Kernel Density and Local Polynomial Regression Methods

Estimation, inference, bandwidth selection, and graphical procedures for kernel density and local polynomial regression methods, including robust bias-corrected confidence intervals as described in Calonico, Cattaneo and Farrell (2018, ). The package includes 'lprobust()' for local polynomial point estimation and robust bias-corrected inference, 'lpbwselect()' for local polynomial bandwidth selection, 'kdrobust()' for kernel density point estimation and robust bias-corrected inference, 'kdbwselect()' for kernel density bandwidth selection, and 'nprobust.plot()' for plotting results. The main methodological and numerical features are described in Calonico, Cattaneo and Farrell (2019, ).

Kernel Density and Local Polynomial Regression Methods

The package nprobust implements estimation, inference, bandwidth selection, and graphical procedures for kernel density and local polynomial regression methods, including robust bias-corrected confidence intervals.

• lprobust(): local polynomial point estimation and robust bias-corrected inference.
• lpbwselect(): data-driven bandwidth selection for local polynomial regression.
• kdrobust(): kernel density point estimation and robust bias-corrected inference.
• kdbwselect(): data-driven bandwidth selection for kernel density estimation.
• nprobust.plot(): graphical presentation of lprobust() and kdrobust() results.

See references for methodological and practical details.

Website: https://nppackages.github.io/.

Source code: https://github.com/nppackages/nprobust.

Authors

Sebastian Calonico ([email protected])

Matias D. Cattaneo ([email protected])

Max H. Farrell ([email protected])

Installation

To install/update use R:

install.packages("nprobust")

Usage

library(nprobust)

# Cholesterol trial data used by the Python and Stata examples.
data <- read.csv("../nprobust_data.csv")
control <- data$t == 0

# Local polynomial regression with robust bias-corrected confidence intervals.
result <- lprobust(data$cholf[control], data$chol1[control])
summary(result)

# Data-driven bandwidth selection.
bw <- lpbwselect(data$cholf[control], data$chol1[control],
                 bwselect = "mse-dpi", neval = 7)
summary(bw)

# Kernel density estimation.
density <- kdrobust(data$chol1[control], neval = 30)
summary(density)

# Kernel density bandwidth selection.
summary(kdbwselect(data$chol1[control], bwselect = "imse-dpi"))

# Plot a local polynomial fit.
nprobust.plot(result, xlabel = "chol1", ylabel = "cholf")

• Replication: nprobust illustration, nprobust data.

Dependencies

• ggplot2

References

For overviews and introductions, see nppackages website.

Software and Implementation

• Calonico, Cattaneo and Farrell (2019): nprobust: Nonparametric Kernel-Based Estimation and Robust Bias-Corrected Inference.
  Journal of Statistical Software 91(8): 1-33.

Technical and Methodological

• Calonico, Cattaneo and Farrell (2018): On the Effect of Bias Estimation on Coverage Accuracy in Nonparametric Inference.
  Journal of the American Statistical Association 113(522): 767-779.

• Calonico, Cattaneo and Farrell (2022): Coverage Error Optimal Confidence Intervals for Local Polynomial Regression.
  Bernoulli 28(4): 2998-3022.