This paper presents a precise characterization of the bias of least squares in two limited dependent variable models, the Tobit model and the truncated regression model. For the cases considered, the method of moments can be used to correct the bias of OLS. For more general cases, the results provide approximations which appear to be relatively robust. 13, and o, . In this paper we present a precise characterization of that bias for the particular case in which xt, as well as -,, is normally distributed. We also show that the bias of the OLS slope estimator can be corrected by dividing each estimate by the sample proportion of nonlimit observations. Other structural parameters can be consistently estimated in a similar fashion. We present some evidence on the effect of nonnormality with respect to the predictions obtained in the normal model. The case in which the sample contains only nonlimit observation (the truncated regression model) is considered elsewhere (Olsen (7)). We analyze the relationship between his results and ours, and derive some predictions of the normal model with respect to the seriousness of "truncation bias."
[1]
Arthur S. Goldberger,et al.
Linear regression after selection
,
1981
.
[2]
R. Olsen,et al.
Approximating a Truncated Normal Regression with the Method of Moments
,
1980
.
[3]
Ray C. Fair,et al.
A Note on the Computation of the Tobit Estimator
,
1977
.
[4]
W. R. Buckland,et al.
Distributions in Statistics: Continuous Multivariate Distributions
,
1974
.
[5]
Takeshi Amemiya,et al.
Regression Analysis when the Dependent Variable is Truncated Normal
,
1973
.
[6]
J. Tobin.
Estimation of Relationships for Limited Dependent Variables
,
1958
.
[7]
K. Pearson,et al.
ON THE GENERALISED PROBABLE ERROR IN MULTIPLE NORMAL CORRELATION
,
1908
.