Grant McDermott bio photo

Grant McDermott

Data. Economics. Environment.

Email Twitter LinkedIn   Scholar   ORCID GitHub

Below is a recent article that I wrote for BizNews. The tl;dr version is that – surprise! – correlation doesn’t imply causation. Unfortunately, this simple rule continues to be ignored by many economic and financial pundits in the popular press.

Imagine that you are interested in determining the effect that policing has on crime. Say, does deploying more police officers lead to less crime? You might be tempted to answer this question as follows: 1) Gather up all the data that you can find on police numbers and crime statistics from different parts of the country. 2) Do some formal test (calculate a correlation coefficient, run a regression, etc.) and see what comes out of it.

Unfortunately, this simplistic approach would be unmistakably and irrevocably wrong. For one thing, I can virtually guarantee that it would yield a positive relationship. (The headline writes itself: “Police cause crime!”) The reason – and also the reason why our naive approach is fundamentally flawed – is straightforward: More police are deployed in areas with higher crime rates. Our headline has it backwards and we have made the classic error of mistaking correlation for causation.

Once you really start to think about it, this seemingly simple (and oft-repeated) admonition of correlation-vs-causation has powerful implications. Almost every empirical question becomes more complicated than it first appears. Another example: Does sending your kids to a good high school improve their grades and increase their chances of attending university? Probably, but how can we truly be sure when good high schools already tend to attract students that have higher IQs, strong work ethics and/or wealthy parents – all of which contribute significantly toward university attendance in of themselves? The question we are asking fundamentally requires us to think about a counterfactual world where smart, hard-working kids with wealthy parents are equally likely to enroll in impoverished high schools with a poor track records, as they are good ones. (Ditto for children with less fortunate backgrounds and all those in between.) Only then can we compare these counterfactual worlds to our own and attribute any residual difference to the “causal effect” of good schooling. Notice that I haven’t even touched on the issue of cost and whether it is actually worth shelling out all that extra cash on a posh boarding school…

Overcoming these sorts of problems is largely what modern-day economics and econometrics is all about. Economists even have a fancy word for it: “identification”. (As in, identifying true causal effects.) If you were to attend a presentation or seminar at any economics department around the world, I can assure you that much of the discussion would be given over to identification strategies, plausible counterfactuals, and the like. It can be hard and tedious work addressing these concerns, but it is always necessary. Similarly, it’s easy to run regressions. The tricky part is identifying meaningful relationships.

I was reminded of such issues as I read a recent article in this very publication: “Ugly secret – Weak Rand hit SA trade, economy by R150bn since 2010.” By now, you will hopefully see where I am going with this, but if not: The article (and intro from our magnanimous editor, Alec Hogg) undeniably mistake correlation for causation. For example, how can we be sure that causation doesn’t run the other way around? I.e. That the weak Rand is not a response – preemptive by the SARB or otherwise – to deteriorating economic conditions? Or, that the Rand and trade surplus aren’t both being driven by some other unobserved and unaccounted factor(s)? What the economist Don Cox memorably called “the dreaded third thing”.

The answer is we can’t. Certainly not with the information presented in the article alone. A priori, it would be equally plausible to claim that a strong Rand would have made the trade surplus even worse than it currently is. (Note that this latter argument at least appeals to the correct definition of a counterfactual.)

To end on a conciliatory note: The Maynard article does raise several thoughtful points. The fact that demand for exported South African goods may be more elastic than previously thought, for instance, constitutes a perfectly valid and insightful economic argument. I should also state for the record that I am in no way, shape or form arguing in favour of a weak Rand policy. Among other things, it imports inflation at a time when citizens and businesses are already struggling to keep up with sharply rising prices. (I suspect that corruption and mismanagement in the presidency are at least as much to blame as SARB policy, but that is a discussion for another day.)

Still, this is not the first time that I have become embroiled in discussions about the specific effects of a weak Rand. Then, as now, my overriding point is that we should be cautious of headline-grabbing statements regarding the direction and magnitude of such effects. A reliable rule of thumb is that arguments resting on simple correlation coefficients, or a single figure depicting two economic time series against each other, should be consumed with a healthy serving of salt. They are best viewed as inspiration for exploring deeper correlations and (potential) causation with more sophisticated methods, but no more.

We all know that correlation doesn’t imply causation. The real trick is in realising just how often this rule applies to the world around us.

P.S. - Fortunately, economists do have a number of empirical tools and tricks to tease out reliable causal effects. I won’t discuss these in detail here, but just in case you’re wondering: 1) Deploying extra police does reduce crime (link); 2) Going to a good school doesn’t necessarily increase your grades, although it does improve behavioural outcomes (link); 3) Currency devaluations can be either contractionary or expansionary depending on the circumstances (link).