So after gwern pointed out that there is a transcript and I have read it I made a back of the envelope calculation.
Assumptions: According to Wikipedia people with a bachelor's degree or higher make $56078 per year, so about $27 per hour. Learning German increases yearly income by 4% and learning it takes about 750 class hours, according to the foreign service institute. Learning Spanish increases income by 1.4% and takes 600 class hours to learn. If we assume that one class hour costs 11.25$ per hour (by glancing at various prices posted on different sites) we can make a calculation.
Assuming the language is learnt instead of working and that the foregone hours have no impact on later earning and that the foregone hours are paid at average salary, the student incurs opportunity cost in addition to the pure class cost. Ignoring all other effects, learning German costs $28657 with return of 2243 p.a. and learning Spanish costs $22926 with return $841 p.a. This works out to 7.8% and 3.6% on initial investment respectively.
So after 13 years learning German pays off, after 28 years learning Spanish pay off. Assuming the language is learnt at young age, at least learning German can be worthwhile. More benign assumptions, such as learning outside of class with some kind of program like Duolingo, will increase the return further, making learning even more worthwhile.
Of course I did not consider learning something else in those hundreds of hours that could have an even greater effect on income but for high income earners language learning is a very plausible way to increase their income. I assume this goes especially for people that have more obvious use of learning an additional language like translaters or investors.
You're assuming that the correlation is purely causal and none of the increased income correlating with language learning is due to confounds; this is never true and so your ROI is going to be overstated.
This works out to 7.8% and 3.6% on initial investment respectively.
Most people have discount rates >4%, which excludes the latter. Throw in some sort of penalty (50% would not be amiss, given how many correlations crash and burn when treated as causal), and that gets rid of the former.
Language learning for Americans just doesn't work out unless one has a special reason.
(Unless, of course, it's a computer language.)
You know the drill - If it's worth saying, but not worth its own post (even in Discussion), then it goes here.