Why is not the Dickey-Fuller test applied directly on :

$Y_t = \rho Y_{t-1} + u_t$

instead of :

$\Delta Y_t = (\rho-1) Y_{t-1} + u_t$.

Many papers apply the Dicker-Fuller on the first difference operator and later use Augmented Dickey Fuller also with difference operators but not explaining why not dealing directly with $Y_t$.


You are right that this is pure convention.

The only practical reason (I am aware of) is that the unit root null hypothesis $H_0:\rho=1$ is evidently equivalent to $H_0:\rho-1=0$. Hence, the default t-statistic produced by standard software packages directly yields the Dickey-Fuller statistic.

Recall, however, that the critical values and p-values supplied by standard packages are typically taken from the standard normal or t-distribution, which are not the right ones for the Dickey-Fuller test.

  • 1
    $\begingroup$ Thanks. I'm missing something however. When applied on an Autoregressive model of order 1, $\Delta S_t = \alpha + \beta t + (\phi -1) S_{t-1} + e_t$ we have the following hypothesis : Null Hypothesis: A unit root is present in the autoregressive model (of order 1), the process is non stationary. Alternative Hypothesis: The model is stationary. Here what is stationary ? $\Delta S_t$ or $S_t$ or both ? $\endgroup$ – Bastiat Oct 16 '18 at 11:49
  • $\begingroup$ In $S_t$, as that is what you are testing the unit root null hypothesis on, and taking differences and tesing a zero coefficient in the transformed model is just an equivalent reformulation of the null. $\endgroup$ – Christoph Hanck Oct 16 '18 at 12:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.