Основные понятия
Exploiting information in non-standard spaces through Generalized Stochastic Dominance (GSD).
Аннотация
The article discusses statistical comparison of random variables in spaces with locally varying scale of measurement. It introduces the concept of Generalized Stochastic Dominance (GSD) to address the challenge of utilizing information encoded in such spaces. The study proposes a statistical test for GSD, operationalized through linear optimization and robustified using imprecise probability models. Applications in multidimensional poverty measurement, finance, and medicine are illustrated.
- Introduction
- Challenges in statistics and machine learning involve comparing random variables mapping between measurable spaces.
- Attention is given to stochastic orderings for comparison.
- Regularization
- Regularization aims to increase the test statistic's sensitivity.
- Two approaches for regularization: order-theoretic and parameter-driven.
- Generalized Dominance
- Introduces a stochastic order that optimally exploits partial cardinal information in multidimensional spaces.
- Extends beyond traditional stochastic dominance.
- Testing for Dominance
- Statistical testing for GSD using i.i.d. samples.
- Formulation of hypotheses and test statistic computation.
- Robustified Testing Using IP
- Robustification of the test towards deviations in assumptions.
- Utilizes imprecise probabilities to account for biased samples.
- Multidimensional Spaces with Differently Scaled Dimensions
- Application of GSD to multidimensional poverty measurement.
- Comparison of subgroups in poverty analysis.
- Applications
- Implementation details for computing test statistics.
- Example application in poverty analysis.
Статистика
"Our findings are illustrated with data from multidimensional poverty measurement, finance, and medicine."
"The proofs of Propositions 1 to 8, and Corollary 1 can be found in the supplementary material."
Цитаты
"Our paper addresses all situations where, in addition, epistemic uncertainty has to be taken into account."
"Our contribution considers generalized stochastic dominance (GSD) that ensures exploiting the entire information encoded in data with locally varying scale of measurement."
"Our framework allows handling multidimensional structures with differently scaled dimensions in an information-efficient way."