Self-controlled and family-based designs
The person (or family) acts as their own control - so all time-invariant/shared confounding disappears without you measuring it
Under development. This page is an orientation to when these designs are relevant, not a full how-to. The sibling section will be expanded with a code example.
So far, comparison has been about finding a group to compare against - a comparison cohort or matched controls - and then adjusting for confounding. Two other designs turn that around: they let the person themselves (or the family) be the control, so all time-invariant confounding (what does not change over time: genes, fixed socioeconomic status (SES), unmeasured lifestyle) disappears without you having to measure it.
What is it called? In Danish the correct form is selvkontrolleret (not selvkontrollerende, which means exercising self-discipline). In practice, though, Danish researchers almost always use the English names, self-controlled case series (SCCS) and case-crossover - so those are what we use here.
Self-controlled designs (SCCS and case-crossover)
Idea. The person is their own control: you look at whether the outcome happens more often in a short period right after a transient exposure than in the person’s other time. Because it is the same person, all confounding that does not change over time (genes, sex, fixed SES, unmeasured lifestyle) disappears, without you measuring it.
When. Transient exposure + acute outcome. Classic: a drug or vaccine and an acute adverse event (the design was originally developed for vaccine safety).
Two variants:
- SCCS (self-controlled case series) models the rate of the outcome within each person’s follow-up and compares exposed vs. unexposed periods.
- Case-crossover compares the exposure in a “hazard window” just before the event with the exposure in earlier reference windows in the same person.
Important assumptions (otherwise the estimate is biased):
- the exposure is transient (not permanent);
- the event does not affect the probability of future exposure (and preferably not the length of follow-up either, e.g. if the outcome is death);
- the outcome is acute and well-defined in time.
R. The SCCS package for self-controlled case series; case-crossover can be done as conditional logistic regression (survival::clogit with a strata() per person, see Regression).
Family- and sibling-based designs
Under development - a code example is coming from a colleague. For now this section is only an orientation.
Idea. Compare siblings (or twins) within the same family. That holds everything siblings share - genes and upbringing environment - constant, and you isolate the effect of what differs between them. Families are linked via the parents’ CPR.
Analysis. Conditional logistic regression or stratified Cox on the family ID - the same strata idea as matched case-control, just with the family as the matched set.
Not a magic bullet. A sibling design only removes shared confounding. Non-shared confounding - what differs between siblings (e.g. different birth weight or differences in upbringing over time) - remains. And if the exposure is strongly correlated between siblings, the design can even amplify bias from remaining measurement error. Use it deliberately, and report both the ordinary and the sibling-controlled analysis.
Remember: anything leaving DST must go through output control - no small cells, only aggregated results. See Phase 14 - Export and repatriation.
- Petersen, Douglas & Whitaker, “Self controlled case series methods: an alternative to standard epidemiological study designs”, BMJ 2016;354:i4515 - an introduction to SCCS.
- The
SCCSpackage - self-controlled case series in R. - Frisell, “Invited Commentary: Sibling-Comparison Designs, Are They Worth the Effort?”, Am J Epidemiol 2021 - what sibling designs can and cannot do.