observation_processes.RmdUsers can optionally request for afscOM to simulate observations from fishery catch or scientific surveys, by specifying simulate_observations=TRUE in the model_options list object. The observation submodel is currently developed based on surveys used by NOAA AFSC assessment scientists, but may be generalized in the future.
NOTE: The observation submodel has not been tested when the model contains multiple spatial regions.
The observation submodel generates relative population number (RPN) abundance indices, relative population weight (RPW) abundance indices, and age composition data from two scientific surveys: the Alaska longline survey [citation] and the NOAA bottom trawl survey [citation]. Data generated from each survey is representative of the entire coastwide population (even though some regions of Alaska are only sampled by the longline survey in alternating years). Surveys are assumed to occur halfay through the projection year, and survey catches are not removed from the population.
Abundance indices (RPNs and RPWs) are log-normally distributed. For RPNs, the predicted abundance index is:
\(I_y = q^{ll}\sum{N_{y,a,s}\exp{(-Z_{y,a,s}/2})v^{ll}_{y,a,s}}\)
And for RPWs is:
\(I_y = q^{trwl}\sum{N_{y,a,s}\exp{(-Z_{y,a,s}/2})v^{trwl}_{y,a,s}w_{y,a,s}}\).
For both surveys \(q\) represents the catchability coefficient for the specific survey and \(v\) represents the survey selectivity. For RPWs, \(w\) represents weight-at-age-and-sex. R functions for computing RPNs and RPWs according to the above equations are available as simulate_rpn() and simulate_rpw().
Observations are lognorally distributed around the predicted log abundance index (i.e. \(log(I_y)\)). Lognormal observations are bias-corrected to ensure that the log-mean and arithemtic-mean are consistent. Users must specify a coefficient of variation (CV) for abundance index observations for each survey. The underlying function that yields lognormal observations is as follows:
# pred is the predicted abudnance index as described above
# cv is a user input model_option
simulate_lognormal_obs <- function(pred, cv){
sds <- sqrt(log(cv^2 + 1))
return(rlnorm(1, meanlog=log(pred)-(sds^2)/2, sdlog = sds))
}Age composition data is multinomially distributed. Predicted age compositions are calculated as the proportion of the population that is each age (e.g. \(p_y = N_{y,a,s}/\sum_a{N_{y,a,s}}\)). Users can, optionally, specify whether age composition data should be generate in a sex-aggregated or sex-disaggregated (the default) fashion. Is disaggregated comps are requested, prediction compoistions represent the proportion of population in each age, of each sex. An R function for computing age composition predictions is available as simulate_ac().
Observations are multinomially distributed about the prediction age compositions. A user supplied number of multinomially samples (\(n\)) regulates how many random draws are taken from the multinomial distribution. Users can specify if the returned observations are in proportions, or integer format (which is necesarry for some estimation models), through the as_integers model option. As with the predictions, observations can be either sex-aggregated or sex-disaggregated (the default). An optional aging error matrix can also be applied to observation (though this hs not been thoroughly tested). The underlying function that yields multinomial observations is as follows:
simulate_multinomial_obs <- function(pred, samp_size, aggregate_sex=FALSE, as_integers=FALSE, age_err=NA){
multi <- array(0, dim=dim(pred), dimnames=dimnames(pred))
p <- pred
if(!all(p == 0)){
# [...]
multi <- apply(p, 3, function(x){
tmp <- rmultinom(1, samp_size, prob=x)
if(!as_integers){
tmp <- (tmp[,1]/sum(tmp[,1]))
}
# Include aging error if available
if(!all(is.na(age_err))){
tmp <- tmp %*% age_err
}
return(tmp)
})
# [...]
}
return(multi)
}At present, only catch-at-age data is generated from fisheries catch. The age composition data is generated in same way as for surveys, except that data is representative of catch-at-age from the whole system rather than population numbers-at-age. An R function for computing predicted catch-at-age observations is available as simulate_caa(). Like with survey age compositions, fisheries catch-at-age is multinomially distributed.
The observation submodel currently built-in to afcsOM is designed to emulate common data sources and data types for NOAA stock assessment biologists, but does not cover the full scope of possible observation data types. For example, there is not currently support for generating fishery CPUE indices or egg productions indices. For more complex observation processes, users can author custom observation submodels to generate additional data.
To implement a custom observation submodel, turn off the defalt observation model by setting model_options$simulate_observations=FALSE (this ensures that the project function wont generate its own observations). From there, you can define any type of function necesarry, and provide, as input, outputs from the project() function. For example:
model_options$obs_pars$sex_rat$sd <- 0.05
custom_observation_process <- function(naa, model_options){
# return observations of population sex-ratio
n_males <- sum(naa[,,2,,drop=FALSE])
n_females <- sum(naa[,,1,,drop=FALSE])
true_prop_males <- n_males/(n_males+n_females)
# assume its normally distributed (should be a better assumption, since its a proportion)
obs_prop_male <- rnorm(1, mean=true_prop_males, sd=model_options$obs_pars$sex_rat$sd)
return(obs_prop_males)
}
model_options$simulate_observations = FALSE # turn off default simulations
out <- project(removals, fleet.props, dem_params, prev_naa, recruitment, options=model_options)
naa <- out$naa_tmp
propmale_obs <- custom_observation_process(naa, model_options)If users want to simply augment the current set of observations (as is being done in the above example), setting model_options$simulate_observations=FALSE is not strictly necesarry.
For multiyear simulations, users will need to implement their own multiyear simulation framework with the custom observation model. See the “Multiyear Projections” page for additional information.