Zellner's g-prior alpha value (Bayesian linear regression)

TJ_99

Hello,

I might have noticed a small discrepancy with the g-prior function while conducting a Bayesian linear regression exercise. I was originally working in an older version of JASP 2020 (version 14) and when I went to the advanced options and selected the g-prior option the nearby alpha value box next to the Hyper-g option was greyed out.

However, when I updated to JASP version 18, and I selected g-prior, there is now an active alpha value box immediately adjacent to the g-prior selection. The position of this alpha box seems to have shifted slightly in the newer update.

The alpha value box has a restriction on the inputs from 2 to 4 and a recommended value of 3, which is what I thought Liang et al. (2008) recommended for the Hyper-g prior option.

I have seen on (https://cran.r-project.org/web/packages/BAS/BAS.pdf) some text that pertains - "g-prior", Zellner’s g prior where ‘g‘ is specified using the argument ‘alpha‘. Does this mean that the alpha value box in the 2024 JASP version 18 is really stating a g input? and if so why is it limited to only a range of 2-4? or is the alpha some sort of weighting factor? I thought the value of g was usually determined by the references shown below from Consonni et al. (2018).

In summary, did the newer release of JASP get an update to adjust the g-prior distribution manually (i.e., what is the alpha box, and what is its function?) or did my code glitch?

Thanks,

TJ

EJ

Thanks for this detailed report. I've notified the team and we'll respond shortly (I hope :-))

EJ

vandenman

Hi TJ,

Thanks for bringing this up. Your report is spot on, we're using the same input field for the g of the g-prior and the alpha for the hyper-g priors. However, this is indeed overly restrictive. I'll update this for the next release so that the g-prior gets a separate input field without any restrictions.

Cheers,

Don