Recruitment deviation penalties

Penalties on ln_RecDevs by population and region, independent or as a process over time. Called from get_recruitment_penalty.

Usage

get_rec_devs_penalty(
  n_pop,
  n_regions,
  n_est_rec_devs,
  rec_region_prop_spec,
  rec_region_prop,
  ln_sigmaR,
  bias_ramp,
  sigmaR_switch,
  ln_RecDevs,
  sigmaR2_early,
  sigmaR2_late,
  do_rec_bias_ramp,
  map_ln_RecDevs = NULL,
  RecDevs_model = 1,
  RecDevs_rho = NULL,
  RecDevs_rw_init_sigma = 5,
  RecDevs_pen_center = 0
)

Arguments

n_pop, n_regions, n_est_rec_devs

Dimension sizes.

rec_region_prop_spec

Integer switch; 1 skips populations and regions with a fixed zero recruitment proportion.

rec_region_prop

Array [pop, region] of the regional apportionment.

ln_sigmaR

Array [early/late, pop, region] of log sigma.

bias_ramp

Numeric vector [year] of bias ramp factors, between 0 and 1.

sigmaR_switch

Integer year index the deviations switch from the early to the late sigma at.

ln_RecDevs

Array [pop, region, year] of recruitment deviations.

sigmaR2_early, sigmaR2_late

Arrays [pop, region] of squared sigma, used for the bias-corrected mean.

do_rec_bias_ramp

Integer switch enabling the bias ramp log sigma term.

map_ln_RecDevs

Array [pop, region, year] mirroring map$ln_RecDevs. NA cells are fixed rather than estimated and go unpenalized, and cells sharing a level split one penalty. NULL penalizes every cell in full.

RecDevs_model

Integer process error structure: 1 independent, 2 random walk, 3 AR1.

RecDevs_rho

Array [pop, region] of unconstrained AR1 correlations, transformed to \((-1, 1)\) here. Read under RecDevs_model = 3.

RecDevs_rw_init_sigma

Standard deviation given to year one of a random walk. Default 5, which leaves the level of the series effectively free; NA starts the walk at zero under its own sigma. Read under RecDevs_model = 2.

RecDevs_pen_center

Integer. 1 centers on the deviations' own weighted mean, 0 on the bias-corrected mean. Read under RecDevs_model = 1 only.

Value

Array [pop, region, year] of negative log-likelihood penalties, zero where nothing is penalized.

Details

Under RecDevs_model = 1 the deviations are independent, on an early and a late sigma either side of sigmaR_switch. Both regimes are the same normal penalty read at a different sigma over different years, with an extra log sigma term when the bias ramp is on; that term is what makes the sigma estimable, since without it a larger sigma always reduces the penalty. Under RecDevs_model = 2 or 3 each year is centered on the one before it, so neither the bias ramp nor the own-mean center applies. The equations are in the model equations vignette.