lcl-choice 0.1.14

A high-performance Python package for estimating latent-class conditional logit models.

LCL: Latent-Class Conditional Logit Estimation in Python

LCL is a high-performance Python package for estimating latent-class conditional logit models.

Built for researchers and econometricians handling large discrete choice datasets, LCL employs JAX for GPU-accelerated gradient descent and Just-In-Time (JIT) compilation, alongside Polars for lightning-fast data management.

🚀 Features

• Blazing Fast Estimation: Core likelihood functions are written in pure JAX, allowing for seamless hardware acceleration (CPU/GPU/TPU) and automatic differentiation.
• Modern Data Handling: Native support for Polars DataFrames, avoiding the memory overhead and bottlenecking of traditional pandas pipelines.
• Fail-Fast Type Checking: Powered by jaxtyping and beartype, LCL strictly enforces tensor shapes and data types at runtime. If you add an unsupported dimension to your design matrix, LCL catches it immediately with a readable error—no more cryptic JAX compilation tracebacks!

📦 Installation

Although the package is imported as lcl, it is hosted on PyPI as lcl-choice.

pip install lcl-choice

Note: If you plan to run LCL on a GPU, ensure you install the correct GPU-enabled version of JAX for your system.

💡 Quickstart

Here is a minimal example of estimating a basic latent-class logit model using synthetic discrete choice data.

import polars as pl
import jax.numpy as jnp
import lcl

# 1. Load your choice data
df = pl.DataFrame({
    "chooser_id": [1, 1, 2, 2],
    "alt_id": [1, 2, 1, 2],
    "choice": [1, 0, 0, 1],
    "price": [10.5, 12.0, 9.5, 11.0],
    "quality": [4, 5, 3, 5]
})

# 2. Format the data into JAX arrays
# (Assuming a utility function where users choose between alternatives based on price and quality)
X = jnp.array(df.select(["price", "quality"]).to_numpy())
choices = jnp.array(df["choice"].to_numpy())

# 3. Initialize and fit the model
# Estimate a model with 2 distinct latent consumer classes
model = lcl.LatentClassConditionalLogit(n_classes=2)
results = model.fit(X, choices)

print(results.summary())

🗺️ Roadmap & Future Developments

LCL is under active development. Although the core estimation engine is functional, we are actively working on expanding the package's accessibility and feature set. Upcoming milestones include:

• Model Selection: How many latent classes genuinely reflect your data? We are developing a blocked cross-validation utility to let your data speak for themselves.
• Comprehensive Documentation: Detailed tutorials and mathematical appendices are in the works.
• Companion Paper: A scholarly working paper detailing the econometric framework, hardware benchmarking, and Monte Carlo simulations is currently in preparation.

Feature Requests: If there are specific constraints, optimization routines, or post-estimation tools you would like to see, please feel free to open a Feature Request on our GitHub Issues page!

🛠️ Development & Contributing

We welcome contributions! LCL uses uv for modern, isolated dependency management.

# Clone the repository
git clone https://github.com/zeyveld/latent-class-conditional-logit.git
cd latent-class-conditional-logit

# Sync the virtual environment and install dev dependencies
uv sync --all-extras --dev

# Run the test suite
uv run pytest tests/

🤝 Acknowledgments

In addition to the developers behind JAX, Polars, Beartype, and Jaxtyping, we are especially grateful to the creators of the xlogit package (Cristian Arteaga, JeeWoong Park, Prithvi Bhat Beeramoole, and Alexander Paz). Their highly efficient conditional logit logic profoundly influenced this package.

📝 Citation

If you use LCL in your research or publications, please consider citing it:

@software{lcl_2026,
  author = {Jeffries, Anna and Zeyveld, Andrew},
  title = {LCL: Latent-Class Conditional Logit Estimation in Python},
  year = {2026},
  url = {https://github.com/zeyveld/latent-class-conditional-logit}
}