You'd need to do it on enough projects to move folks around to different teams, and you'd need to do it for long enough on each project to average out the effects of everything from interpersonal differences to Mark going through a divorce, and hope that everyone is employed there long enough not to invalidate huge pieces of your data.

To the point of the parent, IMO speed of project execution is often an absolutely terrible metric to optimize for, especially if the parallel teams know that that's the goal. If you can find a generalizable objective metric that works across many different dev projects, then you've found something magical.