Greetings all, With all the modesty at my disposal, please may I make the offering that integer differencing is a very blunt way to make a data-set, stationary. Furthermore any analysis done upon the differenced data, is simply that; namely, an analysis conducted upon data that is not the original data-set. Integer differencing induces corruption to data. Finally, GPH works very very selectively. By this I mean that GPH only gives an agreeable result with data that GPH is amenable to. If you get me. If anyone is interested further, I'll try to clarify further... Meanwhile thankyou for reading so far. (Tanaka is very good reading. Thankyou for pointing me to the paper). Best regards from Richard Hudson Dr RJF Hudson Qld Australia rjfhud(a)powerup.com.au ----- Original Message ----- From: savos schmagges To: Gretl Users Forum Sent: Tuesday, August 10, 2010 9:04 PM Subject: Re: [Gretl-users] LW estimator and the GPH test Reading "Tanaka - The nonstationary fractional unit root" (1999) will help you --> see http://hermes-ir.lib.hit-u.ac.jp/rs/bitstream/10086/13409/1/0100701301.pdf But the power of the LWE and GPH is very low in comparison to Robinson's Lagrange Multiplier Test or Lobato's Efficient Wald Test... don't forget to difference the time series because the tests in gretl are only valid for stationary data. Regards ------------------------------------------------------------------------------ _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users

The Problem is that Robinson's LWE is only valid for stationary data. It depends on the approximation of the spectrum of an ARFIMA-process and only the spectrum of an stationary time series make sense. --> see Robinson's paper (1995, Semiparametric Gaussian Estimation of Long-range Dependance). So you have to force the time series to be stationary - of course integer differencing is an harsh way to achieve it, but it is necessary as long as you don't use Shimotsu's LWE. You only have to add 1 to your estimate for getting the differencing parameter. For the GPH-Estimator it is the same issue... I am always interested in discussing the properties of long-range time series ;-) cheers