hyemin-han
commented
1 year ago In my opinion, that is because how the null interval was set in ttestBF. Because this code, TWOT, is developed for a very special purpose upon a request, comparing two groups to see whether A > B, I had to set that specific null interval parameter accordingly. As a result, non-positive effect sizes are likely to be not well treated. So, depending on your interest, you need to adjust the null interval properly. You may refer to this stackexchange post for further information:
https://stats.stackexchange.com/questions/262605/understanding-bayes-factor-in-bayesfactor-package
One additional caveat: I have never tested whether ttestBF-based bayesian testing code behaves properly when a null hypothesis becomes an interest. So, you need to double check that. At the beginning, BayesFactorFMRI has been developed for the most common test in fMRI analysis, A>B (or B<A).
Hi. I'm trying to use your software to test the evidence FOR the null hypothesis. In this context, I tried to see if I could fine very low BFs, but I could not find any BFs below 0.9. To investigate that further, I checked the cohen's ds, but surprisingly, none was below 0. I think this should not be the case, because, checking the image of my parameter estimates at the 2nd level (conducted with fsl), I found a lot of (highly) negative values, which should convert to negative ds. They instead correspond to very low but positive ds in the output of your software. The activations found with fsl and your software are otherwise consistent (which is nice :) ).
So, I am wondering:
why negative parameter estimates don't become negative ds? they should be
Does this impact the BFs? and especially, is this why I don't find any very low BFs? Normally, I should expect very low BFs (much lower than 0.9) for highly negative betas.
what do you think about the possibility to use your software to test the null hypothesis?