The use of uncertain exposure-A method to define switching and add-on in pharmacoepidemiology

When defining exposure to pharmacological treatments in pharmacoepidemiology, register data often do not provide information regarding if a pharmacological treatment is a switch or an add-on. This study aims to compare two methods defining switching and add-on therapies and their impact on exposure-outcome associations. Additionally, to guide bias reduction, it aims to describe how the methods relate to immortal time bias and selection bias.

Cohort study using Swedish population-based health registers to identify antidepressant (AD) prescriptions as exposures while hospitalizations for psychiatric reasons were used as an empirical outcome example. The first method for exposure definition used conditioning on future exposure (FE), the second used the concept of uncertain exposure (UE). To estimate associations between outcome and exposure categories "Use of one AD," "Use of two or more ADs", and "UE" compared to "Unexposed," hazard ratios (HRs) and 95% confidence intervals were estimated using Cox regression adjusted for age and sex.

Using the UE method, 7.2% of time periods were classified as "UE" with a notable proportion of psychiatric hospitalizations (7.7%) occurring during this time, while when using the FE method these hospitalizations were distributed over unexposed time and AD use time. The FE method resulted in slightly higher associations than the UE method. The highest HR was found during "UE": HR (95% CI) 5.54 (5.06-6.07).

This study suggests that to reduce the potential immortal time bias, selection bias, and exposure misclassification inherent to the FE method, the UE method could be used for identifying switching and add-on therapies. If not used as a main exposure definition, the UE method may be used to investigate the impact of UE time in a sensitivity analysis.