Abstracts Category : Other

New statistical methods for using confounders measured in a validation sample to enhance time-to-event analyses of large databases

by Rebecca Burne

MOTIVATION Unmeasured confounding is a common, and important, problem in many observational studies. Pharmacoepidemiological studies, which typically rely on large administrative databases, are especially prone to this source of bias. Often, however, smaller validation samples are available, such as clinical datasets collected as part of a study into a particular disease or drug, which provide information on additional confounders, not measured in the large 'main' database. In the last decade, several methods have been proposed to utilize such validation samples in order to correct for unmeasured confounding bias in the analyses of large main databases. However, most recent methods do not extend to time-to-event analyses, and none have as yet been developed which deal with time-varying unmeasured confounding and mediation. OBJECTIVES, METHODS and RESULTS The over-arching aim of this thesis is, therefore, to develop and validate new methods for enhancing time-to-event analysis in the presence of unmeasured confounders available only in a small validation sample. The original contributions of the thesis consist of three manuscripts. In the first manuscript, I rely on comprehensive simulations to assess the performance, in time-to-event analyses, of an existing method for utilizing validation sample data to reduce unmeasured confounding bias, propensity score calibration (PSC), (Strmer et al., 2005). The second manuscript proposes a novel, martingale residual (MR)-based imputation approach to this problem for time-to-event analysis, and validates it in simulations. In a real-life application, with an external validation sample, the method suggested a substantially weaker association between glucocorticoid therapy and risk of type II diabetes mellitus than conventional analysis that adjusted for only fully measured confounders. In the third manuscript, I extend the MR-based imputation to Cox Marginal Structural Models (MSM), with time-varying confounders/mediators, and demonstrate that it yields more accurate estimates than alternative methods based on regression calibration. I then apply the proposed method in Cox MSM analyses to re-assess the association between DPP-4 inhibitor use and risk of hypoglycemia, using a large database where HbA1c results, a potential time-varying confounder/mediator, is available only for a small subset of study subjects. MOTIVATION Les facteurs de confusion non mesurs sont un problme commun, et important, dans de nombreuses tudes d'observation. Les tudes pharmacopidmiologiques, qui reposent gnralement sur de grandes bases de donnes administratives, sont particulirement sujettes cette source de biais. Souvent, cependant, des chantillons de validation plus petits sont disponibles, tels que des bases de donnes cliniques recueillies dans le cadre d'une tude sur une maladie ou un mdicament particulier, ce qui fournit des informations sur des facteurs de confusion supplmentaires, non mesurs dans la grande base de donnes principale. Dans la dernire dcennie, plusieursAdvisors/Committee Members: Michal Abrahamowicz (Supervisor).