21.3 Statistical vs.  Computational Efficiency

There is a difference between pure computational efficiency and statistical efficiency for Stan programs fit with sampling-based algorithms. Computational efficiency measures the amount of time or memory required for a given step in a calculation, such as an evaluation of a log posterior or penalized likelihood.

Statistical efficiency typically involves requiring fewer steps in algorithms by making the statistical formulation of a model better behaved. The typical way to do this is by applying a change of variables (i.e., reparameterization) so that sampling algorithms mix better or optimization algorithms require less adaptation.