Global dynamics of a system governing an algorithm for regression with censored and non-censored data under general errors. (Q1428475)

The global dynamics of an iteration scheme for performing regression with grouped data is analyzed. An iterative procedure for estimating the parameters of the regression model is proposed, which is a modification of the well-known EM algorithm of \textit{A. P. Dempster, N. M. Laired} and \textit{D. B. Rubin} [J. Royal Statist Soc., B 39, 1--38 (1977; Zbl 0364.62022)]. It, first, imputes the censored data by a conditional modal step and, second, the estimate is updated by simple least squares. Thus the awkward integration involved in the E-step is avoided by the algorithm. It is shown that the algorithm is characterized as a multidimensional piecewise linear system. Based on a general theorem on the stability of piecewise linear contractive dynamic systems, it is proved that the system is globally convergent to a single equilibrium point regardless the selected initial point. Furthermore, it is shown that the piecewise characterization suggests an alternative algorithm to speed up the search of the unique equilibrium point. Numerical results illustrate the convergence rate and the precision of the estimate.