4

Over 60 years ago Durbin and Watson suggested a testing procedure for assessing autocorrelation in regression relationships.The test is known to not work in the presence of lagged dependent variables, in which case Breusch-Godfrey test from the late 1970s applies. Any other known situations where the Durbin-Watson test should not be used?

Hibernating
  • 3,723
  • 2
  • 21
  • 34
  • 1
    Section 12.2 Testing for Serial Correlation of Wooldridge's Introductory Econometrics has a good discussion and has an implicit answer to the question: http://www.eco.uc3m.es/~jgonzalo/teaching/EconometriaII/SerialCorrelation.pdf this is close to what I could tick as the accepted answer. Anyone willing to summarize? – Hibernating Feb 08 '14 at 03:24

1 Answers1

3

The Durbin Watson test or dwtest is useful for checking the presence of first-order autocorrelation only. However, your time series might have higher-order autocorrelations as well. In that case Breusch-Godfrey test is used.

A similar argument holds true when you are working on a regression problem that doesn't involve a time series. In that case we can apply the same test to check the correlation in residuals.

Nick Stauner
  • 11,558
  • 5
  • 47
  • 105
NG_21
  • 1,436
  • 3
  • 17
  • 25
  • lmtest::dwtest and car::durbinWatsonTest are particular implementations of the test that has been generalized to apply for lags of any order. For example, in JBES, April 1987, Vol. 5, No. 2 you may find this: "Following Vinod (1973), the Durbin-Watson(DW) test for autocorrelation at lag k is defined by..." – Hibernating Feb 08 '14 at 01:26