Working Paper Department of Economics ^simheteicallt Trimmki) Least Squares Estimation Foe Tobit Models

This papjer proposes alternatives to maximum likelihood estimation of the censored and truncated regression models (known to economists as "Tobit" models) . The proposed estimators are based on symmetric censoring or truncation of the upper tail of the distribution of the dependent variable. Unlike methods based on the assumption of identically distributed Gaussian errors/ the estimators are consistent and asymptotically normal for a wide class of error distributions and for heteroscedasticity of unknown form. The paper gives the regularity conditions and proofs of these large sample results, demonstrates how to construct consistent estimators of the asymptotic covariance matrices, and presents the results of a simulation study for the censored case. Extensions and limitations of the approach are also considered. SYMMETRICALLY TRIMMED LEAST SQUARES ESTIMATION FOR TOBIT MODELsl/ by James L. Powell 1 . Introduction Linear regression models with nonnegativity of the dependent variable—known to economists as "Tobit" models, due to Tobin's [1958] early investigation—are commonly used in empirical economics, where the nonnegativity constraint is often binding for prices and quantities studied. When the underlying error terms for these models are known to be normally distributed (or, more generally, have distribution functions with a known parametric form), maximum likelihood estimation and other likelihood-based procedures yield estimators which are consistent and asymptotically normally distributed (as shown by Amemiya [1973 J "and Heckman [I979j)» However, such results are quite sensitive to the "specification of the error distribution; as demonstrated by Goldberger [198OJ and Arabmazar and Schmidt [1982 J, likelihood-based estimators are in general inconsistent when the assumed parametric form of the likelihood function is incorrect. Furthermore, heteroscedasticity of the error terms can also cause inconsistency of the parameter estimates even when the shape of the error density is correctly specified, as Hurd [1979J, Maddala and Kelson [1975J, and Arabmazar and Schmidt [1981 J have shown. Since economic reasoning generally yields no restrictions concerning the form of the error distribution or homoscedasticity of the residuals, the sensititivy of likelihood-based procedures to such assumptions is a serious concern; thus, several strategies for consistent estimation without these assumptions have been investigated. One class of estimation methods combines the likelihood-based approaches with nonparametric estimation of the distribution function of the residuals; among these are the proposals of Miller [1976 J, Buckley and James [1979], Duncan [1983], and Fernandez [1985 J. At present, the large-sample distributions of these estimators are not established, and the methods appear sensitive to the assumption of identically distributed error terms. Another approach was adopted in Powell [1984 J, which used a least absolute deviations criterion to obtain a consistent and asymptotically normal estimator of the regression coefficient vector. As with the methods cited above, this approach applies only to the "censored" version of the Tobit model (defined below), and, like some of the previous methods, involves estimation of the density function of the error terms (which introduces some ambiguity concerning 2/ the appropriate amount of "smoothing").— In this study, estimators for the "censored" and "truncated" versions of the Tobit model are proposed, and are shown to be consistent and asymptotically normally distributed under suitable conditions. Ihe approach is not completely general, since it is based upon the assumption of symmetrically (and independently) distributed error terms; however, this assumption is often reasonable if the data have been appropriately transformed, and is far more general than the assumption of normality usually imposed. The estimators will also be consistent if the residuals are not identically distributed; hence, they are robust to (bounded) heteroscedasticity of unknovm form. In the following section, the censored and truncated regression models are defined, and the corresponding "symmetrically trimmed" estimators are proposed and discussed. Section 3 gives sufficient conditions for strong consistency and asymptotic normality of the estimators, and proposes consistent estimators of the asymptotic covariance matrices for purposes of inference, while section 4 presents the results of a small scale simulation study of the censored "symmetrically trimmed" estimator. The final section discusses some generalizations and limitations of the approach adopted here. Proofs of the main results are given in an appendix. 2. Definitions and Motivation In order to make the distinction between a censored and a truncated regression model, it is convenient to formulate both of these models as appropriately restricted version of a "true" underlying regression equation (2.1) y* = x'^p^ + u^, t = 1, ..., T, where x. is a vector of predetermined regressors, p is a conformable vector of 4/ O unknown parameters, and u, is a scalar error term. In this context a censored regression model will apply to a sample for which only the values of x, and y = max{0, y*} are observed, i.e., (2.2) y^ = max{0, x^p^ + u^} ,