21.5 Well-Specified Models

Model misspecification, which roughly speaking means using a model that doesn’t match the data, can be a major source of slow code. This can be seen in cases where simulated data according to the model runs robustly and efficiently, whereas the real data for which it was intended runs slowly or may even have convergence and mixing issues. While some of the techniques recommended in the remaining sections of this chapter may mitigate the problem, the best remedy is a better model specification.

Counterintuitively, more complicated models often run faster than simpler models. One common pattern is with a group of parameters with a wide fixed prior such as normal(0, 1000)). This can fit slowly due to the mismatch between prior and posterior (the prior has support for values in the hundreds or even thousands, whereas the posterior may be concentrated near zero). In such cases, replacing the fixed prior with a hierarchical prior such as normal(mu, sigma), where mu and sigma are new parameters, with their own hyperpriors.