Compute FAB (Frequentist and Bayes) Conformal Prediction Intervals.
fabPrediction
An R package for constructing frequentist prediction regions using indirect information.
References
Bersson and Hoff (2023). Frequentist Prediction Sets for Species Abundance using Indirect Information.
Bersson and Hoff (2022). Optimal Conformal Prediction for Small Areas.
Installation
install_packages("fabPrediction")
library(fabPrediction)
Usage
To load the package:
library(fabPrediction)
The two main functions are
predictionInterval, which constructs prediction intervals for a continuous response. This function can be used to construct nonparametric FAB or distance-to-average conformal intervals, or parametric normal or Bayesian intervals.predictionSet, which constructs prediction sets for a categorical counts response. This function can be used to construct nonparametric FAB or direct sets, or a parametric Bayesian set.
Construction of basic FAB prediction regions are demonstrated below. Please see the vignette for full package capabilities, including empirical Bayes procedures to obtain estimates of prior hyperparameters based on auxiliary data.
Continuous Response
We wlil demonstrate usage on a random normal sample of length 10.
y = rnorm(10)
A FAB prediction interval with 1-alpha coverage can be constructed for these data based on a prior parameters mu and tau2 from a Normal-Normal working model:
y_PI = predictionInterval(y, method = "FAB",
alpha = .05,
mu = 0, tau2 = 1/2)
and plotted:
plot(y_PI)
Categorical Response
We wlil demonstrate usage on a random multinomial sample for 10 categories based on a heterogeneous prior concentration gamma.
gamma = c(10:1)
y = c(rmultinom(1,20,rdirichlet(gamma)))
A FAB prediction set with 1-alpha coverage can be constructed for these data based on an estimate of the prior parameter gamma from a Multinomial-Dirichlet working model:
y_PS = predictionSet(y, method = "FAB",
alpha = .15,
gamma = gamma)
And this prediction set can be plotted:
plot(y_PS)