predict.fit_model: Predict density for new samples (_Beta version; may change...

In James-Thorson/VAST: Vector-Autoregressive Spatio-Temporal (VAST) Model

Predict density for new samples (Beta version; may change without notice)

Description

predict.fit_model calculates predictions given new data

Usage

## S3 method for class 'fit_model'
predict(
  x,
  what = "D_i",
  Lat_i,
  Lon_i,
  t_i,
  a_i,
  c_iz = rep(0, length(t_i)),
  v_i = rep(0, length(t_i)),
  new_covariate_data = NULL,
  new_catchability_data = NULL,
  do_checks = TRUE,
  working_dir = getwd()
)

Arguments

x	Output from fit_model
what	Which output from fit$Report should be extracted; default is predicted density
t_i	Vector of integers, providing the time (e.g., calendar year) for each observation i. The first modeled interval will be min(t_i), and last will be max(t_i), and every interval corresponding to integers between min(t_i) and max(t_i) will also be modeled.
a_i	Vector containing values greater than zero, providing sampled area for each observation i, e.g., as_units(x, "km^2") for area swept in square-kilometers. use as_units(1, unitless) for observations without a natural area measurement (noting that resulting densities no longer have interpretable units in that case), or for observations where the response b_i already standardizes for area, e.g., b_i = as_units(x, "kg/km^2")
c_iz	Vector of integers ranging from 0 to the number of variables minus 1, providing the category (e.g., species, length-bin) for each observation i. This can be specified as a matrix, such that each observation is associated with multiple categories. Such specification treats samples as arising from the sum across multiple categories, e.g., to account for unlabeled multispecies data.
v_i	Vector of integers ranging from 0 to the number of vessels minus 1, providing sampling category (e.g., vessel or tow) associated with overdispersed variation for each observation i (by default v_i=0 for all samples, which will not affect things given the default values for OverdispersionConfig). In some cases a portion of observations have a overdispersion-effect, but not others, and in this case the observations without are specified as v_i=NA
keep_old_covariates	Whether to add new_covariate_data to existing data. This is useful when predicting values at new locations, but does not work when predicting data are locations with existing data (because the interpolation of covariate values will conflict for existing and new covariate values), e.g., when calculating partial dependence plots for existing data.

Details

predict.fit_model is designed with two purposes in mind:

1. If new_covariate_data=NULL as by default, then the model uses the covariate values supplied during original model fits, and interpolates as needed from those supplied values to new predicted locations. This then uses *exactly* the same information as was available during model fitting.

2. If new_covariate_data is supplied with new values (e.g., at locations for predictions), then these values are used in combination with original covariate values when interpolating to new values. However, supplying new_oovariate_data at the same Lat-Lon-Year combination as any original covariate value will delete those matches in the latter, such that originally fitted data can be predicted using alternative values for covariates (e.g., when calculating partial dependence plots)

Examples

## Not run: 

# Showing use of package pdp for partial dependence plots
pred.fun = function( object, newdata ){
  predict( x=object,
    Lat_i = object$data_frame$Lat_i,
    Lon_i = object$data_frame$Lon_i,
    t_i = object$data_frame$t_i,
    a_i = object$data_frame$a_i,
    what = "P1_iz",
    new_covariate_data = newdata,
    do_checks = FALSE )
}

library(ggplot2)
library(pdp)
Partial = partial( object = fit,
                   pred.var = "BOT_DEPTH",
                   pred.fun = pred.fun,
                   train = fit$covariate_data )
autoplot(Partial)


## End(Not run)

James-Thorson/VAST documentation built on Feb. 9, 2025, 9:05 a.m.