Big data in official statistics

Big data come in high volume, high velocity and high variety. Their high volume may lead to better accuracy and more details, their high velocity may lead to more frequent and more timely statistical estimates, and their high variety may give opportunities for statistics in new areas. But there are also many challenges: there are uncontrolled changes in sources that threaten continuity and comparability, and data that refer only indirectly to phenomena of statistical interest. Furthermore, big data may be highly volatile and selective: the coverage of the population to which they refer may change from day to day, leading to inexplicable jumps in time-series. And very often, the individual observations in these big data sets lack variables that allow them to be linked to other datasets or population frames. This severely limits the possibilities for correction of selectivity and volatility. Also, with the advance of big data and open data, there is much more scope for disclosure of individual data, and this poses new problems for statistical institutes.

So, big data may be regarded as so-called nonprobability samples. The use of such sources in official statistics requires other approaches than the traditional one based on surveys and censuses. A first approach is to accept the big data just for what they are: an imperfect, yet very timely, indicator of developments in society. In a sense, this is what national statistical institutes (NSIs) often do: we collect data that have been assembled by the respondents and the reason why, and even just the fact that they have been assembled is very much the same reason why they are interesting for society and thus for an NSI to collect. In short, we might argue: these data exist and that’s why they are interesting.

A second approach is to use formal models and extract information from these data. In recent years, many new methods for dealing with big data have been developed by mathematical and applied statisticians. New methods like machine-learning techniques can be considered alongside more traditional methods like Bayesian techniques. National statistical institutes have always been reluctant to use models, apart from specific cases like small-area estimates. Based on experience at Statistics Netherlands, we argue that NSIs should not be afraid to use models, provided that their use is documented and made transparent to users. On the other hand, in official statistics, models should not be used for all kinds of purposes.