Which occurs when/because for whatever reason, samples with two negative traits are excluded, the example demonstrates it well: if P(a, b), P(~a, b), P(a, ~b), P(~b, ~a) all equal 0.25:

If you exclude the (~a, ~b) sample then within the population that remains it looks like a and b are negatively correlated. (p(a|~b) = 1, p(a|b)=0.5)

It's interesting because this is something that happens a lot => when dating, looks negatively correlate with niceness within people you date because you don't date people who are neither. Diseases in hospital populations are negatively correlated because if you don't have anything you're not in hospital.