Analysis was performed on 100,000 python repositories. Code analysis was performed on the go using custom scripts, github API, and t2.micro instances. Read a little more on approach here.
In the past, most studies have used package download data from pypi and related sites. Whilst this has been helpful, downloaded packages may not be used at all in scripts. This necessitated the exploration of the libraries that actually get imported. If a library was imported in a script, there was a higher likelihood of being used, than if it was just downloaded from pypi.