The choicedata package simplifies working with choice data in R.
Package design
The package breaks choice-data modeling into a series of objects. Each object contains the information needed for the next step.
choice_formula: the choice model formula.choice_alternatives: the set of choice alternatives.choice_effects: the choice effects, defined bychoice_alternativesandchoice_formula.choice_parameters: the model parameters, determined bychoice_effectsand estimated viachoice_likelihood.choice_identifiers: the identifiers for deciders and choice occasions.choice_preferences: the choice preferences of the deciders, identified bychoice_identifiers.choice_responses: the choice responses, influenced bychoice_preferences.choice_covariates: the choice covariates.choice_data: the choice data, built bychoice_covariatesandchoice_responses.choice_probabilities: the choice probabilities, computed fromchoice_dataandchoice_parameters.choice_likelihood: the likelihood of the choice model, formed bychoice_probabilities.
The objects are designed to be modular and can be combined in various ways to create a range of modeling workflows.
Examples
Empirical data
The TravelMode data set of the AER package (Kleiber and Zeileis 2008) contains the revealed preferences of 210 travelers choosing between air, train, bus, and car:
library("choicedata")
data("TravelMode", package = "AER")
TravelMode$choice <- TravelMode$choice == "yes"
head(TravelMode)
#> individual mode choice wait vcost travel gcost income size
#> 1 1 air FALSE 69 59 100 70 35 1
#> 2 1 train FALSE 34 31 372 71 35 1
#> 3 1 bus FALSE 35 25 417 70 35 1
#> 4 1 car TRUE 0 10 180 30 35 1
#> 5 2 air FALSE 64 58 68 68 30 2
#> 6 2 train FALSE 44 31 354 84 30 2We can transform the data from long (one row per choice alternative) to wide format (one row per choice occasion):
long_to_wide(
data_frame = TravelMode,
column_alternative = "mode",
column_decider = "individual"
)
#> # A tibble: 210 × 20
#> individual income size wait_air wait_train wait_bus wait_car vcost_air
#> <fct> <int> <int> <int> <int> <int> <int> <int>
#> 1 1 35 1 69 34 35 0 59
#> 2 2 30 2 64 44 53 0 58
#> 3 3 40 1 69 34 35 0 115
#> 4 4 70 3 64 44 53 0 49
#> 5 5 45 2 64 44 53 0 60
#> 6 6 20 1 69 40 35 0 59
#> 7 7 45 1 45 34 35 0 148
#> 8 8 12 1 69 34 35 0 121
#> 9 9 40 1 69 34 35 0 59
#> 10 10 70 2 69 34 35 0 58
#> # ℹ 200 more rows
#> # ℹ 12 more variables: vcost_train <int>, vcost_bus <int>, vcost_car <int>,
#> # travel_air <int>, travel_train <int>, travel_bus <int>, travel_car <int>,
#> # gcost_air <int>, gcost_train <int>, gcost_bus <int>, gcost_car <int>,
#> # choice <fct>We can construct model design matrices:
mode_data <- choice_data(
data_frame = TravelMode,
format = "long",
column_choice = "choice",
column_decider = "individual",
column_alternative = "mode",
column_ac_covariates = c("income", "size"),
column_as_covariates = c("wait", "vcost", "travel", "gcost")
)
mode_effects <- choice_effects(
choice_formula = choice_formula(
formula = choice ~ vcost | income | wait,
error_term = "probit"
),
choice_alternatives = choice_alternatives(
alternatives = levels(TravelMode$mode)
),
choice_data = mode_data
)
mode_design <- design_matrices(mode_data, mode_effects)
mode_design[[1]] |> round()
#> vcost income_bus income_car income_train ASC_bus ASC_car ASC_train
#> air 59 0 0 0 0 0 0
#> bus 25 35 0 0 1 0 0
#> car 10 0 35 0 0 1 0
#> train 31 0 0 35 0 0 1
#> wait_air wait_bus wait_car wait_train
#> air 69 0 0 0
#> bus 0 35 0 0
#> car 0 0 0 0
#> train 0 0 0 34Simulated choice
generate_choice_data() makes it straightforward to simulate choice data. The example below simulates 200 ranking tasks with three alternatives and recovers the data-generating parameters by optimizing the likelihood:
library("choicedata")
set.seed(1)
sim_effects <- choice_effects(
choice_formula = choice_formula(
formula = choice ~ x | y + 0 | z,
error_term = "logit"
),
choice_alternatives = choice_alternatives(
J = 3,
alternatives = c("A", "B", "C")
)
)
sim_parameters <- generate_choice_parameters(sim_effects)
(sim_data <- generate_choice_data(
choice_effects = sim_effects,
choice_identifiers = generate_choice_identifiers(N = 200),
choice_parameters = sim_parameters,
choice_type = "ranked"
))
#> # A tibble: 200 × 13
#> deciderID occasionID choice y x_A x_B x_C z_A z_B
#> * <chr> <chr> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 1 1 B -0.621 0.487 0.738 0.576 -0.305 1.51
#> 2 2 1 A 0.594 -2.21 1.12 -0.0449 -0.0162 0.944
#> 3 3 1 C -0.156 0.919 0.782 0.0746 -1.99 0.620
#> 4 4 1 A -0.0538 -1.47 -0.478 0.418 1.36 -0.103
#> 5 5 1 B -0.165 -1.38 -0.415 -0.394 -0.0593 1.10
#> 6 6 1 A 0.769 -0.253 0.697 0.557 -0.689 -0.707
#> 7 7 1 B 1.43 -0.112 0.881 0.398 -0.612 0.341
#> 8 8 1 B -0.0392 1.98 -0.367 -1.04 0.570 -0.135
#> 9 9 1 A 0.153 0.690 0.0280 -0.743 0.189 -1.80
#> 10 10 1 C 0.291 2.17 0.476 -0.710 0.611 -0.934
#> # ℹ 190 more rows
#> # ℹ 4 more variables: z_C <dbl>, choice_A <int>, choice_B <int>, choice_C <int>
sim_likelihood <- choice_likelihood(
choice_data = sim_data,
choice_effects = sim_effects
)
true_vector <- switch_parameter_space(
choice_parameters = sim_parameters,
choice_effects = sim_effects
)
fit <- stats::optim(
par = stats::rnorm(length(true_vector)),
fn = function(par) {
compute_choice_likelihood(
choice_parameters = par,
choice_likelihood = sim_likelihood,
logarithm = TRUE,
negative = TRUE
)
}
)
estimated_parameters <- switch_parameter_space(
choice_parameters = fit$par,
choice_effects = sim_effects
)
data.frame(dgp = true_vector, estimated = fit$par) |> round(2)
#> dgp estimated
#> beta_1 -1.98 -2.36
#> beta_2 0.58 0.53
#> beta_3 -2.64 -2.83
#> beta_4 5.04 5.64
#> beta_5 1.04 1.18
#> beta_6 -2.59 -3.03Related work
{Rprobit} (Bauer et al. 2023) provides maximum approximated composite marginal likelihood estimation for efficient probit choice modeling.
RprobitB (Oelschläger and Bauer 2025) provides Bayesian tools for estimating probit models.
Contact
You have a question, found a bug, or want to contribute? Please file an issue on GitHub.