihood framework that conditions out the individual baseline and adjusts for age, making it possible to estimate transient exposure effects from cases alone. Over the following decade Farrington and colleagues elaborated the theory and its assumptions, culminating in the widely used 2006 Statistics in Medicine tutorial by Whitaker, Farrington, Spiessens and Musonda, which provided worked examples, code, and guidance on assumptions. The SCCS has since become a mainstay of pharmacoepidemiology and vaccine safety surveillance, with active methodological extensions for event-dependent exposure and observation, semiparametric age modeling, and applications well beyond vaccines.
When to use it
Use the self-controlled case series when you are studying an acute, well-defined event in relation to a transient exposure, and you are worried about confounding by stable individual characteristics that are hard to measure — the canonical example being vaccine or drug safety, where exposed and unexposed people differ systematically. It is especially attractive when you have good data on cases and their exposure timing but lack a clean comparison group or reliable denominators for the unexposed. The design requires that events be relatively rare over the observation window, that the event not change the chance of future exposure, and that the event not terminate follow-up; when a fatal or recurrence-altering outcome violates these, modified SCCS variants or alternative designs are needed. It is less suitable for chronic exposures with no clear on-off timing, for outcomes that develop slowly, or when time-varying confounders other than age cannot be modeled.
Limitations and considerations
- Controls only time-invariant confounders; time-varying confounders other than age must be modeled explicitly or they bias the estimate.
- Assumes the event is rare within the observation period, so that an event does not appreciably deplete subsequent time at risk.
- Standard SCCS assumes the event does not alter the probability of later exposure and does not end observation, which fails for some fatal or recurrence-blocking outcomes.
- Results are sensitive to the chosen risk-window definition, which must be pre-specified on mechanistic grounds rather than fitted to the data.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Automatically controls all fixed within-person confounders, measured or not, because each case is their own control.
- Requires data on cases only, which is efficient and avoids the need to sample or measure a separate control group.
- Yields an easily interpreted within-person relative incidence comparing exposed to unexposed time.
- Has a well-developed likelihood framework with extensions for age adjustment, event-dependent exposure, and event-dependent observation.
Applications
- Evaluating whether vaccines transiently raise the risk of acute adverse events such as seizures or intussusception.
- Assessing short-term risks of medications, for example whether a drug elevates the rate of myocardial infarction in the days after initiation.
- Studying triggers of acute events, such as infections preceding stroke or trauma preceding venous thromboembolism.
- Active drug and vaccine safety surveillance where rapid, confounder-robust signal evaluation is needed.
- Any setting with rich case data and exposure timing but no clean unexposed comparison group.
How it works
How it works
The analysis begins by defining, for each case, an observation period over which they are followed, and then partitioning that period into risk windows and control windows. A risk window is the interval after an exposure during which the event might plausibly be affected — for a vaccine, perhaps the days or weeks following administration — while the remaining time is treated as the unexposed control period. The choice of risk-window length and placement is a substantive decision driven by the hypothesized biological mechanism. Whitaker and colleagues emphasize that careful, pre-specified definition of these windows, including any pre-exposure periods to detect event-dependent exposure, is central to a credible SCCS.
Formula
[a_i,b_i]=\bigcup_k e_{ik},\quad e_{ik}\in\{\text{risk},\text{control}\};\ \ \text{exposure} \rightarrow \text{risk window}Within each individual the event rate is modeled as a log-linear function of an individual-specific baseline term, the exposure indicator that switches on during risk windows, and time-varying covariates such as age. The individual baseline absorbs everything fixed about that person, while the exposure coefficient captures how much the rate changes when the person is in a risk window versus a control window. Farrington's formulation casts this as a Poisson process whose intensity is modulated multiplicatively by exposure and age, so that the parameter of interest is the log relative incidence.
Formula
\lambda_i(t)=\exp\bigl(\phi_i+\beta\, x_i(t)+\gamma\, \mathrm{age}(t)\bigr),\quad x_i(t)=\mathbb{1}[t\in\text{risk}]The key statistical step is conditioning. Because we only sampled cases and we condition on the total number of events each individual experienced, the individual-specific baseline term drops out of the likelihood entirely. What remains is a multinomial-style conditional likelihood in which the probability that an event falls in a given interval depends only on the exposure status and length of that interval relative to the others within the same person. Farrington showed this conditional argument is what makes the design self-controlling: all time-invariant confounders vanish without being measured, because they are part of the eliminated baseline.
Formula
\Pr(\text{event in } e_{ik}\mid n_i)=\dfrac{\exp(\beta x_{ik})\,\tau_{ik}}{\sum_{l}\exp(\beta x_{il})\,\tau_{il}}Maximizing the conditional likelihood produces an estimate of the exposure coefficient, whose exponential is the relative incidence: the ratio of the event rate during risk periods to the rate during control periods, within individuals. A relative incidence above one indicates the exposure transiently raises the event rate, and below one that it lowers it. Because the estimate is built entirely from within-person contrasts, it has a clean interpretation as a within-person rate ratio adjusted for all fixed characteristics and, through the age terms, for the modeled time trends.
Formula
\mathrm{RI}=e^{\hat{\beta}}=\dfrac{\text{rate in risk period}}{\text{rate in control period}}Finally, the analyst must verify the assumptions on which the design rests. The SCCS assumes events are rare enough within the observation period that one event does not materially deplete the time at risk, that occurrence of the event does not alter the probability of subsequent exposure (no event-dependent exposure), and that the event does not curtail or end the observation period (as a fatal outcome would). Whitaker and colleagues describe diagnostics and extensions — pre-risk windows, modified likelihoods for event-dependent observation, and semiparametric age adjustment — for situations where these assumptions are strained, and stress checking them before trusting the relative incidence.
Formula
\Pr(\text{event}\mid \text{at risk})\ \text{small};\quad \text{event}\not\Rightarrow \text{exposure or end of obs.}Common pitfalls
- Defining risk windows post hoc to maximize the apparent effect, which invalidates inference.
- Applying the method to outcomes that are fatal or that prevent recurrence without using the appropriate event-dependent-observation extension.
- Ignoring age or seasonal trends that move together with exposure timing, leaving residual time-varying confounding.
- Using the design for chronic, continuous exposures that lack the transient on-off structure the model assumes.
Frequently asked questions
Why does the SCCS only need data on cases?
The design compares exposed and unexposed time within each person who had the event, so the relevant contrast is internal to cases. By conditioning on the number of events each case experienced, the individual's baseline rate cancels out of the likelihood, and with it every fixed characteristic. Non-cases contribute no information about within-person timing, so they are simply not needed. Farrington's original derivation makes this explicit: the conditional likelihood depends only on the exposure status and length of each person's intervals, which is why the method is so efficient with case-only data.
What confounding does the SCCS handle, and what does it not?
It eliminates all confounders that are constant within a person over the observation period — sex, genetics, chronic comorbidity, stable socioeconomic factors — because these are absorbed into the conditioned-out baseline. It does not automatically handle time-varying confounders. The most important of these is age, which is modeled explicitly with age categories or splines; seasonality and other time trends must likewise be included if they move with exposure timing. Whitaker and colleagues stress that residual time-varying confounding is the main threat once fixed confounders are removed.
When do the SCCS assumptions break down?
Three assumptions matter most. First, events should be rare within the observation window so that having an event does not deplete later time at risk. Second, the event should not change the probability of subsequent exposure — for instance, if a reaction leads clinicians to withhold further doses, this event-dependent exposure biases the standard estimate. Third, the event should not end observation, as a fatal outcome does. Each violation has a tailored extension — pre-exposure windows, modified likelihoods for event-dependent exposure, and methods for event-dependent observation — described in the Whitaker tutorial, but the basic model should not be used when they hold without correction.
Related methods
Sources
Farrington, C. P. (1995). Relative Incidence Estimation from Case Series for Vaccine Safety Evaluation. Biometrics, 51(1), 228-235.
Open DOI referenceCopied.Whitaker, H. J., Farrington, C. P., Spiessens, B., & Musonda, P. (2006). Tutorial in Biostatistics: The Self-Controlled Case Series Method. Statistics in Medicine, 25(10), 1768-1797.
Open DOI referenceCopied.