sup_single_subsample: Supervised single subsampling method

In RobinMDunn/ConformalTwoLayer: Distribution-Free Prediction Sets for Two-Layer Hierarchical Data

Supervised single subsampling method

Description

Construct one augmented subsample containing one observation per subject and (X_k+1, Y_k+1) = (X_new, Y_new). On the augmented subsample, fit model mu.hat. Compute nonconformity scores R_i = |mu.hat(X_i) - Y_i|, i = 1, ..., k+1. The p-value is the proportion of R_i scores greater than or equal to R_k+1. At a given X_new = x, the prediction set is the set of all (x, y) with p-values >= alpha.

Usage

sup_single_subsample(
  xy_data,
  model_formula,
  alpha,
  n_val,
  k_indices,
  grid_values,
  new_xy_data
)

Arguments

xy_data	Data frame containing observations and outcomes for all subjects. Must include a Subject column that identifies subjects.
model_formula	Linear model formula for mu.hat which will be fit on augmented samples of one observation per subject plus hypothetical data on new subject
alpha	Significance level
n_val	Number of observations from each subject. For our examples, we assume this is equal across subjects.
k_indices	Labels of subjects to be treated as observed data
grid_values	Vector of starting values to start for lower and upper bounds of prediction interval. Should contain values across the range of Y.
new_xy_data	Covariate and outcome data for new subject

Value

List containing prediction interval size at new observation's covariate values, prediction interval lower bound at new observation's covariate values, prediction interval upper bound at new observation's covariate values, and whether new observation's outcome is contained inside prediction interval.

RobinMDunn/ConformalTwoLayer documentation built on March 22, 2022, 6:38 p.m.