Diagnostics and priors

Priors are explicit

create_mediation_prior_specification() exposes all default scales. The defaults are weakly informative, not silently data-adaptive. Use prior predictive simulation to determine whether those priors imply plausible mediator and outcome distributions in the declared design.

spec = specify_multilevel_gaze_mediation(
    prepared,
    mediator_family="gaussian",
    outcome_family="bernoulli",
    priors=create_mediation_prior_specification(coefficient_sd=0.75),
    missingness_policy="quality_eligible",
)

prior_predictive_check_mediation(spec, draws=200, seed=2026)

Required sampler diagnostics

Before interpreting an indirect effect, inspect:

• maximum R-hat;
• minimum bulk ESS;
• divergences;
• maximum-tree-depth hits;
• prior predictive behavior;
• posterior predictive behavior.

The standard extraction functions treat critical convergence failure as a blocker rather than a footnote.

Prior sensitivity

Refit the same declared model under at least one defensible narrower and wider coefficient prior. Compare the indirect-effect distribution, direct paths, posterior predictive behavior, and convergence diagnostics. A result that changes materially under modest prior changes should be reported as prior-sensitive.

Random slopes

Random slopes are optional and should reflect the repeated-measures design rather than a maximal-by-default rule. For the stable simple mediation model, the Python implementation recognizes mediator_x, outcome_x, and outcome_m. If both participant-specific a and b slopes are estimated, participant-specific indirect effects can be derived; otherwise the package does not fabricate them. Serial and moderated extensions currently use participant random intercepts only; non-empty random-slope requests are rejected explicitly rather than ignored.

Worked sensitivity example

See Worked example: prior sensitivity for a reproducible narrow/default/wide coefficient-prior workflow and reporting language.