Design and analysis with observational data : Protocols and modeling with the aim of causal inference

Abstract: This thesis consists of six papers that study the design and analysis with observational data.There is a growing interest in using real-world evidence (RWE) for regulatory purposes. The belief is that observational data can make drug developmentmore efficient and speed up patient access to new drugs. Paper I presents a study protocol for a comparative effectiveness evaluation of two recently reimbursed hormonal treatments (NHTs) given to patients with advanced prostate cancer. The study protocol aims to present the study design, which is done without access to outcome data. Paper II presents the results from the same comparative effectiveness evaluation in clinical practice. The study shows the strength of using a matched sample and IV strategies simultaneously, even though a lack of precision using the IV analysis can be noticed.Paper III presents a study protocol from one of a few comparative effectiveness evaluations of the NHTs against Standard of Care (SoC). Almost no patients were prescribed any of the two drugs before June 2015, as the drugs were yet to be reimbursed, creating a possibility of using historical controls. Paper IV presents the results from the comparative effectiveness evaluation. We cannot rule out that the difference in mortality maybe due to confounding. However, using a bounding strategy of the effect, we do not have sufficient evidence to show that NHT reduces mortality compared to SoC.In Paper V, we investigate how high-dimensional data on healthcare consumption can be used when adjusting for imbalances between groups in an observational study. Our method employs a two-level neural attention model, where it is possible to include high-dimensional daily health data.Paper VI introduces the smooth transition duration model. This model allows for analysis of policy changes when the outcome of interest is the duration until some event and when the policy change introduces different regimes, i.e., before and after the change and in the proposed model, we allow for the change between regimes to be gradual.