Can Predictive Modeling Inform the Selection of Time Zero for Target Trial Emulations? An Empirical Study of Atorvastatin Initiation in Medicare Beneficiaries

Purpose: When emulating trials of medication initiation using real-world data, there may be ambiguity regarding the most suitable time zero event for the research question of interest. The time zero event must be strongly associated with the clinical indication for treatment, confer a reasonably high probability of actual treatment initiation, and be measurable with sufficient temporal precision in the source data. When it is uncertain whether a candidate event will satisfy these three conditions simultaneously, empirical identification of predictors of medication initiation can provide valuable guidance. The objective of this study was to empirically identify predictors of incident atorvastatin initiation to inform the definition of time zero for future target trial emulations. Methods: A retrospective cohort study was conducted using Medicare claims data from January 1, 2018 - December 31, 2019 (study period). The study population included beneficiaries: age >= 65 years with >= 12 months of continuous enrollment and no statin medication for >= 12 months as of the study period start date. The index date was the date of the first pharmacy dispensing claim for a newly initiated medication during the study period, and classified as atorvastatin or non-atorvastatin (non-atorvastatin initiators randomly sampled at 10%). Candidate predictor variables were ascertained within 6 months pre-index and included demographics, comorbidities, healthcare utilization, and pharmacotherapy. A prespecified eight-step procedure was used to identify significant, independent predictors of incident atorvastatin initiation (vs initiating a medication other than atorvastatin). Results: The study cohort comprised 117,069 incident atorvastatin initiators and 393,083 non-atorvastatin initiators. The strongest predictors of atorvastatin initiation were having an inpatient admission for cerebral infarction (OR 11.29, 95% CI 10.47-12.17) and myocardial infarction (OR 6.56, 95% CI 5.92-7.27). For example, a White male with a recent hospitalization for cerebral infarction had a predicted probability of atorvastatin initiation of 69% (95% CI 68-70). Conclusion: The empirically identified predictors of atorvastatin initiation among statin naive Medicare beneficiaries align with ACC/AHA guidelines recommending prompt statin therapy for secondary prevention. These predictors satisfy the three key requirements for a valid time zero event and should mitigate channeling bias and residual confounding in future target trial emulations.