strategize: An R package for discovering optimal treatment strategies in high-dimensional data

Software for implementing optimal stochastic intervention analysis. Current implementation handles conjoint data from experiments. Future work may also include text, network, and time series data, with observational designs potentially supported.

Installation

The most recent version of strategize can be installed directly from the repository using the devtools package:

devtools::install_github("cjerzak/strategize-software/strategize")

The package can then be loaded into your R session:

library(strategize)

Package functions can also be accessed as strategize::function_name.

Python Backend Setup

The package uses JAX for automatic differentiation. Set up the Python backend with:

strategize::build_backend(conda_env = "strategize_env")

This creates a conda environment with JAX, numpy, optax, equinox, and numpyro. On Linux with NVIDIA GPUs, it auto-detects the driver version and installs appropriate CUDA wheels.

Tutorial

Below is a minimal example demonstrating how to discover an optimal set of factor‐level probabilities for a simple conjoint design. Because the package does not ship with data, we begin by simulating a small dataset.

set.seed(123)

# Example data with two factors and a binary outcome
n <- 200
W <- data.frame(
  sex   = sample(c("Male", "Female"), n, replace = TRUE),
  party = sample(c("A", "B"),       n, replace = TRUE)
)
Y <- rbinom(n, 1, 0.5)

# Original (uniform) assignment probabilities for each factor
p_list <- list(
  sex   = c(Male = 0.5, Female = 0.5),
  party = c(A = 0.5,   B = 0.5)
)

# Search for a probability distribution that maximizes the expected outcome
library(strategize)
fit <- cv_strategize(
  Y = Y,
  W = W,
  p_list = p_list,
  lambda = 0.1,
  nSGD = 100,
  adversarial = FALSE
)

# Optimized factor-level probabilities and predicted outcome
fit$PiStar_point
fit$Q_point_mEst

This script simulates a two-factor forced-choice design, fits the model, and returns PiStar_point, the recommended distribution over factor levels, along with the expected outcome Q_point_mEst under that distribution.

Key Features

• Adversarial Mode: Set adversarial = TRUE to find Nash equilibrium strategies in two-player zero-sum games (e.g., competing candidates)
• Cross-Validation: Use cv_strategize() to automatically select the regularization parameter lambda
• One-Step Estimation: Use strategize_onestep() for simultaneous outcome modeling and distribution optimization
• Helper Functions: Use create_p_list() to easily create probability lists from your data, and strategize_preset() for quick analysis with sensible defaults

Documentation

• Main Vignette - Comprehensive tutorial and methodology overview
• QuickStart vignette (vignettes/QuickStart.Rmd) - Getting started guide
• Troubleshooting vignette (vignettes/Troubleshooting.Rmd) - Common issues and solutions

License

GPL-3.

References

Jerzak, Connor T., Priyanshi Chandra, and Rishi Hazra. 2025. "Selecting Optimal Candidate Profiles in Adversarial Environments Using Conjoint Analysis and Machine Learning." arXiv preprint arXiv:2504.19043. https://arxiv.org/abs/2504.19043.

@misc{jerzak2025selectingoptimalcandidateprofiles,
      title={Selecting Optimal Candidate Profiles in Adversarial Environments Using Conjoint Analysis and Machine Learning},
      author={Connor T. Jerzak and Priyanshi Chandra and Rishi Hazra},
      year={2025},
      eprint={2504.19043},
      archivePrefix={arXiv},
      primaryClass={stat.ME},
      url={https://arxiv.org/abs/2504.19043}, 
}