Granger causality

General idea:

The cause contains information about the futute of the effect which is not in any other variable (including the past of the effect)

  • the cause happens before to the effect

  • the cause has unique information about the future of the effect

x Granger-causes y if

p(yt+1|Zt)p(yt+1|Ztxt)

where Zt is all information available at t and Ztxt is all information except that contributed by the past and present of x ({xt}t)

x doesn’t Granger-causes y if

p(yt+1|Zt)=p(yt+1|Ztxt)

no information is lost by excluding x

Note:

It is possible that both x Granger-causes y and y Granger-causes x

Theory vs. practice

  • in theory, Zt contains the infinite past of all variables in the universe

  • in practive Zt includes only a few variables, often only y and x, and only a few lags

Testing for (pairwise unconditional) Granger non-causality

yt=α0+i=1pαiyti+j=1qβjxtj+εt

if βj=0 for all j, x doesn’t Granger-cause y.

Testing for (pairwise conditional) Granger non-causality

yt=α0+i=1pαiyti+l=1kγlztj+j=1qβjxtj+εt

if βj=0 for all j, x doesn’t Granger-cause y, conditional on z.

  • interpretation: given the information in the past of y and z, the past of x doesn’t help forecast y

  • intuition for the test: we are asking whether the variance of the forecast errors is significanty smaller when lags of x are included in the predictive regression for y.

Granger causality as transfer of information

  • latent variables (more general than future values of observed variables)

  • how informative (more general than informative/uninformative)

  • frequency-domain formulation (informative frequencies)