Two tests for heterocedasticity in nonparametric regression (Q2430230): Difference between revisions

The nonparametric heteroscedastic regression model \(Y_{t,n}=m(x_{t,n})+s(x_{t,n})\varepsilon_{t,n}\) is considered, where \(m\) and \(s\) are unknown smooth functions, \(x_{t,n}\) are fixed design points on \([0,1]\), and \(\varepsilon_{t,n}\) are independent errors with \(E\varepsilon_{t,n}=0\), \(E(\varepsilon_{t,n})^2=1\) (some extensions for correlated errors are also discussed). Two tests are proposed to test the hypothesis \(H_0\): \(s^2(x)=\)\,const, \(x\in[0,1]\). The first test is based on the classical Fisher test application to squared residuals of locally-polynomial fit for \(Y_{t,n}\). The second one uses a weighted \(L_2\) distance from a nonparametric estimate of \(s^2(x)\) and a constant. A bootstrap-type procedure is used to calculate the critical region for the second fit. Performance of the tests is investigated via simulations.