A0531
Title: Same data meta analysis: Inference on neuroimaging multiverse data
Authors: Jeremy Lefort-Besnard - INRIA Rennes (France)
Camille Maumet - INRIA Rennes (France)
Thomas Nichols - University of Oxford (United Kingdom) [presenting]
Abstract: There is a great diversity of analytical approaches for brain MRI, meaning that there are many possible variations of one result. Integrating a set of such multiverse statistic maps is challenging, as there is strong and possibly complex dependence between the results, violating the independence assumption of traditional meta-analysis. A suite of same-data meta-analysis (SDMA) models are developed that account for dependence among multiple multiverse results from a single dataset. Two main approaches are obtained, one based on the average of inputs ("Stouffer") and one based on generalized least squares (GLS) that optimally combines the correlated data. The validity and accuracy of these models are assessed in a set of simulations as well as on two real-world multiverse outputs originating from the same data: the "NARPS" multiverse analysis and the "HCP Young Adult" multiverse analysis, which generated respectively 70 and 24 different statistical maps. It is shown that all of the methods control false positives, but on real data, the GLS methods can be very unstable if there are complex patterns of correlation among the inputs. These sorts of SDMA approaches are an important tool as more researchers conduct multiverse analyses.
