TESTING FOR SERIAL CORRELATION IN LEAST SQUARES REGRESSION. III.
Abstract
Long ago in the desk computer days, the authors suggested, as a test statistic for serial correlation in the errors u of the regression model, y = X(beta)+u d = Summation ((Z sub t) - (Z sub (t-1))) squared/Summation ((Z sub t) - Z bar) squared. Tables of bounding significance points (d sub L) and (d sub U) were provided and a Beta approximation suggested to resolve the inconclusive result, (d sub L)<(d sub obsd.)<(d sub U). A large literature has grown up; other statistics, with distributions not depending on X, and approximate d distributions have been suggested. It is shown here that d has power advantages over its competitors and that, with modern computers, there is no problem in finding the significance points of d - in fact, the original Beta method is quite satisfactory. In an appendix the basic distribution theory of ratios of quadratic forms is given. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1970
- Accession Number
- AD0707015
Entities
People
- Geoffrey S. Watson
- James Durbin
Organizations
- Johns Hopkins University