Cochrane Orcutt

[tutonic]

Hi guys.

I'm planning on running CO on my time series data to correct for AR(1) and then perform feasible GLS using the estimated rho from the CO iterative procedure.

I've read that you can just regress y on x and include ad AR(1) term in EViews and it will produce results similar to if you manually do CO. Can anyone shed some light on this?

The output that I get when I run "ls y c x ar(1)" is titled ARMA Maximum Likelihood (BFGS) under Method even though I used least squares command. So my question is as follows: can I use this output and interpret coefficients like how I'd interpret the coefficients from the feasible GLS?

[EViews Gareth]

On the Options page of the Estimation dialog, change the ARMA method to CLS

[tutonic]

On the Options page of the Estimation dialog, change the ARMA method to CLS

So if I change the estimation from ARMA to CLS, I can interpret the model like how I'd interpret my feasible GLS model obtained from getting estimated rho via CO procedure?

[startz]

Yes, although EViews does iterated Cochrane-Orcutt.

[tutonic]

Yes, although EViews does iterated Cochrane-Orcutt.

By Iterated CO, you mean the part where it says Convergence achieved after XX iterations, right?

I've changed my estimation method to CLS. One thing worries me. My R squared is extremely high (0.997). Is that normal? Also, seeing as to how I've already corrected for the AR(1) via CO and Feasible GLS, does it make sense to use HAC robust standard errors?

Also, my DW statistic in the ARMA CLS output still points to the existence of AR(1) in the model, since my DW stat is 1.15. Is this normal?

If my AR(1) variable has a coefficient marginally >1, does this suggest that my process is mildly explosive?

It's quite a relief that I can just interpret the ARMA CLS output the same way I would interpret the FGLS output. So, to clarify, the coefficient of my regressors still measure marginal effects of that particular variable on Y while holding all else constant, right?

I'm assuming that the AR(1) variable doesn't require any interpretation with regards to coefficient?

Sorry for the whole host of questions. My prof seems to be taking forever to reply me.

[startz]

By Iterated CO, you mean the part where it says Convergence achieved after XX iterations, right?

right

I've changed my estimation method to CLS. One thing worries me. My R squared is extremely high (0.997). Is that normal? Also, seeing as to how I've already corrected for the AR(1) via CO and Feasible GLS, does it make sense to use HAC robust standard errors?

The high R-square is in part because of the explanatory power of the AR(1) termt

Also, my DW statistic in the ARMA CLS output still points to the existence of AR(1) in the model, since my DW stat is 1.15. Is this normal?

This may suggest second order serial correlation

If my AR(1) variable has a coefficient marginally >1, does this suggest that my process is mildly explosive?

unfortunately, yes

It's quite a relief that I can just interpret the ARMA CLS output the same way I would interpret the FGLS output. So, to clarify, the coefficient of my regressors still measure marginal effects of that particular variable on Y while holding all else constant, right?

I'm assuming that the AR(1) variable doesn't require any interpretation with regards to coefficient?

right

Sorry for the whole host of questions. My prof seems to be taking forever to reply me.

[EViews Glenn]

It's actually not iterated Corchrane-Orcutt. It's straight non-linear least squares. In general, this doesn't matter, but sometimes does for instrumental variables. I believe we have a discussion of this in the manual.

[tutonic]

By Iterated CO, you mean the part where it says Convergence achieved after XX iterations, right?

right

I've changed my estimation method to CLS. One thing worries me. My R squared is extremely high (0.997). Is that normal? Also, seeing as to how I've already corrected for the AR(1) via CO and Feasible GLS, does it make sense to use HAC robust standard errors?

The high R-square is in part because of the explanatory power of the AR(1) termt

Also, my DW statistic in the ARMA CLS output still points to the existence of AR(1) in the model, since my DW stat is 1.15. Is this normal?

This may suggest second order serial correlation

If my AR(1) variable has a coefficient marginally >1, does this suggest that my process is mildly explosive?

unfortunately, yes

It's quite a relief that I can just interpret the ARMA CLS output the same way I would interpret the FGLS output. So, to clarify, the coefficient of my regressors still measure marginal effects of that particular variable on Y while holding all else constant, right?

I'm assuming that the AR(1) variable doesn't require any interpretation with regards to coefficient?

right

Sorry for the whole host of questions. My prof seems to be taking forever to reply me.

That's great. The mildly explosive process is actually very useful in my analysis.

So, should I then correct the standard errors using HAC as a safety measure or is it already redundant now since FGLS theoretically provides BLUE estimates?

If there is a possibility of second order correlation, can I just run "ls y c x AR(1) AR(2)" to fix this? I tweaked with my estimation and added in the AR(2) into my model. DW stat is now 2.53 so no issues there but now my coefficients are all very insignificant, when economic intuition says they ought to be. There shouldn't be a problem with the data since it's coming from a reputable source.

Or is there another way to correct for second order autocorrelation that I'm not aware of?

You're such a great source of help startz. Thanks!

It's actually not iterated Corchrane-Orcutt. It's straight non-linear least squares. In general, this doesn't matter, but sometimes does for instrumental variables. I believe we have a discussion of this in the manual.

I'll take a look at it. Thanks.

[startz]

Unfortunately, having an AR(1) coefficient above 1 means that many things in the regression don't work as it implies the errors do not have a finite variance.

Adding an AR(2) is the right thing to do. The fact that the results aren't as expected doesn't make it wrong.

My guess is that your dependent variable is nonstationary, which may be accounting for a variety of problems.

[tutonic]

Unfortunately, having an AR(1) coefficient above 1 means that many things in the regression don't work as it implies the errors do not have a finite variance.

Adding an AR(2) is the right thing to do. The fact that the results aren't as expected doesn't make it wrong.

My guess is that your dependent variable is nonstationary, which may be accounting for a variety of problems.

It seems like I'm opening up a can of worms.

My estimated AR process was nonstationary when I only included the AR(1) in my previous specification. With the AR(2) term added in, there is no prompt for nonstationarity right now in my EViews output but the terms are still statistically insignificant.

I did the ADF test for all my variables and, when testing for unit root in 1st difference, only 1 of my X variable fails to reject the null, aka is non-stationary and my Y variable is also non-stationary.

However, the X and Y variable in question is stationary when testing unit root in 2nd differences.

Is there any way for me to proceed? Should I re-specify my problem variables as [x-x(-2)] & [y-y(2)]?