Get Started with rbbnp

Overview

The rbbnp package implements a novel bias-bound approach to nonparametric inference developed by Schennach (2020). The key insight is that while we cannot consistently estimate pointwise bias, we can estimate an upper bound on bias magnitude, enabling valid confidence intervals with optimal bandwidths.

The Problem

Traditional nonparametric inference faces a dilemma:

Optimal bandwidths minimize mean squared error but introduce non-negligible bias
Undersmoothing reduces bias but produces inefficient, wider confidence intervals

The Solution

The bias-bound approach constructs confidence intervals that explicitly account for potential bias:

$CI = [\hat{f} - \bar{b} - z_{\alpha/2}\hat{\sigma}, \quad \hat{f} + \bar{b} + z_{\alpha/2}\hat{\sigma}]$

where $\bar{b}$ is the estimated bias bound.

Installation

# Install from CRAN
install.packages("rbbnp")

# Or install development version from GitHub
# install.packages("devtools")
devtools::install_github("xinyu-daidai/rbbnp-dev")

Quick Start

Density Estimation

# Generate sample data
X <- gen_sample_data(size = 500, dgp = "2_fold_uniform", seed = 123456)

# Estimate density with bias-aware confidence intervals
fit <- biasBound_density(X, h = 0.1, kernel.fun = "Schennach2004")

# View summary
fit
#> Bias-Bounded Density Estimation
#> 
#> Call:
#> biasBound_density(X = X, h = 0.1, kernel.fun = "Schennach2004")
#> 
#> Sample size: n = 500
#> Bandwidth:   h = 0.1000 (user-specified) 
#> Kernel:      Schennach2004 
#> 
#> Bias bound parameters:
#>   A = 3.3985, r = 1.8166
#>   bias bound b1x = 0.2063
#> 
#> Evaluation points: 100 (range: [-0.1639, 2.0517])
#> Confidence level: 95%
#> 
#> Use summary() for detailed statistics
#> Use plot() to visualize results

# Visualize results
plot(fit)

The plot shows:

Blue line: Estimated density $\hat{f}(x)$
Orange band: Bias range where true $E[\hat{f}]$ lies
Green band: 95% confidence interval

Conditional Expectation (Regression)

# Generate regression data: Y = -X^2 + 3X + noise
Y <- -X^2 + 3*X + rnorm(500) * X

# Estimate conditional expectation
fit_reg <- biasBound_condExpectation(Y, X, h = 0.1, kernel.fun = "Schennach2004")

# Visualize
plot(fit_reg)

Working with Results

Both functions return S3 objects with standard methods:

# Extract parameters
coef(fit)
#>        A        r        h 
#> 3.398493 1.816558 0.100000

# Get confidence intervals
head(confint(fit))
#>      lower     upper
#> [1,]     0 0.2062650
#> [2,]     0 0.2062650
#> [3,]     0 0.2062650
#> [4,]     0 0.2062650
#> [5,]     0 0.2169229
#> [6,]     0 0.2309202

# For regression: get fitted values
head(fitted(fit_reg))
#> [1] 0.0420927 0.1386661 0.2211023 0.2938845 0.3597238 0.4202832

Next Steps

Density Estimation: Detailed guide to density estimation
Regression: Conditional expectation estimation
Theory: Mathematical background

References

Schennach, S. M. (2020). A Bias Bound Approach to Non-parametric Inference. The Review of Economic Studies, 87(5), 2439-2472. doi:10.1093/restud/rdz065

Xinyu Dai and Susanne M. Schennach

2026-02-14