Interpreting data is hard.

Eg, if you have a bunch of airplanes returning from a war, you might be tempted to armor the areas where they were hit. But you want to armor planes where the holes arent because those are the critical areas.

If you can read literal bullet holes backwards, hundred dimensional preference vectors say something, but we generally have no clue.

Analytics don’t capture the “why” of what they measure, on the user side. Indicators may move in the seemingly right direction for the wrong reasons. Analytics typically can’t tell you what the user wanted to achieve. Knowing the “why” gives a better basis to decide on what changes to try, or to realize what’s actually wrong with the user-facing side of things. Another reason is that they tend to measure an average where in reality there is no average user.

Not GP, but I can humbly share mine.

I think it's because a lot of analytics are data points oriented instead of being workflow oriented. So you can see that feature A is not being used a lot, but it's very important in a particular flow. Feature B may be used a lot, but it can be only important for a particular class of users while very detrimental mentally for another class.

People use software for a needs, but rarely I've seen a need being highlighted when interpreting analytics data.

I think the answer to this is roughly ... the content of an applied statistics degree, or a research methodology practicum. There are so, so many ways that quantitative measurement of human phenomena can mislead, from measurement error to conflation between trends/sample sizes/effect sizes, to p-hacking, to good old bias (in interpretation or in deciding what tools to use to smooth or normalize data), and thousands more.