A Comparison of Several Specification Error Tests for a General Alternative.

This paper compares several tests for specification error in regression models when the alternative is nonspecific. The author considers the Durbin-Watson, RESET, Chow, and differencing tests. Perhaps not surprisingly, each test tends to perform well only against their "designed" alternatives. That is, with only a few exceptions, the differencing tests perform well against omitted dynamics, the Durbin-Watson test performs well when the omitted is autocorrelated, RESET performs well against incorrect functional form, and the Chow test performs well against structural shifts. These results are disappointing in that the tests have been proposed as tests against "general" or "nonconstructive" alternatives. Copyright 1989 by Economics Department of the University of Pennsylvania and the Osaka University Institute of Social and Economic Research Association.