I have recently learned about horseshoe regression as an attractive alternative to PLSR for problems with many predictors. In my case I have a hyperspectral dataset with many (over 2k) spectral bands. I have attempted to start fitting a model using brms
set_prior(horseshoe(df = 1, scale_slab=2.5)) fit <- brm(Y~., family = gaussian() , data =data2)
however in addition to spectral information, I have additional information such as sampling day, as well as the varietal of each plant I sampled and the ID of the sensor used to measure the response I’m modeling. I want to include these variables in the model, but have the following questions (forgive me if they are silly questions):
- Is there a reason to set different priors on my non spectral covariates? e.g normal(0, 1)
- Is there a way to include varietal and machine ID as random effects in such a model? Would it be as simple as adding them to the model as if this were a standard linear regression i.e.
+ (1 | varietal) + ( 1 | machine), or is there a particular principled way to do this in horseshoe regression?