Self-Controlled Case Series

Self-controlled case series (abbreviated SCCS) is a within-individual, cohort-derived statistical method in epidemiology and biostatistics developed by C. Paddy Farrington in 1995 for post-licensure vaccine safety surveillance and pharmacoepidemiology. Unlike traditional cohort studies that follow both cases and non-cases across an entire population, the self-controlled case series samples only individuals who have experienced at least one clinical outcome event (n_i ≥ 1) during a pre-specified observation period [a_i, b_i]. Each patient's observation timeline is partitioned into one or more post-exposure risk windows (when the biological hazard of an adverse event is hypothesized to be elevated following immunization or drug initiation) and unexposed baseline control windows (comprising all remaining person-time within [a_i, b_i] outside the risk windows). By conditioning the Poisson likelihood on the total number of events observed for each individual i and on that individual's exposure history, the SCCS estimator algebraically cancels the individual baseline incidence rate φ_i, thereby eliminating all time-invariant multiplicative confounding—such as sex, genetic predisposition, socioeconomic tier, chronic frailty, and healthy user bias—while retaining virtually the same statistical efficiency as a full population cohort study for rare outcomes.

Overcoming Frailty and Confounding by Indication in Safety Registries

Farrington introduced the SCCS method at the United Kingdom Public Health Laboratory Service to evaluate whether routine childhood immunizations—such as the measles-mumps-rubella (MMR) and diphtheria-tetanus-pertussis (DTP) vaccines—acutely increased the incidence of febrile convulsions, idiopathic thrombocytopenic purpura (ITP), aseptic meningitis, or intussusception. In routine administrative healthcare databases, estimating rare adverse events via standard cohort comparisons faces two formidable obstacles:

1. Unmeasured Baseline Frailty and Confounding by Indication: Children with severe neurological disorders or congenital immunodeficiencies may selectively avoid or delay vaccination (the healthy vaccinee effect), or, conversely, high-risk elderly patients may be preferentially targeted for pneumococcal and influenza immunization.
2. Incomplete Denominator Records: Administrative hospital discharge registries often capture all hospitalized adverse events (cases) with high fidelity, whereas linking immunization history across millions of healthy non-cases in the population denominator is computationally expensive or unavailable.

By restricting data collection to cases alone and comparing event rates across risk versus control person-time within the same individual, the SCCS method requires only case records with accurate dates of event onset and exposure dates.

Underlying Non-Homogeneous Poisson Process

Suppose individual i is observed over a defined age or calendar interval [a_i, b_i], which is partitioned into K disjoint sub-intervals indexed by age group j (j = 1, ..., J) and exposure risk tier k (k = 0 for baseline control time; k = 1, ..., R for distinct post-exposure risk windows, such as days 1–7, 8–14, and 15–28 post-vaccination). Let e_ijk denote the duration of person-time spent by individual i in age group j and exposure state k, and let n_ijk denote the number of clinical events occurring in that interval.

In the underlying cohort model, events for individual i arise from a non-homogeneous Poisson process with instantaneous incidence rate λ_ijk:

λ_ijk = φ_i × exp(α_j + β_k)

where:

• φ_i = exp(γ_i + z_i^Tθ) is the individual-specific baseline incidence rate encapsulating all fixed genetic, demographic, and unmeasured baseline frailty covariates z_i.
• exp(α_j) represents the age-specific (or calendar-season-specific) relative incidence multiplier for age group j (with α₁ = 0 as reference), controlling for time-varying confounding by natural developmental maturation.
• exp(β_k) = RI_k is the Relative Incidence (incidence rate ratio) in exposure risk window k relative to baseline control person-time (k = 0, where β₀ = 0).

Conditioning on Total Individual Event Count

Conditioning on the total number of events n_i.. = ∑_j,k n_ijk ≥ 1 observed for subject i over [a_i, b_i] transforms the unconditional Poisson distribution into a multinomial distribution across the cells (j, k) of subject i's timeline. The individual-specific frailty term φ_i appears identically in both the numerator and the denominator and cancels out completely:

L_i(α, β) = ∏_j,k [ ( e_ijk exp(α_j + β_k) ) / ( ∑_r,s e_irs exp(α_r + β_s) ) ]^n_ijk

Maximizing the product of these multinomial likelihoods across all N cases (conveniently fitted in standard statistical software as a conditional Poisson GLM with individual fixed intercepts and log person-time offset ln(e_ijk)) yields consistent maximum-likelihood estimates of the relative incidence RI_k = exp(β_k).

Three Structural Assumptions of the Standard SCCS

For the standard Farrington SCCS estimator to be unbiased, three conditions must hold:

1. Recurrent Events Must Be Independent (or the Event Must Be Rare): Events must follow a Poisson process conditional on frailty and covariates; if a first event biologically increases the individual's hazard of recurrence (such as a first epileptic seizure or recurrent myocardial infarction), the analysis should be restricted to first events only, which remains valid when the event is rare in the population.
2. Event Occurrence Must Not Alter Subsequent Exposure Probability (Event-Independent Exposure): Because the SCCS conditions on the entire exposure history across [a_i, b_i]—including control person-time occurring after the event—an event that contraindicates, delays, or precipitates subsequent exposure violates strict exogeneity.
3. Event Occurrence Must Not Censor the Observation Period (No Event-Dependent Mortality): If the outcome event carries high acute case fatality (such as sudden cardiac arrest or fatal pulmonary embolism), post-event control person-time is truncated prematurely.

Pre-Exposure Windows and Event-Dependent Extensions

When an acute clinical event (such as a high fever or seizure) causes clinicians to postpone a scheduled immunization for two weeks, the period immediately preceding vaccination will exhibit an artificially depleted rate of events. Including this depleted pre-vaccination interval in the baseline control pool underestimates baseline incidence and biases the post-vaccination relative incidence upward. Including a dedicated pre-exposure washout period (for example, the 14 to 28 days before vaccination) as a separate parameter removes this distortion (a form of reverse causality or protopathic bias).

When an event permanently contraindicates future doses (for example, intussusception after Dose 1 of rotavirus vaccine contraindicating Dose 2, or Guillain-Barré syndrome contraindicating booster immunization) or causes death, standard SCCS is biased. Farrington, Whitaker, and Ghebremichael-Weldeselassie (2009, 2018) developed two rigorous generalizations implemented in the SCCS R package:

• Pseudo-Likelihood SCCS for Event-Dependent Exposure: Redefines the effective observation start time at exposure initiation or weights post-event person-time so that only post-exposure intervals remain in the conditioning set when exposure cannot follow the event.
• SCCS for Event-Dependent Observation Periods (Mortality): Models the joint distribution of event times and post-event survival censoring times to eliminate truncation bias in fatal cardiovascular or thromboembolic safety studies (such as evaluating adenoviral-vectored or mRNA COVID-19 vaccines and myocarditis or thrombosis with thrombocytopenia syndrome).

Why is the SCCS method nearly as statistically powerful as a full cohort study?

In a cohort of 1,000,000 children studying a rare adverse event that occurs in only 500 children, the 999,500 non-cases contribute zero events to the numerator and serve only to estimate the proportion of person-time spent exposed versus unexposed at each age. Because the 500 cases in an SCCS spend the vast majority of their own 365-day observation period in unexposed control person-time and only a small fraction (e.g., 14 to 42 days) in the risk window, within-case person-time partitioning provides a highly precise denominator while completely eliminating between-person variance in baseline risk φ_i.

When should an investigator choose SCCS over a case-crossover study?

The SCCS method is preferred when exposure is a point-in-time medical intervention (such as vaccination) or episodic prescription with well-defined post-exposure risk windows and strong age-related variation in baseline disease incidence (such as pediatric intussusception, which peaks steeply between 4 and 9 months of age). The case-crossover study is preferred when exposure is a continuous, stationary environmental variable (such as daily ambient PM_2.5 or temperature) analyzed across time-stratified calendar windows.

1. Farrington, C. P. (1995). Relative incidence estimation from case series for vaccine safety evaluation. Biometrics, 51(1), 228–235. https://doi.org/10.2307/2533328
2. 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. https://doi.org/10.1002/sim.2302
3. Petersen, I., Douglas, I., & Whitaker, H. (2016). Self controlled case series methods: an alternative to standard epidemiological study designs. BMJ, 354, i4515. https://doi.org/10.1136/bmj.i4515
4. Farrington, P., Whitaker, H., & Ghebremichael-Weldeselassie, Y. (2018). Self-Controlled Case Series Studies: A Modelling Guide with R. Chapman and Hall/CRC. https://doi.org/10.1201/9780429491313