[Gretl-users] Durbin-Watson statistic and Durbin's h

Hi all, from the next version of gretl (that is due in a matter of weeks, if not days), you may find a slight change in the output of OLS with lagged depoendent variables. The gretl output for OLS with time-series data includes by default the venerable DW statistic for 1st order autocorrelation, as most packages do. However, it is well known that that statistic is invalid if the regressors include lags of the dependent variable, so in that case we print Durbin's h statistic instead (again, like most packages do).[*] In finite samples, the h statistic may be not computable, because it involves taking the square of a quantity that could be negative. In those cases, we'd just print out the DW statistic instead of the h statistic. On closer inspection, Allin and I thought it's better to print NA instead. The rationale is: printing NA is more informative than supplying a statistic we know to be invalid in context. So, don't complain about that when you see NA in your OLS output :D It's a feature, not a bug! [*] Fron an econometric point of view, I'd say that the best way to test for autocorrelation is using the Breusch-Godfrey LM test, but I digress. ------------------------------------------------------- Riccardo (Jack) Lucchetti Dipartimento di Scienze Economiche e Sociali (DiSES) Università Politecnica delle Marche (formerly known as Università di Ancona) r.lucchetti(a)univpm.it http://www2.econ.univpm.it/servizi/hpp/lucchetti -------------------------------------------------------