Granger Causality Testing

[entityman]

Hi All,

I'm afraid I'm a bit of an amateur in this game and I would really appreciate some help with them following problem:

I am trying to test for Granger Causality, as I want to (hopefully) demonstrate that one of my independents (P) causes change in the dependent (G) and not vice versa. I have input my data into eViews and clicked on the Granger Causality Test Button under Group Statistics. When I set time lags to just 1, I obtain the following results:

Pairwise Granger Causality Tests
Date: 03/06/09 Time: 16:14
Sample: 1982 2007
Lags: 1

Null Hypothesis: Obs F-Statistic Prob.

P does not Granger Cause G 24 11.2726 0.0030
G does not Granger Cause P 1.76457 0.1983

LP does not Granger Cause G 24 0.83705 0.3706
G does not Granger Cause LP 0.63125 0.4358

LE does not Granger Cause G 24 0.58287 0.4537
G does not Granger Cause LE 0.77417 0.3889

LP does not Granger Cause P 24 3.54079 0.0738
P does not Granger Cause LP 1.49184 0.2355

LE does not Granger Cause P 24 1.65853 0.2118
P does not Granger Cause LE 0.51274 0.4818

LE does not Granger Cause LP 25 1.90781 0.1811
LP does not Granger Cause LE 12.4476 0.0019

I hope this implies that P causes G but G does not cause P. Is this right and is it simply a case of conducting an F test to establish the null hypotheses?

Similarly, could anyone explain why when I run the same test with 4 lags the results are very different:
Pairwise Granger Causality Tests
Date: 03/06/09 Time: 16:17
Sample: 1982 2007
Lags: 4

Null Hypothesis: Obs F-Statistic Prob.

P does not Granger Cause G 21 0.88652 0.5010
G does not Granger Cause P 0.61322 0.6612

LP does not Granger Cause G 21 6.01657 0.0068
G does not Granger Cause LP 0.24897 0.9048

LE does not Granger Cause G 21 2.73040 0.0794
G does not Granger Cause LE 0.63034 0.6502

LP does not Granger Cause P 21 0.85991 0.5151
P does not Granger Cause LP 0.40672 0.8004

LE does not Granger Cause P 21 0.71685 0.5964
P does not Granger Cause LE 1.89032 0.1769

LE does not Granger Cause LP 22 0.42876 0.7854
LP does not Granger Cause LE 3.02119 0.0576

Which, owing to the lack of discrepancy between the aforementioned values suggests to me that there is little indication of either P causing G or G P for this lag number.

Alternatively, does anyone know of some easy-to-follow literature around this subject (I find the eViews help file is somewhat limited in its explanations).

Many Thanks,
Max

[QSnakecharmer]

Yep, the interpretation is good, but the results depend on the number of lags you use in the test.

Remember that the fact that X does not granger-cause Y doesn’t necessarily imply that Y is independent of X, granger causality only refers to the capacity of X to forecast Y, if your reject granger-causality tests, it just means that lead-lags of X could not be used to properly forecast Y.

A good and easy to follow book about it is “Time Series Models for Business and Economic Forecasting” of Philip Hans Franses.

[myersmurumu]

Hello, am also stuck with that problem. What are the optimal lags for testing granger causality? because when they change, the results are different too. Some help

[sntgo11]

The lag length for the Granger causality test should be the same used to estimate the underlying VAR or VEC model

[aadyas]

Regarding lag selection, one will have to use either AIC or SIC values. Calculate AIC or SIC values using VAR for upto 8 lags (depending on the number of observations). Lower the AIC or SIC, better the model.