Your parameters that have lognormal priors need <lower = 0>
bounds, as well as any additional truncation you wish to specify. See here.
The MCMC sampler does not draw from the prior distribution. Additionally, I don’t see how the traceplots that follow demonstrate the latter point?
parameters {
real <lower = 2.2> alpha_8_4;
}
already constraints alpha_8_4 to be greater than 2.2.
parameters {
real <lower = 0, upper = 2.2> alpha_8_1;
}
model {
alpha_8_1 ~ lognormal(mu_alpha, sigma_alpha);
}
The declared bounds will constrain alpha_8_1
to be between [0, 2.2]. Placing the lognormal distribution on alpha_8_1
just changes the shape of the prior from uniform between [0, 2.2] to having truncated lognormal between [0, 2.2].
You shouldn’t need to do this once you change the parameter declarations to have appropriate bounds.
Higher level thoughts: relabeling quickly becomes infeasible, you need to reparametrize / break the symmetry in the model in order to ensure accurate estimation. Also I’m not sure what you mean by rotational invariance?