Causal Inference with Targeted Learning for Producing and Evaluating Real-World Evidence

Targeted Learning (TL) provides a unified framework for generating and evaluating real-world evidence (RWE) and thus can serve as a foundation for principled use of RWE in support of regulatory decision-making. The TL Roadmap is a systematic guide to navigating the study design and analysis challenges inherent in real-world studies, including non-randomized treatment, intercurrent events, and loss to follow up. Steps in the roadmap are as follows: (0) formulate a precise clinical (causal) question of interest and describe the study giving rise to the data; (1) define a statistical model that captures the true data-generating process and a corresponding full data, or causal model; (2) specify a causal estimand (corresponding to step 0) as a parameter of the full data; (3) specify a corresponding statistical estimand in observed data with corresponding identification assumptions; (4) after data acquisition, estimation and inference using targeted minimum loss-based likelihood estimation and super learning; and (5) sensitivity analysis to quantify the robustness of study findings and level of support in the data for the substantive conclusion, including violations of the identification assumptions. We present a case study illustrating how following the roadmap improves clarity and transparency.