PURPOSE: The objective of this paper is to advocate a new way of sampling controls in the case-time-control design in a cohort of drug users when the studied outcome prevents further treatment.

METHODS: Mathematically we demonstrate how a standard sampling of controls, where controls are sampled among all subjects without an event at end-of-study, leads to a biased effect estimate. We propose to add the requirement that controls initiate treatment before the calendar time of event of their matched case to circumvent this. The standard and proposed sampling methods are compared in a simulation study and in an empirical data example examining the effect of nonsteroidal anti-inflammatory drug usage on the risk of upper gastrointestinal bleeding.

RESULTS: When the controls are sampled the standard way, the case-time-control design confers a bias because cases and controls have a different time-trend of exposure. The bias has been upwards in all the scenarios we have investigated. The requirement we add to be a potential control ensures that cases and controls have the same time-trend of exposure when treatment and outcome are independent. The simulation study confirms that the proposed sampling method removes the bias between treatment and outcome. The proposed sampling method lowered the odds-ratio estimate from 3.72 to 3.26 in the data example.

CONCLUSION: The proposed sampling method makes it possible to use the case-time-control design in a cohort of subjects with registered use of a drug when outcome prevents further treatment.