As a former ICPC world finalist who later worked at Google, I would say that practicing for programming contests is almost like cheating as far as Google interviews are concerned. The Google interview format is almost identical to the ICPC practice sessions I did in college.

I've seen this video before, and IMO the extreme similarity between programming contest questions and Google interview questions could explain the negative correlation. Specifically, borderline engineers with programming contest experience are more likely to get hired than borderline engineers without programming contest experience, and therefore the set of Google engineers with programming contest experience includes more borderline people than the set of Google engineers without programming contest experience. Thus the slight negative correlation.

When I was in high school, being good at programming contests meant picking up a bunch of habits that you'd never use at work -- memorizing the same twenty #include lines that include every conceivable STL data struct you'd need; using single letter variable names; no comments whatsoever. If you and the next person come up with the solution at the same time, you might lose simply because the other person could type faster. You have to un-learn these habits for industry.

Just about anyone can get exposure to a set of representative coding questions (see: the USACO training robot, or Cracking the Coding Interview), but training for these contests means spending XX hours a year under a time limit trying to write code from memory (because you don't have time to look things up in the manual).

I expect this to be the true root cause of the negative correlation as well. Technical questions are good at rooting out those not good at data structures and algorithms, and good at promoting those that do. However, you can be excellent at solving ICPC questions and not know anything about documentation, object oriented design, OS, databases, anything web related, etc. It's a very specific skill and practicing programming contests optimizes for it.

This is the best explanation of the phenomenon I've read. I think you're exactly right.

In my opinion, based on my and my friends' experience, being good at programming competitions clearly makes you very good at interviewing.

I don't think that being good at contests makes you bad at the job, though. Consider the following scenario: assume that skill of being good at contests is completely orthogonal to being good at the job (so the real correlation coefficient is 0, instead of negative, observed by Norvig). Then, since it's easier to to get hired if you're good at contests, the observed correlation coefficient inside a single company will be negative, due to selection bias. Depending on the effect sizes, the selection bias might even make observed correlation coefficient negative even if it's positive in reality (which I think it is).

For a more intuitive explanation, consider a used car market buyer who has preference for black cars (this corresponds to algorithm-based interview process of companies like Google or Facebook). Even though the color of the car should have no correlation with its quality, the black cars owned by that buyer will more likely be worse quality than non-black ones.

> Depending on the effect sizes, the selection bias might even make observed correlation coefficient negative even if it's positive in reality (which I think it is).

This is well known in SAT scores. SAT math and verbal scores have a very strong positive correlation in reality; inside almost all colleges they are negatively correlated, because colleges impose threshold effects below and above. (That is, colleges admit students in a narrow band of SAT scores, rather than admitting students who are above a minimum threshold.) You can stylize that into the idea that total SAT score is constant for the students in any given college, which easily explains why component scores would be negatively correlated within colleges.