2
$\begingroup$

I have a series for Portuguese GDP from 1995 to 2015 (quarterly data).

The plot of GDP over time shows clearly a trend. I ran a Dickey Fuller and the series turned out to be stationary by using the option drift (p_value is close to zero).

How shall I interpret this result? There is clearly a trend so the series is (at least)not mean stationary. Running the Dickey Fuller,by adding the trend option, the series results non-stationary. When I notice a trend behaviour shall I always include a trend in my regression in order to make my analysis valid? Including more lags the situation doesn't improve. Can you help me with this puzzling outcome? Shall I consider the series non stationary and take the first difference to solve it?(I include stata outputs below)

enter image description here

$\endgroup$
0
$\begingroup$

Normally you would expect $\operatorname{log}(GDP)$ to be integrated of order one and $\Delta \operatorname{log}(GDP)$ (the growth rate of $GDP$) to be stationary. The long-run average growth rate will be a couple of percent (rather than zero) per year so it is reasonable to include a drift term when testing for a unit root in $\operatorname{log}(GDP)$. This is the same as to say that $\operatorname{log}(GDP)$ has a linear time trend, or that $GDP$ is growing exponentially.

Meanwhile, if you allow for a linear trend in the augmented Dickey-Fuller test specification, that means you allow for a linear trend in the growth rate of $GDP$, or a quadratic trend in $\operatorname{log}(GDP)$. That would be difficult to interpret from the subject-matter perspective. Would you expect the growth rate to have a linear trend? I doubt that.

I doubt it makes sense testing for a unit root in the level of $GDP$, as the level is growing or diminishing exponentially, if we follow the ideas in the first paragraph. So if you conduct the test on $\operatorname{log}(GDP)$, perhaps your results will match the expected findings.

If they do not, then you have to make a choice: (1) believe the scheme above and blame the short sample period, a structural change or yet something else for a "weird" test result, or (2) ditch the scheme (say that it is it that is "weird") and follow the test result.

$\endgroup$
  • $\begingroup$ Is it possible that the initial ADF test rejected a unit root because the Portuguese GDP data were characterized by a deterministic, rather than stochastic trend? $\endgroup$ – Amaziah Oct 20 '17 at 17:58
  • $\begingroup$ @Amaziah, Yes, if the ADF test specification allowed for the deterministic trend. If the deterministic trend was not allowed for, the null hypothesis would have not been rejected in presence of a deterministic trend (because a deterministic trend is even stronger than a stochastic trend). $\endgroup$ – Richard Hardy Oct 20 '17 at 18:25

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.