One of the better sources for these differences and their magnitude is Peopleware. Instead of citing vague research "somewhere", they did their own primary research. More specifically they set up coding wars where different programmers at different companies assigned to the same task independently.
They found 10x differences in time spent on the average coding war (which was typically a pretty small sample) with those finishing faster typically producing programs that worked better. So by any reasonable measure, at least 10x productivity.
There is your 10x, right? Wrong. They found that the best predictor of programmer performance was the performance of another programmer at the same company. Furthermore they managed to correlate a lot of that performance factor to specific factors, such as having a phone that turned off, a private room, adequate desk space, and so on. There was still something like an unexplained factor of 3 left over, but they didn't know whether it was environment variables they had not looked at (eg training), individual ability (which might be correlated across institutions), etc.
It is true that this research happened in the 80s, before the Internet was widespread. I would love to see it replicated again.
(Note, I'm aware of the existence of many other poorly controlled studies, see http://www.flownet.com/gat/papers/lisp-java.pdf for an example, but what is critical in the Peopleware study is that they acquired information both about productivity and about factors that might cause productivity differences.)
In Peopleware, they also observed that some programmers did not themselves appear to produce much, but their presence on a team made others more productive. They were not managers, they were not leads, they just helped the team work better together, in some fashion, "managing" out or across rather than up or down.
I will jump the shark and speculate that these gel-members tend to be female more often than male, and the gelling-like effect comes from how their presence and patterns of behavior influence the group dynamic. I will further speculate that the dearth of women in development incurs a hidden cost to productivity, since statistically, fewer teams will have women, and consequently, the incidence of gel-members will be lower, and teams more fragmented.
I think there is a business opportunity here somewhere, or maybe it has already been realized in places, but the players have not been eager to trumpet their results.
The thing about these studies, which I do appreciate, is that they are measuring a programmer's ability to work on something:
1) Small.
2) By themselves.
3) Well understood and documented.
4) Doesn't have any threat of changing out from under them.
In the real world, the bottleneck usually isn't writing code, it's writing code that does the right thing and making the right people happy in the right way.
In some ways, trying to apply the lessons from Peopleware in the real world is like using your knowledge of unicorn anatomy to predict the Kentucky Derby.
They found 10x differences in time spent on the average coding war (which was typically a pretty small sample) with those finishing faster typically producing programs that worked better. So by any reasonable measure, at least 10x productivity.
There is your 10x, right? Wrong. They found that the best predictor of programmer performance was the performance of another programmer at the same company. Furthermore they managed to correlate a lot of that performance factor to specific factors, such as having a phone that turned off, a private room, adequate desk space, and so on. There was still something like an unexplained factor of 3 left over, but they didn't know whether it was environment variables they had not looked at (eg training), individual ability (which might be correlated across institutions), etc.
It is true that this research happened in the 80s, before the Internet was widespread. I would love to see it replicated again.
(Note, I'm aware of the existence of many other poorly controlled studies, see http://www.flownet.com/gat/papers/lisp-java.pdf for an example, but what is critical in the Peopleware study is that they acquired information both about productivity and about factors that might cause productivity differences.)