This thesis develops techniques for adjusting for selection bias using Gaussian process models. Selection bias is a key issue both in sample surveys and in observational studies for causal inference. Despite recently emerged techniques for dealing with selection bias in high-dimensional or complex situations, use of Gaussian process models and Bayesian hierarchical models in general has not been explored. Three approaches are developed for using Gaussian process models to estimate the population mean of a response variable with binary selection mechanism. The first approach models only the response with the selection probability being ignored. The second approach incorporates the selection probability when modeling the response using dependent Gaussian process priors. The third approach uses the selection probability as an additional covariate when modeling the response. The third approach requires knowledge of the selection probability, while the second approach can be used even when the selection probability is not available. In addition to these Gaussian process approaches, a new version of the Horvitz-Thompson estimator is also developed, which follows the conditionality principle and relates to importance sampling for Monte Carlo simulations. Simulation studies and the analysis of an example due to Kang and Schafer show that the Gaussian process approaches that consider the selection probability are able to not only correct selection bias effectively, but also control the sampling errors well, and therefore can often provide more efficient estimates than the methods tested that are not based on Gaussian process models, in both simple and complex situations. Even the Gaussian process approach that ignores the selection probability often, though not always, performs well when some selection bias is present. These results demonstrate the strength of Gaussian process models in dealing with selection bias, especially in high-dimensional or complex situations. These results also demonstrate that Gaussian process models can be implemented rather effectively so that the benefits of using Gaussian process models can be realized in practice, contrary to the common belief that highly flexible models are too complex to use practically for dealing with selection bias.

No description available.

No description available.

No description available.

No description available.

No description available.

La randomisation mendélienne est une méthode d’instrumentation utilisant des instruments de nature génétique afin d’estimer, via par exemple la régression des moindres carrés en deux étapes, une relation de causalité entre un facteur d’exposition et une réponse lorsque celle-ci est confondue par une ou plusieurs variables de confusion non mesurées. La randomisation mendélienne est en mesure de gérer le biais de confusion à condition que les instruments utilisés soient valides, c’est-à-dire qu’ils respectent trois hypothèses clés. On peut généralement se convaincre que deux des trois hypothèses sont satisfaites alors qu’un phénomène génétique, la pléiotropie, peut parfois rendre la troisième hypothèse invalide. En présence d’invalidité, l’estimation de l’effet causal de l’exposition sur la réponse peut être sévèrement biaisée. Afin d’évaluer la potentielle présence d’invalidité lorsqu’un seul instrument est utilisé, Glymour et al. (2012) ont proposé une méthode qu’on dénomme ici l’approche de la différence simple qui utilise le signe de la différence entre l’estimateur des moindres carrés ordinaires de la réponse sur l’exposition et l’estimateur des moindres carrés en deux étapes calculé à partir de l’instrument pour juger de l’invalidité de l’instrument. Ce mémoire introduit trois méthodes qui s’inspirent de cette approche, mais qui sont applicables à la randomisation mendélienne à instruments multiples. D’abord, on introduit l’approche de la différence globale, une simple généralisation de l’approche de la différence simple au cas des instruments multiples qui a comme objectif de détecter si un ou plusieurs instruments utilisés sont invalides. Ensuite, on introduit les approches des différences individuelles et des différences groupées, deux méthodes qui généralisent les outils de détection de l’invalidité de l’approche de la différence simple afin d’identifier des instruments potentiellement problématiques et proposent une nouvelle estimation de l’effet causal de l’exposition sur la réponse. L’évaluation des méthodes passe par une étude théorique de l’impact de l’invalidité sur la convergence des estimateurs des moindres carrés ordinaires et des moindres carrés en deux étapes et une simulation qui compare la précision des estimateurs résultant des différentes méthodes et leur capacité à détecter l’invalidité des instruments. / Mendelian randomization is an instrumentation method that uses genetic instruments to estimate, via two-stage least squares regression for example, a causal relationship between an exposure and an outcome when the relationship is confounded by one or more unmeasured confounders. Mendelian randomization can handle confounding bias provided that the instruments are valid, i.e., that they meet three key assumptions. While two of the three assumptions can usually be satisfied, the third assumption is often invalidated by a genetic phenomenon called pleiotropy. In the presence of invalid instruments, the estimate of the causal effect of exposure on the outcome may be severely biased. To assess the potential presence of an invalid instrument in single-instrument studies, Glymour et al. (2012) proposed a method, hereinafter referred to as the simple difference approach, which uses the sign of the difference between the ordinary least squares estimator of the outcome on the exposure and the two-stage least squares estimator calculated using the instrument. Based on this approach, we introduce three methods applicable to Mendelian randomization with multiple instruments. The first method is the global difference approach and corresponds to a simple generalization of the simple difference approach to the case of multiple instruments that aims to detect whether one or more instruments are invalid. Next, we introduce the individual differences and the grouped differences approaches, two methods that generalize the simple difference approach to identify potentially invalid instruments and provide new estimates of the causal effect of the exposure on the outcome. The methods are evaluated using a theoretical investigation of the impact that invalid instruments have on the convergence of the ordinary least squares and two-stage least squares estimators as well as with a simulation study that compares the accuracy of the respective estimators and the ability of the corresponding methods to detect invalid instruments.