Package {pstratreg}

Type: Package
Title: Estimate Bounds via Principal Stratification
Version: 1.0.2
Description: When studying the causal effect of a treatment on an outcome, post-treatment events may render the outcome undefined. For example, the effect of job training on hourly wage is ill-defined for those who would not be employed without job training. This package estimates average causal effects among those whose outcome would be defined regardless of the intervention, producing a set-identified result. For more details, see Lundberg and Cho (2025) <doi:10.48550/arXiv.2508.14770>.
License: MIT + file LICENSE
Encoding: UTF-8
Imports: dplyr, foreach, methods, purrr, stats, tidyr, tidyselect
Config/roxygen2/version: 8.1.0
NeedsCompilation: no
Packaged: 2026-10-23 23:52:02 UTC; ilundberg
Author: Ian Lundberg [aut, cre], Soonhong Cho [aut]
Maintainer: Ian Lundberg <ianlundberg@ucla.edu>
Depends: R (≥ 4.1.0)
Repository: CRAN
Date/Publication: 2026-11-07 17:20:02 UTC

pstratreg: Estimate Bounds via Principal Stratification

Description

When studying the causal effect of a treatment on an outcome, post-treatment events may render the outcome undefined. For example, the effect of job training on hourly wage is ill-defined for those who would not be employed without job training. This package estimates average causal effects among those whose outcome would be defined regardless of the intervention, producing a set-identified result. For more details, see Lundberg and Cho (2025) doi:10.48550/arXiv.2508.14770.

Author(s)

Maintainer: Ian Lundberg ianlundberg@ucla.edu

Authors:

Print

Description

Print results from principal stratification regression.

Usage

## S3 method for class 'pstratreg'
print(x, ...)

Arguments

x

An object of class pstratreg, such as the output of a call to pstratreg()

...

Other arguments to print commands

Value

No return value, called for side effects.

Estimate Principal Stratification Regression Bounds

Description

Uses principal stratification and parametric models to bound the average causal effect among those who would have a valid outcome under either treatment condition

Usage

pstratreg(
  formula_y,
  formula_s,
  formula_sq_resid = NULL,
  family_y = "gaussian",
  homoskedastic = TRUE,
  data,
  weights = NULL,
  treatment_name,
  monotonicity_positive = FALSE,
  monotonicity_negative = FALSE,
  mean_dominance_y1_positive = FALSE,
  mean_dominance_y1_negative = FALSE,
  mean_dominance_y0_positive = FALSE,
  mean_dominance_y0_negative = FALSE,
  aggregate = TRUE,
  group_vars = NULL
)

Arguments

formula_y

A model formula for the outcome

formula_s

A model formula for the binary indicator for whether the outcome exists

formula_sq_resid

A model formula for the log mean of squared residuals for a variance function regression. Should only use variables that are used in formula_y. Because the Gamma regression for squared residuals is more computationally demanding, one might want a simpler formula_sq_resid than formula_y, perhaps to allow only highly structured heteroskedasticity. Defaults to match formula_y. No outcome need be specified in this formula. Can be NULL if 'homoskedastic = TRUE' because then there is no model for the conditional variance, which is assumed to be constant.

family_y

Character family of the outcome, either "gaussian" or "binomial"

homoskedastic

A logical for whether homoskedasticity of the outcome should be assumed. If FALSE, estimates by variance function regression for log squared residuals with the same formula as in formula_y.

data

A data frame

weights

A numeric vector of weights, of length nrow(data)

treatment_name

A character string for the name of the treatment variable in data,

monotonicity_positive

A logical. Whether to assume M1 >= M0, so that treatment never causes the outcome to be undefined

monotonicity_negative

A logical. Whether to assume M1 <= M0, so that lack of treatment never causes the outcome to be undefined

mean_dominance_y1_positive

A logical. Whether to assume E(Y1 | S1 = 1, S0 = 1) >= E(Y1 | S1 = 1, S0 = 0), so that the always-survivor treated units have mean outcomes at least as great as the observed-survivor treated units

mean_dominance_y1_negative

A logical. Whether to assume E(Y1 | S1 = 1, S0 = 1) <= E(Y1 | S1 = 1, S0 = 0), so that the always-survivor treated units have mean outcomes no greater the observed-survivor treated units

mean_dominance_y0_positive

A logical. Whether to assume E(Y0 | S1 = 1, S0 = 1) >= E(Y0 | S1 = 0, S0 = 1), so that the always-survivor untreated units have mean outcomes at least as great as the observed-survivor untreated units

mean_dominance_y0_negative

A logical. Whether to assume E(Y0 | S1 = 1, S0 = 1) <= E(Y0 | S1 = 0, S0 = 1), so that the always-survivor untreated units have mean outcomes no greater the observed-survivor untreated units

aggregate

A logical. Whether to aggregate results or return one estimate per row of data

group_vars

Character vector of names of variables in data by which to group when aggregating results. Only relevant when aggregate = TRUE

Value

An object of class pstratreg, which is a list containing:

estimates_y

A data frame or tibble containing point bounds for the outcome values among those whose outcome exists regardless.

estimates_s

A data frame or tibble containing point estimates for outcome existence.

fit_s

The glm object for the model of outcome existence.

fit_y

The glm or lm object for the model of outcome values conditional on existence.

fit_sq_resid

The glm object for the model of squared outcome residuals under heteroskedasticity (if homoskedastic = FALSE).

call

A list of arguments passed to the function.

Examples

sim <- pstratreg_sim(n = 100)
result <- pstratreg(
     formula_y = formula(y ~ x + a),
     formula_s = formula(s ~ x + a),
     data = sim,
     treatment_name = "a"
)
result <- pstratreg(
     formula_y = formula(y ~ x + a),
     formula_s = formula(s ~ x + a),
     data = sim,
     treatment_name = "a",
     aggregate = TRUE,
     homoskedastic = FALSE,
     monotonicity_negative = TRUE
)

Standard errors for pstratreg

Description

Calculate standard errors by a nonparametric bootstrap, and 95

Usage

pstratreg_se(x, r = 100)

Arguments

x

An object of class pstratreg, such as the output of a call to pstratreg()

r

Number of bootstrap replicates

Value

Data frame with estimates, standard errors, and confidence intervals for many quantities of interest.

Examples

sim <- pstratreg_sim(n = 100)
result <- pstratreg(
     formula_y = formula(y ~ x + a),
     formula_s = formula(s ~ x + a),
     data = sim,
     treatment_name = "a"
)
se <- pstratreg_se(result, r = 10)

Simulate Data

Description

Simulate data to illustrate principal stratification regression bounding.

Usage

pstratreg_sim(n = 100)

Arguments

n

Integer number of cases

Value

Simulated data frame

Examples

pstratreg_sim(n = 100)

Summary

Description

Summarize results from principal stratification regression.

Usage

## S3 method for class 'pstratreg'
summary(object, ...)

Arguments

object

An object of class pstratreg, such as the output of a call to pstratreg()

...

Other arguments to summary commands

Value

No return value, called for side effects.