Come analyze HEASARC, IRSA, and MAST data in the cloud! The Fornax Initiative is now welcoming all interested beta users.
Common pitfalls
- Forgetting bayes on. Setting prior
expressions with bayes <par> <type> declares
them, but the samplers do not consult them until inference is
switched on.
- Improper priors with nested sampling. A flat prior on
a parameter that ranges over many decades produces a prior with
effectively no normalization. nest run will refuse to start;
the fix is to set jeffreys on the parameter and push the
hard lower limit above zero with newpar.
- Stale fit covariance. For MH the proposal covariance
is taken from the current fit. If the fit converged to a basin
far from the posterior mode of interest — for example because of
“temporarily frozen due to zero norm” parameters — the
proposal will reject almost everything. Refit, or supply a
diagonal proposal with proposal gaussian deltas.
- Running HMC without warm-up. The step size and mass
matrix must adapt before the production samples are taken; running
with hmc warmup
on a non-trivial model
guarantees divergences. The default of 
warm-up iterations
is sensible for most problems.
- Comparing chains of unequal length. Rubin–Gelman
assumes the chains being compared have run for the same number of
iterations. Pad the shorter chain or discard the excess from the
longer one before invoking the statistic.
- Reporting frequentist confidence intervals from Bayesian samples. Posterior credible intervals are not frequentist confidence intervals, and they should not be reported as such. Be explicit in the text about which kind of interval is quoted.