I have some experience in that area. If I'm reading this correctly, they don't know race, only ZIP code. They are extrapolating race from zip code. They also seem to be deriving results for extrapolated black/hispanic populations based on using the other populations as a control.

Its really just a question of whether their sample sizes are large enough to justify a conclusion about the overall population. They are reporting a 95% confidence interval. Real back-of-the-napkin math (read: don't take this to the bank) would say that a sample size of 9,604 could get you a 95% confidence interval with a 1% margin of error on a population of 100 million. So it seems they are doing ok, but this is not to say there are not a bunch of other possible issues or confounds to consider if someone dug into the study.

It depends on the inference you want to make. Average treatment effects are integrated over the entire sample. But that means it can “miss” or be biased for a subsample. You can then explicitly model that subsample, but sub samples by definition have less power. It’s not just black or minority, it’s also any subsample, like a history of hear disease, or height.

For further reading google: “Gelman there is never enough data”

This is actually slowing down some trials in the USA as they don't have enough non-white participants. I don't know if that's because they aren't trying hard enough or ethnic groups just don't trust the studies/corps running them. Probably a mix.