Protocol Generator

Which design does my question need?

A good protocol names the design it did not choose, and says why.

1. 1

   Do you have an active comparator — another drug, not 'nobody'?

   That means a drug used for the same indication, at the same point in the treatment pathway, with no a-priori link to your outcome. If one exists, a cohort design is available and is usually the better choice: people who get drug A and people who get drug B are far more alike than people who get a drug and people who get nothing, so confounding by indication is much smaller to begin with. If no such comparator exists, a self-controlled or calendar-time design is the usual destination — but not the only one, because a comparator arm need not be a comparator drug.

   If yes → ACNUCCWST

   If no → SCCSCCOITSTIT and questions 2 and 6 still apply. Two of the cohort designs can run against an untreated arm rather than a comparator drug: cloning, where the strategies only declare themselves later — “start within 90 days” — which is question 2; and sequential trial emulation, where eligibility itself varies over time and the not-yet-treated at each origin form the control arm. A rare outcome or a purely descriptive aim does not need a comparator either.

2. 2

   Can you tell the two strategies apart on day one?

   If your strategies are 'start within 90 days', 'stay on treatment', or 'treat to target', then on day one nobody has yet declared which arm they are in. Sorting patients by what they later did credits one arm with follow-up during which they could not yet have had the event — immortal time — and that alone can manufacture a protective effect. Cloning every patient into both strategies, or emulating a fresh trial at each eligible date, is what prevents it.

   If yes — each arm is declared at time zero → this question does not narrow anything on its own: an ordinary new-user cohort is available, but only if question 1 found you a comparator. A sequential trial emulation is also available here, and does not need one — it is triggered by eligibility varying over time rather than by the arms being indistinguishable on day one, and at each origin the initiators and the not-yet-treated are perfectly distinguishable.

   If no — they only diverge later → CCWST

3. 3

   Is the exposure short-lived and the outcome sudden?

   A transient exposure plus an acute, datable onset is what makes a self-controlled design work: exposed and unexposed time both exist within the same person, so the comparison can be made inside them. Choose SCCS if you can lay out each case's whole observation period as exposed and unexposed person-time; choose case-crossover if you can only look back a short way from the event.

   If yes, and you can lay out the whole observation period → SCCS but only if the event neither ends observation nor changes later exposure — a fatal outcome, or a drug that gets stopped once the event happens, rules out the standard SCCS however complete your data are. Extensions exist for both, and the SCCS template here offers the Farrington one; neither is the model you get by default.

   If yes, but you can only look back a short way → CCO

   If no → no self-controlled design applies — the other questions are the ones that route you.

4. 4

   Is the confounding you are worried about fixed, or does it change over time?

   Genetics, lifelong habits, stable frailty — self-controlled designs remove all of these by construction, and trend-in-trend removes them through the trend contrast, in both cases without your ever having to measure them. That is a genuinely strong property. It does not extend to time-varying confounding: ageing, season, worsening disease and calendar trends in prescribing all survive, and must still be modelled explicitly.

   If it is fixed → SCCSCCOTIT and this branch is about confounding you cannot measure — removing it unmeasured is what these three do. A fixed confounder you did record is ordinary work for any of the cohort designs, so answering “fixed” is not on its own a reason to leave them.

   If it changes over time → ACNUCCWST and you must have measured it — none of these removes a time-varying confounder you did not record. Where the confounder is itself affected by earlier treatment, a baseline propensity score is not enough and a g-method is needed.

5. 5

   Was there a date — or a gradual change over the years?

   A policy, reimbursement or label change with a clear start date makes interrupted time series available: the outcome's own pre-intervention trend, extrapolated forward, becomes the counterfactual. If the change was gradual rather than a step, there is no date to interrupt at, and trend-in-trend uses the gradient instead — a drug's use falling away after a safety warning serves as well as one climbing.

   If there was a date → ITS

   If use rose or fell gradually → TIT

   If neither — no date, and no marked trend either way → this question does not route you: both calendar-time designs read the effect off a change over calendar time, and with neither there is nothing for either to read. Almost every drug's use moves over a decade, so answer this on whether that movement is the contrast you mean to rely on — not on whether it happened.

6. 6

   Is the outcome rare — or are you not really comparing anything?

   These are two separate questions that happen to point the same way. For a rare outcome, following an entire cohort spends most of its effort on people who never have the event; sampling controls from the source population instead is far more efficient for the same precision. And if what you want is a rate, a trend or a utilisation pattern rather than an effect, that is a descriptive analysis — legitimate and often the most useful output, provided it is not written up as though it were causal.

   If the outcome is rare → CC though a rare outcome with a good active comparator is often still a cohort study, with case-control sampling inside it for efficiency.

   If no causal contrast is intended → DA

   If the outcome is common and you do intend a causal contrast → neither half of this question applies, which is the ordinary case rather than a problem — questions 1 to 5 are the ones that route you.

One caution. Every design here removes some confounding and leaves the rest. The self-controlled designs remove time-invariant confounding without measuring it, but that does not make the two interchangeable: SCCS uses the whole observation period, so it is unusable if the event changes later exposure, while a case-crossover looking only backwards never samples post-event exposure and is untouched by it — the bidirectional referent offered as a sensitivity analysis gives that immunity up. What breaks the case-crossover is a trend in exposure prevalence, and chronic exposure or insidious onset breaks it outright. The cohort designs handle time-varying confounding, but only for what you actually measured — and measuring it is not always enough: where a confounder is itself affected by earlier treatment, the estimate a baseline propensity score gives is biased however well that confounder was recorded, and a g-method is needed — clone-censor-weight's inverse-probability-of-censoring weighting is the one on this page. Choose on the assumption you can defend, not on the estimate you would prefer — and pre-specify the analysis that would show you were wrong.