psmqte

psmqte estimates treatment effects by propensity-score matching in R. It is designed for empirical workflows where users want quantile treatment effects as well as the more familiar average treatment effects from the same matching design.

The package currently supports:

• quantile treatment effects (QTE),
• quantile treatment effects on the treated (QTT),
• average treatment effects (ATE), and
• average treatment effects on the treated (ATT).

Matching is nearest-neighbor matching with replacement on an estimated or user-supplied propensity score. The default propensity score is estimated by logistic regression.

Installation

Install the development version from GitHub:

install.packages("remotes")
remotes::install_github("Li-Lu-SoB/psmqte")

Quick Start

The package includes a self-contained synthetic demo data set. It is useful for checking installation, learning the function interface, and reproducing examples without downloading external data.

library(psmqte)

dat <- psmqte_example_data()

fit <- psmqte(
  Y ~ X1 + X2 + X3 + age + size,
  treatment = "D",
  data = dat,
  taus = seq(0.1, 0.9, by = 0.1),
  M = c(1, 2, 4, 8),
  se = FALSE
)

head(summary(fit))
plot(fit, estimand = "qte")
plot(fit, estimand = "qtt")

The default plot method uses a monochrome, print-friendly style with serif text, distinct line types and point shapes for each value of M, and confidence intervals when they are available.

Set se = TRUE to compute feasible influence-function standard errors and pointwise confidence intervals:

fit_se <- psmqte(
  Y ~ X1 + X2 + X3 + age + size,
  treatment = "D",
  data = dat,
  taus = c(0.25, 0.5, 0.75),
  M = c(1, 4),
  se = TRUE
)

summary(fit_se)

Average Effects

Use psmate() to estimate ATE and ATT with the same style of formula interface:

avg_fit <- psmate(
  Y ~ X1 + X2 + X3 + age + size,
  treatment = "D",
  data = dat,
  M = c(1, 2, 4, 8)
)

summary(avg_fit)

Using Your Own Data

Your data should contain:

• one numeric outcome variable;
• one binary treatment variable coded as 0/1, TRUE/FALSE, or control/treated;
• observed covariates used to estimate the propensity score.

Example:

fit <- psmqte(
  outcome ~ x1 + x2 + x3,
  treatment = "treated",
  data = my_data,
  taus = seq(0.1, 0.9, by = 0.1),
  M = c(1, 2, 4),
  se = TRUE
)

summary(fit)

To use a different propensity-score specification from the outcome formula, provide ps_formula:

fit <- psmqte(
  outcome ~ x1 + x2,
  treatment = "treated",
  data = my_data,
  ps_formula = treated ~ x1 + x2 + x3 + x4,
  M = c(1, 2, 4)
)

To match on an externally estimated propensity score, provide ps_score:

my_data$ps <- fitted(glm(treated ~ x1 + x2 + x3, data = my_data,
                         family = binomial()))

fit <- psmqte(
  outcome ~ x1 + x2 + x3,
  treatment = "treated",
  data = my_data,
  ps_score = my_data$ps,
  M = c(1, 2, 4)
)

Diagnostics

The fitted object stores propensity-score overlap and match-reuse diagnostics:

fit$diagnostics$overlap
fit$diagnostics$reuse

For covariate balance, use psm_balance():

psm_balance(dat, treatment = "D", covariates = c("X1", "X2", "X3", "age"))

The matching matrix and reuse counts are also available directly:

match_obj <- psm_match(dat$p_true, dat$D, M = 4)
head(match_obj$match_matrix)
head(match_obj$K_all)

Example Data

psmqte_example_data() loads the package's built-in demo data:

demo <- psmqte_example_data()

The demo data are synthetic and generated specifically for this package. They include observed variables (Y, D, X1, X2, X3, age, size) and simulation-only variables (Y0, Y1, p_true) that are useful for teaching and checking examples.

Fresh simulated data can also be generated:

sim <- psmqte_example_data("simulated", n = 2000, seed = 123)

An optional real-data example is available through the causaldata package:

install.packages("causaldata")
nhefs <- psmqte_example_data("nhefs")

Main Functions

• psmqte() estimates QTE and QTT curves.
• psmate() estimates ATE and ATT.
• psm_match() constructs nearest-neighbor matches and reuse counts.
• psm_balance() computes standardized mean differences.
• m_grid_diverging() creates a diverging-M grid.
• psmqte_example_data() loads example data.
• simulate_psmqte_data() generates synthetic data.

Inference

For QTE and QTT, psmqte() can compute feasible influence-function standard errors using local same-arm nearest-neighbor estimates of conditional CDF terms. When the propensity score is estimated internally by logistic regression, the default uses a first-step score adjustment.

For QTT, the estimated-propensity-score adjustment uses a practical finite-difference approximation to the target-drift term. For ATE and ATT, psmate() reports large-sample matching standard errors.

References

Abadie, A. and Imbens, G. W. (2016). Matching on the estimated propensity score. Econometrica, 84(2), 781-807.

He, Y. and Han, F. (2024). On propensity score matching with a diverging number of matches. Biometrika, 111(4), 1421-1428.