I agree, but there are two problems. First; a point made in the article - the breadth of data available in many domains (economics, business, medicine, psychol

I agree, but there are two problems. First; a point made in the article - the breadth of data available in many domains (economics, business, medicine, psychology, linguistics, biology, and more) is vast and unmanageable. An a-priori selection made by the attention of the investigator misdirects us. Additionally some data is not captured due to process problems, a classic is missed opportunity in business. You advocate focus on new business opportunities, I advocate focus on our old markets - I show that by treating our current customers differently a 5% reduction in cost is possible. You claim a 50% revenue opportunity, but how do you know? I can claim a causal link, you are claiming hypotheticals. I win the argument and then a year later we examine the data - I have achieved the cost saving, and there is no record of all the business that we didn't get. The opportunity loss is not on the spreadsheet. Conversely if you win and we look a year later I can show that our customers still cost us what they did, and I can show (if you were successful) that my initiative would have saved 7.5% of the spend one year ago (+50% thanks to you). Problem two; outside physics and chemistry not many things are comparable in the sense that statistics expects them to be. This is closely related to the scope of data problem, but hides it. Widgets made by machines or seeds carefully selected and screened can be made to share (approximately) an identity so that they can be statistically reasoned about in bulk. Other things can't and medicine is really running up against this now. Pain killers work differently on people with different genetics and different genders (also different ages, contexts...) Cancer evolves differently from case to case. The assumption of similarity and category membership is often broken and we often do not notice before doing a t-test.