Read on Twitter
"differences are reliable if the 95% credible interval
for the comparison does not contain zero (Nicenboim & Vasishth, 2016)" Please no, don't cite us anymore, this was a mistake
. At that time, reviewers were obsessed with H0 rejection, and we were trying to appease them 1/3
for the comparison does not contain zero (Nicenboim & Vasishth, 2016)" Please no, don't cite us anymore, this was a mistake
. At that time, reviewers were obsessed with H0 rejection, and we were trying to appease them 1/3
In 2013-2016, Savage-dickey method was available for BF, but it wasn't trustworthy, and loo was either spitting large k-hats, and in general is undecided for nested comparisons. So every time reviewers wanted closure, we just said that. 2/3
But it's not right, I'm sorry! Please either just report the estimate of the effect and live with the uncertainty, or if you or the reviewer really want to quantify evidence against (or in favor of) the null, just run 80000 samples and use Bayes factor from bridgesampling 3/3
We are writing a paper saying among other things that we were wrong, I hope people read that 
4/3
4/3
