Generalized Estimation of Error Components Models with a Serially Correlated Temporal Effect
提出两种广义最小二乘估计量，用于估计误差由截面特定分量和序列相关时间分量组成的线性面板回归模型，无需参数化时间序列误差的协方差函数，并在频域中计算以简化处理。
For over a decade, econometricians have shown continuing interest in error components models and in the use of panel data. Early contributions to this literature, notably Balestra and Nerlove [1966] and Wallace and Hussain [1969], developed methods for analyzing models in which the error is decomposed into two or more independent, spherical effects. Several useful papers in the early 1970s, Nerlove [1971, 1.971a], Amemiya [1971], and Fuller and Battese [1974], for example, elaborated the properties of error components models and offered alternative methods of estimation. A potentially serious drawback to many previously discussed estimation methods is the tendency to ignore problems typically associated with analysis of time series data. Limited extensions in this direction have been offered by Da Silva [1975], Lillard [1978], Revankar [1980] and Kiefer [1980]. Given these initial efforts, it is important to investigate a more general approach to dealing with time series stochastics in analysis of panel data. This paper examines methods for estimating linear pooled regression models in which the error can be expressed as the sum of a cross section specific component and a serially correlated cross section-time series component. Specifically, we propose two GLS regression estimators, neither of which requires parameterization of the covariance function of the time series errors. There is no unique way to compute these estimates. However, for computational ease, both estimators are computed in the frequency domain. In the case we consider, the problem of serially correlated errors in the time domain is reduced to one of heteroscedasticity in the frequency domain. The only requirements of the data are that time series errors for each cross section be stationary and that errors be stochastically similar across cross sections. Both estimators may be shown to be equivalent to the standard time domain Aitken estimator when error variances and autocovariances are known. When these are unknown, proposed two-step estimators prove equivalent to the Aitken result asymptotically.