1

Factor in the information from the question "Does causation imply correlation?", "Mathematical Definition of Causality", and "How to formally tell is one time series affects another".

Let $y = f(x)$, where $f$ is continuous and differentiable. Does this guarantee that $x$ will Granger-cause $y$?

Given that $f$ can be horribly nonlinear, it seems that there are a limited number of tests and methods for conditioning the data in such a way that the granger causality test can be applied effectively. Even something like $y=e^x + \epsilon$, where $\epsilon$ is white noise, can cause problems with estimating the order of integration (thus testing for cointegration, and then Granger-causality according to the Toda and Yamamoto method as described in Dave Giles' blog post).

Richard Hardy
  • 54,375
  • 10
  • 95
  • 219
Eric Inclan
  • 33
  • 4
  • Could you clarify the sense in which a function $f$ could be considered as related to Granger causality, which concerns predictability in a time series? – whuber Sep 16 '20 at 16:52
  • 1
    See also my comment under the answer. – Richard Hardy Sep 16 '20 at 17:08
  • @whuber - Take the equation for force, mass, and acceleration F = ma, where mass is constant. I'm under the impression that if F = ma, then F should granger-cause a (because it actually causes a). Is that misguided? – Eric Inclan Sep 16 '20 at 17:16
  • 1
    @whuber - Thinking about this harder, I will post a separate question. – Eric Inclan Sep 16 '20 at 17:26

1 Answers1

1

No.

Consider $x(t) \sim \operatorname{Uniform}(-1,1)$. For any function $f$, including $f(x)=x$ and $f(x)=0$ it is obvious to see that past values of $x$ don't help in predicting $y$.

Firebug
  • 15,262
  • 5
  • 60
  • 127
  • 1
    Granger causality requires past values of $x$ to help in predicting $y$ *once the predictive information contained in past values of $y$ has been accounted for*. Forgetting the latter clause seems to be a common mistake. – Richard Hardy Sep 16 '20 at 17:07
  • @RichardHardy - I'll replace my previous question with this one: doesn't that imply that for test for Granger-Causality, y must be a function of at least 2 variables with (each with their own auto-regressions)? – Eric Inclan Sep 16 '20 at 17:23