Files in this item
|(no description provided)|
|Title:||Tests for parameter constancy in dynamic and non-Gaussian models|
|Doctoral Committee Chair(s):||Kuan, Chung-Ming|
|Department / Program:||Economics|
|Degree Granting Institution:||University of Illinois at Urbana-Champaign|
|Abstract:||In this dissertation various tests for parameter constancy in dynamic and non-Gaussian econometric models are studied. The emphasis of this dissertation is given to the moving-estimates (ME) and recursive-estimates (RE) tests because these two tests can be easily extended to models with non-Gaussian errors.
In chapter 2, tests for parameter constancy of a regression model are reviewed and two new tests, curve-CUSUM and front-window moving-estimates tests, are proposed. Asymptotic critical values of the ME test with window bandwidth smaller than 0.5 are obtained from simulations; finite-sample critical values of the one-dimensional ME test are also obtained from response surfaces. It is found that there is no one test that uniformly dominates the other tests and that those tests under consideration usually have better power in detecting a change in the intercept term.
In chapter 3, we observe that the ME and RE tests suffer severe size distortion in dynamic models. As these size distortions are resulted from serial correlation of data, two modifications of the ME and RE tests are proposed to correct the finite-sample bias. It is shown by simulations that these modified versions yield more accurate empirical sizes in finite samples.
Chapter 4 concerns extensions of the ME and RE tests to models with I(1) processes. The limiting distributions of the ME and RE tests are derived and power performance is studied. It is also found that in models with I(1) variables the limiting distributions of the type 2 modified ME and RE tests are similar to those of the original tests in stationary models. Therefore, these modified versions are applicable in detecting a structural change in models with I(0) variables and robust to models with I(1) variables without a change. In particular, these modified versions perform reasonably well in detecting a structural change in a cointegrating regression.
Chapter 5 deals with tests for parameter constancy in a model with non-Gaussian errors. It is shown that the ME and RE tests based on ordinary least squares estimates have no power in models with infinite variance errors. The principles of the ME and RE tests are then extended to non-Gaussian models by replacing the OLS estimates with robust estimates. It is found that the robustified ME and RE tests with least absolute deviation estimates have better power performance than those with trimmed regression quantile estimates.
|Rights Information:||Copyright 1994 Chen, Mei-Yuan|
|Date Available in IDEALS:||2011-05-07|
|Identifier in Online Catalog:||AAI9512328|