Robust tests for heteroscedasticity in a general framework
Résumé
In this paper, we suggest two heteroscedasticity tests that require little knowledge of the functional relationship determining the variance. The first one is based on a Taylor series expansion of the unknown scedastic function and the second one is based on artificial neural networks. These tests are easy to apply and perform well in our small-sample simulations, but they possess asymptotically incorrect sizes except in the case of normal errors. Therefore, we propose a simple modification in order to correct this non-robustness property. We investigate the size and the power of these tests by Monte Carlo experiments by comparing them to well-known heteroscedasticity tests.