Length Composition Likelihood — get_length

Computes likelihood for observed length compositions using a probability-of-length-at-age (PLA) matrix. Gradients propagate through to growth parameters when they are estimated.

Usage

get_length_like(
  lf_obs_flat,
  lf_obs_ints,
  lf_obs_prop,
  catch_pred_fya,
  pla,
  lf_n_f,
  lf_fishery_f,
  lf_year_fi,
  lf_n_fi,
  lf_minbin,
  lf_maxbin,
  removal_switch_f,
  lf_switch,
  n_len,
  n_lf,
  log_lf_tau,
  lf_addtocomp = 1e-08,
  lf_n_int_fi = NULL
)

Arguments

lf_obs_flat: numeric vector of unrounded length composition counts (used when lf_switch=1).
lf_obs_ints: integer vector of integer length composition counts (used when lf_switch=3).
lf_obs_prop: numeric vector of length composition proportions (used when lf_switch=2).
catch_pred_fya: 3D array [n_fishery, n_year, n_age] of predicted catch-at-age from do_dynamics().
pla: matrix [n_len, n_age] probability-of-length-at-age from get_pla(). On the AD tape when growth parameters are estimated.
lf_n_f: integer vector [n_fishery] of number of observations per fishery.
lf_fishery_f: integer vector of fishery indices for each observation group.
lf_year_fi: list of integer vectors of year indices for each observation.
lf_n_fi: list of integer vectors of sample sizes for each observation.
lf_minbin: integer vector [n_fishery] of minimum bin index per fishery.
lf_maxbin: integer vector [n_fishery] of maximum bin index per fishery.
removal_switch_f: integer vector [n_fishery] indicating if fishery is removed (0=included, 1=removed).
lf_switch: integer likelihood type selector (1=multinomial, 2=dirichlet, 3=dirichlet-multinomial).
n_len: integer number of length bins.
n_lf: integer total number of length composition observations.
log_lf_tau: numeric vector [n_fishery] of log-scale variance adjustment parameters.
lf_addtocomp: small non-negative numeric constant added to predicted proportions before normalisation to robustify zero bins. Default 1e-08.
lf_n_int_fi: optional list of integer Dirichlet-multinomial sample sizes.

Value

a numeric vector of negative log-likelihood contributions, one per observation.

Details

For Dirichlet, concentration is pred * n_i * exp(log_lf_tau[f]); for Dirichlet-multinomial, it is pred * exp(log_lf_tau[f]). The multinomial branch uses RTMB's integer-rounded count behavior. The Dirichlet-multinomial is intended for real counts with variance adjustment equal to one; use Dirichlet for fractional effective sample sizes.