Beginner - Different multiple regression problems

[90frmat]

Hello, i'm new at eviews and i've encountered a few problems.

I'm doing a multiple regression and the first question i have is regarding when to log your variables. My Y is the gold price and the variables are gdp, cpi and the interest rate.

Another problem is that the Durbin-Watson test is really low, like 0,02 or something. That suggests i have autocorrelation and i wonder what i can do about it?

I'm also doing a unit root test, the dickey-fuller test. I'm wondering if i should test it in "level" or "1st difference". If i choose level all of em are non-stationary and with the 1st difference they are fine. If the are non-stationary, how do i make it stationary?

Hope you can answer these questions, and as i mentioned i'm new with eviews so if you could explain it as simple as possible that would be very much appreciated. tell me if i should copy in some numbers or anything to make it easier for you to answer my questions.

thanks in advance

[NicV]

Regarding logs: in its basic form it depends on what kind of (interpretation of) output you want.
Using logs both sides will give you elasticities, which may be more realistic than levels both sides as log-log would e.g. say that if GDP grows with 1%, the gold price increases with 0.5% per ounce as opposed to level-level which would say e.g. every €1 billion in GDP increases the gold price with €5 per ounce, be it from a GDP of 1 billion to 2 billion or from a GDP of 500 to 501 billion.
In the case of CPI, interest rates and such you could take a log on the left while levels on the right, such that a percentage point increase in inflation has an x% effect on the gold price.

About the autocorrelation: I thought one way to control for this is using AR terms in your equation, e.g. AR(1) if you think there is a one period serial correlation (you can put in more AR terms, AR(2), AR(3) etc. to control/check for serial correlation going back more than one period). Note though, that Richard Startz expressly mentions in his EViews illustrated manual that the DW-statistic is not reliable if you include lagged terms of the dependent variable. I'm not sure if it could not still serve as a very rough indicator when it's very off, despite having a lagged dependent variable term(?)

Regarding unit roots: a way to fix unit roots (i.e. making non-stationary data stationary) is to difference them. Your unit root test thus showed you that your data is first-difference stationary.

Most introductory econometrics textbooks cover these topics.

Please, someone correct me if I've made mistakes!

[econometrics]

Hello, i'm new at eviews and i've encountered a few problems.

I'm doing a multiple regression and the first question i have is regarding when to log your variables. My Y is the gold price and the variables are gdp, cpi and the interest rate.

Another problem is that the Durbin-Watson test is really low, like 0,02 or something. That suggests i have autocorrelation and i wonder what i can do about it?

I'm also doing a unit root test, the dickey-fuller test. I'm wondering if i should test it in "level" or "1st difference". If i choose level all of em are non-stationary and with the 1st difference they are fine. If the are non-stationary, how do i make it stationary?

Hope you can answer these questions, and as i mentioned i'm new with eviews so if you could explain it as simple as possible that would be very much appreciated. tell me if i should copy in some numbers or anything to make it easier for you to answer my questions.

thanks in advance

If there is autocorrelation in ur data... write the equation in OLS and add "AR(1)" if not solved then "AR(2)"..... and ur autocorrelation will be solved. to confirm it go (view/residual diagnostics/correlation LM stat) of OLs and check that Pro chisq is bigger than LM stat. meaning no autocorrelation.
When start applying the test for stationarity and it is not stationary in first attempt, follow steps one by one;;; 1. (level + trend and intercept) 2. (level + intercept) 3. (level + none) still not stationary? go for 1.(1st difference + trend and intercept) and so on ........ once u find all data is stationary, make a note that it is all at same level be it "i(0)" or "i(1) or "i(2)". if zero then no need for more test. If all are integrated order 1 meaning "i(1)" then apply (VAR, Johansen cointegration and VECM) tests.
Good Luck

[startz]

The response about linking the AR command to the frequency of the data is wrong. The Durbin-Watson is a test for first order serial correlation. So you should always start the AR(1). If that's not enough then you add both AR(1) and AR(2). Etc. You very rarely leave out an AR order in the list.