Just to add a note here: using the "height" example, suppose I told you that research has shown black people were on average shorter than white people. Then, it turns out, that my sample of "blacks" was from the area with the pigmy ethnicities, and if I excluded those from my definition of "blacks" then they were on average taller than white people. This is an extreme example, but here the statement "the mean height of black people is less than the mean height of white people" might be TRUE, but it won't be USEFUL.

This is because your sample of black people contains within it many separate distributions for the same attribute, and simply taking their mean is not helpful. I'm merely saying that in an unhomogeneous group, averages are more likely to be misleading.

Sure, a set of equal numbers of mice and elephants are on average bigger than a set of guinea pigs, but that's not a useful statement. And a generalisation from that particular sample to "short-haired animals are bigger than long-haired animals" would be outrightly unjustifiable from your data.