Tuesday, May 12, 2015

Alternative Tests for Serial Independence

The following question arose in a (fairly) recent email from Daumantas:
"I wonder if you could give any references -- or perhaps make a new blog post -- about testing for serial correlation: Breusch-Godfrey versus Ljung-Box test. I have no problem finding material on the two tests (separately), but I am interested in a comparison of the two. Under what conditions should one test be favoured over the other? What pitfalls should one be aware of before choosing one or the other test? Or perhaps both of them should be put to rest in favour of some new, more general, more robust or more powerful test?"
Daumantas apparently raised the same question on stackexchange, and got some sensible responses.

If this interests you, the response there that refers to Chapter 2 of Fumio Hayashi's, Econometrics, is right on target. There's no point in me repeating it here.

Rob Hyndman also had an interesting and useful post about the L-B test.

My recommendation - stick with the Breusch-Godfrey test if you're testing regression residuals.

© 2015, David E. Giles