There is some worthwhile discussion of this in a small subreddit: https://www.reddit.com/r/dataisugly/comments/1cbivz3/cant_fi...

My highlight from that discussion is from someone writing a dissertation on a similar topic "that there's a kind of hill-shaped relationship. Expertise diversity tends to have the greatest benefit at moderate values rather than very high or very low values", which throws off this paper which tries to measure a linear correlation.