I ran a ADF test in gretl and in R. (The regression was no constant, LNEG lagged once, d_LNEG lag1 and d_LNEG lag2) In gretl and R, the results were for most purposes identical for all numbers for Z and d_LNEG lag 1 In gretl the results for d_LNEG lag2 were coefficient std. error t-ratio p-value d_LNEG_2 -0.470623 0.126295 -3.726 0.2006 But in R the results were z.diff.lag2 -0.47062 0.12629 -3.726 0.00053 Which is correct?

I recomend use Eviews, stata or other program for this, or use tables of critical values of MacKinnon, see Wooldridge "Introductory Econometrics. Greetings Hebert Suarez Peru On Sun, 30 Nov 2008 06:59:46 -0500, Sven Schreiber <svetosch(a)gmx.net&gt; wrote: > Am 30.11.2008 12:27, Charles Ward schrieb: >> I ran a ADF test in gretl and in R. >> >> (The regression was no constant, LNEG lagged once, d_LNEG lag1 and >> d_LNEG lag2) >> In gretl and R, the results were for most purposes identical for all >> numbers for Z and d_LNEG lag 1 >> >> In gretl the results for d_LNEG lag2 were >> coefficient std. error t-ratio p-value >> d_LNEG_2 -0.470623 0.126295 -3.726 0.2006 >> But in R the results were >> z.diff.lag2 -0.47062 0.12629 -3.726 0.00053 >> >> Which is correct? >> > > If I understand correctly, you implemented the ADF test manually by > running the appropriate OLS regression? > > Then, unless I'm missing something, the distribution for the lagged > differences would be standard and a t-ratio of -3.7 most definitely > wouldn't give you a p-value of 0.20. So gretl would be wrong here, but I > don't know the reason. > > good luck, > sven > _______________________________________________ > Gretl-users mailing list > Gretl-users(a)lists.wfu.edu > http://lists.wfu.edu/mailman/listinfo/gretl-users > >

On Thu, 4 Dec 2008, Sven Schreiber wrote: > I just noticed that in the ADF test specification and output > it's not 100% obvious (at least to me) whether the lag order > refers to right-hand levels or differences. I would suggest that > at least in the output this should be clarified. (It could be in > the manual, I haven't checked right now, but in any case it > should be self-explanatory IMHO.) It's explicit in section 20.3 of the manual. The wording in the command help refers to "lags of the dependent variable", and this phrase appears under an equation which shows (1 - L)y_t as the dependent variable in the test regression, which in turn is prefaced by the statement, "In all cases the dependent variable is the first difference of the specified variable, y". So I'd say it's reasonably clear -- given, also, that surely that's what everyone would expect?

The basic DF test involves a regression of \delta y_t on y_{t-1}. The ADF test then adds k lags of \delta y_t to this regression to remove serial correlation from the disturbance. If one considers the DF test as a regression of y_t on y_{t-1} then one would need to also add the same number of additional lags to the equation to remove the serial correlation. Thus in a sense the number of additional lags required is the same regardless of which approach one takes. I would tend to use the first always. Best Regards John 2008/12/4 Sven Schreiber <svetosch(a)gmx.net&gt;: > Am 04.12.2008 21:28, Allin Cottrell schrieb: >> On Thu, 4 Dec 2008, Sven Schreiber wrote: >> >>> I just noticed that in the ADF test specification and output >>> it's not 100% obvious (at least to me) whether the lag order >>> refers to right-hand levels or differences. I would suggest that >>> at least in the output this should be clarified. (It could be in >>> the manual, I haven't checked right now, but in any case it >>> should be self-explanatory IMHO.) >> It's explicit in section 20.3 of the manual. The wording in the >> command help refers to "lags of the dependent variable", and this >> phrase appears under an equation which shows (1 - L)y_t as the >> dependent variable in the test regression, which in turn is >> prefaced by the statement, "In all cases the dependent variable is >> the first difference of the specified variable, y". >> >> So I'd say it's reasonably clear -- given, also, that surely >> that's what everyone would expect? >> > It's absolutely clear in the help and manual, no doubt. That lag orders > always refer to the dependent variable in unit-root contexts however is > _not_ a universal convention. Quite often the description starts with > levels and only after assuming unit roots differences are taken. In that > case the lag order of the differences are given as k-1 for example > (where k is the lag order of levels). > > It's no big deal; if I really need to know I would look it up in the > help and I would find it. However, there is already a model description > line in the standard ADF test output exactly so that people don't always > have to read the manual. All I'm saying is that it might be useful to > change that line so that the lag order thing is clarified, too. > > thanks, > sven > _______________________________________________ > Gretl-users mailing list > Gretl-users(a)lists.wfu.edu > http://lists.wfu.edu/mailman/listinfo/gretl-users >

On Thu, 4 Dec 2008, Sven Schreiber wrote: > Please guys, there are different conventions out there about > whether lag order refers to levels or differences in the > unit-root/cointegration context. Believe me, I know what I'm > talking about here... I believe you! Sorry to make a meal out of this, but is this the question that needs clarification, then: When gretl says the lag order is, say, 3, does that mean the ADF regression contains 3 or 2 lags of \tilde y, where \tilde y is the dependent var in the DF regression?

On Thu, 4 Dec 2008, Sven Schreiber wrote: > Ok, let's get explicit then... Thanks, Sven, I think I finally get it. We advertise an "ADF test with lag order p". Do we mean: A. A test regression that includes p lags of the dependent variable in that regression (namely, the first difference of the variable to be tested), OR B. A test appropriate for the alternative hypothesis that the variable in question follows an AR(p) process in its level rather than a random walk (so that the test regression should contain p-1 lags of the regressand) ? If I'm reading you right, you know that for gretl the answer is A (as indeed it is) but you doubt whether this is stated clearly enough.

On Thu, 4 Dec 2008, Allin Cottrell wrote: > On Thu, 4 Dec 2008, Sven Schreiber wrote: > >> Ok, let's get explicit then... > > Thanks, Sven, I think I finally get it. We advertise an "ADF test > with lag order p". Do we mean: > > A. A test regression that includes p lags of the dependent > variable in that regression (namely, the first difference of the > variable to be tested), OR > > B. A test appropriate for the alternative hypothesis that the > variable in question follows an AR(p) process in its level rather > than a random walk (so that the test regression should contain p-1 > lags of the regressand) ? > > If I'm reading you right, you know that for gretl the answer is A > (as indeed it is) but you doubt whether this is stated clearly > enough. We free software buffs do have a talent for creating looong mailing list threads out of the simplest things. Since it looks like this thread is dying out, I'll put my 2 cents in to revive it: shouldn't we be consistent with the Johansen tests (\lambda-max and trace)? Those are just the multivariate generalisation of ADF, and we use B there. I know, I should have stayed quiet :-)

I recomend use Eviews, stata or other program for this, or use tables of critical values of MacKinnon, see Wooldridge "Introductory Econometrics.

On Sun, 30 Nov 2008, Charles Ward wrote: > In gretl the results for d_LNEG lag2 were > coefficient std. error t-ratio p-value > d_LNEG_2 -0.470623 0.126295 -3.726 0.2006 > But in R the results were > z.diff.lag2 -0.47062 0.12629 -3.726 0.00053 > > Which is correct? Thank you for spotting this. The p-value given by gretl for the second lag of LNEG was incorrect...

Am 30.11.2008 19:48, Allin Cottrell schrieb: >> In gretl the results for d_LNEG lag2 were >> coefficient std. error t-ratio p-value >> d_LNEG_2 -0.470623 0.126295 -3.726 0.2006 >> But in R the results were >> z.diff.lag2 -0.47062 0.12629 -3.726 0.00053 > > Thank you for spotting this. The p-value given by gretl for the > second lag of LNEG was incorrect: in fact it was the Dickey-Fuller > p-value for the first lag of LNEG, printed in the wrong place. > But for the case of ADF without constant IIRC the p-value for the ADF-t=-3.7 would be much lower than 0.20, no?

From the output above, we don't know what the adf tau value was --

On Sun, 30 Nov 2008, Sven Schreiber wrote: > Am 30.11.2008 19:48, Allin Cottrell schrieb: > >> In gretl the results for d_LNEG lag2 were > >> coefficient std. error t-ratio p-value > >> d_LNEG_2 -0.470623 0.126295 -3.726 0.2006 > >> But in R the results were > >> z.diff.lag2 -0.47062 0.12629 -3.726 0.00053 > > > > Thank you for spotting this. The p-value given by gretl for the > > second lag of LNEG was incorrect: in fact it was the Dickey-Fuller > > p-value for the first lag of LNEG, printed in the wrong place. > > > > But for the case of ADF without constant IIRC the p-value for the > ADF-t=-3.7 would be much lower than 0.20, no? >From the output above, we don't know what the adf tau value was -- all we have is the t-stat for d_LNEG_2. That is, it was only the p-value that was misplaced in the printout, the other components on the d_LNEG_2 line actually belonged there (and agreed with R). Allin. _______________________________________________ Gretl-users mailing list Gretl-users(a)lists.wfu.edu http://lists.wfu.edu/mailman/listinfo/gretl-users