Emulating a target trial to avoid immortal time-bias in the estimation of PCSK9 inhibitor effects post myocardial infarction: Methodological opportunities and challenges

Background: Observational data may enable estimation of real-world causal effects. However, there are often challenges related to the absence of randomization, confounding and immortal time-bias. One potential solution is “target trial emulation” with a clone-censor-weight estimation approach, which minimizes immortal time bias in settings where the treatment decision and the treatment initiation do not coincide. The strengths and limitations of this approach need to be explored.

Objectives: To emulate a target trial using the clone-censor-weight method to compare negative control outcomes (NCOs) and cardiovascular outcomes of treatment strategies with proprotein convertase subtilisin/kexin type 9 inhibitor monoclonal antibodies (PCSK9i) after a myocardial infarction (MI).

Methods: Using Swedish national health registries, we identified patients that had an MI and a pharmacy dispensation of a PCSK9i between 2015 and 2021. A trial emulation framework was established to evaluate two treatment strategies; (a) initiation of PCSK9i within 12 months after MI and, (b) no initiation of PCSK9i within 12 months. Each MI was cloned into both treatment strategies, and patients were censored when the observed treatment became inconsistent with the assigned treatment, possible informative censoring was addressed by applying inverse probability weights (IPW) of remaining uncensored, given covariates, for each treatment strategy. Covariate balance (defined as a standardized mean difference, SMD of ± 0.1) was assessed at 12 months (i.e., at the time where all informative censoring has occurred in the two treatment strategies). IPWs were applied in the outcome model (weighted Cox proportional hazard model). NCOs were used to assess the presence of residual bias.

Results: In total, 47,346 MI patients were cloned into both treatment strategies. Among those, 561 (1.3%) patients initiated PCSK9i treatment. Whilst, the hazard ratio for most of the NCOs was around 1 (fracture: 1.01, 95% confidence interval [CI] 0.87-1.14; glaucoma: 1.04, 95%CI 0.70-1.45; arthroplasty: 1.14, 95%CI 0.71-1.80), imbalance remained on some covariates (e.g., ezetimibe SMD=0.27). Untrimmed IPWs improved balance at the cost of unreasonable effect estimates. Trimmed IPWs gave more stable effect estimates but at the cost of some covariate imbalance.

Conclusions: The clone-censor-weight approach has potential to handle immortal time bias. However, we noted difficulties in situations with few patients following a treatment strategy (which led to extreme weights), and large baseline covariate differences between treatment strategies. Further research into the implementation of this approach is warranted.