DataTaunew | comments | leaders | submitlogin
Statistics Done Wrong - The woefully complete guide (refsmmat.com)
22 points by while 3754 days ago | 10 comments


5 points by hmswaffels 3754 days ago | link

A friendly, informative, and light guide to some of the common issues in frequentist statistics/NHST. However, Baysian methods aren't even touched upon, which would address the whole, you know, P-Value issue.

From this, I conclude that the Author doesn't live in the Baye Area.

-----

3 points by capnrefsmmat 3753 days ago | link

The Author (me) learned intro stats from a Bayesian, but wasn't sure how to introduce the concepts at a low level.

I may have to add a section on Bayes factors for hypothesis testing, though:

https://www.andrew.cmu.edu/user/kk3n/simplicity/KassRaftery1...

-----

1 point by hmswaffels 3752 days ago | link

I liked the examples you used in your articles. [I had seen the dead salmon paper before, but your way of presenting it was funny.]

Not sure how one present practical bayes stats at the low level either- Kruschke does a great job, but it takes him most of a text book, and a whole lot of different coins from different mints to do it.

-----

2 points by Tomrod 3754 days ago | link

Just out of curiosity, what do Bayesians use for p-vals and other fitness measures?

-----

4 points by while 3754 days ago | link

Posterior distributions. From them you can derive confidence intervals (or credible intervals as called within Bayesian stats) for your estimates and also set up tests. I'm not that familiar with testing within the Bayesian regime though so if someone else could add on this I would also be interested.

-----

6 points by capnrefsmmat 3753 days ago | link

Bayes factors are the rough equivalent of hypothesis tests:

https://www.andrew.cmu.edu/user/kk3n/simplicity/KassRaftery1...

Alternately you can directly compute posterior odds.

-----

2 points by Tomrod 3753 days ago | link

Great read. Thanks.

-----

3 points by hmswaffels 3752 days ago | link

For an overview, check out: http://www.indiana.edu/~kruschke/articles/KruschkeAJ2012.pdf

Kruschke goes over decision rules for dealing with the null hypothesis within the context of a 95% highest probability density interval, among other things.

-----

2 points by Tomrod 3753 days ago | link

Sounds firmer than frequentist methodology. Thou almost convinceth me to become a Bayesian.

-----

2 points by while 3753 days ago | link

well, there are other problems with the Bayesian approach. Look into uninformative priors for example. Bayesian stats are really powerful for some tasks and frequentist in others. I think it is always good to know a bit of both and when to use what.

-----




RSS | Announcements