unit-averaging 1.0.0

Unit averaging in Python: efficient ensemble estimation of unit-specific parameters

unit_averaging

Optimal Estimation for Heterogeneous Data

A Python package for estimating unit-specific parameters in heterogeneous data settings (panel data and meta-analysis). Implements unit averaging: an ensemble method that efficiently combines information across multiple units (e.g., countries, firms, or studies) while accounting for their individual differences.

Key Features

• Optimal weighting: Automatically computes weights that balance bias and variance
• Flexible workflow: Works with panel data, meta-analysis, and other heterogeneous datasets
• Versatility: Can be used on top of various standard estimation packages
• Customizable: Implement your own weighting schemes by subclassing base classes
• Theoretically grounded: Based on statistical theory with proven performance

Installation

Install the package with pip:

pip install unit_averaging

Documentation

• Tutorials: Step-by-step guides to get started
• API Reference: Detailed documentation for all classes/functions
• Theory: Mathematical foundations of unit averaging
• Original Paper: Complete theoretical treatment

Quick Start

The core workflow of the package in a small synthetic example:

import numpy as np
from unit_averaging import OptimalUnitAverager, InlineFocusFunction

# Example: Forecasting with a linear focus function
x_value = 1.0  # Your covariate (e.g., lagged value)
estimates = {   # Dict of unit-specific coefficient estimates
    "unit1": np.array([0.5, 0.3]),
    "unit2": np.array([0.7, 0.1])
}
covariances = { # Dict of unit-specific covariance matrices
    "unit1": np.array([[0.1, 0.0], [0.0, 0.1]]),
    "unit2": np.array([[0.1, 0.0], [0.0, 0.1]])
}

# Define focus function: e.g., μ(θ) = θ₀ + θ₁ * x
focus = InlineFocusFunction(
    focus_function=lambda coef: coef[0] + coef[1] * x_value,
    gradient=lambda coef: np.array([1, x_value])
)

# Create and fit averager
averager = OptimalUnitAverager(
    focus_function=focus,
    ind_estimates=estimates,
    ind_covar_ests=covariances
)
averager.fit(target_id="unit1")

Citation

If you use unit_averaging in your research, please cite:

Brownlees, C. T., & Morozov, V. (2024). Unit Averaging for Heterogeneous Panels:

@misc{Brownlees2024UnitAveragingHeterogeneous,
   title = {Unit Averaging for Heterogeneous Panels},
   author = {Brownlees, Christian and Morozov, Vladislav},
   year = {2024},
   month = may,
   number = {arXiv:2210.14205},
   eprint = {2210.14205},
   primaryclass = {econ},
   publisher = {arXiv},
   doi = {10.48550/arXiv.2210.14205},
   archiveprefix = {arXiv},
}

Support

For questions, issues, or contributions:

• Report bugs on our GitHub issues
• Contribute via pull requests