Goodness-of-fit testing the error distribution in multivariate indirect regression (Q2323938)

The authors study goodness-of-fit tests for model assumptions on the distribution function of the error \(\epsilon_j\) in the indirect regression model \[ Y_j = \int_{\left[0,1\right]^m} \theta\left(\mathbf{u}\right) \psi\left(X_j-\mathbf{u}\right)\,d \mathbf{u} + \epsilon_j, \qquad 1 \leq j \leq n \] with known \(\psi\), the Fourier coefficients of which decay polynomially. Throughout the whole paper it is assumed that \(\theta\) is periodic and weakly differentiable, \(\epsilon_1, \dots, \epsilon_n\) are i.i.d. with mean zero and common Lebesgue density \(f\), and the covariates \(X_1, \dots, X_n\) are also i.i.d. and independent of \(\epsilon_1, \dots, \epsilon_n\). Based on a nonparametric Fourier series estimator, the authors consider a Kolmogorov-Smirnov test statistic, the critical values of which are derived by an explicit asymptotic expansion of the corresponding empirical distribution function. After a Khmaladze transformation, this yields an invariance principle allowing to simulate the limiting distribution universally. Finally, the finite sample performance of the test is investigated in a simulation study.