Exgaussian fitting for reaction-times: prios and link functions

Hi All,
I’m fitting reaction-time (RTs) with exGaussian using brms and would appriciate your help on two basic issues:

  1. Estimates: What I need is to examine effects on the shape of the RT distributions (i.e., mu an tau parameter in the exGaus) in natural scale of second or milliseconds. Yet, since brms is using a ‘log’ link by default its a bit tricky to convert interaction effects, etc and I think this might be a place where students will make code errors. What is the cost of sampling using an ‘identity’ link? Is there a reason we should avoid using ‘identity’ link by default?

  2. Priors: The default priors for the intercept in brms are very specific (and with large values). Any idea where this came from and way? Is the numbers trying to reflect milliseconds? I want to set the prior myself - but I’m sure there is a good reason for these numbers - I just couldn’t figure out what.

Thank you all!

    rt    ~ 1,
    sigma ~ 1,
    beta  ~ 1
  data = df,
  warmup = 1,
  iter = 2,  
  cores =1, 
  family = exgaussian(link = "identity", 
                      link_sigma = "log", 
                      link_beta = "log"))
                      prior     class coef group resp  dpar nlpar lb ub  source
 student_t(3, 523.1, 125.1) Intercept                                   default
           normal(1.7, 1.3) Intercept                  beta             default
       student_t(3, 0, 2.5) Intercept                 sigma             default

I got a few answers by now. Posting them in case they will help others.

  1. To get estimates back in natural scale (e.g., ms) we can use emmeans
 emmeans(model,  ~x1* x2, dpar = "beta", regrid = "response")

  1. We need the log link otherwise sampling will go to negative tau values which are bad for like estimation and not plausible.

Note: exGaus mcmc sampling seem to have a tendency to get stack on some chains, and setting initial values really helps Don't forget your inits | A. Solomon Kurz