Reviewer wants justification for the default prior
Hi,
I Have conducted a repeated measures Bayesian ANOVA and stated the default prior: r scale fixed effects =0.5; r scale random effects = 1). I have also cited and read the appropriate articles (e.g., Rouder et al., 2012). But the reviewer wants a specific justification. I do not expect a big effects from the experimental condition when I compare the models.
Any idea how to formulate the justification this without going to the math? The point is that it is a default prior that should fit most cases in experimental psychology. But the reviewer wants more than that.
Best regards,
Ester
Comments
Dear Ester,
Perhaps you only have fixed effects, in which case I'd just report those. The ANOVA priors were proposed by analogy to the t-test; if you conduct a between-subjects t-test with the default r=.707 setting you ought to get the same result as for a one-way ANOVA with two levels. You could conduct a robustness analysis and examine the extent to which you get qualitatively similar results if you change the settings somewhat. I think this is more compelling than philosophical argument.
Cheers
E.J.
We have the same problem.
This is not a very satisfying answer :-(
Hi Pete,
Some additional thoughts:
1. The default priors were chosen to meet formal desiderata, see @ARTICLE{BayarriEtAl2012,
AUTHOR = {Bayarri, M. J. and Berger, J. O. and Forte, A. and {Garc\'{\i}a-Donato}, G.},
TITLE = {Criteria for {B}ayesian Model Choice With Application to Variable Selection},
JOURNAL = {The Annals of Statistics},
YEAR = {2012},
volume = {40},
pages = {1550--1577},
}
2. One can promote subjective or informed priors, and that's fine and useful, but (a) default priors still provide a good reference point; (b) for complicated models the subjective approach becomes practically very difficult.
3. A robustness analysis is always useful.
4. If others have more information they can use a different prior -- as long as the data are available then anyone is free to apply any model they like.
5. At APS, Julia Haaf presented work that showed how one can add theoretically-motivated order-constraints. This is not yet possible in JASP, but it's on our agenda.
6. It is always a good idea to acknowledge uncertainty; there is no "ultimate" or "correct" prior distribution. But the defaults seem to work well enough for the applications that have been encountered, and provide a useful alternative solution to the "p<.05" summary.
Hope this helps.
Cheers,
E.J.