VAR: Granger casuality test

[Qi Ng]

When using Granger casuality test, should I use absolute values or first differences of the variables I have?
When I use absolute values I get p-values <0.05, but first differences yield more sensible answers where some are statistically significant while others are not.

I used the procedures from http://blog.stata.com/2016/08/09/vec...ions-in-stata/

[Anat Tchetchik]

Qi, what do you mean by absolute values? do you mean that you have used the variables at levels? I'm not sure but I think that before conducting the Granger casuality test, the data should be stationary (and not unit root).

[Qi Ng]

Qi, what do you mean by absolute values? do you mean that you have used the variables at levels? I'm not sure but I think that before conducting the Granger casuality test, the data should be stationary (and not unit root).

Hi Anat. I am talking about the raw data, such as stock exchange indexes or stock prices. Do I have to transform the data?

[Nick Cox]

Let's please standardise on causality as the standard spelling. (It's a difficult word for many people with English as their first language too, and frequently I read of a yearning for casual inferences.)

Sorry, I can't contribute beyond that.

[Anat Tchetchik]

Nick Cox Sorry, it was a typo.. lucky enough I didn't type casualty .. Qi Ng, I re-checked, your variables should be unit-roots for the Engle-Granger causality test. After verifying that they are unit root processes (using -dfuller- -pperron- or many other of the many unit-root tests available) you can run the command (after running the -var- command.) you do not need to transform the variables but use them at level ("raw" data).

[Qi Ng]

Nick Cox Sorry, it was a typo.. lucky enough I didn't type casualty .. Qi Ng, I re-checked, your variables should be unit-roots for the Engle-Granger causality test. After verifying that they are unit root processes (using -dfuller- -pperron- or many other of the many unit-root tests available) you can run the command (after running the -var- command.) you do not need to transform the variables but use them at level ("raw" data).

I have multiple variables. Do I need to test all of them if there is a casuality between my dependent variable and independent variables?

I tested for unit root (at the untransformed "raw" data) using -dfuller-, and found that the dependent variable has a unit process, i.e. the p-value was greater than 0.05 and test-statistic was larger than at 1%, 5% and 10% critical value. Does this mean that we fail to reject the null hypothesis has a unit root (that it might have a unit root)?

However, when I used the first difference, the result was the opposite. Now the p-value was <0.05, and we reject the null hypothesis that it has a unit root.
And this means that I reject the null hypothesis that the data has a random walk with drift, but cannot reject that the alternative hypothesis that is just a random walk.

What does this mean? Am I not able to use the Var Granger casuality test because the untransformed data had a unit process? What's the difference between using the first difference and untransformed "raw" data for the -dfuller- test?

Thanks for the help, Anat. Much appreciated.

[Anat Tchetchik]

Qi Ng your inference is correct, if the p-value is greater than 0.05 (from what you wrote I understand it is even greater than 0.1) than you cannot reject -dfuller- null hypothesis that the variable contains a unit root, and yes, if you test the first difference and reject the null, it means that the variable is I(1) (to be on the safe side, I would also employ the -kpss- test whose null is the opposite- stationarity). There is no problem with having a drift in your variable they are still unit-root. As far as I understand, all variables in your -vargranger- command should be unit root process since your are fitting var prior to -vargranger- but again I'm not sure. Maybe this doc. will help

http://www.stata.com/manuals13/tsvargranger.pdf

Goodluck