Say I've got a building full of people. 48% are male. 35% have blue eyes, 50% are over 50, and 22% are balding.

You can start speculating on the relationship between age and hair status because they're both based on the same population: the inhabitants. Older people are more likely to be bald is a fair hypothesis.

But if you start trying to make guesses about how many bald men are in the building, you have a problem, because you're making a common bad assumption that all bald people are male, but also now you're comparing a stat about the whole population to one about 48% of the population, which you don't do.

When you're doing scientific experiments and you want to know these things, you set up the cohorts ahead of time to test your hypotheses. For instance a test group that's 100% men or 100% women. Or only people with blue eyes. That way any two correlations you're looking at are based on the exact same number of samples.

And in this case as someone else already explained, the % is relative to total bankruptcies not the entire population, and the other is about the entire population.

It does, indirectly. The percentage of people going bankrupt is X/100, but a percentage increase in that percentage is (X/100)*(Y/100).