I'm estimating a simultaneous equations model using TSLS (or 3SLS). However, my residuals are clearly serially correlated, so I would like to run it in AR(1) specification. It's not apparent to me how to do that in gretl. In EViews, one simply adds the regression term 'AR(1)' and it does some .magic. I believe it just adds lags of the endogenous RHS and LHS variables to the instruments list, but I'm unable to recreate eviews results with gretl using the "lags" options in the TSLS dialogue.

On 07/20/2012 08:22 AM, Riccardo (Jack) Lucchetti wrote: > On Thu, 19 Jul 2012, Trevor Zink wrote: > >> I'm estimating a simultaneous equations model using TSLS (or 3SLS). >> However, >> my residuals are clearly serially correlated, so I would like to run >> it in >> AR(1) specification. It's not apparent to me how to do that in gretl. >> >> >> >> In EViews, one simply adds the regression term 'AR(1)' and it does some >> .magic. I believe it just adds lags of the endogenous RHS and LHS >> variables >> to the instruments list, but I'm unable to recreate eviews results with >> gretl using the "lags" options in the TSLS dialogue. I don't think gretl has a super-quick way of imitating this exact specification. If that's really what you need exactly (and chances are it is not, see below), you (roughly, and no guarantees) would need to add the first lag of all the variables to the list of explanatory variables (and instruments), and restrict the coefficients of these lags to have the same proportion to their non-lagged coefficients. See the 'restrict' command, but I'm actually not sure right now if this type of restriction is supported, especially in a system context. Just adding the lags without restricting them is fine, too, but of course it adds more parameters to estimate and is not equivalent. > > > I don't know what Eviews does. It'd be interesting to know: however, a > perfectly sane strategy (which may well be what Eviews in fact does) > would be including one or more lags of the LHS variable among the > exogenous variables of your model, that is both as explanatory variables > and instruments. > > I agree with Jack, this is probably enough to deal with the autocorrelation and also computationally more robust. I'm pretty sure Eviews actually does what it says, it models the errors as an AR(1) process: (1-\rho L) u_t = e_t, and so multiplies through the entire equation with (1-\rho L), not just the LHS. In the system case I'm not sure if it uses/imposes the same \rho for all equations. BTW, I have to say that the flexibility of EViews in the simultaneous equation area is impressive. For a recently published paper we needed non-linear 3SLS and Eviews had it. Granted, we discovered a pretty obvious bug there (had nothing to do with the numerical maximization), which seems to indicate that nobody is using that stuff, otherwise the bug probably would have been discovered before. (The bug was quickly fixed by their staff.) So that doesn't mean gretl should invest much effort there, but there are cases where even Eviews comes in handy. cheers, sven

On Fri, 20 Jul 2012, Rodrigo Alfaro Arancibia wrote: > Please note from EViews' manual: "EViews estimates the model as a > nonlinear regression model (...), require specification of additional > instruments to satisfy the instrument order condition for the transformed > specification. By default, the first-stage instruments employed in TSLS > are > formed as if one were running Cochrane-Orcutt using Fair’s prescription." > > Therefore you cannot replicate the results using 2SLS in gretl or > other software given that 2SLS are linear models. The fact that EViews > does a nonlinear regression is rooted in the Cochrane-Orcutt procedure > which, in the standard LS context, imposes constraints on the > coefficients. You _can_ replicate them in gretl: actually, from what you describe you have the choice of using a) the nls command b) the mle command c) the gmm command Only, it's a lot more work. But if you had the absolute necessity of replicating Eviews' algorithm, it's perfectly possible.

On 07/20/2012 04:32 PM, Rodrigo Alfaro Arancibia wrote: > 2012/7/20 Riccardo (Jack) Lucchetti <r.lucchetti(a)univpm.it&gt;: >>> Therefore you cannot replicate the results using 2SLS in gretl or >>> other software given that 2SLS are linear models. The fact that EViews >>> does a nonlinear regression is rooted in the Cochrane-Orcutt procedure >>> which, in the standard LS context, imposes constraints on the >>> coefficients. >> >> >> You _can_ replicate them in gretl: actually, from what you describe you have >> the choice of using >> >> a) the nls command >> b) the mle command >> c) the gmm command >> >> Only, it's a lot more work. But if you had the absolute necessity of >> replicating Eviews' algorithm, it's perfectly possible. > > Ok, you can do it, but it requires more than using the tsls command. > Jack tends to get nervous when he reads "cannot be done in gretl" so he also tends to skip over the fine print (which in this case was "... using 2SLS...") :-)