Sure, there is some difference between Gretl and Stata's user-written ado. I think that the problem is how time-dummies are used as instruments or so. But my point is the following: you have 3 chi2(100) tests at hand, which are well-above 100, and therefore it rejects strongly the null. I don't see how you could have a p-value of 20%!! 2013/5/20 Pindar <pindar777(a)gmail.com&gt; > Am 20.05.2013 11:56, schrieb Sven Schreiber: >> Am 20.05.2013 11:52, schrieb Pindar: >> >>> Hola Rodrigo, >>> >>> The p-value for Hansen test is reported as " 0.218". >>> But with the output in the paper and gretl there are 3 different test >>> statistics for chi2(100): >>> >>> Sargan_xtabond2: 186.90 >>> Sargan_gretl: 154.81 >>> Hansen_xtabond2: 110.70 >>> >>> I would like to be sure how to interpret differences in the diagnostic >>> checks between gretl and stata. >>> >> Yes, a useful question I think. But are the coeff estimates always the >> same, are you absolutely sure you are comparing identical >> specificiations? In panel settings and GMM settings there can be subtle >> differences. > > Thanks Sven for pointing me to the 'always': The coefficients for the const and the time dummies differ! > > Trying to change the setting for the time dummies leads to 'completely different' coefficients while > it does not alter the Sargan test statistic. I obviously failed in replicating the time dummy instruments: > > <hansl> > open abdata.gdt > genr time > genr timedum > list TD_roodman = dt_2 dt_3 dt_4 dt_5 dt_6 dt_7 dt_8 > > dpanel 1; n const w w(-1) k k(-1) TD_roodman ; \ > GMM(n,2,8) GMM(w,2,8) GMM(k,2,8) \ > GMMlevel(w,1,1) GMMlevel(k,1,1) TD_roodman --sys > > # This estimation gives a 'Sargan test' > # Sargan over-identification test: Chi-square(100) = 154.367 [0.0004] > <hansl> > > Best > Leon > > Here the Roodman stata output of coefficients: > <djfegaef.png> > > >> cheers, >> sven >> _______________________________________________ >> Gretl-users mailing list >> Gretl-users(a)lists.wfu.edu >> http://lists.wfu.edu/mailman/listinfo/gretl-users > > > > > > > _______________________________________________ > Gretl-users mailing list > Gretl-users(a)lists.wfu.edu > http://lists.wfu.edu/mailman/listinfo/gretl-users _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users

On Wed, 22 May 2013, Pindar wrote: > Ok, I see, the 20% cannot be the right p- value. Does someone use Stata to check what the newest xtabond2 says? > > U seem to have some experience with dpanel estimation. What do u think about that: If I have a 12x412 panel is it then ok to use 225 instruments or should one stay near 100 instruments? If you ask me, 225 orthogonality conditions is a sure recipe for disaster, unless you have 10 billion observations or more. The conventional wisdom among practitioners is that should never use all available instruments when doing panel GMM, unless your T is very small. You may want to read Roodman (2009), Oxford Bulletin. ------------------------------------------------------- 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 ------------------------------------------------------- _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users