Today's talk: A Bayesian t test using PyMC, 1 February 2013

[Andrew Straw]

A Bayesian t test using PyMC

Andrew Straw, IMP

Friday, 1 February 2013, 12:30pm, GMI Orange Seminar Room (room 9.36)

Across many branches of biology, null hypothesis significance testing is the tried-and-true method of establishing whether an effect is “real”. Nevertheless, many pitfalls must be avoided to correctly evaluate statistical significance and it is easy to make mistakes that render the analysis invalid. I will summarize a recent paper Bayesian estimation supersedes the t test (“BEST”) by John Kruschke (2012, Journal of Experimental Psychology: General.). Kruschke’s Bayesian approach purports to acheive the same goals as the t test with fewer potential pitfalls. Furthermore, his approach has several advantages, such as the ability to accept the null hypothesis. Computationally, credible intervals of important parameters such population means and effect size are found using Monte-Carlo techniques such as the Metropolis-Hastings algorithm. I will discuss a Python implementation of the BEST algorithm written using PyMC.