ISPE Self-Controlled Designs Guidance (Cadarette 2021)

What it is

"Control yourself: ISPE-endorsed guidance in the application of self-controlled study designs in pharmacoepidemiology" (Cadarette et al., Pharmacoepidemiology and Drug Safety, 2021), endorsed by the International Society for Pharmacoepidemiology (ISPE), is a methodological reference — not a one-page reporting checklist — that establishes common language, a structured worksheet, and best-practice recommendations for the family of self-controlled (within-person) designs: the self-controlled case series (SCCS), the case-crossover (CCO) design, the case-time-control (CTC) design, and the case-case-time-control (CCTC) design. Its organizing insight is that these designs make the case its own control, so every time-invariant confounder — measured or unmeasured, including stable genetics, chronic comorbidity, sex, and durable lifestyle — is eliminated by the design itself rather than by adjustment. The guidance walks the analyst through the assumptions each design depends on (transient exposure, abrupt outcome, exposure independent of the event, event independent of observation) and how the choice among the four designs follows from which of those assumptions hold. It is maintained by ISPE as part of its methods-guidance output and is widely cited as the field's reference statement on when and how to use self-controlled designs. (This catalog refers to it descriptively as the ISPE self-controlled-designs guidance; "SCOPE" is not an official name or acronym for it and is avoided to prevent confusion with the EU SCOPE Joint Action in pharmacovigilance.)

When to use

Reach for this guidance whenever the scientific question concerns the transient effect of an intermittent or short-duration exposure on the risk of an abrupt, well-dated acute event — vaccine safety (seizures, intussusception, myocarditis after a dated dose), acute drug toxicity (NSAIDs and GI bleed, antibiotics and arrhythmia), or any "did exposure raise risk in the hours-to-weeks after it occurred" question — using claims, EHR, registry, or linked routinely collected data destined for an FDA/EMA submission, an imposed or voluntary PASS, an HTA/payer safety narrative, or a peer-reviewed journal. The within-person comparison is most valuable precisely when between-person confounding by indication or frailty is severe and hard to measure, because it sidesteps the need for a comparator cohort entirely. Decision rules among the siblings: use SCCS when complete exposure and event histories are observable across a defined window and the rate of events is the estimand; use case-crossover when you have cases only and want to compare each case's exposure in a hazard window against earlier referent window(s); add case-time-control when an underlying time trend in exposure prevalence would bias the case-crossover (a separate control series estimates and removes that trend); and use case-case-time-control when that exposure-trend control itself risks overadjustment because the future-case control series shares the cases' characteristics. This is the design family to consult instead of a cohort/active-comparator guidance when the comparison is within-person.

What it requires

The guidance enforces the substantive elements that determine whether a self-controlled analysis is valid, framed for real-world data: (1) a clear, dated, intermittent exposure definition with explicit start/stop, and an induction/latency and risk (hazard) window justified on pharmacology rather than convenience; (2) an abrupt, accurately dated outcome with a validated phenotype/algorithm and a defined deduplication rule so recurrent or carried-forward codes are not double-counted; (3) explicit treatment of the observation period and of person-time partitioned into risk versus control windows; (4) a stated and defended position on the two design- breaking dependencies — event-dependent exposure (does the event itself change future exposure probability, as a stroke stops anticoagulation?) and event-dependent observation (is the event fatal or does it truncate follow-up?), invoking the appropriate SCCS extension when either holds; (5) handling of time-varying confounders within the person (age, calendar season, disease progression) by modeling or by escalating to CTC/CCTC; and (6) pre-specified sensitivity and bias diagnostics — alternative risk-window lengths, negative-control exposures, and washout/clean-period definitions. It is fundamentally a transparency-and-assumptions instrument: it forces the analyst to state which design was chosen and why, and to demonstrate that the design's assumptions are met.

When NOT to use — limitations and common misapplications

This is methodological guidance for within-person designs, not a universal RWE quality score, and it does not certify causality; meeting its recommendations does not rescue a study whose design assumptions are violated. Concrete failure modes: (a) the outcome is cumulative or chronic (cancer, atherosclerosis, fibrosis) rather than abrupt — the transient-effect logic does not hold and a self-controlled design is the wrong tool; (b) event-dependent exposure — the acute event changes the chance of future exposure, biasing a standard SCCS unless the event-dependent-exposure extension is used; (c) event-dependent observation — the outcome is fatal or terminates observation, requiring Farrington's extension or a different design; (d) uncontrolled within-person time trends in exposure or background risk, which a plain case-crossover will absorb as spurious association unless escalated to case-time-control / case-case-time-control; (e) induction or latency longer than the observation window, so the relevant exposure cannot be captured; and (f) using a self- controlled design when the real question is a between-person comparison (drug A vs drug B effectiveness, absolute risk in the population) — for that, a cohort/active-comparator design is required and this guidance does not apply. Picking the wrong sibling (e.g., case-crossover where a time trend mandates CTC) is itself a misapplication this guidance is written to prevent.

How it maps to this catalog

The four designs this guidance governs are implemented directly by `self-controlled-case-series`, `case-crossover`, `case-time-control`, and `case-case-time-control` — consult those for estimators, code, and worked examples. The central risk-window requirement is implemented by `exposure-lag-induction-latency-window-rwe` (induction and latency) and `as-treated-risk-window-construction-rwe` (hazard-window construction). Event handling maps to `acute-event-deduplication-window-rwe` (one-event-per-episode rules) and `recurrent-events-analysis-rwe`. The dated, validated outcome that self-controlled designs depend on is delivered by `diagnosis-phenotype-algorithm-1ip-2op-time-window-rwe` and `claims-outcome-algorithm-ppv-sensitivity-rwe`. Bias diagnostics map to `negative-control-exposures-rwe`, and the clean-period logic to `washout-clean-lookback-period-rwe`. For the operational substrate, `claims-analysis` covers the routinely collected data these designs run on. Applied note (claims/EHR/registry RWE): in claims, the exposure date is the pharmacy fill or administration date and the hazard window is built in `days_supply` space, so deduplicate acute outcome codes (e.g., a single seizure may generate multiple inpatient + outpatient claims within days) before partitioning person-time; in vaccine-safety EHR/registry work, the dated dose makes SCCS especially clean, but confirm the event did not stop or alter subsequent dosing (event-dependent exposure) and that fatal events are handled with the appropriate extension rather than silently dropped.