The Earth is Round (p < 0.05) (1994) | Hacker News Reader

The Earth is Round (p < 0.05) (1994)

ist-socrates.berkeley.edu

45 pointsnathell13 comments

Side note: while trying to submit this, I've noticed that the submission form doesn't escape less-than signs automatically. I had to manually escape the entity.

Why 0.05? because 1 in 20~22: two standard deviations. so if your distribution isn't normal Gaussian you don't even have tradition on your side, (or perhaps 1 in 20 doesn't fit your risk model for say, testing pharmaceuticals?)

http://www.jerrydallal.com/lhsp/p05.htm

"Still, why should the value 0.05 be adopted as the universally accepted value for statistical significance? ... It was Fisher who suggested giving 0.05 its special status... The value for which P=0.05, or 1 in 20, is 1.96 or nearly 2 ...Deviations exceeding twice the standard deviation are thus formally regarded as significant. Using this criterion we should be led to follow up a false indication only once in 22 trials"

It's a good article, but I was really hoping they'd prove the earth round (or not). Sadly the only mention is the title.

There's a very easy way to do null hypothesis testing.

Take the sha256 sum of all your experimental data. If the first fives bits of the hash are 0, then you can reject the null hypothesis with p < 3%

What is a sha256 SUM?

a cryptogtaphic digest, http://linux.die.net/man/1/sha256sum

The message is a joke intended to make people think about limitations of null hypothesis techniques. On the one hand, the technique I describe is statistically rigorous, on the other hand, it is stupid because it has no power and is always random.

However, in the Popperian framework of null hypothesis testing, there is no way to even express the fact that this test is worse than any other.

Relevant XKCD http://xkcd.com/1478/

Also http://xkcd.com/882/

Hypothesis: the submitter either came across the link while researching this xkcd, or remembered about the link after seeing the xkcd.

Unfortunately I just realized I have no idea how to do statistics anymore, so we'll have to wait for the submitter to confirm.

Submitter here. The latter is correct.

From 1994... that's depressing.

The discussions over the problems of frequentist statistics are old, old problems in which people have grown and died without settling. It seems like we're getting close, finally, now that Bayesian methods are becoming easier and easier.

Mind you, there are also a number of old, old discussions of the problems of Bayesian statistics, which I fully expect to grow and die without the world having resolved (or, even worse: considered resolved too glibly!).

Reply on news.ycombinator.com