Monte Carlo evaluation of classical heteroscedasticity test in linear regression

Abstract

In order to assess and compare the finite-sample performance of five widely used tests for identifying heteroscedasticity in linear regression models—the Park, Breusch–Pagan, Goldfeld–Quandt, Glejser, and White tests— this study uses a Monte Carlo simulation framework. Both homoscedastic and heteroscedastic error structures are used to evaluate the tests' statistical power and empirical Type I error rates. Each experimental situation is duplicated 1,000 times at significance levels of 5% and 1%, and simulations are run for sample sizes ranging from 20 to 100 observations. Significant variations in the competing processes' performance are revealed by the simulation findings. In the majority of heteroscedasticity cases, the Glejser test typically exhibits the best statistical power, demonstrating strong sensitivity to deviations from constant error variance. Nevertheless, in homoscedastic conditions, this increased sensitivity is accompanied by rather high Type I error rates. On the other hand, as sample size and heteroscedasticity severity rise, the Breusch– Pagan and White tests show greater detection power and better control over the nominal Type I error rate. When the degree of heteroscedasticity is quite low, the Goldfeld– Quandt test typically records the lowest performance. Overall, the results indicate that no test predominates under all experimental circumstances. The Breusch-Pagan and White tests offer a better balance between Type I error control and statistical power than the Glejser test, which is extremely susceptible to heteroscedasticity. Therefore, the study offers helpful empirical data to help researchers choose the best heteroscedasticity diagnostic techniques for linear regression analysis.