Pseudo-likelihood inference underestimates model uncertainty: Evidence from Bayesian nearest neighbours