Case-Control protocol

In plain terms. You find the people who got the disease, then you find comparable people who did not, and you look backwards to see who had the exposure. It is efficient because you only have to look at a sample of the people who stayed well. Nearly everything that goes wrong with a case-control study goes wrong in how that sample is chosen.

The one rule that decides whether the study is valid

Controls must represent the source population — the person-time — that produced the cases. Controls are not a "healthy comparison group": they are a sample of the denominator. If a person in your control sample had developed the outcome, they would have been captured as one of your cases. If that is not true, the sample is not of the study base and the odds ratio is biased no matter how the analysis is done (Wacholder 1992; Vandenbroucke & Pearce 2012).

How you sample decides what the odds ratio means, and it is not always an approximation:

Incidence-density (risk-set) sampling — a control is drawn from those still at risk at the moment each case occurs. The odds ratio estimates the incidence-rate ratio, with no rare-disease assumption. Controls may later become cases, and a person may be sampled more than once; that is correct, not an error.
Cumulative (survivor) sampling — controls are drawn from those still outcome-free at the end of follow-up. The odds ratio approximates the risk ratio only if the outcome is rare; otherwise it overstates it.
Case-cohort (case-base) sampling — controls are a random subcohort drawn at baseline, and one subcohort can serve several outcomes. This scheme names two estimands and you have to say which is yours. The cross-product against the subcohort estimates the risk ratio with no rare-disease assumption — but over a fixed cohort with complete follow-up, which censoring and competing risks break. The usual analysis where follow-up is incomplete is Prentice-weighted (pseudo-likelihood) estimation, which is a weighted Cox partial likelihood and therefore estimates a hazard ratio; its robust variance is needed because the cases and the subcohort overlap. Either is defensible. Reporting one under the other's name is not.

Knol et al (2008) surveyed 150 published case-control studies and reported that 90% gave only an "odds ratio", though most of the designs they used estimated a rate ratio or a risk ratio. Note that the entitlement was usually conditional — of the 125 studies with incident cases, 57 needed a stable source population for the rate-ratio reading and 16 of the 17 risk-ratio cases needed the rare-disease assumption, while only 48 needed no assumption at all, by virtue of time-matching or concurrent sampling. So say which measure you are estimating and what it rests on.

← All designs

Study design diagram

Exposure and covariates are assessed in look-back windows before the index date. Drag the ● handles to adjust the windows.

Phenotype library

Pick a phenotype

Each entry says whether a study has measured how well it identifies what it names. Nothing here is added to your protocol's reference list — this builder has no mechanism for that, so cite what you use yourself.

Consistency checks · case-control 防呆

Study design diagram

Pick a phenotype