Towards an efficient use of available data in clinical research: Development, validation, and implementation of innovative statistical approaches in causal inference

Evidence-based medicine requires the best available research information for decision-making. While randomized trials are considered the gold standard for causal inference, they can be resource-intensive or raise ethical concerns. Robin's generalized methods, or "G-methods", are a family of three methods that combine longitudinal observational data with complex statistical methods to appropriately adjust for time-varying treatments and time-varying confounders. G-methods can support decision-making by emulating hypothetical trials and, under some assumptions, produce comparable causal estimates. However, G-methods' full potential in observational studies remains underexplored. This project develops existing tools and proposes novel and innovative strategies for an efficient use of available data in clinical research by following the common thread of the G-methods. It contributes to the field of G-methods across four distinct levels. First, the project aims to emulate five trials using one of the three G-methods; the parametric g-formula. It will serve as a use-case to demonstrate how causal inference can generate evidence for currently debated clinical topics in medicine, including HIV, transplant medicine and post-surgical management, using well-established, high-quality cohort databases. When possible, dynamic interventions that allow treatment switch reflecting the reality will be developed and serve as pragmatic control. Second, the project contributes to the advancement of G-methods by developing and validating the parametric g-formula in a Bayesian framework. This allows flexible modeling of spatially correlated data that cannot be modeled in a frequentist way due to complex correlation structure. Coupling G-methods with Bayesian geospatial statistics is innovative and opens new horizons for the field. This methodological work will improve the validity of causal inference estimates by correctly specifying models in the presence of spatially correlated data. Implementing the developed methods in a malaria-endemic setting will generate more evidence of the benefit of sleeping under an insecticide-treated bed net on the risk of dying in under five years old children. Rising insecurities related to climate change, war situations, and pandemics contribute to developing spatial inequalities in health. In such a context, the project anticipates future needs for methods that better account for spatial variation. Third, the project contributes to the development of g-estimation (another G-methods) with a Bayesian geostatistical formulation within the framework of analyses of real randomized clinical trials that suffer from interferences due to spatially-correlated inter-current events. These are post-randomization events that can affect study outcomes and modify the initial description of the estimated treatment effect of a trial. Developed methods will be applied to VITAL, an ongoing trial that assesses the effect of a new differentiated service delivery model for HIV care in Lesotho. Deviation from study protocol regarding drug refills might have happened during the COVID-19 pandemic. The project will look at a hypothetical strategy that prevents COVID-19-related inter-current events from having occurred by de-mediating their effect from the direct controlled effect of the planned intervention. Nowadays, considerable attention is being paid to the estimand, a systematic description of the treatment effect to be quantified to answer the trial's research question. The novel methods developed in this project have the potential to influence methodological guidelines that will support the implementation of the estimand.Last, the methods developed will be disseminated to facilitate their understanding and application, using multiple approaches to reach a wide audience from different background.