Xkcd: Statistics Tip | Hacker News Reader

Xkcd: Statistics Tip

xkcd.com

14 pointsgnufx7 comments

Not to denigrate statistics at all, but this rings true to a physicist. In my field we won by increasing resolution and lowering background so the data were essentially obvious at some level, with small error bars from fitting peak areas. I think it's from the humour of "A Random Walk in Science": "The experiment failed, so we had to use statistics."

I feel that's not a good example. Parameter identification is a standard tool to try to figure out what factors might have been at play.

I feel that's not a good understanding of experimental physics. Why do you think I didn't understand the rather limited independent variables in my measurements or not understand the experimental setup well enough to debug experiments which I hadn't been involved in? Make good measurements of the right things. That's what advances experimental science. Obviously have a basic understanding of statistics for guidance.

> I feel that's not a good understanding of experimental physics. Why do you think I didn't understand the rather limited independent variables in my measurements or not understand the experimental setup well enough to debug experiments which I hadn't been involved in?

Your example is not good because increasing resolution is a "let them eat cake" copout, and ignores the importance of grasping basic parameter identification techniques to figure out key parameters from real data that is available instead of hypothetical data whose recording might even be unrealistic.

Suggesting that you should start by looking at the right place and employ real or fictional equipment to get enough resolution to make the data obvious is both unrealistic and unhelpful.

Any graph can look conclusive if you play enough with it. That's why you need to add tables

> That's why you need to add tables

I would say that that's why you need statistics. It's literally the field dedicated to reliably make sense out of data.

The important thing is to add error bars, which has probably been in XKCD passim. Yes, you want the numbers, but I've quite often just digitized plots from papers. I wrote a program for inverse plotting with an HP flatbed sometime in the '80s.

Reply on news.ycombinator.com