Title: Distributional information criteria for Irpino-Verde two-component regression Authors:  Andrej Srakar - Institute for Economic Research Ljubljana (Slovenia) [presenting]
Abstract: Information criteria are standard tools in model selection. Yet they have not been developed so far in a distributional symbolic data context. Distributional Akaike and Bayesian information criteria are developed using likelihood functions for symbolic data previously generated. Our analysis takes place in a standard two-component Irpino-Verde framework using the Wasserstein metric to assess the distance of distributions and very general assumptions on model errors. Augmented leveraged bootstrap is used to derive confidence intervals for model parameters and statistics. It is studied the asymptotic consistency of the measures in a mathematical and simulation context and it is observed their computational efficiency (a problem of the original Irpino-Verde regression context). The aim is to observe the performance of the approach in a regression of long-term care provision on stringency measures in times of the COVID-19 pandemic using Corona 2 survey data of Survey of Health, Ageing and Retirement in Europe (SHARE). Multiple extensions of the approach are discussed to different distributional divergence metrics, likelihood function specifications, model selection criteria, inference possibilities and extensions to a possible Bayesian distributional context.