Hacker News Reader

Comment by knuthsat | Hacker News Reader

The researchers would say that they adjusted the statistics to eliminate that variable but these statistical tricks work only in physics yet somehow they find it acceptable to use for such small N.

View on HN

Small N point might be true, but “these statistical tricks work only in physics” is false on numerous spots. Things like statistical methods for controlling for certain associations aren’t “tricks” (they use GLM here) and they do not “only” work in physics. Without bringing up the whole replication crisis, there’s a lot of good science in other fields being done with these methods.

But what if the relationship with obesity has nonlinearities, how good is this control?

There are studies with millions of participants and by controlling for smoking, obesity, they still find increased risk and associations that are still very probably noise.fitting.

I’m not disagreeing with you that proper statistical research especially in a clinical setting is hard. It was the “tricks” and hint of non-physics : no-science I was responding to. I do think we should give doctors the benefit of the doubt. Just imagine the work involved in research with 100 real patients, let alone larger numbers. Then the next point is: where to get your research subjects. If you want to study non-linearities you’d want both patients and healthy subjects (and ideally people who switch categories). That’s several orders of magnitude more complex to get going. Not every subject gets funding for that scale.

Do you mean that the GLM wouldn't properly account for obesity's relationship with testosterone if the two have a nonlinear relationship? Is this something that could be accounted for by splines? I'm still trying to learn this stuff.

Reply on news.ycombinator.com