Not about randomness - but about curve fitting. It is actually very difficult to verify non-linear effects -- or maybe I should say the opposite. The statistical tools we use to identify non-linearities are prone to be very noisy, so even in the subset of data including the no-variation responses I am quite skeptical that downward increase is real, or just due to variance in the tails of the data.

So a common social science finding is `* graded effects` where you might not by default expect them, and that is the main headline of the paper. I think plateau effects are reasonable in many situations, https://blogs.sas.com/content/iml/2020/12/14/segmented-regre..., but the noisy data itself often won't be able to clearly differentiate between different curve shapes.