Skip to contents

Iteratively reweight composition data (McAllister-Ianelli)

Source: R/6-reweight.R
reweight_comps.Rd

Tunes multinomial composition weights by refitting: each pass sets a fleet's weight to the McAllister & Ianelli (1997) weight implied by the previous fit, and stops once the weights settle. The weight multiplies the input sample size, so changing it changes the fit and hence the next implied weight, which is why the weights are tuned iteratively rather than in one step.

Usage

reweight_comps(
  object = NULL,
  n_iter = 10,
  tol = 0.01,
  fleets = NULL,
  verbose = TRUE,
  fit = NULL
)

Arguments

object

An Rceattle model fitted at estimateMode "Estimate" (0) or "Hindcast" (1), the modes that optimize the hindcast.

n_iter

Maximum number of iterations (default 10).

tol

Relative change in the weights below which to stop (default 0.01).

fleets

Fleets to tune, as Fleet_code values or as names matching fleet_control$Fleet_name (as in linkage_spec()). Give ids or names, not a mix, R coerces c(1, "Survey") to c("1", "Survey"). Default NULL tunes every eligible fleet.

verbose

Print the weights each iteration (default TRUE).

fit

deprecated name for object, still accepted so existing scripts keep working. Supplying both is an error.

Value

The model refitted with the final weights, holding a reweight element: history (one row per fleet per iteration), iterations, and converged. The last rows of history are the weights the returned model was fitted with.

Details

Only multinomial fleets are tuned. A Dirichlet-multinomial fleet estimates its own weight within the likelihood, so tuning it externally would compete with that estimate; those fleets are named and skipped, as are any requested through fleets that have no fitted composition data.

Only the age/length composition weight (Comp_weights) is tuned. The conditional age-at-length and diet weights have McAllister-Ianelli analogues of their own (CAAL_weights, Diet_comp_weights), but they are not tuned here; set them by hand if needed.

Tuning stops when the largest relative change in any weight falls below tol. Hitting n_iter first gives a warning, not an error, the partly tuned fit is returned either way, and $reweight$history shows whether the weights were still settling.

The weight is a parameter, not a data input, so each pass holds it to the next through inits; fleet_control$Comp_weights supplies the value only when a model is built from scratch, and editing that column then refitting from an existing fit has no effect (fit_mod() warns when it detects this). Each pass writes the new weight to both, so the returned model's data list rebuilds the tuned model from scratch.

References

McAllister, M.K. and Ianelli, J.N. (1997) Bayesian stock assessment using catch-age data and the sampling-importance resampling algorithm. Canadian Journal of Fisheries and Aquatic Sciences 54:284-300.

Francis, R.I.C.C. (2011) Data weighting in statistical fisheries stock assessment models. Canadian Journal of Fisheries and Aquatic Sciences 68:1124-1138. (The main alternative; applied by hand rather than by this function.)

Examples

if (FALSE) { # \dontrun{
fit  <- fit_mod(BS2017SS, estimateMode = "Hindcast")
tuned <- reweight_comps(fit)
tuned$reweight$history      # weight per fleet, per iteration
} # }