The cause and effect paradigm underlying medical research has led to an enhanced etiological understanding of many diseases and the development of many lifesaving drugs, but the paradigm does not always include an understanding of the pathways involved. Causal mediation analysis extends the cause and effect relationship to the cause and effect through a mediator, an intermediate variable on the causal pathway. The total effect of an exposure on an outcome is decomposed into two parts: 1) the indirect effect of the exposure on the outcome through the mediator and 2) the direct effect of the exposure on the outcome through all other pathways. In this dissertation, I describe various counterfactual causal mediation frameworks with identifiability assumptions that all lead to the Mediation Formula. The indirect and direct effects can be estimated from observable data using a semi-parametric algorithm derived from the Mediation Formula that I generalize to different types of mediators and outcomes. With an increased interest in causal mediation analysis, thoughtful consideration is necessary in the application of the Mediation Formula to real-world data challenges. Here, I consider three motivating causal mediation questions in the areas of HIV curative research and cardio-respiratory fitness. HIV curative treatments typically target the viral reservoir, cells infected with latent HIV. Quantifying the effect of an HIV curative treatment on viral rebound over a set time horizon mediated by reductions in the viral reservoir can inform future directions for improving curative treatments. In cardiorespiratory fitness research, metabolites, molecules involved with cellular respiration, are believed to mediate the effect of physical activity on cardiorespiratory fitness. I propose three novel adaptations to the semi-parametric estimation algorithm to address three data challenges: 1) Numerical integration and optimization of the observed data likelihood for mediators with an assay lower limit (left-censored mediators); 2) Pseudo-value approach for time-to-event outcomes on a restricted mean survival time scale; 3) Elastic net regression for high-dimensional mediators. My novel approaches provide estimation frameworks that can be applied to a broad spectrum of research questions. I provide simulation studies to assess the properties of the estimators and applications of the methodologies to the motivating data. / 2025-06-16T00:00:00Z

No description available.

Le bénéfice de survie restreint à un horizon temporel donné a été proposé comme mesure alternative aux mesures relatives de l’effet d’un traitement notamment dans le cas de non proportionnalité des risques de décès. Le bénéfice de survie restreint correspond à la différence des survies moyennes entre deux bras de traitement et s’exprime en nombre d’années de vie gagnées. Dans la littérature, cette mesure est présentée comme plus intuitive que le hazard ratio et plusieurs auteurs ont proposé son utilisation pour le design et l’analyse d’un essai clinique. Toutefois, ce n’est pas actuellement la mesure qui est utilisée de manière courante dans les essais randomisés. Cette mesure s’applique quelle que soit la distribution des temps de survie et est adaptée si l’hypothèse des risques proportionnels n’est pas respectée. De plus, le bénéfice de survie restreint peut être utilisé en évaluation médico-économique où la mesure d’un effet absolu est nécessaire (nombre d’années de vie gagnées pondérées ou non par la qualité de vie). Si l’on souhaite estimer le bénéfice de survie restreint à partir d’une méta-analyse sur données individuelles, se pose alors la question de prendre en compte l’effet essai dû à la structure hiérarchique des données. L’objectif de cette thèse était de comparer des méthodes statistiques d’estimation du bénéfice de survie restreint à partir de données individuelles d’une méta-analyse d’essais cliniques randomisés. Le point de départ a été une étude de cas (étude coût-efficacité) réalisée à partir des données de la Meta-Analysis of Radiotherapy in Lung Cancer. Cette étude a montré que les cinq méthodes d’estimation étudiées conduisaient à des estimations différentes du bénéfice de survie et de son intervalle de confiance. Le choix de la méthode d’estimation avait également un impact sur les résultats de l’analyse coût-efficacité. Un second travail a consisté à mener une étude de simulation pour mieux comprendre les propriétés des méthodes d’estimation considérées en termes de biais moyen et d’erreur-type. Enfin, la dernière partie de la thèse a mis en application les enseignements de cette étude de simulation au travers de trois méta-analyses sur données individuelles dans le cancer du nasopharynx et dans le cancer du poumon à petites cellules. / The survival benefit restricted up to a certain time horizon has been suggested as an alternative measure to the common relative measures used to estimate the treatment effect, especially in case of non-proportional hazards of death. The restricted survival benefit corresponds to the difference of the two restricted mean survival times estimated for each treatment arm, and is expressed in terms of life years gained. In the literature, this measure is considered as more intuitive than the hazard ratio and many authors have suggested its use for the design and the analysis of clinical trials. However, it is not currently the most used measure in randomized trials. This measure is valid under any distribution of the survival times and is adapted if the proportional hazards assumption does not hold. In addition, the restricted survival benefit can be used in medico-economic evaluation where an absolute measure of the treatment effect is needed (number of [quality adjusted] life years gained). If one wants to estimate the restricted survival benefit from an individual participant data meta-analysis, there is a need to take into account the trial effect due to the hierarchical structure of the data. The aim of this thesis was to compare statistical methods to estimate the restricted survival benefit from an individual participant data meta-analysis of randomized trials. The starting point was a case study (cost-effectiveness analysis) using data from the Meta-Analysis of Radiotherapy in Lung Cancer. This study showed that the five investigated methods yielded different estimates for the restricted survival benefit and its confidence interval. The choice of a method to estimate the survival benefit also impacted on cost-effectiveness results. Our second project consisted in a simulation study to have a better understanding of the properties of the investigated methods in terms of bias and standard error. Finally, the last part of the thesis illustrated the lessons learned from the simulation study through three examples of individual participant data meta-analysis in nasopharynx cancer and in small cell lung cancer.