Investigating Alternative Estimation Methods for LDV Econometric Models that Normally Involve Multidimensional Integrals

一大类有限因变量(LDV)模型的估计虽然具有理想的理论性质，但由于需要计算多维积分，用传统方法是不可行的。例如，具有几个同时相关的备选方案的离散选择模型和具有时间相关误差的LDV模型。最近提出的一种方法(McFadden(1986))，模拟矩方法(MSM)，通过避免数值计算多重积分的需要，使得某些离散选择模型的估计是可行的。Hajivassiliou和McFadden(1987)对该方法进行了扩展，并提供了有限的计算实验。然而，在当前的大型计算机技术下，不可能将这种新方法(MSM)与经典最大似然估计(MLE)的理论基准进行对比，因为MLE需要难以处理的积分。鉴于这些估计问题的性质，超级计算机矢量化技术的应用将提供一套全面的实验来比较MSM和其他经典的估计方法。此外，超级计算机的使用将使经典的最大似然估计方法适用于其中几个模型，考虑到所涉及的积分，这一方法目前超出了大型计算机的能力。这项奖励将提供匹兹堡中心所需的计算能力。

The estimation of a large class of limited dependent variables (LDV) models, though possessing desirable theoretical properties, is infeasible by conventional methods because it requires the evaluation of multidimensional integrals. Examples are discrete choice models with several contemporaneously correlated alternatives and LDV models with temporally correlated errors. A recently proposed method (McFadden (1986)), the Method of Simulated Moments (MSM), was shown to make the estimation of certain discrete choice models feasible, by avoiding the need for numerically calculating multiple integrals. Hajivassiliou and McFadden (1987) extended the method and presented limited computational experiments on it. With current mainframe computer technology, however, it is not possible to investigate the performance of this novel method (MSM) against the theoretical benchmark of classical maximum likelihood estimation (MLE) which requires the intractable integrals. Given the nature of these estimation problems, the application of supercomputer vectorization techniques will afford a comprehensive set of experiments to compare the MSM against the other classical estimation methods. Moreover, the use of a supercomputer will make feasible the classical method of MLE for several of these models, which, in view of the integrals involved, is currently beyond the capacity of mainframe computers. This award will provide access to the requisite computing capacity at the Pittsburgh Center.