Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Supported by

Two questions about cauchy prior

edited April 5 in JASP & BayesFactor

Hi all! This is a bit out of place, but I am hoping that I can find an answer here because there is likely a great deal of overlap.

I am using the advice from Rouder and Morey (2012) regarding a default prior for regression (i.e., a Cauchy distribution with a scale of 1). I'm applying this in the computation of Bayes Factors in R using the "brms" package.

I have two specific questions.

  1. The paper recommends making the prior location and scale invariant by standardizing predictors and outcome variables. I take this to mean standardizing the predictors and outcomes such that they have a mean of 0 and a standard deviation of 1. However, I'm unclear what to do in the case of a categorical predictor with two levels. It's important to note that the input variable in "brms" is treated as a numeric variable, not a factor, even in the case of a categorical variable. Thus, the actual values chosen to represent group 1 and group 2 are important. Elsewhere (i.e., Gelman et al., 2008) I have seen the recommendation to scale categorical predictors in the same way as continuous predictors (in that case, to have a mean of 0 and sd of 0.5). Would I apply the same logic here and scale my categorical predictor to have a mean of 0 and sd of 1?
  2. The paper is talking about linear regression. Is there a comparable default prior for logistic regression?

Thanks very much! Like I said, I know this is a bit off topic as I'm not talking about JASP or BayesFactor. However, I hope that the overlap in subject matter (i.e., priors recommended by Rouder and Morey) makes it okay to post this here.

Comments

  • Dear dsidhu,

    1. I recall that Richard explicitly discusses the case of categorical predictors mixed with continuous predictors...ah yes, here: http://bayesfactorpcl.r-forge.r-project.org/#glm You could consult the documentation for further details, or send Richard an Email. But I do think this ought to be treated as a factor, not a number.
    2. There have been some suggestions for logistic regression. We are currently looking into this more deeply. If you Google this you'll see some suggestions.

    Cheers,
    E.J.

Sign In or Register to comment.