Hi Allin Cottrell, Thank a lot for your suggestion. It works! My data are not "perfect" due to missing values, and some zeros seem to cause some trouble when taking the natural log of the dependent variable. The data are defined as time series data when uploaded into Gretl, but missing values cause trouble when running an ols regression. The HAC seems to be deactivated when the data are cleaned (when the rows of missing values are deleted from the data-file). When these data are deleted, the time interval between the observations is not consistent. However, when redefining the new, cleaned data, and defining them (data) as time series, the HAC-option function well. Maybe the developers of Gretl can look into this problem - or lack of flexibility. R-does handle it, and also Eviews. Kind regards, Tobben

Hi Sven, Thank for your comments. I agree with you that “gappy” time series clearly corrupt a valid evaluation of the model, and the estimates – depending on the amount of data gabs (systematically or non-systematically). But what shall we do when we only have access to incomplete data? My ad-hoc point is that if we detect heterosc/autocorr by using the incomplete time series, it is necessary to correct the st.dev. of the coefficient estimate. I experienced that it was difficult to activate the HAC in Gretl when the data are incomplete, and to what I could see, the HAC-option was deactivated when I used “Drop observations with missing values”. I’m a new user of Gretl. I’m impressed over the job the developers have done, and I look forward using the program 😊 Regards, Torbj. Fra: Sven Schreiber <sven.schreiber(a)fu-berlin.de&gt; Sendt: 8. november 2022 10:48 Til: gretl-users(a)gretlml.univpm.it Emne: [Gretl-users] Re: How to apply the HAC option in Gretl? Am 07.11.2022 um 22:23 schrieb tobbenlorentzen(a)gmail.com <mailto:tobbenlorentzen@gmail.com> : Hi Allin Cottrell, Thank a lot for your suggestion. It works! My data are not "perfect" due to missing values, and some zeros seem to cause some trouble when taking the natural log of the dependent variable. The data are defined as time series data when uploaded into Gretl, but missing values cause trouble when running an ols regression. The HAC seems to be deactivated when the data are cleaned (when the rows of missing values are deleted from the data-file). When these data are deleted, the time interval between the observations is not consistent. However, when redefining the new, cleaned data, and defining them (data) as time series, the HAC-option function well. Maybe the developers of Gretl can look into this problem - or lack of flexibility. R-does handle it, and also Eviews. Hi, don't know about R, but what you're saying about Eviews is only true up to a certain extent. I just checked on Eviews 11 (not current, I know, anybody please speak up if the situation has changed since then); yes, Eviews automatically removes the interior missing values for the OLS estimate, which saves you the hassle of removing them manually. (Note that if you have a bunch of lags, this can lead to a substantial "hole" in your estimation sample, but I guess the user is supposed to be aware of that.) However, if you then let Eviews run diagnostic tests on the estimated residual "gappy" time series, it does carry them out, but tells you: "interior missing value lagged residuals set to zero". This is of course a pragmatic approach (and asymptotically justified), but will bias the results in finite samples. This can also affect the estimates of the HAC standard errors. Maybe this way of handling it is preferable, given that other alternatives will also be imperfect. But there is some arbitrariness involved in the approach of Eviews. cheers sven

Am 09.11.2022 um 13:30 schrieb Riccardo (Jack) Lucchetti: No, not quite. Even if I force the bandwidth to 2 in Eviews (it would automatically choose 4 here), and switch off the d.f. adjustment, the standard errors do not match. (Coefficients do match.) I get 0.133099 and 0.140418 instead. This is still with Eviews 11. cheers sven

Am 10.11.2022 um 18:32 schrieb Cottrell, Allin: Yeah, good point of course. No, they don't. For some newly drawn sample data it's 0.125 vs. 0.114 (Eviews/gretl) and 0.1034 vs. 0.0889. Results are much closer when I let Eviews do "Newey-West fixed bandwidth = 4.0000". Could it be that gretl's parameter terminology uses the square root of what Eviews does? (Bartlett kernel in both cases.) thanks sven

Am 10.11.2022 um 22:07 schrieb Riccardo (Jack) Lucchetti: Ah, I got it now. The Eviews documentation says "the bandwidth is equal to one plus what Stock and Watson term the 'truncation parameter' m". So I set it to 3 in Eviews (given the 2 in gretl), and still no d.f. correction. Then the results are the same for Eviews and gretl, without missing data. With the gappy data I still have  mismatch even under the new bandwidth parameter. I now have with Eviews 0.1274 and 0.1444. (Where Jack's results were 0.1226 and 0.1756.) But I haven't tried to understand what you, Jack, did - did you also try to replace the missing residuals with zeros? cheers sven

Hi all, I think it is not consistent to substitute missing values with zeros, because zeros are time-dated value. Missing values are genuine non-information, and it could be anything (except if the data have some structure). Regards, Torbjørn Fra: Sven Schreiber <sven.schreiber(a)fu-berlin.de&gt; Sendt: 10. november 2022 22:38 Til: gretl-users(a)gretlml.univpm.it Emne: [Gretl-users] Re: How to apply the HAC option in Gretl? Am 10.11.2022 um 22:07 schrieb Riccardo (Jack) Lucchetti: On Thu, 10 Nov 2022, Sven Schreiber wrote: Results are much closer when I let Eviews do "Newey-West fixed bandwidth = 4.0000". Could it be that gretl's parameter terminology uses the square root of what Eviews does? (Bartlett kernel in both cases.) More likely, the two-sided bandwidth (but this is just a wild guess). Ah, I got it now. The Eviews documentation says "the bandwidth is equal to one plus what Stock and Watson term the 'truncation parameter' m". So I set it to 3 in Eviews (given the 2 in gretl), and still no d.f. correction. Then the results are the same for Eviews and gretl, without missing data. With the gappy data I still have mismatch even under the new bandwidth parameter. I now have with Eviews 0.1274 and 0.1444. (Where Jack's results were 0.1226 and 0.1756.) But I haven't tried to understand what you, Jack, did - did you also try to replace the missing residuals with zeros? cheers sven

On Thu, Nov 10, 2022 at 12:03 PM Sven Schreiber <sven.schreiber(a)fu-berlin.de&gt; wrote: I can't do the Eviews comparison, but I've looked into R: lm() for estimation and NeweyWest() from the sandwich package for HAC standard errors. Using the new HAC_ALLOW_MISSING setting, here's what I found: * With no missing values, gretl and R produce HAC results identical to machine precision (given the same HAC parameters). * With missing values, the HAC results differ noticeably, and it turns out that R is ignoring the missing values altogether. I can replicate the R results in gretl if after sprinkling the NAs I do "smpl --no-missing" then reestablish the time-series property with "setobs 1 1 --special-time-series". I wonder if that could be the case with Eviews too. Allin

Am 10.11.2022 um 20:57 schrieb Cottrell, Allin: Just as a general update to all users out there, Allin has worked hard (as usual) to implement this functionality in gretl, and in the current development source code it is there and working. (The original problem was that when the time series contain gaps and thus observations are not fully contiguous, computing Newey-West robust standard errors/covariance estimates is not directly possible, so some more or less arbitrary but justifiable workarounds have to be used.) As always with gretl, this will be duly documented. It is not yet in the current snapshots, I believe. We will probably ask you guys to test the new feature when updated snapshots for Mac and Windows are available. thanks sven

Am 11.01.2023 um 20:24 schrieb tobbenlorentzen(a)gmail.com: I can't take credit for the programming work here; in any case, gretl 2022c is the current version and does not have it yet. But 2023a will. In the meantime, the recent snapshots also have it, so you can try out the feature there. For example, let's take again Jack's little script: <hansl> nulldata 64 set seed 20221109 setobs 1 1 --special-time-series x = uniform() < 0.9 ? normal() : NA y = uniform() < 0.9 ? normal() : NA ols y const x --robust </hansl> Here's the output I'm getting with a snapshot (on Windows) from Jan 7th (sorry for the German, but you get the point): <output> Modell 1: KQ, benutze die Beobachtungen 1-64 (T = 53) Fehlende oder unvollständige Beobachtungen entfernt: 11 Abhängige Variable: y HAC Standardfehler, Bandbreite 2, Bartlett-Kern Observations not contiguous: ES method used              Koeffizient   Std.-fehler   t-Quotient    p-Wert   ------------------------------------------------------------   const      −0.536282      0.114281      −4.693      2.06e-05 ***   x          −0.0521759     0.151110      −0.3453     0.7313 Mittel abhängige Var.  −0.529945   Stdabw. abhängige Var. 1.023859 Summe quad. Residuen    54.38531   Stdfehler Regression 1.032656 R-Quadrat               0.002306   Korrigiertes R-Quadrat -0.017257 F(1, 51)                0.119221   P-Wert(F) 0.731301 Log-Likelihood         −75.88750   Akaike-Kriterium 155.7750 Schwarz-Kriterium       159.7156   Hannan-Quinn-Kriterium 157.2904 </output> But this whole thing also works in the GUI. cheers sven