Robust estimation based on grouped-adjusted data in censored regression models (Q2640285)

The censored regression model is defined as an ordinary multiple linear regression model, but the dependent variable \(y^*_ j\) can not be observed if \(y_ j^*<0\). Thus, observations are made on \(y_ j=\max (y^*_ j,0)\). Several estimators of the parameter vector have been proposed. One is the maximum likelihood estimator, also known as the Tobit MLE. In this article, an alternative estimator is proposed. The merits of the new estimator compared to the older are that the new estimator is consistent and asymptotically normally distributed even under heteroscedasticity and nonnormally distributed error terms. The computational efforts seem also to be reduced for the new estimator compared to some of the older. The estimation procedure consists of two stages. In the first stage the sample space of the independent variables is divided into a number of cells and the median of the dependent variable is computed for each cell. A special procedure described is used if more than 50 \% of the observations are censored in a cell. In the second stage a weighted Tobit MLE is applied to the medians from the first stage. Under a set of assumptions consistency and asymptotic normality are proved. A Monte Carlo experiment is also conducted in order to illustrate the effects of nonnormality and heteroscedasticity. A simple linear model with fixed values on the independent variable and a fixed value on the slope parameter is used in the simulation. For studying the effects of nonnormality three different error distributions are used: the Cauchy, Laplace, and normal distribution. Two degrees of censoring are used: 25 \% and 50 \%. The intercept is set equal to -10.0 when the censoring is 50 \%, while in the 25 \% censoring case the intercept is set equal to - 1.6 when the error term distribution is Cauchy, equal to -2.09 for the Laplace distribution, and -2.94 for the normal distribution. Unfortunately, this design opens up the possibility of contamination between some of the ``treatments'' in the experiment. A similar contamination problem is also apparent in the design for studying the effects of heteroscedasticity. Therefore, the results from the experiments must be considered to be of very limited value. The situation does not become clearer when we learn from the article that the design is styled after previously published studies.