Maximin Effects and Neuronal Activity

Specialeforsvar ved Søren Wengel Mogensen

Titel: Maximin Effects and Neuronal Activity

 Abstract: Large-scale heterogeneous data sets are intimately related to the advent of big data and pose both methodological and computational challenges. We define and characterize maximin effects in the context of a linear mixture model. The motivating ideas behind these effects resemble those behind classical minimax estimators in decision theory. In a statistical framework, the maximin effects can be thought of as a it common component that is present for all parameter values in the mixture model. Thus, they offer a parsi-monious way of summarizing an involved model. Maximin aggregation (magging) is a computationally efficient plug-in estimator of the maximin effects. A key property of this estimator is its robustness to outlying data points. This property also means that noisy data can result in an utterly uninformative magging estimate. We apply these methods to a rich and complex data set on neuronal activity in ferrets and suggest some heuristics for better analyzing the output of the magging procedure. We also discuss sampling, misalignment, normalization and computational issues in the context of the ferret data  

Vejleder: Niels Richard Hansen
Censor:   Karl Bang Christensen, Inst. for Folkesundhed