Efficient Estimation of the Causal Effects of Stochastic Interventions
Authors: Nima Hejazi and David Benkeser
What’s txshift?
The txshift R package is designed to provide facilities for the construction of efficient estimators of the counterfactual mean of an outcome under stochastic interventions that depend on the natural value of treatment (Dı́az and van der Laan 2012; Haneuse and Rotnitzky 2013). txshiftimplements and builds upon a simplified algorithm for the targeted maximum likelihood (TML) estimator of such a causal parameter, originally proposed by Dı́az and van der Laan (2018), and makes use of analogous machinery to compute an efficient one-step estimator (Pfanzagl and Wefelmeyer 1985). txshift integrates with the sl3 package (Coyle, Hejazi, Malenica, et al. 2022) to allow for ensemble machine learning to be leveraged in the estimation procedure.
For many practical applications (e.g., vaccine efficacy trials), observed data is often subject to a two-phase sampling mechanism (i.e., through the use of a two-stage design). In such cases, efficient estimators (of both varieties) must be augmented to construct unbiased estimates of the population-level causal parameter. Rose and van der Laan (2011) first introduced an augmentation procedure that relies on introducing inverse probability of censoring (IPC) weights directly to an appropriate loss function or to the efficient influence function estimating equation. txshift extends this approach to compute IPC-weighted one-step and TML estimators of the counterfactual mean outcome under a shift stochastic treatment regime. The package is designed to implement the statistical methodology described in Hejazi et al. (2020) and extensions thereof.
Installation
For standard use, we recommend installing the package from CRAN via
install.packages("txshift")Note: If txshift is installed from CRAN, the sl3, an enhancing dependency that allows ensemble machine learning to be used for nuisance parameter estimation, won’t be included. We highly recommend additionally installing sl3 from GitHub via remotes:
remotes::install_github("tlverse/sl3@master")For the latest features, install the most recent stable version of txshift from GitHub via remotes:
remotes::install_github("nhejazi/txshift@master")To contribute, install the development version of txshift from GitHub via remotes:
remotes::install_github("nhejazi/txshift@devel")Example
To illustrate how txshift may be used to ascertain the effect of a treatment, consider the following example:
library(txshift)
#> txshift v0.3.9: Efficient Estimation of the Causal Effects of Stochastic
#> Interventions
library(sl3)
set.seed(429153)
# simulate simple data
n_obs <- 500
W <- replicate(2, rbinom(n_obs, 1, 0.5))
A <- rnorm(n_obs, mean = 2 * W, sd = 1)
Y <- rbinom(n_obs, 1, plogis(A + W + rnorm(n_obs, mean = 0, sd = 1)))
# now, let's introduce a a two-stage sampling process
C_samp <- rbinom(n_obs, 1, plogis(W + Y))
# fit the full-data TMLE (ignoring two-phase sampling)
tmle <- txshift(
W = W, A = A, Y = Y, delta = 0.5,
estimator = "tmle",
g_exp_fit_args = list(
fit_type = "sl",
sl_learners_density = Lrnr_density_hse$new(Lrnr_hal9001$new())
),
Q_fit_args = list(fit_type = "glm", glm_formula = "Y ~ .")
)
tmle
#> Counterfactual Mean of Shifted Treatment
#> Intervention: Treatment + 0.5
#> txshift Estimator: tmle
#> Estimate: 0.7688
#> Std. Error: 0.0189
#> 95% CI: [0.7296, 0.8038]
# fit a full-data one-step estimator for comparison (again, no sampling)
os <- txshift(
W = W, A = A, Y = Y, delta = 0.5,
estimator = "onestep",
g_exp_fit_args = list(
fit_type = "sl",
sl_learners_density = Lrnr_density_hse$new(Lrnr_hal9001$new())
),
Q_fit_args = list(fit_type = "glm", glm_formula = "Y ~ .")
)
os
#> Counterfactual Mean of Shifted Treatment
#> Intervention: Treatment + 0.5
#> txshift Estimator: onestep
#> Estimate: 0.7671
#> Std. Error: 0.0192
#> 95% CI: [0.7273, 0.8027]
# fit an IPCW-TMLE to account for the two-phase sampling process
tmle_ipcw <- txshift(
W = W, A = A, Y = Y, delta = 0.5, C_samp = C_samp, V = c("W", "Y"),
estimator = "tmle", max_iter = 5, eif_reg_type = "glm",
samp_fit_args = list(fit_type = "glm"),
g_exp_fit_args = list(
fit_type = "sl",
sl_learners_density = Lrnr_density_hse$new(Lrnr_hal9001$new())
),
Q_fit_args = list(fit_type = "glm", glm_formula = "Y ~ .")
)
tmle_ipcw
#> Counterfactual Mean of Shifted Treatment
#> Intervention: Treatment + 0.5
#> txshift Estimator: tmle
#> Estimate: 0.76
#> Std. Error: 0.0205
#> 95% CI: [0.7176, 0.7978]
# compare with an IPCW-agumented one-step estimator under two-phase sampling
os_ipcw <- txshift(
W = W, A = A, Y = Y, delta = 0.5, C_samp = C_samp, V = c("W", "Y"),
estimator = "onestep", eif_reg_type = "glm",
samp_fit_args = list(fit_type = "glm"),
g_exp_fit_args = list(
fit_type = "sl",
sl_learners_density = Lrnr_density_hse$new(Lrnr_hal9001$new())
),
Q_fit_args = list(fit_type = "glm", glm_formula = "Y ~ .")
)
os_ipcw
#> Counterfactual Mean of Shifted Treatment
#> Intervention: Treatment + 0.5
#> txshift Estimator: onestep
#> Estimate: 0.76
#> Std. Error: 0.0204
#> 95% CI: [0.7177, 0.7978]Issues
If you encounter any bugs or have any specific feature requests, please file an issue. Further details on filing issues are provided in our contribution guidelines.
Contributions
Contributions are very welcome. Interested contributors should consult our contribution guidelines prior to submitting a pull request.
Citation
After using the txshift R package, please cite the following:
@article{hejazi2020efficient,
author = {Hejazi, Nima S and {van der Laan}, Mark J and Janes, Holly
E and Gilbert, Peter B and Benkeser, David C},
title = {Efficient nonparametric inference on the effects of
stochastic interventions under two-phase sampling, with
applications to vaccine efficacy trials},
year = {2021},
doi = {10.1111/biom.13375},
url = {https://doi.org/10.1111/biom.13375},
journal = {Biometrics},
publisher = {Wiley Online Library}
}
@article{hejazi2020txshift-joss,
author = {Hejazi, Nima S and Benkeser, David C},
title = {{txshift}: Efficient estimation of the causal effects of
stochastic interventions in {R}},
year = {2020},
doi = {10.21105/joss.02447},
url = {https://doi.org/10.21105/joss.02447},
journal = {Journal of Open Source Software},
publisher = {The Open Journal}
}
@software{hejazi2022txshift-rpkg,
author = {Hejazi, Nima S and Benkeser, David C},
title = {{txshift}: Efficient Estimation of the Causal Effects of
Stochastic Interventions},
year = {2022},
doi = {10.5281/zenodo.4070042},
url = {https://CRAN.R-project.org/package=txshift},
note = {R package version 0.3.9}
}Related
R/
tmle3shift- An R package that is an independent implementation of the same core methodology for TML estimation as provided here but written based on thetmle3engine of thetlverseecosystem. Unliketxshift, this package does not provide tools for estimation under two-phase sampling designs.R/
medshift- An experimental R package for estimating causal mediation effects with stochastic interventions, including via inverse probability weighted and asymptotically efficient one-step estimators, as first described in Dı́az and Hejazi (2020).R/
haldensify- An R package for estimating the generalized propensity score (conditional density) nuisance parameter using the highly adaptive lasso (Coyle, Hejazi, Phillips, et al. 2022; Hejazi, Coyle, and van der Laan 2020) via an application of pooled hazard regression (Dı́az and van der Laan 2011).R/
lmtp- An R package for estimating the causal effects of longitudinal modified treatment policies, which are a generalization of the type of effect considered in this package. The LMTP framework was first introduced in Dı́az et al. (2021) and thelmtppackage is described in Williams and Dı́az (2023).
Funding
The development of this software was supported in part through grants from the National Library of Medicine (award no. T32 LM012417), the National Institute of Allergy and Infectious Diseases (award no. R01 AI074345), and the National Science Foundation (award no. DMS 2102840).
License
© 2017-2024 Nima S. Hejazi
The contents of this repository are distributed under the MIT license. See below for details:
MIT License
Copyright (c) 2017-2024 Nima S. Hejazi
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
