Retrospective Nested Case-Control Study

In a standard retrospective case-control study, cases are identified from a clinical or registry source and controls are drawn from the general population or a hospital population, which creates potential for selection bias because cases and controls may not arise from the same underlying population. In a retrospective nested case-control study, both cases and controls are drawn from the same pre-defined retrospective cohort, so they share the same eligibility criteria and follow-up period. This eliminates many sources of selection bias and ensures the odds ratio estimates the incidence rate ratio for that cohort.

Risk-set sampling means controls are selected from participants who were event-free and still under follow-up at the exact time each case experienced the outcome. This time-matched sampling ensures controls were genuinely at risk of being cases at that moment, so the comparison is between the case's exposure history and the contemporaneous exposure history of those who had not yet experienced the event. Without risk-set sampling, the odds ratio would not accurately estimate the incidence rate ratio.

One to five controls per case is the standard range. Statistical efficiency increases sharply from 1 to 4 controls per case, but gains beyond 4 are small. If the cohort is large and exposure ascertainment is cheap, 4–5 controls per case is common. If exposure retrieval is expensive (e.g., laboratory assays on stored specimens), 1–2 controls per case may be the practical limit.

As an observational design it cannot establish causality on its own, but it can provide strong evidence for or against a causal association when the cohort has high-quality data, exposure precedes outcome by a meaningful lag period, dose-response relationships are present, and findings are consistent across sensitivity analyses. Residual confounding from unmeasured variables remains the main threat to causal inference.

Conditional logistic regression is the correct primary method because it accounts for the matched structure of the case-control sets. Each matched set (one case, one or more controls) is treated as a stratum. Confounders not used in matching are entered as covariates. The resulting odds ratio is an unbiased estimate of the incidence rate ratio from the full cohort under risk-set sampling.