Inference And Decision Making With Heterogeneous Information

Author : Imke Mayer
Publisher :
ISBN 13 :
Total Pages : 0 pages
Book Rating : 4.:/5 (128 download)

DOWNLOAD NOW!

Download or read book Causal Inference from Heterogeneous Data with Missing Data written by Imke Mayer and published by . This book was released on 2021 with total page 0 pages. Available in PDF, EPUB and Kindle. Book excerpt: The problem of missing data is unavoidable in statistical practice, most analysis methods cannot be implemented directly from incomplete data. This domain is expanding rapidly within the statistical community, as the problem of missing data is exacerbated by the multiplicity of data collected, often from different sources of information. It is therefore crucial to identify effective methodologies for carrying out (causal) analyses in the presence of incomplete data, and to know how much confidence can be placed in the results obtained from incomplete data.The subject of this dissertation is to propose new methods in the context of causal inference, adapted to some of the challenges of modern data collection processes, namely missingness and heterogeneity; and to develop practical methodologies suited to assess questions of medical relevance and support decision making in a context of time and resource constraints, as it is the case for instance in critical care management.We take the perspective of causal inference and treatment effect estimation to address these challenges. Theory and methodologies for treatment effect estimation are well understood especially in the experimental study case in randomized controlled trials, the textit{gold standard} to assess treatment and intervention effects. However, for observational data causal inference methods and results are not widely applied, the number of successful and accepted examples in applied domains still remain quite low currently and only cover few domains of science. Because classical statistical frameworks can only derive limited causal knowledge from observational data without access to treatment randomization, observational studies are rarely considered acceptable as a valid tool for analyzing causality. Additionally, other challenges that arise in practice, limit the leveraging of observational data. For instance the presence of missing values does not only lead to violations of underlying data generating assumptions, it also leads to practical limitations that make it difficult to (implicitly) ignore such missing values, especially in the era of ``big data'' and high-dimensional data. The contributions of this thesis consist of three main parts. The first part covers the case of missing values in observational studies and their impact on causal analyses, namely identifiability and estimation issues. We propose to explicitly integrate the missing values in the classical causal inference framework, allowing to define identifiability assumptions of treatment effects in the presence of missing values and we provide a generic and flexible estimation approach leveraging recent results from semi-parametric statistics. In the second part we consider a different set of problems, arising in the case of simultaneous availability of experimental and observational studies for the same question of interest the issue of how to relate, such studies, how to leverage their respective advantages and how to overcome their shortcomings is a relevant research topic studied by various fields, ranging from social and economic sciences, to biomedical and pharmaceutical research, as well as among the broad computer science community. We begin by reviewing the current state of the art addressing the question of how to generalize or transport results from experimental studies to more representative and general populations. We then address the question of how these results and methods are altered by the presence of missing values in either data source. To generalize effects of any treatment in such cases, we propose an estimation strategy that is based on multiple imputations. Finally, the third part of this manuscript focuses on utilization: We describe the concrete application, communication and implementation of the developed methodologies to critical care management and other relevant fields. Necessary code resources and instructional tutorials are made available as open source material.