Protocol templates → Self-controlled case series

Multi-national SCCS protocol builder

Self-controlled case series. Each case acts as their own control; all time-invariant between-person confounding (genetics, lifelong habits) is automatically removed. AsPEN convention: separate a short pretreatment window (e.g., 14 days) before the first dispensing to absorb confounding by indication and benchmark the underlying-disease risk before therapy (see Wang GH et al BMJ 2021; PMID 34503972).

← All designs

Before you use this design

An SCCS that breaks one of these four assumptions still returns an incidence rate ratio with a confidence interval. Nothing warns you. Read the plain line first; the paragraph under it is the version for a methods section or a reviewer.

1. Having the event must not change the chance of later exposure

If the event — or the illness that produced it — makes a clinician start, stop, or delay the drug, the design is biased.

Why it matters, and what to do

Why · The SCCS likelihood conditions on each case's whole exposure history being fixed in advance. If exposure after the event differs from what it would have been, that condition fails. The mirror image is what you can actually see in the data: an excess (or dearth) of exposures after an event is the same thing as an excess (or dearth) of events just before an exposure, which is why a pre-exposure risk window measures it directly.

What to do · Put a short pre-exposure (pretreatment) window immediately before exposure start and report its incidence rate ratio as a separate parameter, never folded into the baseline. A raised pre-exposure IRR is a finding, not a nuisance: in the AsPEN antipsychotic-falls series a 14-day pretreatment window carried an IRR of 6.17, which the authors read as the underlying disease rather than the drug. [Whitaker 2006; Whitaker 2018; Wang GH BMJ 2021]
2. The length of observation must not depend on the event

If having the event can end the person's follow-up, the design is biased. Death is the obvious case.

Why it matters, and what to do

Why · The standard model requires each case's observation period to be fixed independently of when their events occur. Censoring at death breaks this whenever the outcome itself raises mortality; the extreme case is an outcome that is death, for which the standard SCCS is simply not applicable.

What to do · If the outcome is fatal or raises short-term mortality, say so and either restrict to non-fatal events or use a method for event-dependent observation periods (Farrington et al, JASA 2011). Do not silently keep the standard model. [Farrington 2011; Petersen 2016]
3. Repeat events in the same person must be independent

If one event makes the next one more likely, counting them all makes the study look more precise than it is.

Why it matters, and what to do

Why · The Poisson likelihood treats a person's events as independent given their exposure and age. For clustered outcomes — repeat falls in the same frail patient, repeat admissions for the same episode — that fails, and confidence intervals are too narrow.

What to do · Either restrict to the first event per person (the Event handling field, under the case definition) or justify independence explicitly. If the outcome can only happen once, first-event handling is the only coherent choice. [Whitaker 2006]
4. Only time-invariant confounding is removed; time-varying confounding is not

Being your own control removes anything about you that does not change — genes, sex, lifelong habits. It does nothing about things that change while you are being observed.

Why it matters, and what to do

Why · Age and season are the two that change in every study and are handled by splitting person-time into age bands and calendar strata inside the same conditional Poisson model. Anything else that varies within a person over the observation period — an intercurrent illness, a concomitant drug, a change in care setting — remains a confounder unless it is modelled as time-varying.

What to do · Set age bands and season in the time-varying confounders section, and list any other time-varying covariate explicitly. If you leave season unadjusted for a seasonal outcome, state why. [Whitaker 2006; Petersen 2016]

1 · Study identification

Study title Short ID Version · Date Lead investigator(s)

2 · Research question

Population Exposure Outcome

3 · Sites

Tick every database (not just country). Feature chips indicate what each database can support; only tick a site whose features cover what your case definition / exposure / time-varying covariates need.

Legend Rx Prescriptions Dx Diagnoses Lab Lab results Img Imaging Mor Mortality linkage Can Cancer-registry linkage Preg Pregnancy / birth linkage Gen Genomics / biobank linkage

Taiwan

CGRD RxDxLabImg NHIRD RxDxLabImgMorCanPreg TWB LabGen

South Korea

HIRA RxDx NHIS-NHID RxDxLabMorCan

Japan

JMDC Claims Database RxDxLab Medical Data Vision (MDV) Database RxDxLab NDB RxDx

Hong Kong

CDARS RxDxLabImgMorCan

Australia

PBS Rx VLHD RxDxMor

Finland

FinReg RxDxMorCanPreg

United Kingdom

CPRD RxDxMorCanPreg ISD RxDxLabMorCanPreg THIN RxDx UKB LabImgMorCanGen

Netherlands

PHARMO Database Network RxDxLabMorCan

United States

20% sample RxDxMorCan Medicaid RxDxMor

Others (one per line)

⚠ Feasibility heads-up; not a blocker

Your study question implies data features that some selected sites don't carry. Multi-country SCCS protocols often accommodate this; just be explicit in the per-site analytic plan.

4 · Case definition (outcome)

Validation citation auto-appends to References on export.

Operational definition Code set Event handling

5 · Exposure definition

Exposure; name Exposure; code set Exposure period definition Days of supply default (days) Grace period (days)

6 · Design variant

SCCS variant Choose Standard when most cases have substantial unexposed time. Choose Extended when within-person unexposed time is sparse and you want to borrow strength from a never-exposed cohort. Choose Active-comparator when patients are continuously on some therapy of the same indication and the relevant question is 'drug A vs drug B'.

7 · Risk windows

Timeline anchor — how is time measured & where does baseline come from?

Exposure-anchored (per person)
Day 0 = each case's own first exposure. Periods are counted in days before/after that day (baseline 90 d, risk 30 d…) — like a cohort look-back. Different cases sit at different calendar dates. Use when what matters is time since exposure. Calendar-anchored (shared dates)
Everyone is followed from the same calendar date (e.g. 1 Jan 2012), and baseline is a fixed prior calendar period (e.g. all of 2011 = 1 year). Use when you want a common calendar reference window, not per-person exposure timing.

Common follow-up start (calendar date) Baseline before start (years)

Risk / reference periods — one continuous strip (Type | days, in order)

The case's observation time as one continuous strip of consecutive periods — each line is a segment (Type | days), in order. Add any period anywhere (chips append; edit the text to insert/reorder), or drag a segment's right edge in the diagram. Day 0 = the start of the first Exposure risk segment. Types are auto-coloured (reference / exposure / pre / post risk). Risk window definition (justification) Baseline (control) period — definition / justification Covariate / eligibility look-back windows (measurement)

Measure baseline covariates / eligibility before exposure start (day 0) — each is a draggable window, exactly like the cohort builder (Name | days). This is the measurement layer; the risk-vs-reference windows above are SCCS's within-person estimand, not a cohort follow-up.
These windows do not adjust anything. A covariate measured once before day 0 is constant within a person, and the conditional Poisson likelihood conditions on the person — so it drops out of the model entirely. Use these to decide who is eligible, to describe the cases, and to define subgroups; anything that must be adjusted for has to vary within the observation period, and belongs under time-varying confounders below.

8 · Observation period

Period starts at Period ends at Study calendar period

9 · Time-varying confounders

Age band width Adjusts for within-person ageing within the observation period. Season adjustment Other time-varying covariates (one per line)

10 · Statistical analysis

Primary model Effect estimate Meta-analysis

Subgroup analyses; tick those to perform

Age (≥65 vs <65 years) Age (≥75 vs <75 years) Sex (male vs female) Country / site First-ever vs repeated exposure Index calendar year stratum Baseline CV disease (yes vs no) Baseline CKD (eGFR <60 vs ≥60) Frailty / Charlson Comorbidity Index (high vs low) Concomitant exposure (yes vs no)

Custom subgroups (one per line)

11 · Sensitivity analyses

Alternative risk window length (e.g., 7-30, 0-90, 0-180 d) Alternative pre-risk window (0, 30, 60 d); checks for event-dependent exposure Alternative post-risk window (washout) (0, 30, 90 d) Restrict to first event only (vs all events) Method of Farrington for event-dependent exposure (SCCS-EDE) Negative-control outcome; should show null Exclude seasonal months (if outcome is seasonal) Restrict to patients with ≥1 year of observation before first exposure Sub-divide risk window into early (0-14 d) vs late (15+ d) to look at risk attenuation

Custom sensitivity analyses (one per line)

12 · Execution & timeline

Architecture

Timeline preset

Milestones (one per line)

13 · Ethics, dissemination, references (optional)

Ethics Dissemination Key references (one per line)

Whitaker HJ, Farrington CP, Spiessens B, Musonda P. Tutorial in biostatistics: the self-controlled case series method. Stat Med 2006;25(10):1768-97. doi:10.1002/sim.2302
Farrington CP. Relative incidence estimation from case series for vaccine safety evaluation. Biometrics 1995;51(1):228-35. doi:10.2307/2533328
Petersen I, Douglas I, Whitaker H. Self controlled case series methods: an alternative to standard epidemiological study designs. BMJ 2016;354:i4515. doi:10.1136/bmj.i4515
Whitaker HJ, Ghebremichael-Weldeselassie Y, Douglas IJ, Smeeth L, Farrington CP. Investigating the assumptions of the self-controlled case series method. Stat Med 2018;37(4):643-58. doi:10.1002/sim.7536 (tests and graphical checks for the assumptions above.)
Farrington CP, Anaya-Izquierdo K, Whitaker HJ, Hocine MN, Douglas I, Smeeth L. Self-controlled case series analysis with event-dependent observation periods. J Am Stat Assoc 2011;106(494):417-26. doi:10.1198/jasa.2011.ap10108 (the method to use when the event curtails observation, e.g. a fatal outcome.)
Musonda P, Farrington CP, Whitaker HJ. Sample sizes for self-controlled case series studies. Stat Med 2006;25(15):2618-31. doi:10.1002/sim.2477 (sample-size formulae for this likelihood, including the age-adjusted case.)
Wang GH, Man KKC, Chang WH, Liao TC, Lai EC. Use of antipsychotic drugs and cholinesterase inhibitors and risk of falls and fractures: self-controlled case series. BMJ 2021;374:n1925. doi:10.1136/bmj.n1925; PMID 34503972. (AsPEN exemplar; 14-day pretreatment window, pretreatment IRR 6.17.)
Hsieh CY, Su CC, Shao SC, Sung SF, Lin SJ, Kao Yang YH, Lai EC. Taiwan's NHIRD: past and future. Clin Epidemiol 2019;11:349-358.
Shao SC, Chan YY, Kao Yang YH, Lin SJ, Hung MJ, Chien RN, Lai CC, Lai EC. The Chang Gung Research Database. Pharmacoepidemiol Drug Saf 2019;28(5):593-600.

Live preview

Auto-saved to this browser

Study design diagram

Pick the timeline anchor above (exposure-anchored vs calendar-anchored). Two layers: (1) covariate / eligibility look-back before the anchor — the same measurement windows as the cohort builder; (2) the case's observation time as one continuous strip of consecutive periods (reference / pre-exposure risk / exposure / post-exposure), which you edit freely — add any period, reorder in the text, or drag a segment's ○ edge. This is a within-person time partition, not a cohort follow-up.

Phenotype library