GDP Quarterly National Accounts: Quality and Methodology Information

Accuracy and reliability

The degree of closeness between an estimate and the true value.

The accuracy of GDP (i.e. how close is an estimate to the true value) is almost impossible to assess because it is based on a large number of data sources which each have their own margins of error, and there is no single ‘true’ data source for GDP. Instead, other dimensions of statistical quality are usually monitored to assess the statistics. In particular, timeliness and reliability are highly valued.

The reliability of GDP statistics is usually determined by how much initial estimates are revised over time. If revisions are too large then the value of the statistics is reduced. The timeliness of GDP statistics is particularly important for guiding and evaluating economic policy and the behaviour of people and businesses. There is therefore a careful balance which must be achieved between ensuring that the results published as early as possible with adequate reliability.

Expectations of accuracy and reliability in early estimates are often too high, and revisions are an inevitable consequence of this trade-off between timeliness and accuracy.

Analysis published alongside Scotland’s GDP in the past showed that that headline growth rates in the statistics are typically revised, on average, by around plus or minus 0.1 percentage points in the quarter following their initial release, and by around plus or minus 0.2 percentage points over the course of the next year. Occasionally there have been much larger revisions introduced to the statistics due to methodology improvements or changes to data sources.

Similar statistics for the UK and OECD members demonstrate that this scale of revisions is normal by international standards. Internationally, the OECD publishes similar analysis of revisions of quarterly GDP in selected countries. This analysis shows that, on average amongst the 18 countries considered, GDP growth rates are revised by +/- 0.18 percentage points between first estimates and the result published five months later, and by +/- 0.3 percentage points by one year later. There is considerable variation in the scale of revisions between the countries selected.

Very few statistical revisions arise as a result of errors in the popular sense of the word. All estimates, by definition, are subject to statistical "error". In this context the word refers to the uncertainty inherent in any process or calculation that uses sampling, estimation or modelling. Most revisions reflect either the adoption of new statistical techniques or the incorporation of new information, which allows the statistical error of previous estimates to be reduced. Only rarely are there avoidable errors such as human or system failures and such mistakes are made clear when they do occur.