Inferencedata Handling

InferenceData Handling

Use this skill when you need to construct an ArviZ InferenceData object manually from a CmdStanPy fit. Most workflows do not need to do this directly — the fit-pipeline reference scripts construct InferenceData with the right groups. This skill documents (a) the conventions those scripts follow, so downstream code can rely on them, and (b) the manual conversion pattern when you need it.

Conversion pattern

Never use bare az.from_cmdstanpy(fit). It dumps everything into the posterior group, breaking downstream PPC, LOO, and prior workflows.

import arviz as az

idata = az.from_cmdstanpy(
    fit,
    posterior_predictive=["y_rep"],         # variables ending in _rep
    log_likelihood="log_lik",               # per-obs log-lik for LOO
    observed_data={"y": y_obs},             # original data
    coords={"obs_id": np.arange(N)},        # dimension labels
    dims={"y_rep": ["obs_id"], "log_lik": ["obs_id"]},
)

Variable naming conventions

Use y_obs for observed data and y_rep for replications throughout the workflow. Mixing y, y_pred, y_sim causes cascading KeyErrors across downstream agents.

Case sensitivity

Stan lowercases all variable names internally. Y1_rep in Stan → y1_rep in CmdStanPy output. Always use lowercase when referencing variables in Python.

Validation

When you construct InferenceData manually, sanity-check the required groups before saving:

assert "posterior" in idata.groups()
assert "posterior_predictive" in idata.groups()
assert "y_rep" in idata.posterior_predictive
assert "log_likelihood" in idata.groups()
idata.to_netcdf("posterior.nc")

fit-pipeline > references/posterior_fit.py writes posterior.nc with these groups; the repository's tests assert them.

Common failures

Loading

idata = az.from_netcdf("posterior.nc")
print("Groups:", idata.groups())

Related skills

• fit-pipeline — the reference scripts construct InferenceData for the posterior fit (posterior_fit.py) and the prior predictive run (prior_predictive.py: prior + prior_predictive groups from a GQ-only prior_model.stan fit).
• stan > Generated-Quantities-Only Programs — how to write the GQ-only programs those runs consume.