Defended theses

In a cluster randomized trial, randomization units are groups of individuals, rather than individuals themselves. Nursing homes are facilities for older adults in need of care. Regarding the type of intervention assessed in such settings, cluster randomized trial is as a well-adapted design. With the global ageing of the population, trials in nursing homes are required but still underrepresented and the reasons are, among others, methodological issues such as the high risk of attrition, essentially due to death. The objective of this PhD thesis was to provide a validated approach to estimate an intervention effect when a cluster randomized trial is planned in nursing homes and faces the risk of a high rate of discontinuation due to death. In the first part of this work we investigated, through a methodological review, the strategies used to deal with that attrition. The review was based on reports of cluster randomized trials planned in nursing homes and published between 2005 and 2020 in selected general medicine and geriatric journals with high impact factors. In the second part of this work, we focused on the closed-cohort recruitment strategy, the most frequently used design but also the most exposed to the risk of attrition. The aim of the second part was to assess how an open-cohort design could have been considered as a relevant alternative to a closed-cohort design. The last part of this work was to assess through a Monte Carlo simulation study how bias can be reduced when estimating an intervention effect using an open-cohort design as compared to a closed cohort, in the context of cluster randomized trial in nursing home with not at random missing data.
Most of the interventions assessed in nursing homes are at cluster level, making the open-cohort a well-adapted design. Individual attrition is no longer an issue and it provides low biased estimates of the intervention effect. Open-cohort must be considered more often when cluster randomized trials are planned in nursing homes.

A cluster randomized trial (CRT) is a study design in which groups of individuals, called clusters, are randomized rather than the individuals themselves. Data then have a hierarchical structure: participants are nested within clusters, which are randomized into intervention arms. When the outcome is binary, the statistical models estimate an odds ratio (OR), while international reporting recommendations also call for the estimation of an absolute effect, i.e. a risk difference (RD). The aim of this thesis was to evaluate analytical methods for estimating a RD in ERCs via simulation studies.

For the two works of this thesis, a single simulation plan was developed to simulate a CRT with a binary outcome, individual- and cluster-level covariates, including confounding in individual-level covariates. The first work focused on individual-level analyses. Conditional (via generalized linear mixed models, GLMM) and marginal (via generalized estimating equations, GEE) approaches were compared using different link functions (identity, logit and log). For the two latter, G-computation was then used to estimate the RD. The marginal approach offered the best performance measures,
except for the convergence rate. The recommendation was therefore to use the GEE approach with an identity link function, or the GLMM approach with the identity link in case of non-convergence. A second work focused on clusterlevel analyses with adjustment on individual and cluster-level covariates. A two-stage procedure (TSP), G-computation and Targeted Maximum Likelihood Estimation (TMLE) were compared, along with an unadjusted method acting as a control method. Once there was adjustment, there was no longer any bias, whatever the method. The TMLE method with adjustment only for individual-level covariates (confounders) showed the best performance, particularly in terms of
estimation precision.

At the end of this work, recommendations on the statistical approach to be used to estimate an absolute intervention effect in the context of an CRT were made, to follow the reporting recommendations of the CONSORT Statement.

Today, therapeutic evaluation primarily relies on the randomized controlled trial with a two-parallel group design. However, for rare diseases, this standard is challenging due to the required sample sizes. Thus, this thesis set three objectives, corresponding to three distinct studies. The first objective was to review the study designs used to evaluate therapies in a given group of rare diseases, specifically rare superficial vascular anomalies (SVA), which particularly affect the pediatric population. A systematic literature review analyzing trials conducted between January 2000 and January 2021 on the treatment of rare SVA was carried out. This review helped to compile an inventory of the study designs used as well as the justifications underlying the choice of these methodological designs. The second objective aimed to deepen the analysis of a methodological design identified for the therapeutic evaluation of rare SVA, namely the "individual stepped-wedge randomized trial," in order to facilitate its implementation by determining an appropriate sample size calculation. The validity of this sample size calculation formula has then been assessed using a Monte Carlo simulation approach. Finally, the third objective of the thesis was to identify the methodological design considered most relevant by a panel of experts, to conduct clinical trials on rare SVA. This work used the international Delphi consensus method, bringing together medical experts and patient association representatives, by presenting them with various clinical and therapeutic situations. This approach led to a consensus on several specific clinical situations, thus providing a basis to guide future therapeutic studies in rare SVA. 

In a cluster randomized trial, randomization units are groups of individuals, rather than individuals themselves. Nursing homes are facilities for older adults in need of care. Regarding the type of intervention assessed in such settings, cluster randomized trial is as a well-adapted design. With the global ageing of the population, trials in nursing homes are required but still underrepresented and the reasons are, among others, methodological issues such as the high risk of attrition, essentially due to death. The objective of this PhD thesis was to provide a validated approach to estimate an intervention effect when a cluster randomized trial is planned in nursing homes and faces the risk of a high rate of discontinuation due to death. In the first part of this work we investigated, through a methodological review, the strategies used to deal with that attrition. The review was based on reports of cluster randomized trials planned in nursing homes and published between 2005 and 2020 in selected general medicine and geriatric journals with high impact factors. In the second part of this work, we focused on the closed-cohort recruitment strategy, the most frequently used design but also the most exposed to the risk of attrition. The aim of the second part was to assess how an open-cohort design could have been considered as a relevant alternative to a closed-cohort design. The last part of this work was to assess through a Monte Carlo simulation study how bias can be reduced when estimating an intervention effect using an open-cohort design as compared to a closed cohort, in the context of cluster randomized trial in nursing home with not at random missing data.
Most of the interventions assessed in nursing homes are at cluster level, making the open-cohort a well-adapted design. Individual attrition is no longer an issue and it provides low biased estimates of the intervention effect. Open-cohort must be considered more often when cluster randomized trials are planned in nursing homes.