PosiR provides tools for post-selection inference (PoSI) in linear regression models. Post-Selection Inference addresses the challenge of performing valid statistical inference after model selection, ensuring that confidence intervals maintain their nominal coverage probability (e.g., 95%) even when the model is chosen based on the data. The package implements simultaneous confidence intervals using bootstrap-based max-t statistics, following Algorithm 1 from Kuchibhotla, Kolassa, and Kuffner (2022).
You can install the development version of PosiR
from
GitHub:
# Install devtools if not already installed
if (!requireNamespace("devtools", quietly = TRUE)) {
install.packages("devtools")
}# Install PosiR
::install()
devtools
# Optional dependencies for vignette and examples
install.packages(c("dplyr", "pbapply"))
This example demonstrates how to use simultaneous_ci()
to compute simultaneous confidence intervals for regression coefficients
across a set of models:
library(PosiR)
# Simulate data
set.seed(123)
<- matrix(rnorm(100 * 3), 100, 3)
X colnames(X) <- c("X1", "X2", "X3")
<- 1 + X[, "X1"] * 0.5 + rnorm(100) # True intercept = 1, X1 coefficient = 0.5
y
# Define model universe (column indices of X)
<- list(
Q model1 = 1:2, # Model with X1, X2
model2 = 1:3 # Model with X1, X2, X3
)
# Compute simultaneous confidence intervals
<- simultaneous_ci(X, y, Q, B = 500, verbose = FALSE)
result
# View results
print(result$intervals)
#> model_id coefficient_name estimate lower upper psi_hat_nqj
#> 1 model1 (Intercept) 0.96831201 0.7198033 1.2168207 1.084196
#> 2 model1 X1 0.44983825 0.2037940 0.6958825 1.062799
#> 3 model2 (Intercept) 0.97292290 0.7230406 1.2228052 1.096215
#> 4 model2 X1 0.45219170 0.2012421 0.7031413 1.105600
#> 5 model2 X2 0.04485171 -0.1971332 0.2868366 1.028019
#> se_nqj
#> 1 0.1041248
#> 2 0.1030922
#> 3 0.1047003
#> 4 0.1051475
#> 5 0.1013913
# Plot the intervals
plot(result, main = "Simultaneous Confidence Intervals", las.labels = 1)
##
Interpretation
The output result$intervals
provides the coefficient
estimates and simultaneous 95% confidence intervals for each model in
Q
. For example:
The (Intercept)
and X1
intervals in
model1
should contain their true values (1 and 0.5,
respectively).
The intervals are wider than naive intervals to account for model
selection uncertainty, ensuring valid coverage across all models in
Q
.
Vignette: Run vignette(“Vignette”).
Source Paper: Kuchibhotla, A., Kolassa, J., & Kuffner, T. (2022). Post-selection inference. Annual Review of Statistics and Its Application, 9(1), 505–527. DOI: 10.1146/annurev-statistics-100421-044639.