Why though?
Yes, you are not controlling for the amount of coffee consumed. But that should cancel out in a population based study. The null hypothesis should be that humans drink coffee in a gaussian distributed way (not like with alcohol where the consumption is exponentially distributed). If you then randomly sample from that coffee drinker population (again, gaussian distributed), you can then control for variations in the coffee consumption by comparing the averages of coffee drinkers to non-drinkers.
But, yes, this does center around the hypothesis that coffee consumption is normally distributed. To confirm that I'd assume that there has been some research done on coffee consumption. Research that I have not done, to be clear. But I with no priors, I'd assume a normal distribution, because, well, darn near everything in bio is normally distributed.