'Re-make/Re-model': Should big data change the modelling paradigm in official statistics?

摘要

Big data offers many opportunities for official statistics: for example increased resolution, better timeliness, and new statistical outputs. But there are also many challenges: uncontrolled changes in sources that threaten continuity, lack of identifiers that impedes linking to population frames, and data that refers only indirectly to phenomena of statistical interest. We discuss two approaches to deal with these challenges and opportunities. First, we may accept big data for what they are: an imperfect, yet timely, indicator of phenomena in society. These data exist and that's why they are interesting. Secondly, we may extend this approach by explicit modelling. 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 the 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. Moreover, the primary purpose of an NSI is to describe society; we should refrain from making forecasts. The models used should therefore rely on actually observed data and they should be validated extensively.