fairlex: Leximin Calibration for Survey Weights

fairlex calibrates survey weights to population targets. When the targets cannot all be hit, because sources disagree or weights are capped, it shares the misses out as evenly as possible and tells you which targets are the problem.

Features

• Leximin calibration: the largest relative miss across your targets is as small as the weight bounds allow, then the next largest, and so on

• Weights move as little and as evenly as possible among those with the best misses

• A report naming the worst-missed targets and why they miss

• Per-variable priorities (importance=), targets as totals or shares

• Replicate weights for standard errors (calibrate_replicates)

Installation

pip install fairlex

Quick Start

import pandas as pd
from fairlex import calibrate

df = pd.DataFrame({
    "sex": ["f", "m", "f", "m", "f"],
    "age": ["young", "young", "old", "old", "young"],
})
targets = {"sex": {"f": 5, "m": 5}, "age": {"young": 7, "old": 3}}

report = calibrate(df, targets, total=10, bounds=(0.5, 2.0))
print(report)          # misses, weight changes, design effect
report.weights         # calibrated weights, indexed like df

Contents

• Why leximin
  • Setup
  • Misses bound the bias of every estimate
  • What importance means
  • Other properties
  • What this does not cover
• Standard errors
  • Match the bootstrap to the design
  • Evidence
• Examples
  • Basic Example
• API Reference
  • calibrate()
  • calibrate_replicates()
  • CalibrationReport

Indices and tables

• Index

• Module Index

• Search Page