Robust inference for ordinal response models (Q2408243)

Consider an ordinal response model of the form \[ Y_i = j \Longleftrightarrow \alpha_{j-1} <Y^\ast_i \leq \alpha_j, \quad j=1,2,\ldots, m, \] where \(-\infty = \alpha_0 < \alpha_1 < \ldots <\alpha_m = +\infty\) and, for \(n\) observations, the latent variable \(Y_i^{\ast}\) follows a regression model of the form \[ Y_i^\ast = \beta_1 X_{i1} + \beta_2 X_{i2} + \ldots + \beta_{ip} X_{ip} + \varepsilon_i, \quad i=1,2,\ldots, n. \] Here, \((Y_i, X_{i1}, \ldots, X_{ip})\), \(i=1,\ldots, n,\) are observed and one is interested in estimating \(\theta = (\alpha_1,\ldots, \alpha_{m-1}, \beta_1,\ldots, \beta_p)\). In the paper, first the classical maximum likelihood estimator inference for this problem is reviewed. Then the robustness of current methods for such models is analysed, with special emphasis on the behaviour of the generalized residuals for some popular link functions. After that, a new robust estimator based on \(M\)-estimation is proposed for this model, and its asymptotic behaviour is analysed analytically and numerically. The results are also used to derive robust testing procedures for this model. The results are illustrated with many examples and simulations.