Considering the advantages of bayesian mixed-effects models with informative priors over non-bayesian approaches, I am attempting to figure out the best/most accurate priors to define for my models.
There is ample info on human performance in lie detection research, and I am attempting to use it, but I am less knowledgable in how to properly define priors in brms.
The info I would like to use is as follows:
- most studies use a bounded/truncated scale, 0-100%, which typically follows a normal distribution
- the average expected mean is 50% (SD = 15%) (although, the research always finds the population average to be 54%; which reflects the high/biased truth detection (62%) and low lie detection (46%))
- effect sizes (from interventions) are typically small-to-moderate (d = 0.1 - 0.5)
Considering the above, would something like this be reasonable:
brm( Accuracy | trunc(lb = 0, ub = 1) ~ veracity * training + (1 + veracity | subject),
data = data, family = [???] #not sure which; normal, beta?
prior = c(set_prior("normal(.5,.5)", class = "b"),
set_prior("cauchy(0,5)", class = "sd"),
set_prior("lkj(2)", class = "cor")))
Any suggestions are welcomed.