Regression- Based Inference in Linear Time Series Models with Incomplete Dynamics

) Regression-based heteroskedasticity and serial correlation robust standard errors and specification tests are proposed for linear models that may not represent an expectation conditional on all past information. The statistics are computable via a sequence of linear regressions, and the procedures apply to models estimated by ordinary least squares or two stage least squares. Examples of the specification tests include tests for nonlinearities in static models, exclusion restriction tests in finite distributed lag models, heteroskedasticity/serial correlation-robust Chow tests, tests for endogeneity, and tests of overidentifying restrictions. Some new tests of the assumptions underlying Cochrane-Orcutt estimation are also proposed, and some considerations when applying the various robust tests are discussed.