Issues for the Calculation of Bayes Factors #225
Replies: 1 comment 2 replies
-
Again, this is an issue of scope. The "reasoning" behind the use of normal priors is... low friction for the reader, as those are not really the focus of the vignettes. To be clear, the vignettes are not fully fledged guides for model fitting with Stan or BayesFactor, but are guides for the functionality offered by bayestestR. You seem to be under the impression that the difference between our examples and |
Beta Was this translation helpful? Give feedback.
-
What has guided the choice of a normal prior as the default, mostly, in the vignettes? Why prefer it to use of a Cauchy prior? As implemented in the BayesFactor package (Morey and Rouder 2018), this uses a numerical approximation that avoids the convergence issues that arise for Markov Chain Monte Carlo. I cannot see why one would want to use Cauchy as implemented in bayestestR (Makowski, Ben-Shachar, and Lüdecke 2019). I would use bayestestR with a Normal prior only if provided with a good reason to prefer, in a specific analysis, that to Cauchy. What reasons might be adduced? The choice of prior, and issues with convergence, appear to me to trump differences that arise from different approaches to scaling. Accordingly, I regard these as secondary.
The attached pdf pursues these and related issues.
Issues for the Calculation of Bayes Factors.pdf
.
Beta Was this translation helpful? Give feedback.
All reactions