Hacker News Reader

Comment by Dylan16807 | Hacker News Reader

There's good reason to believe tons of sociological theories. The problem is figuring out whether the pro or con factors win out, and figuring out if you've nailed down every single factor.

Being simple, clear, and well-reasoned in the realm of people's behavior is not enough to give you very good accuracy.

The kind of priors you get by reasoning in a vacuum just aren't good enough in some realms. Bayesianly combining those priors with anecdotes doesn't really help.

Some ex-post theories are fine but I don't think you've hit upon the way to distinguish them.

View on HN

>There's good reason to believe tons of sociological theories. The problem is figuring out whether the pro or con factors win out, and figuring out if you've nailed down every single factor. Being simple, clear, and well-reasoned in the realm of people's behavior is not enough to give you very good accuracy.

This is all argument in favor of a weak prior, not to ignore the existence and utility of a prior completely. Ideally, you have solid (relevant) empirical data that informs your prior but if you're seriously evaluating "advice" then I'd venture to say that you're probably lacking such data.

Imo, the critique to be had of what I said above is that there's no clear general way to form good priors and thus the advice is somewhat empty. But that's just life -- there's no magic bullet algorithm to discover truth and one has to use his/her brain to do the best they can. Many who seem to well in this area tend to like the Bayesian framing of beliefs, and personally I've found it useful to guide my own thinking as well.

You can reframe almost any statistics as Bayesian if you try hard enough. That doesn't mean it's actually helping you. If you have a method for generating 'good' priors and it doesn't work, it does far more harm than good when you use that to analyze this kind of anecdote. Anecdotes are such weak evidence that the chance it pushes your priors into the correct result is super low. You'd be better off not trying. Just accept the need to gather proper amounts of evidence.

Or in other words: Some undescribed system of priors could help, but the system you actually described doesn't help. So "use priors" is a bad answer.

Reply on news.ycombinator.com