Moneybox

Arin Dube Demolishes Reinhart/Rogoff Causal Argument

This is not as funny as en Excel error, but Arindrajit Dube’s re-analysis of the Reinhart/Rogoff statistical work on debt:GDP ratios and growth is actually the most important yet. What Dube does is directly investigate the causal element of their argument, which is the one that’s relevant for policy purposes. He confirms what seems to be the common ground of everyone in this debate, namely that there’s a statistical correlation between high debt:GDP ratios and slow GDP growth. But is that because a high ratio causes a low denominator, or because a low denominator causes a high ratio? The theoretical argument for the latter is strong whereas the former causal interpretation relied on some kind of unknown dark matter. The empricial evidence for the dark matter was supposed to be the existence of a “tipping point” when the debt:GDP ratio reaches 90% but we’ve now seen that there is no such tipping point. What Dube has done is find empirical evidence for causation running in the reverse direction.

He did this “by regressing current year’s GDP on (1) the next 3 years’ average GDP growth, and (2) last three years’ average GDP growth.” You can see the plots above. The outcome is that the backward-looking correlation is stronger than the forward-looking one.

Long story short, there is no empirical evidence for the existence of the relevant kind of macroeconomic dark matter and there’s some empirical confirmation of the basic theory that high-debt episodes are largely caused by slow-growth episodes.

Stepping back I want to make the point that it’s striking that R&R didn’t even check this. I don’t begrudge any academic’s right to rush into publication with an interesting empirical finding based on the assembly of a novel and useful dataset. I don’t even begrudge them the right to keep their dataset private for a little while so they can internalize more of the benefits. But Reinhart and especially Rogoff have spent years now engaged in a high-profile political advocacy campaign grounded in a causal interpretation of their empirical work that both of them knew perfectly well was not in fact supported by their analysis. The natural step between interesting correlation that doesn’t establish causation and political advocacy campaign based on an unsupported causal inference is do some further statistical work to test different causal theories.

They just didn’t do it. And as soon as the data was available, naturally enough other people decided to try methods appropriate to interrogating the causal issue and the causal inference R&R plainly preferred was demolished within 24 hours.