Skip to contents

Posterior choice probabilities and shares for hierarchical Bayes fits

Source: R/hb_postest.R
predict.choicer_hb.Rd

Computes counterfactual choice probabilities, integrating over the posterior draws and (at the population level) over the random-coefficient distribution: for each kept draw \((b_r, W_r, \delta_r, \ldots)\) one \(\beta \sim N(b_r, W_r)\) is drawn and the model probabilities are averaged. Alternatives in newdata that were not in the estimation sample receive a posterior-predictive \(\delta_{new} \sim N(z_{new}'\theta_r, \sigma_{d,r}^2)\) — the entry counterfactual unlocked by the random-effects \(\delta\). Price or subsidy counterfactuals are just modified covariate columns in newdata.

Usage

# S3 method for class 'choicer_hb'
predict(
  object,
  newdata = NULL,
  level = c("population", "individual"),
  n_draws = 200L,
  aggregate = TRUE,
  ...
)

Arguments

object

A choicer_hmnl or choicer_hmnp fit.

newdata

Data frame with the estimation columns (choice column not required). NULL (default) predicts on the estimation data.

level

"population" (default) integrates over \(N(b, W)\); "individual" uses the respondent-level \(\beta_i\) (requires the prediction rows to belong to estimation respondents; fully posterior-integrated when the fit kept keep_beta_i = "draws").

n_draws

Number of posterior draws to integrate over (thinned evenly from the kept draws; default 200).

aggregate

If TRUE (default) return a per-alternative posterior share table (including the outside option); if FALSE return the posterior-mean probability per row of the prediction data.

...

Ignored.

Value

With aggregate = TRUE, a data.table with columns alternative, share (posterior mean), sd, lower, upper (95% equal-tailed interval); the posterior share draws are attached as attr(, "draws"). With aggregate = FALSE, a numeric vector of posterior-mean choice probabilities, one per prediction row.

Details

HMNL probabilities are closed-form logit; HMNP probabilities use the 1-D Gauss-Hermite representation of the iid-probit integral \(P(j) = \int \phi(u) \prod_{k \ne j} \Phi(V_j - V_k + u) du\).

Examples

# \donttest{
sim <- simulate_hmnl_data(N = 100, T = 3, J = 4, seed = 42)
fit <- suppressWarnings(run_hmnlogit(sim$data, "task", "alt", "choice", c("x1", "x2"),
                    person_col = "pid", alt_covariate_cols = "z1",
                    mcmc = list(R = 500, burn = 200)))
#> MCMC run time 0h:0m:0.04s
predict(fit)                       # posterior shares, estimation data
#>    alternative     share         sd     lower     upper
#>         <char>     <num>      <num>     <num>     <num>
#> 1:           1 0.1861200 0.01664956 0.1560522 0.2199947
#> 2:           2 0.1895338 0.02042708 0.1565832 0.2316484
#> 3:           3 0.2841487 0.02256055 0.2359523 0.3257916
#> 4:           4 0.1968078 0.01553763 0.1642162 0.2280392
#> 5:   (outside) 0.1433896 0.02301090 0.1042867 0.1899193
cf <- sim$data
cf$x1 <- cf$x1 + 0.5               # a counterfactual attribute change
predict(fit, newdata = cf)
#>    alternative     share         sd      lower     upper
#>         <char>     <num>      <num>      <num>     <num>
#> 1:           1 0.1935651 0.02206788 0.15326894 0.2375603
#> 2:           2 0.1947846 0.02428569 0.15318034 0.2462673
#> 3:           3 0.2949920 0.02519159 0.23801711 0.3437924
#> 4:           4 0.2049632 0.01821599 0.17148835 0.2399649
#> 5:   (outside) 0.1116951 0.04918297 0.03281954 0.2079229
# }