In allen-chen-noaa-gov/barebones.FishSET: Barebones discrete choice modeling

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

Runs discrete choice models. If you run into problems you can contact [e-mail].

• User supplies catch, choice, and distance data, plus any other data otherdat they need to run their chosen likelihood.
• Currently ships with a conditional logit function logit_c, an average catch logit function logit_avgcat, a full information likelihood with correction function logit_correction_polyint, an expected profit model with a normally distributed catch function epm_normal, an expected profit model with a log-normally distributed catch function epm_lognormal, and an expected profit model with a weibull distributed catch function epm_weibull.

Data

The user supplies catch, choice, and distance data. The data catch and choice should be dimensions (number of observations) x 1.

zi <- read.csv("zi.csv", header = FALSE)
zi$V1 <- as.numeric(zi$V1)

catch <- read.csv("catch.csv", header = FALSE)

choice <- read.csv("choice.csv", header = FALSE)
choice$V1 <- as.numeric(choice$V1)

predicted_catch <- read.csv("predicted_catch.csv", header = FALSE)
predicted_catch$V3 <- as.numeric(predicted_catch$V3)

distance <- read.csv("distance.csv", header = FALSE)

otherdat <- list(griddat = list(si = as.matrix(predicted_catch)),
    intdat = list(zi = as.matrix(zi)))

library(barebones.FishSET)

(Simulated catch and choice data are included in this package i.e. data(catch).)

str(catch)
str(choice)

The distance data should be dimensions (number of observations) x (number of alternatives). (Simulated distance data is included in this package i.e. data(distance).)

str(distance)

Other data may be something like predicted_catch for each alternative (e.g. with dimensions (number of observations) x (number of alternatives)). This data could be constructed by the user before estimation of the discrete choice model, for example by looking at a moving average of historical catches. Below is an example of predicted catches, as well as a data frame zi representing harvester characteristics. (Simulated harvester characteristics and predicted data are included in this package i.e. data(predicted_catch).)

str(otherdat)

For the logit_c function, any number of grid-specific variables (e.g. expected catch that varies by location) or interaction variables (e.g. vessel characteristics that affect how much disutility is suffered by traveling a greater distance) are allowed. However, the user must place these in otherdat as list objects named griddat and intdat respectively. Note the variables within griddat and intdat have no naming restrictions. Also note that for this likelihood griddat variables are dimension (number of observations) x (number of alternatives), while intdat variables are dimension (number of observations) x 1, to be interacted with the distance to each alternative.

If there are no other data, the user can set griddat as ones with dimension (number of observations) x (number of alternatives) and intdat variables as ones with dimension (number of observations) x 1. Finally, users can write their own likelihoods, and this example is specific to the package-supplied conditional logit (logit_c) function (see documentation for other likelihoods).

To run models call the discretefish_subroutine function

The user supplies initial parameters, optimization options, the likelihood function name, and the optimization method for a total of 8 inputs. For example:

initparams <- c(2.5, -0.8)
#Initial paramters for revenue then cost.

optimOpt <- c(1000, 1.00000000000000e-08, 1, 1)
#Optimization options for the maximum iterations, the relative tolerance of x,
    #report frequency, and whether to report iterations.

func <- logit_c
#The conditional logit likelihood function.

methodname <- "BFGS"
#The optimization method chosen, which must be one of the base R `optim`
    #options.

The subroutine function takes in 8 inputs and outputs model results in a list that can be summarized as:

errorExplain: If it exists, a description of the model error.
OutLogit: A matrix of coefficients, standard errors, and t-statistics
optoutput: Optimization information (such as number of function iterations)
seoutmat2: Standard errors
MCM: Model comparison metrics (e.g. AIC, BIC)
H1: The inverse hessian

results <- discretefish_subroutine(catch, choice, distance, otherdat,
  initparams, optimOpt, func, methodname)
results

The true marginal utility from catch in this data-generating process was equal to 3, and the true disutility from distance was equal to -1. The model estimates are correct relative to some unknown scale parameter:

results$OutLogit[1, 1]/results$OutLogit[2, 1]

allen-chen-noaa-gov/barebones.FishSET documentation built on March 1, 2024, 8:19 a.m.