Date of Award


Document Type

Open Access Dissertation

Degree Name

Doctor of Philosophy in Statistics (PhD)

Administrative Home Department

Department of Mathematical Sciences

Advisor 1

Yeonwoo Rho

Committee Member 1

Hie Joo Ahn

Committee Member 2

Chee-Wooi Ten

Committee Member 3

Kui Zhang


The MIDAS models are developed to handle different sampling frequencies in one regression model, preserving information in the higher sampling frequency. Time averaging has been the traditional parametric approach to handle mixed sampling frequencies. However, it ignores information potentially embedded in high frequency. MIDAS regression models provide a concise way to utilize additional information in HF variables. While a parametric MIDAS model provides a parsimonious way to summarize information in HF data, nonparametric models would maintain more flexibility at the expense of the computational complexity. Moreover, one parametric form may not necessarily be appropriate for all cross-sectional subjects. This thesis proposes two new methods designed for mixed frequency data.

First part of this thesis proposes a specification test to choose between time averaging and MIDAS models. If time averaging is enough for given mixed frequency data, there is no need to use complicated nonlinear mixed frequency models. In such case, a specification test that justifies the use of the the simplest model, time averaging, is useful. We propose a specification test revising from a DWH type test. In particular, a set of instrumental variables is proposed and theoretically validated when the frequency ratio is large. As a result, our method tends to be more powerful than existing methods, as reconfirmed through the simulations.

The second part of the thesis provides a new way to identify groups in a panel data setting involving mixed frequencies. A flexible MIDAS model is proposed using a nonparametric approach. This nonparametric MIDAS model is further extended to a panel setting using a penalized regression idea. The estimated parameters can then be clustered using traditional clustering methods. The proposed clustering algorithm delivers reasonable clustering results both in theory and in simulations, without requiring prior knowledge about the true group membership information. An empirical application is presented to examine the panel MIDAS model.