DescriptionThis dissertation develops efficient statistical methodologies for combining information
from independent sources. The developments focus on two settings where the
studies are heterogeneous or the studies involve rare events. In these settings, the conventional
combining approaches often lead to inefficient or even invalid statistical inference.
In this dissertation, we propose effective and efficient combining approaches
using confidence distributions. The proposed approaches are justified both theoretically
and numerically. They are also shown to be superior to the conventional approaches.
Combining information from multiple studies, often referred to as meta-analysis
in the literature, has been used extensively in many fields, including health sciences,
social sciences, and others. However, there remain many unresolved problems on how
to effectively and efficiently combine information. For example, • Heterogeneous studies – When the effect of interest is not estimable in heterogeneous studies (e.g., indirect evidence), how can we utilize these studies to perform meta-analysis? • Rare events studies – For clinical trials with rare events, how can we perform meta-analysis and incorporate studies with zero events in the analysis without using artificial continuity corrections or relying on large sample theory? To address these challenging but recurrent problems, this dissertation develops new
meta-analysis approaches based on combining confidence distributions. Roughly speaking,
a confidence distribution refers to a sample-dependent distribution function on
the parameter space with desirable inferential properties in terms of repeated sampling
performance. It can be viewed as a frequentist counterpart to the posterior distribution
in Bayesian inference. In this dissertation, we show the combination of confidence distributions
has desirable properties which are lacking in the conventional approaches. Specifically, 1) in the presence of heterogeneous studies, the proposed approach integrates direct and indirect evidence and achieves asymptotic efficiency; 2) for rare event studies, the proposed approach yields exact inference and incorporates all the studies in the analysis without using artificial continuity corrections for zero events. These properties are demonstrated numerically in simulation studies and real data examples, including flight landing safety data collected by the Federal Aviation Administration and drug safety data collected in diabetes clinical trials.