A0169
Title: Midastar: Threshold autoregression with data sampled at mixed frequencies
Authors: Kaiji Motegi - Kobe University (Japan) [presenting]
John Dennis - Institute for Defense Analyses (United States)
Abstract: The aim is to propose Midastar models by combining the mixed data sampling (MIDAS) approach and the threshold autoregressive (TAR) models. The Midastar model of the first kind is designed for a low-frequency target variable and a high-frequency threshold variable, while the second kind is designed for the reverse case. The Midastar models can detect threshold effects accurately, while the aggregated TAR model has a risk of finding spurious non-threshold effects. The Midastar models have desired asymptotic and finite sample properties. As an empirical application, the Midastar model of the first kind is fitted to Japan's COVID-19 data. The target variable is the growth of weekly hospitalization, and the threshold variable is the growth of daily new confirmed cases in Japan. Statistically significant threshold effects, revealing heterogeneity between the contraction and expansion regimes of the pandemic, are detected. The threshold effects vanish once the daily new confirmed cases are aggregated into the weekly level, a spurious non-threshold effect.
