A self-controlled design still depends on measuring exposure during relevant periods.
The Korean database study compared analyses with inpatient and outpatient prescriptions against analyses using outpatient prescriptions alone. In the mortality example, removing inpatient exposure changed a positive benzodiazepine association into an apparently protective one across three self-controlled designs. The effect was much smaller for the hip-fracture example.
The work demonstrated a data-availability bias in selected examples, not the true causal mortality effect of benzodiazepines. Its magnitude need not transfer to every database or outcome. Hospitalization near death can make missing exposure especially consequential.
For evaluating epidemiological evidence, ask whether hospital-administered medicines appear in the database and how hospitalized days were classified. Within-person comparisons control some stable characteristics but do not repair unobserved exposure. An apparently protective estimate may reflect where data disappear rather than pharmacological benefit.
Source: Publication PMC9794445. The case-review applications above are interpretive analysis.
Related reading: Peruvian ICU Cohort’s Benzodiazepine Dose Association Was Not a Randomized Mortality Effect.