One of the most influential and widely cited economics articles on corruption is Paolo Mauro’s 1995 paper, “Corruption and Growth,” published in the Quarterly Journal of Economics (Vol. 110, No. 3, pp. 681-712). It has become a standard citation for the proposition that corruption is lowers investment, and consequently lowers economic growth. The paper is important because it sparked close to 20 years (and counting) of increasingly sophisticated research on the economic effects of corruption. Furthermore, it leant critical academic support to the emerging anticorruption movement in both civil society and international organizations like the World Bank and IMF. And for those reasons alone, I think one could make a strong case that this paper has had a positive impact on the world.

Yet I come not to praise Mauro’s paper, but to bury it. Though the paper is still widely cited as “proving” or “establishing” that corruption is bad for growth, it does no such thing. I’m happy to put aside (for the moment) objections to the perception-based index that the paper uses to measure corruption. And I’m even willing to put aside the fact that, as Mauro acknowledges (but downplays), although he finds a strong negative correlation between corruption and investment, the relationship he finds between corruption and growth is actually not that robust (that is, it sometimes disappears when certain control variables are added). No, the biggest problem with the article is that, even if it establishes a negative correlation between corruption and growth, it does not establish causation; although his findings are consistent with the claim that corruption causes low growth, they are equally consistent with the claim that low growth causes corruption, or that they are both caused by some third factor.Mauro recognizes this, and claims to identify the direction of causation using a statistical technique called “two stage least squares”, which in turn requires something called an “instrumental variable” (or simply an “instrument”). It’s at this point that those readers without a background in economics or statistics might start skimming, and be inclined just to take the author at his word that he’s used some fancy, semi-magical statistical technique to establish causation. But it’s important even for laypeople interested in corruption research to know what’s going on here, both in order to understand what this particular paper does and does not show, and because this statistical technique is so frequently used in work in the corruption/anticorruption field.

Here’s the basic idea, somewhat simplified, and without the technical jargon. Suppose you have two variables, X and Y (say, “control of corruption” and “growth”). You think that X causes Y, and sure enough, you find a strong correlation in the available data (countries that score high on “control of corruption” also score high on “growth”), even after you control for lots of other factors that might affect both of these variables simultaneously. But you’re still worried: How do you know X is causing Y, and not the other way around? You can solve your problem if you can find a way to isolate variation in your X variable that could not possibly be caused by Y. That is, you want to find a third variable Z—called an “instrument”—that is correlated with X, but could not possibly be correlated with Y except through X. That is, we want to find something that affects “control of corruption” but could not possibly have any direct effect on growth, other than via its effect on corruption. If we find such a variable, we can isolate the causal effect of X on Y, without having to worry about reverse causation from Y to X.

But to do this, we need a valid instrument. What makes an instrument valid?  Two things.  First, it has to be correlated with “control of corruption” (that’s pretty easy to check). Second, it cannot be correlated with “growth”, other than through the channel of corruption control. That’s much harder, and impossible to verify statistically. But we have to ask whether it seems plausible that the chosen instrument satisfies this second condition. That’s not a statistical question, it’s a substantive question; it can’t be answered with data analysis, only with common sense and knowledge of the relevant field.

So, what’s Mauro’s instrument? It’s “ethno-linguistic fractionalization” (ELF), which is a quantitative measure of ethnic/linguistic diversity. The crucial assumptions, for this instrument to be valid, are, first, that ethno-linguistic fractionalization is associated with higher corruption (which it seems to be, according to Mauro’s data), and second – and here comes the crucial bit – that ethno-linguistic fractionalization could not possibly have any negative effect on economic growth, except by increasing corruption levels.

Is that second assumption plausible? In a word, no. Indeed, it seems patently absurd, at least to me. There are all sorts of reasons that highly fractionalized societies might have lower economic growth rates, even if corruption had no effect whatsoever on growth. More fragmented societies may be less politically stable (in ways not captured by control variables), or more prone to violence, or have lower levels of interpersonal trust, or have other institutional features (say, particular voting or lawmaking systems) that affect economic performance. Linguistic diversity might also be a drag on growth because it entails additional communication costs. And higher levels of fractionalization may derive from other causes, like country size or terrain or climate, that also affect growth.

Or maybe not. Maybe none of those factors matters, and Mauro is right that ELF only affects growth by affecting corruption. But the point is we don’t know, and we can’t test this statistically. If you can’t say, with a straight face, that it seems impossible to you that ethno-linguistic fractionalization could affect economic growth through any channel other than government corruption, then you can’t accept that Mauro’s paper establishes that corruption causes (as opposed to correlates with) low economic growth.

Now, just to be clear, I don’t mean to say that I disbelieve the bottom line result. In fact, I think there are a lot of good reasons to think that corruption is indeed bad for the economy, and indeed I have devoted a recent post to making that case. And, to be fair, Mauro was one of the first scholars to try to figure this all out; the more recent research, which I view as more sophisticated and reliable, might not exist if he hadn’t made a first, valiant attempt to try to evaluate the economic impact of corruption using quantitative data. So why am I using a blog post to beat up on the article (and, derivatively, those that continue to cite it for the claim that corruption is bad for growth)?  Because it’s important that those of us who care about basing anticorruption policy on the best available information be scrupulous about the evidence on which we choose to rely, and understand the assumptions of the articles we cite. At this point, no sensible person should be citing Mauro 1995 as establishing that corruption has a negative effect on growth.