This highlights the major downside to "data-driven" policy and decisions.
Data can "lie". What is observed is not always reality, simply what we can see of it.
Consider auctions. You never actually "see" the bidder's demand or utility. Yes, there are some ways to structure auctions that in theory show willingness to pay and such (ignoring confounding factors and irrationality), but you don't actually observe anything beyond the bid.
Similarly, on websites, you don't always know the causal reasons people click here or there. You know perhaps enough to predict a step-wise behavior, but don't (usually) understand the full behavioral lifecycle -- especially if a metric improves but at the hidden cost of decrements to conversion and similar.
Data can "lie". What is observed is not always reality, simply what we can see of it.
Consider auctions. You never actually "see" the bidder's demand or utility. Yes, there are some ways to structure auctions that in theory show willingness to pay and such (ignoring confounding factors and irrationality), but you don't actually observe anything beyond the bid.
Similarly, on websites, you don't always know the causal reasons people click here or there. You know perhaps enough to predict a step-wise behavior, but don't (usually) understand the full behavioral lifecycle -- especially if a metric improves but at the hidden cost of decrements to conversion and similar.