Calculate one-step-ahead observation residuals

Adds CPUE, length-composition, and weight-composition OSA diagnostics to an Opal object. Use plot_osa_sdnr() for the dataset-level SDNR comparison.

Usage

opal_osa(x, seed = 123L, conf = 0.95, name = "osa")

Arguments

x

A fitted opal_obj.

seed

Random seed for discrete residuals; restored on exit. NULL uses and advances the current random-number stream.

conf

Confidence level for approximate chi-squared SDNR intervals.

name

Name under x$derived for the diagnostics.

Value

An updated object. opal_derived() returns its residual table, dataset summaries, and calculation settings.

Details

For fixed-effect fits, CPUE uses its lognormal CDF, and composition residuals use exact sequential binomial, beta, or beta-binomial CDFs for multinomial, Dirichlet, or Dirichlet-multinomial observations. Parameters are held at their fitted values. For random-effect models, RTMB's oneStepPredict() integrates latent states using its Laplace-based OSA approximation. Other observation streams remain conditioned on. There is no silent substitution of conditional residuals for marginal residuals.

Observation order is the stored input order, then ascending composition bin. The final composition bin is constrained and omitted. Removed fisheries, zero sample sizes, exhausted counts, and numerical failures remain in the table with explicit exclusion reasons. Multinomial counts are rounded as in the fitted likelihood; no new pseudocounts are added. SDNR should be near one, but its interval is approximate because fitted parameters and finite samples affect calibration. Catch is conditioned on and has no observation likelihood; process priors are not observations.

See also

Other assessment diagnostics: opal_diagnose(), plot_composition(), plot_osa_residuals(), plot_osa_sdnr(), plot_prior_posterior()