Abstract
In this article, we share some reflections on the state of statistical science and its evolution in the production systems of official statistics. We first try to make a synthesis of the evolution of statistical thinking. We then examine the evolution of practices in official statistics, which had to face very early on a diversification of sou rces: first with the use of censuses, then sample surveys and finally administrative files. At each stage, a profound revision of methods was necessary. We show that since the middle of the 20th century, one of the major challenges of statistics has been to produce estimates from a variety of sources. To do this, a large number of methods have been proposed which are based on very different f oundations. The term “big data” encompasses a set of sources and new statistical methods. We first examine the potential of valorization of big data in official statistics. Some applications such as image analysis for agricultural prediction are very old and will be further developed. However, we report our skepticism towards web-scrapping methods. Then we examine the use of new deep learning methods. With access to more and more sources, the great challenge will remain the valorization and harmonization of these sources.