[luu]

They've shown that random selection is good when there's zero correlation between performance at level N and level N+1.

Personally, I'm more interested in what happens if the correlation is non-zero, and performance at level N+1 has more impact than performance at level N? How does the best promotion strategy change as the level of correlation and impact of being higher up the ladder changes?

[roundsquare]

They didn't give the details of the model. But I don't think you need the model to make some reasonable statements.. Let r = correlation between level N and and N + 1

r = -1 ==> promote worst performers r = 0 ==> promote at random or based on completely independent criteria. r = 1 ==> promote best performers

And you can fill in everything in between. Putting it this way almost makes it seem obvious. In fact though, the hardest part is judging r.

impact of being higher up the ladder changes?

They did do this:

measured as a weighted average of employee competence, with higher-level positions counting for more

Edit: Missed a word.