WORKSHOP 6: The impact of classical statistics versus multi-variate pattern analysis (MVPA) on neuroscientific interpretation
21/02/2017 @ 09:00 - 16:00 UTC+0
Prof. Dr. Danilo Bzdok
Neuroimaging datasets are constantly increasing in resolution, sample size, multi-modality, and meta-information complexity. This opens the brain imaging field to a more data-driven machine-learning regime (e.g., classification algorithms, pattern recognition, cross-validation procedures) , while analysis methods from the domain of classical statistics remain dominant (e.g., ANOVA, Pearson correlation, Student’s t-test). Special interest lies in the interpretational opportunities that open when asking the same neurobiological question using analysis methods from different statistical cultures.
We will discuss 3 everyday scenarios of statistical analysis in functional neuroimaging:
1) In TMS intervention during semantic experiments in humans, we will discuss how standard GLM analyses emphasize notions of “neural activity extent” whereas newer support-vector-machine searchlight MVPA analyses emphasize notions of “activity pattern similarity”. This frequently yields opposing brain effects that tell complementary parts of a same neuroscientific story.
2) Why is there no significance test for brain atlasing using parcellation methods? We will discuss the interpretational differences between dividing the brain into separate regions (clustering methods like k-means and hierarchical clustering) and into overlapping, distributed patterns (decomposition methods like principal component analysis and independent component analysis), and how it is possible to perform forms of “statistical inference” on the obtain brain parcellations.
3) How are the data analysis practices likely to change in the face of currently growing data repositories (e.g., UKBiobank and Precision Medicine Inititative), and how will this affect the type of neuroscientific judgments that we can make based on the data (explained variance versus prediction, extrapolation to the general population versus out-of-sample generalization, etc.).
These general principles considered in 3 case studies are important in many different types of data analysis pipelines, including those in behavioral (non-imaging) psychology, genetics and molecular neurosciences in humans and animals. The successful statistical extraction and neuroscientific interpretation of structured knowledge from current and future neuroimaging datasets will be a critical prerequisite for our understanding of human brain organization in healthy populations and psychiatric disease.