Comparing the effectiveness of multiple dynamic treatment strategies using target trial emulation A counterfactual approach to causal inference from real-world data

Background. Treatment decisions that are dependent on if-then rules on disease status or prior treatment information are dynamic treatment decisions. The effectiveness of dynamic treatment strategies is often investigated with real-world data (RWD). As many different therapy strategies can be observed in routine practice, RWD offer great potential. However, RWD are always associated with risks for several biases including immortal time and selection bias. Objectives. This article shows how to adequately compare dynamic treatment strategies and identify the optimal strategy. A case study is used to illustrate the causal approach described above. Materials and methods. We describe how the combination of three counterfactual approaches allows causal interpretation of results. We describe causal diagrams, target trial emulation, and g-methods. The described causal approach is illustrated by a case study examining when antiviral therapy should be initiated in treatment-naive patients with human immunodeficiency virus (HIV) infection. Results. Causal diagrams visualize underlying causal processes. They help to identify parameters that need to be considered in the analysis. Target trial emulation simulates a randomized clinical trial by defining all possible dynamic strategies, copying ("cloning") patient data, and assigning each patient to each treatment arm. In a causal per protocol analysis, all patients violating the protocol of a given treatment strategy are censored. Informative censoring is adjusted by g-methods. The expected outcomes of each treatment strategy are simulated and compared. Conclusions. Dynamic treatment strategies can be adequately compared using RWD when three causal approaches are combined, and the necessary data are available. These approaches are (1) causal diagrams, (2) target trial emulation, and (3) statistical g-methods.