autocorrelation

[majalyka]

How do you cure autocorrelation? Thanks in advance.

[Zeno]

I used Cochrane-Orcutt or Prais-Winsten Regression for that. However, you have to do the two steps (or more if you estimate iteratively) by hand. Interestingly, the EViews Users Guide II points to a drawback in both methods in the presence of lagged dependent variables and suggests estimating rho in a nonlinear way (see equation 26.10 in the Users Guide II on page 70) instead. However, it is not clear to me how to use this method and an example is unfortunately not provided. If someone has done this before it would be great to share his/her experience.

Zeno

[startz]

EViews has had a standard "fix" for autocorrelation for a couple of decades now. Look at "autoregressive error" in the help file.

[Zeno]

Dear Startz,

this is an interesting point. As you said, one cold use the Newey-West (1987) standard error implemented in EViews. However the OLS estimator is inefficient and FGLS methods such as Cochrane-Orcutt or Prais-Winston are efficient under the strict exogeneity assumption. Furthermore, autocorrelation robust standard errors can be poorly behaved in small samples. It seems to be the general tendency to use autocorrelation robust standard errors instead of FGLS, but it is not clear a propri which one is better in practice.

Zeno

[hum2004]

While all of the above mentioned arguments do have their merits, one has to keep in mind, what is the origin of AR in resids. If it is the consequence of omitted variables all it migth be better to think what other variables might be includes.
Another "cure" is to include more lags (what might come close to include another (the omitted) variable

[thunguyen2210]

I having an Eco project and autocorrelation with my model. can you show me how to use Cochrane-Orcutt as a remedy?
I am an amateur in Eview and an non-native English so pls be detail.
I really need your help. It's urgent
Tks in advance

[startz]

Loosely speaking, include "AR(1)" in the list of regressors. See the help file for more detail.

[gobbble]

Is it also appropriate to include the AR(1) term, if just few, say 2 out of 4 indeoendent variables suffer from autocorrelation?

And also, if individual variables in isolation show autocorrelation, but the overall Durbin Watson statistic of the multivariate regression indicates no autocorrelation, is the regression blue regrading the autocorrelation assumption, so that I do not have to care about autocorrelation at individual level?

I appreciate your help

gobbble

[startz]

Autocorrelation in a regression refers to a property of the error term. Whether the independent variables are autocorrelated is largely irrelevant.