Experiments with autoregressive error estimation
Autocorrelated errors are recognized as potentially troublesome in regression analysis. Because of the computational problems encountered, however, few economists have estimated equations under the assumption of autocorrelated errors. Recently, relatively economical procedures have been developed for estimating equations containing autocorrelated errors. In this study, one of these procedures-autoregressive least squares (A.L.S.) -is applied to equations describing the behavior of various economic agents, by using different unit observation periods-year, quarter and month. Some of the results have been published elsewhere; some are published here. In addition to presenting some results of autoregressive error estimation, this report summarizes experience with the use of AL.S. Some equations presented here were estimated by a simultaneous equations method under the assumption of autocorrelated errors.
The results of four different tests for autocorrelation in errors were compared: Durbin-Watson d statistic, Theil-Nagar d, Hart-von Neumann ratio and A.L.S. Essentially, the Theil-Nagar d test classes as significant those values of d that are significant or inconclusive in the Durbin-Watson test. The Theil-Nagard yielded evidence of autocorrelated errors most frequently; A.L.S., second most frequently; Hart-von Neumann ratio, third most frequently; and Durbin-Watson test, least frequently. The proportions of the equations in which each test provided significant evidence of autocorrelated errors are: Theil-Nagard, 66 percent; autoregressive least squares, 51 percent; Hart-von Neumann ratio, 37 percent; Durbin-Watson test, 26 percent.