Search

This manuscript addresses three new practical methodologies for topics on Bayesian analysis regarding sampling designs and robustness on network data: In the first part of this thesis we propose a general approach for comparing sampling designs. The approach is based on the concept of data compression from information theory. The criterion for comparing sampling designs is formulated so that the results prove to be robust with respect to some of the most widely used loss functions for point estimation and prediction. The rationale behind the proposed approach is to find sampling designs such that preserve the largest amount of information possible from the original data generating mechanism. The approach is inspired by the same principle as the reference prior, with the difference that, for the proposed approach, the argument of the optimization is the sampling design rather than the prior. The information contained in the data generating mechanism can be encoded in a distribution defined either in parameter's space (posterior distribution) or in the space of observables (predictive distribution). The results obtained in this part enable us to relate statements about a feature of an observed subgraph and a feature of a full graph. It is proven that such statements can not be connected by invoking conditional statements only; it is necessary to specify a joint distribution for the random graph model and the sampling design for all values of fully and partially observed random network features. We use this rationale to formulate statements at the level of the sampling graph that help to make non-trivial statements about the full network. The joint distribution of the underlying network and the sampling mechanism enable the statistician to relate both type of conditional statements. Thus, for random network partially and fully observed features joint distribution is considered and useful statements for practitioners are provided. The second general theme of this thesis is robustness on networks. A method for robustness on exchangeable random networks is developed. The approach is inspired by the concept of graphon approximation through a stochastic block model. An exchangeable model is assumed to infer a feature of a random networks with the objective to see how the quality of that inference gets degraded if the model is slightly modified. Decision theory methods are considered under model misspecification by quantifying stability of optimal actions to perturbations to the approximating model within a well defined neighborhood of model space. The approach is inspired by all recent developments across the context of robustness in recent research in the robust control, macroeconomics and financial mathematics literature. In all topics, simulation analysis is complemented with comprehensive experimental studies, which show the benefits of our modeling and estimation methods.