> Restriction of range, such as floor and ceiling effects or selection effects, reduce the power of the experiment, and increase the chance of a type II error.[5] This is because correlations are attenuated (weakened) by reduced variability
If you want to try this for yourself, try generating some fake SAT scores: (a,b) pairs where a and b are normally distributed with mean 500 and sd 100, and where a and b correlate at about 0.8.
Then, from your unbiased data set, select all the pairs for which 1600 >= a + b >= 1300, and check the correlation of a with b in the new, smaller data set.
Just examining this as a thought experiment, you can see that someone with an 800 math score might have a verbal score anywhere from 500 to 800. If we lower the math score, our expectation for the verbal score rises -- someone with a 600 math score will have a verbal score between 700 and 800. That is the basis of the negative correlation in the subgroup, but it's purely an artifact of the way we chose to label data points as being in or out of the subgroup.
This means that, given high IQ, it's not necessary to have good chess skills to be in the elite group, whereas given low IQ, it is necessary to have good chess skills to be in the elite group. This will establish a negative correlation between "chess skills" and IQ within the elite group even though they are positively correlated overall.
Similarly, SAT math and verbal scores tend to be negatively correlated within the student population of any particular university, because universities will reject applicants whose total scores are low, despite the fact that the math and verbal scores are very strongly positively correlated in the larger population.