I do not think EA is about things that are relatively easy to measure. It is about doing things with the highest expected value. It is just that due partly to regression to the mean things with measurably high values should have among the highest expected values. See Adam Caseys posts on 80 000 Hours.

Goodhart, of course: after a short time, only the metric counts.

The solution is obvious: I create enough simulations that are good enough to constitute sentient beings, and make them all happy, that this adds up to MUCH more goodness than my present day job running a highly profitable baby mulching operation to fund it all. Like buying "asshole offsets".

Some policy changes are hard to measure. Some are controversial to measure — you can measure them, but people will call you nasty names for doing so.

I expect that anyone who measured and forecast the health effects of reduction in lead pollution, back in the days of lead paint and leaded gasoline, was probably called "anti-business" or worse. Fortunately, they won anyway, and the effects are indeed measurable — in reduced cases of lead poisoning, and apparently in increased IQs of city residents.