Would it not be useful to have an option to initialize parameters by drawing from the prior? According to this answer
the reason for initializing uniformly on [-2,2] instead, is to avoid numerical errors if a user specifies a flat prior. But sometimes I would like to specify a tight prior around let’s say
theta0, in which case the drawn initial value can have a really small prior probability and extra work is needed to get close to
theta0. Of course I could draw or set the initial values myself and then give as input to Stan, but having an option to do it in Stan would make things easier.