Monday, June 13, 2011

Testing for OPEC-Causality

Last Friday, the OPEC meeting being held in Vienna broke up in what has been widely described as "disarray". A group of OPEC members, led by Iran and Venezuela, vetoed plans for an increase in the world production of oil. Think of the words "supply", "demand", and "elasticity". and you can guess what they were thinking.

Saudi Arabia then announced that it would step up to the plate and increase its production to meet world demand. It'll no doubt be assisted by some of its Gulf allies.

OPEC was founded in September 1960, with 5 member countries. Today, this cartel has 12 members - Algeria, Angola, Ecuador, Iran, Iraq, Kuwait, Libya, Nigeria, Qatar, Saudi Arabia, United Arab Emirates, and Venezuela. Yes, make no mistake about it - OPEC is literally a cartel. It determines the supply and price of crude oil. Let's take a closer look.

I obtained data for the world's consumption and production of crude oil from  BP's Statistical Review of World Energy, 2011. They're in the Excel workbook on the Data page of this blog. The data are also available by region, if you're interested.

We can see that production exceeded consumption from 1965 until about 1981-82, but the converse has been true since then. Keep in mind that quite sizeable inventories of oil are held around the world.

The upward trends in both series are hardly surprising. The other noteworthy feature of the data is the obvious break in the trends and levels of production and consumption in 1979. This was due to the "second" oil crisis, following the Iranian revolution. (You'll also see a smaller such break in the data in 1973 at the time of the "first" oil crisis.) 

As for the price of crude oil itself (from the same source as the consumption and production data), here is what has been happening (in real terms) since 1861 - note the action since 1970:

If OPEC  is a cartel, then you might think that the production of crude oil Granger-causes consumption, and not vice versa. But then there's Saudi Arabia's recent decision. The data in Fig. 1 provide a nice opportunity to check this out.

First, let's do some unit root testing. The EViews workfile for what follows is available on the Code page for this blog. Applying the augmented Dickey-Fuller (ADF) test to the consumption and production data (with drift and trend in the ADF equations), I get test statistics of  -3.129 (p = 0.112) and -2.800 (p = 0.205) respectively. On this basis we might conclude that the production series is I(1), and it's a close call for the consumption series.

But what about the structural break in 1979? We know that level and trend breaks affect the ADF test in the direction of "discovering" unit roots that aren't really there, so let's do a better job and use Perron's (1989) modified ADF test. We have his Case 'C'. I've constructed a dummy variable, DU, which is zero up to and including 1979, and unity thereafer. I've also created another artificial variable, DT, which is zero up to and including 1979, and then it's a linear trend from 1980 to 2010.

To use Perron's test we go through two steps. First, we regress the series (either consumption or production) against an intercept, linear time-trend, DU, and DT. Second, we apply the usual ADF test to the residuals from this regression, using the no-drfit/no-trend option for the ADF regression. The resulting ADF test statistic has a different asymptotic distribution from that for the usual ADF statistic, and the critical values depend on the proportion of the way through the sample that the break occurs. In our case, this proportion is 0.33.

Applying this procedure to the consumption and production data, the ADF test statistics are -3.885 and -4.118 respectively. Perron's 5% (10%) critical values are approximately -4.2 (-3.9). This is spooky! Too close to call, so I'll be careful and presume that at the 10% significance level there may be a unit root in the consumption series. I have to be cautious because Perron's test has only asymptotic validity, and the sample size here is just 46.

To test for Granger-causality between oil production and consumption, I'll use the Toda and Yamamoto (1995) approach to ensure that the Wald tests have their usual asymptotic chi-square distribution, even though the data may be non-stationary. (Even if both series were I(1) and cointegrated, this wouldn't alter what I do from here on.) I've discussed the Toda-Yamamoto procedure in detail in an previous post, so I won't repeat all of the details here.

I've estimated a 2-equation VAR model for oil consumption and production, with an intercept, linear trend, and the 2 artifical variables (DU and DT) in each equation. The usual information criteria suggest that the maximum lag length should be 2. With this choice the residuals are serially independent and normally distributed by the usual tests. To allow for the possibility that at least one of the series may be I(1), I then add a third lag of consumption and production to each equation, but I don't include the coefficients of these extra lags when I test for Granger non-causality.

Testing the hypothesis that oil consumption does not Granger-cause oil production, the Wald statistic is 16.59 (p = 0.0002). I'd reject the null hypothesis in this case. Conversely, testing the hypothesis that oil production does not Granger-cause oil consumption, the Wald statistic is 2.34 (p = 0.3108). I wouldn't reject the null hypothesis in this case.

(Incidentally, if we were to conclude that both of the series are I(0), the Granger-causality results come out the same. The p-values for the two Wald tests in that case are 0.0001 and 0.2729.)

So, I'd conclude from this little exercise that oil production is Granger-caused by oil consumption, and not vice versa. I guess that's what we're seeing right now with Saudi Arabia's reaction to the latest round of OPEC meetings.

Note: The links to the following references will be helpful only if your computer's IP address gives you access to the electronic versions of the publication in question. That's why a written References section is provided.


Perron, P. (1989). The great crash, the oil price shock and the unit root hypothesis. Econometrica, 57, 1361-1401.

Toda and Yamamoto (1995). Statistical inferences in vector autoregressions with possibly integrated processes. Journal of Econometrics, 66, 225-250.

© 2011, David E. Giles


  1. This comment has been removed by a blog administrator.

  2. Would you include lags of the dummy variables?

    1. No. The dummies are just shifting the intercept.