Conditioning on future exposure to define study cohorts can induce bias: the case of low-dose acetylsalicylic acid and risk of major bleeding

Clin Epidemiol. 2017 Nov 23;9:611-626. doi: 10.2147/CLEP.S147175. eCollection 2017.


Background: A principle of cohort studies is that cohort membership is defined by current rather than future exposure information. Pharmacoepidemiologic studies using existing databases are vulnerable to violation of this principle. We evaluated the impact of using data on future redemption of prescriptions to determine cohort membership, motivated by a published example seeking to emulate a "per-protocol" association between continuous versus never use of low-dose acetylsalicylic acid (ASA) and major bleeding (e.g., cerebral hemorrhage or gastrointestinal bleeding).

Materials and methods: Danish medical registry data from 2004 to 2011 were used to construct two analytic cohorts. In Cohort 1, we used information about future redemption of low-dose ASA prescriptions to identify cohorts of continuous and never-ASA users. In Cohort 2, we identified ASA initiators and non-initiators using only contemporaneous data and censored follow-up for changes in use over time. We implemented propensity score-matched Poisson regression to evaluate associations between ASA use and major bleeding and estimated adjusted incidence rate differences (IRDs) per 1,000 person-years and ratios (IRRs) overall and stratified by time since initiation.

Results: Among >6 million eligible Danish adults, we identified 403,693 low-dose ASA initiators (Cohort 2), of whom 189,150 were defined as continuous users (Cohort 1). Overall, IRDs and IRRs were similar across cohorts. However, the IRD for major bleeding in the first 90 days was substantially larger in Cohort 1 (IRD=25 per 1,000 person-years) compared with Cohort 2 (IRD=10 per 1,000 person-years).

Conclusion: Using future medication redemption data to define baseline cohorts violates basic epidemiologic principles. Compared with an approach using only contemporaneous data to define cohorts, the approach based on future redemption data generated a substantially higher short-term association between low-dose ASA use and major bleeding on the absolute, but not the relative, scale possibly due to selection and immortal time biases.

Keywords: cohort studies; immortal time bias; pharmacoepidemiology; selection bias.