Target trial emulation, proposed by Hernan and Robins (Hernan, et al., 2016) is a way to mimic a randomized controlled trial using observational data. It offers different benefits such as accelerating the conclusion of a clinical research question, including large number of observations and the ability to assess “real world” effectiveness. However, they are vulnerable to biases such as confounding biases but also to immortal time bias (ITB) (Hernan, et al., 2016) when the interest is in the survival time. Immortal time arises when participants are considered “immortal” with respect to the observed outcome during a specific period. For instance, a patient in the treatment group must survive till the treatment initiation to be included in the treatment group. Consequently, exposure is not determined at study entry. If that period is not handled correctly, bias will rise in the estimations. We here propose a literature review and comparison of some methods to handle ITB in the context of emulated trials, following the work of Maringe et al. (2020) and Wang et al. (2022).

From the literature review, three methods were identified: a time-varying Cox model (Karim, et al., 2016) to incorporate changes in exposure status over time, a sequential Cox (Gran, et al., 2010) model which decomposes study period in intervals where treatment group is re-evaluated based on treatment initiation during each interval, and a cloning-censoring-weighted approach (Robins, et al., 2000) which consists in creating a clone of all patients in group and censoring it when it doesn’t follow the assigned treatment strategy anymore to handle both confounding and ITB. To evaluate those different approaches, we conducted a simulation study under multiple scenarios and compared their respective performance. To simulate data in the presence of ITB, we used the method proposed by Abrahamowicz et al. (2016) based on a permutation algorithm. We also considered qualitative and quantitative covariates to introduce confounding bias. Methods were compared using the restricted mean survival time (RMST), providing a unified comparative measure that bypasses the limitations of parameter or bias specific comparisons (Royston, et al., 2013). We also used a conventional Cox model (Cox, 1972) to act as a reference model.

While the conventional Cox model was easy to implement, simulations produced biased estimations of treatment effects because it failed to handle ITB. In contrast, both the time-varying Cox and sequential Cox models appropriately handled time-varying treatment and provided non-biased estimations of dynamic treatment effects. However, they both require additional adjustments for confounding factors . Particularly, for the sequential Cox model, more complex implementation and inverse probability of censoring weighting (IPCW) are needed. Finally, although the cloning-censoring-weighting approach was specifically designed to handle ITB and other time-related confounding biases, it introduced informative censoring (dealt with IPCW) and some bias in the estimated treatment effects.

The selected methods yield accurate results; however, further investigation is required to ascertain the most suitable approach for specific clinical research questions. These methods will also be employed in an emulated trial to assess the efficacy of administering furosemide prior to extubation in mechanically ventilated, intubated patients, with a focus on survival and extubation failure.