|M.Sc Student||Baumer Gili|
|Subject||Discovering Replicated Findings Acros Several Studies|
of High Dimension
|Department||Department of Industrial Engineering and Management||Supervisor||Dr. Marina Bogomolov|
|Full Thesis text|
The aim of replicability analysis is to identify the findings that replicate across independent studies that examine the same features, and quantify the strength of replication. These features can be single-nucleotide polymorphisms (SNPs) examined for association with disease, genes examined for differential expression, etc. The importance of replicability analysis is well recognized in many fields, where the intention is to show that the result is consistent over different studies which are different with respect to some aspects (e.g. the considered population, laboratory, measurement technique), and is not unique to a specific study or setting.
We introduce new replicability analysis procedures. One type of procedures extends the partial conjunction approach to replicability analysis, by incorporating an assumed lower bound for the fraction of hypotheses that are null in all the studies. The other type of procedures includes those that select the most promising features from each study and examine only them for replicability.
We study the performance of our methods in numerical simulations, considering different configurations of signals and different dependencies within studies. We also illustrate the performance of our procedures on real data from independent studies of different psychiatric disorders, where the aim is to find single-nucleotide polymorphisms associated with two or more disorders. We conclude with different directions for future work and possible extensions of the proposed procedures.