>> We test these conflicting theories of change in a corpus of 10,830 works of fiction from 1880 to 1999 and find that slightly more than half (54.7 percent) of the variance explained by time is explained better by an author’s year of birth than by a book’s year of publication.
Ten thousand works! Wow, the authors did some reading!
Oh, no, wait. They used topic modelling. Basically, two people manually labelled a corpus with 10 categories of topic and then they trained a text classifier on that data. Then they used the classification to draw conclusions about the texts in the corpus.
I'm so sorry to be the typical disparaging HN commenter, but trying to understand literature through text modelling is like trying to enjoy nature by counting the trees in the forest. And yes, I'm aware it's common practice. That doesn't make it any more sensible.
Probably my favourite XKCD of all time:
"If I assign numbers to each of these things, then it becomes data and I can do maths on it".