Cointegration analysis - DOLS and ECM

[ftc1000]

Hi. I’m using Eviews 7 and have a couple of questions pertaining to my work with cointegrated series. First, is there any sort of Eviews functionality for estimating error correction models, either in a single equation setting or using pooled data? I have noticed in the past that when I try to estimate an ECM by typing the full equation directly into the equation box, it sometimes “blows up”. For example, I’ll get results that look reasonable one time, but the next time the output looks like gibberish (after having changed nothing). Is it better to proceed in two steps, first estimating the long-term relationship in levels and then estimating the error-correction equation using lagged residuals from the first step?

Second, I notice that Eviews 7 has some nice functionality for DOLS estimation in a single equation setting. However, I am doing some work with pooled data. There does not appear to be DOLS functionality in this case, but wanted to verify that is so. How would you suggest I proceed…include the lead/lag terms as 'common coefficients' in the regular pooled equation dialogue? Many thanks.

Steve

[EViews Glenn]

You don't say how you are estimating the ECM but from the sounds of it, you're using the nonlinear ECM specification. Phillips and Loretan (Review of Economic Studies, May 1991) have a discussion of the estimation of this model, noting that it is often quite touchy, with convergence problems in roughly 5-10 percent of their simulations (p. 427-428).

The reason you are getting different results when you estimate the model is that EViews uses the current coefficient values in as starting values for the nonlinear estimation. Nonlinear estimation is often sensitive to starting values, and given the results from Phillips and Loretan, it is not surprising that this model exhibits such sensitivity.

With regard to your question about estimation strategies, Phillips and Loretan is a very good reference on alternative methods.

As to pooled/panel estimation of DOLS, EViews currently does not have built in support for the panel versions of these estimators. Our new support for FMOLS, CCR, and DOLS is provided with an eye toward providing full implementation of the panel versions of these estimators in the near future. In principle, you could perform this estimation using the current tools, but it would require a bit of work to do the estimation for some of the estimators.

One setting in which crafting your own estimator is not particularly different are the between dimension group-mean estimators described by Pedroni (Review of Economics and Statistics, November 2001). These estimators are simply averages of the relevant estimators computed for each cross-section. Consequently computation of these panel estimators should be straightforward given the current tools.

[ftc1000]

Thanks. That's helpful. I was able to pull the article you suggest. I have a couple of follow up questions. First, when you say "nonlinear ECM specification" are you simply referring to entering the full ECM specification in equation form with direct references to the coefficients, with "LS" chosen as the estimation method?

Second, is there any harm done in estimating the ECM in two separate stages (i.e. collecting the residuals from the regression of the long-term relationship and then using them as the independent variable in the ECM stage)? Thanks.

[EViews Glenn]

The two-step method you describe is the standard Engle-Granger approach. There are countless references in the literature for you to consult. See also

http://forums.eviews.com/viewtopic.php?f=18&t=1150

(Since you have the paper)...the nonlinear estimation I am referring to is NLS on equation (36) of Phillips and Loretan. And yes, I am referring to explicit specification of this equation with estimation using EViews LS.