Follow-up on a question I posted on that other website (https://x.com/adam_t_clark/status/1829883123431325947).
I ended up having to directly model latent states, i.e.:
xT ~ logN(x_mu, x_sd)
xOi ~ N(x_Ti, b0+x_Ti*b1)
I was initially sampling xT ~ N(), using a flat prior to limit the sampler to positive values. That ended up biasing the estimates and decreasing power, since it didn't let the sampler explore negative parameter space. Solution was to use a strictly positive distribution.
Thanks, @brosenbaum, for the help!