Examining the Use of Case-Crossover Design with Non-Transient Medication Exposures

Observational studies are an important study design in pharmacoepidemiology research especially when examining the safety of medications. The case-crossover design is a self-controlled study design that only include patients with outcome. This design has advantages like controlling for time-invariants confounders and avoiding biases from inappropriate control selection. By comparing the probability of exposure between different time points, the case-crossover determines whether patients were more likely to be exposed to a drug at a specific time period. The case-crossover design was first designed for transient exposures; however, recent literatures have showed the increase usage of case-crossover design on non-transient exposures. Non-transient exposures can reduce the precision and introduce biases to the effect estimates of case-crossover design. Recommendations to reduce the impact of these biases include expanding the risk and control window and adding a representative control group. With increased use of the case-crossover design, it is not clear how often it is used with non-transient exposures, how frequently the recommendations were applied, and the magnitude of bias introduced by different medication patterns. The objective of this dissertation was to better understand the impact of applying the case-crossover design to non-transient exposures and to examine ways to correct for biases that might be introduced. We conducted three studies to address this objective. The first study was a systematic review on published case-crossover studies from 2015 to 2021. Our study found that using case-crossover design on non-transient medication remains prevalent and around 40% of them did not apply recommended methods to account for biases. The second study used group-based trajectory model to identify common adherence trajectory patterns of non-transient medication using Sodium-Glucose Transporter 2 (SGLT2) as example. The third study used simulation study to quantify bias from different non-transient medication patterns individually and combined. Moreover, we used a real-world example of SGLT2 and outcomes to further validate our results from simulation study. We found that case-crossover produced unbiased results from transient exposures. When evaluating prolonged exposures or a mixture of different medication patterns, case-time-control with longer risk and control windows resulted in unbiased estimates. The real-world example findings were consistent with the simulation. In this dissertation, we provided evidence that there is a lack of implementation for recommended methods on published case-crossover so that can lead to misguided results. Our simulation showed that it is problematic to use case-crossover design on non-transient exposures. If or when need to use case-crossover on non-transient exposures, researchers should apply appropriate methods.