Asymptotics of M-estimators in two-phase linear regression models. (Q2574536): Difference between revisions

This paper deals with the two-phase linear regression model \[ m(x,\vartheta )=m(x,(\vartheta _1,s))=(\alpha _0+\alpha _1x)I(x\leq s)+(\beta _0+\beta _1x)I(x>s), \] in which \(x\) is the design variable, \(\vartheta _1\in K\), a compact in \(\mathbb R^4\), is a vector of regression coefficients with true value \(\theta _1\), and \(s\in \bar {\mathbb R}\) is a change point with true value \(r\). The paper extends the recent asymptotic results of the first two authors on the maximum likelihood estimator (MLE) for the model [\textit{H.L. Koul} and \textit{L. Qian}, J. Stat. Plann. Inference 108, 99-119 (2002; Zbl 1016.62017)] to the \(M\)-estimator \(\hat {\theta }_n=(\vartheta _{1n}(\hat {r}_n),\hat {r}_n)\), where for \(s\in \bar {\mathbb R}\), \(\vartheta _{1n}(s)\) is the minimizer of the \(M\)-process \(M_n(\vartheta _1,s)\) with respect to \(\vartheta _1\) over \(K\), and \(\hat {r}_n\) is the smallest minimizer of \(M_n(\vartheta _{1n}(s),s)\) over \(\bar {\mathbb R}\). Here, the \(M\)-process \(M_n(\vartheta _1,s)\) is for the random sample of observations \((X_1,Y_1),\dots ,(X_n,Y_n)\) defined \(M_n(\vartheta _1,s)=\sum _{i=1}^n\rho (Y_i-m(X_i,(\vartheta _1,s)))\), with a convex differentiable \(\rho :\mathbb R\to [0,+\infty )\). However, the assumptions considered in the present paper are in several important respects weaker than those for which the MLE results have been obtained. In particular, the derivative of \(\rho \) does not need to be smooth, and the design variables \(X_i\) need to have finite moments only up to second order. The paper consists of two main parts. In the first part, the strong consistency of the \(M\)-estimator is proved, and its consistency rate is shown to be \(n\) for \(\hat {r}_n\) and \(n^{1/2}\) for \(\vartheta _{1n}(\hat {r}_n)\). The proof of the consistency rates substantially relies a recent monograph of the first author [\textit{H.L. Koul}, Weighted empirical processes in dynamic nonlinear models. (2002; Zbl 1007.62047)]. In the second part, the \(M\)-process normalized with respect to \((\theta _1,r)\) and the consistency rate is shown to be uniformly asymptotically equivalent, in probability, to the sum of a quadratic form in the regression coefficients and a jump point process in the change point parameter, which is then used to establish the joint weak convergence of the \(M\)-estimators. In particular, \(n(\hat {r}_n-r)\) is shown to converge weakly to a random variable that minimizes a compound Poisson process, whereas \(n^{1/2}(\vartheta _{1n}(\hat {r}_n)-\theta _1)\) is shown to be asymptotically normal and independent of \(n(\hat {r}_n-r)\). The paper concludes with a simulation study for two different functions \(\rho \) and three different error distributions, and with an application to the investigation of the dependence of gasoline mileage on car weight for data on 38 automobile models, originally reported by \textit{H.\ V.\ Henderson} and \textit{P.\ F.\ Velleman} [Biometrics 37, 391-411 (1981; Zbl 0473.62003)]. The unquestionable value of this paper consists in a rigorous mathematical justification, through the proof of asymptotic properties, for using \(M\)-estimators, and for using them under weaker conditions than under which MLEs has been justified. However, the paper seems to be written more for those readers who only want to know the exact formulation of the proved results and exact conditions on which they hold than for those who would like to follow the proofs in detail. The latter have namely to face not only comparatively many misprints in the proofs, especially in the second part of the paper, but also a highly nonsequential way of reasoning that the paper uses, frequently referring to results established only later (in particular, the proof of the main theorem of the first part relies on two results from the second part, the proofs of which refer back to results established during the proof of that theorem, thus requiring a disentanglement of quite complicated logical relationships), and sometimes also imprecisely described principles of the proofs in comments to them (e.g., the introductory comment to the last two lemmas of the second part states that the first of them is used to prove the second, whereas the exact opposite reveals to be the case). Nevertheless, the key ideas and main steps of the proofs are clearly rendered and well understandable.