Saturday, February 22, 2014

Bayesian Tests to Quantify the Result of a Replication Attempt

This manuscript looks like a nice use of Bayes factors to assess replication results. I have not read it yet in detail, but the idea sounds right on target. From the abstract:
To quantify replication outcomes we propose a novel Bayesian replication test that compares the adequacy of two competing hypotheses. The first hypothesis is that of the skeptic and holds that the effect is spurious; this is the null hypothesis that postulates a zero effect size, H0: δ = 0. The second hypothesis is that of the proponent and holds that the effect is consistent with the one found in the original study, an effect that can be quantified by a posterior distribution. Hence, the second hypothesis --the replication hypothesis-- is given by Hr : δ ~ "posterior distribution from original study".
Here is the link: Bayesian Tests to Quantify the Result of a Replication Attempt; Josine Verhagen and Eric-Jan Wagenmakers, University of Amsterdam.

Friday, February 14, 2014

Improved icons for Bayesian and frequentist analysis


This post presents icons that attempt to capture the essence of Bayesian and frequentist analysis. There are four icons: Bayesian and frequentist approaches to decisions about null values, and Bayesian and frequentist approaches to parameter estimation. This post is an update of a previous post, motivated by many helpful comments from readers. For an explanation of what I mean by the "essence" of the approaches, and what I hope to achieve from this exercise, please see the previous post. Without further ado, the icons are presented below, first in a 2x2 grid, then one at a time with explanations in the captions.

Bayesian Frequentist
Null value
assessment
Estimation

Bayesian null value assessment: The light-blue lines indicate the posterior distribution of credible lines. The dark-pink line marks the null value (zero slope). The null value falls far outside any credible value. [Added Feb 16, 2014: Of course, the full decision rule involves a ROPE around the null value. The ROPE is not displayed here, just to keep the icon uncluttered.]

Frequentist null value assessment: The dark-blue line marks the best fit. The dark-pink line marks the null hypothesis.The light-pink lines show the sampling distribution from the null hypothesis. The best fit falls far outside any null-sampled line.

Bayesian estimation: The light-blue lines indicate the posterior distribution of credible lines. There is an explicit distribution of credibilities (i.e., posterior probabilities) across possibilities (possible slopes etc.).

Frequentist estimation: The dark-blue line marks the best fit. The two dark-pink lines mark the limits of the confidence interval. The light-pink lines show the sampling distributions around each of the confidence-interval limits; notice that the best-fit line falls at the extreme of each sampling distribution. There is no distribution of probabilities across possibilities; there are only three point values: the best fit and the two CI limits.
Creative Commons license appended Feb 18, 2014, as suggested by reader comment:
Creative Commons License The four iconic images for Bayesian and frequentist data analysis by John K. Kruschke are licensed under a Creative Commons Attribution 4.0 International License. Based on a work at http://doingbayesiandataanalysis.blogspot.com/2014/02/improved-icons-for-bayesian-and.html.