Stan Tricks and Tips

Wisharts

Wisharts are stubborn, and often not the right prior for covariance matrices — they were a popular choice for being a conjugate prior for the MVN. They are numerically unstable, however, and I've had numerous issues with this:

Rejecting initial value:
  Error evaluating the log probability at the initial value.
Exception: wishart_lpdf: LDLT_Factor of random variable is not positive de
finite.  last conditional variance is 0.00509742.  (in 'model2fcb3ae9e570_
964295899df0e36ebb89bdbe2adb540e' at line 23)