Revised, Adjusted, and Believed Anyway

The figures that economic argument runs on are estimates. They arrive early, from partial information, under a deadline, and they are revised for years afterwards, frequently by more than the change that was originally reported. This is not a scandal about any particular agency. It is a structural property of trying to measure an economy fast enough for the measurement to be useful.

Consider how a quarterly growth figure is built. Some survey returns are in, most tax records are not, entire sectors are estimated from partial indicators. The first estimate is, in effect, a forecast of the final one, and it is published as a fact and headlined as a fact. As company accounts, tax data and annual benchmarks arrive, the picture changes. The revision distribution is wide enough that the sign of a small quarterly movement is not reliable at first publication.

The consequences are not academic. The opening quarter of 2008 in the United States was first published as modest growth and eventually restated as a contraction. Policy in the intervening months was set against a picture that did not exist. Recessions are routinely identified only in revised data, after the decisions that mattered have been taken.

Seasonal adjustment introduces a second layer of modelling. Retail spending rises every December, construction falls every winter, and without correction every series would be dominated by a pattern nobody needs telling about. The correction is produced by a model fitted to previous years, which works well while the pattern holds and fails when it does not. The pandemic broke the estimated seasonality of several major series badly enough that agencies published explicit warnings about them.

Then there is sampling error, which is estimable and almost never printed. A monthly employment change reported as a precise figure frequently carries an error band wider than the change itself, meaning the movement being discussed on the news is statistically indistinguishable from nothing at all. The agencies publish the uncertainty in their technical documentation. It does not survive the journey to the headline.

None of this argues for ignoring the numbers. It argues for treating a single release as weak evidence, watching several indicators, and being suspicious of any narrative built on one quarter's movement in one series. Professional forecasters do this routinely. Public argument almost never does, because a single number is a story and a range is not.

The final element is institutional and it is the one that matters most. Statistics grade the government that produces them. Where ministers can influence what is published, everyone discounts the output: investors, lenders, and eventually the government itself, which loses its own ability to see what is happening. Greece falsified its deficit figures for years before 2009, and the restatement, roughly doubling the reported number, was the immediate trigger for what followed. The crisis is filed under sovereign debt; the proximate mechanism was a statistical office that could be leant on by the ministry it reported to.

The reforms that followed accordingly included statistical independence. It looks like an administrative detail and it is closer to a precondition for anything else working.