Heterogeneous Peer Effect Package provides two-step Generalized Method of Moments (GMM) estimators for heterogeneous peer effects in group-level treatment models developed by Pasquier, Rossi and Wang (2026) < https://crest.science/wp-content/uploads/2026/09/2026-11.pdf>. The package separates the direct effect of treatment from within-group and between-group spillover effects, using a cross-fitted, semiparametric approach that leaves the propensity score unspecified and estimates it nonparametrically. Two identification settings are implemented: one in which eligibility for treatment coincides with group identity, and one in which identity is orthogonal to eligibility, allowing peer effects to differ across subgroups (e.g. by gender). Point estimates, standard errors, and test statistics are returned for the direct effect and for each within- and between-group peer effect.
The goal of heterogeneouspeereffects is to estimate through 2 step GMM methods the peer effect within groups (theta_within) and peer effect inter groups(theta_between). This package enables to have heterogeneous peer effects.
You can install the development version of heterogenuouspeereffects like so:
devtools::install_github("ton-user-github/heterogeneouspeereffects")
TThis example simulates data consistent with the model and estimates the direct effect, the within-group peer effect, and the between-group peer effect using heter_endo_gmm():
library(heterogeneouspeereffects)
# True parameters
delta <- -3
thetaW <- 0.7
thetaB <- 0.3
beta_NE <- -2
beta_E <- -1
G <- 1000
# Individual heterogeneity (alpha)
mu_alphaX <- c(8, 4, -3, 2)
Sigma_alphaX <- matrix(c(4, 1, 0, 0,
1, 4, 0, 0,
0, 0, 4, -4.2,
0, 0, -4.2, 9), nrow = 4)
alphaX <- mvtnorm::rmvnorm(n = G, mean = mu_alphaX, sigma = Sigma_alphaX)
# Share of eligible individuals, correlated with treatment D
logistic <- function(x) 1 / (1 + exp(-x))
s <- runif(n = G)
scaled_s <- (s - mean(s)) / sd(s)
prob_D <- logistic(scaled_s)
D <- rbinom(n = G, size = 1, prob = prob_D)
# Simulate group-level outcomes
YN <- ((1 - thetaW * s) * (alphaX[,1] + beta_NE * alphaX[,3]) +
thetaB * s * (alphaX[,2] + beta_E * alphaX[,4]) +
delta * thetaB * s * D) /
(1 - thetaW + s * (1 - s) * (thetaW^2 - thetaB^2))
YE <- ((1 - thetaW * s) * (1 - s) * thetaB * (alphaX[,1] + beta_NE * alphaX[,3]) +
(1 + thetaW * (s * (1 - s) * thetaW - 1)) * (alphaX[,2] + beta_E * alphaX[,4]) +
delta * (1 + thetaW * (s * (1 - s) * thetaW - 1)) * D) /
((1 - thetaW * s) * (1 - thetaW + s * (1 - s) * (thetaW^2 - thetaB^2)))
# Estimate the model
result <- heter_endo_gmm(YE, YN, D, s)
result
The output is a data frame with the estimated direct effect (delta), the within-group peer effect (theta_within), the between-group peer effect (theta_between), and their standard errors.
For the version with identity orthogonal to eligibility (e.g. gender-based groups), see ortho_heter_endo_gmm() and the corresponding vignette: vignette("Intro_to_ortho_heter_endo_gmm", package = "heterogeneouspeereffects").