🥮 mooncakes.io

moon-causal

A practical MoonBit causal-inference toolkit with propensity modeling, weighting, matching, diagnostics, uncertainty, simulation, and reproducible benchmarks.

causal-inference
statistics
math
moon add qjgqjg/moon-causal@0.2.3
Download zip
Author
Version
0.2.3
License
Apache-2.0
Repository
Last updated
4 hours ago
Downloads
8
README

#moon-causal

moon-causal is a pure MoonBit toolkit for causal inference on observational and experimental data. It provides composable data contracts, propensity models, treatment-effect estimators, balance diagnostics, uncertainty quantification, and deterministic simulation utilities for applications such as A/B testing, policy evaluation, and medical research.

#Core capabilities

  • Data preparation: rectangular dataset validation, summaries, scaling, transformations, missing-value handling, and train/validation splits.
  • Modeling: standardized logistic propensity regression, linear outcome regression, design matrices, polynomial terms, and treatment interactions.
  • Estimation: IPW, stabilized IPW, ATT, ATC, overlap weighting, g-computation, doubly robust estimation, stratification, and nearest-neighbor matching.
  • Diagnostics: standardized mean differences, variance ratios, overlap and positivity checks, calibration, effective sample size, residual diagnostics, and quality reports.
  • Extended methods: subgroup effects, policy value, multi-arm treatments, continuous-dose response, difference-in-differences, event studies, synthetic controls, mediation, and survival curves.
  • Uncertainty and evaluation: bootstrap intervals, permutation inference, jackknife influence values, cross-validation, classification metrics, and deterministic synthetic benchmarks.

#Quick start

Requirements: MoonBit stable toolchain.

moon check --deny-warn
moon test --deny-warn
moon run cmd/main

The library has no runtime dependency outside the MoonBit toolchain. The command-line example runs a deterministic observational benchmark and prints the estimated effect, uncertainty, effective sample size, and propensity AUC.

#CLI

Run the included example from the repository root:

moon run cmd/main

Expected output for the default configuration:

moon-causal deterministic benchmark
sample_size=200
true_ate=1.75
estimated_ate=1.695501947510638
absolute_error=0.0544980524893619
standard_error=0.15562533888456542
effective_sample_size=177.61530018366125
propensity_auc=0.6822682268226823

The benchmark uses 200 observations, 4 covariates, treatment effect 1.75, and seed 20260818UL. It is intended as a reproducible regression and integration fixture, not as a claim about performance on a particular real-world dataset.

#Architecture

The root package is organized by responsibility:

  • core_types.mbt, validation.mbt, and preprocessing.mbt define data contracts and transformations.
  • linear_algebra.mbt, models.mbt, formula.mbt, and metrics.mbt provide numerical foundations and predictive models.
  • estimators.mbt, matching.mbt, causal_methods.mbt, and advanced_causal.mbt implement treatment-effect estimators.
  • balance.mbt, diagnostics.mbt, audit.mbt, and sensitivity.mbt provide diagnostics and robustness analysis.
  • heterogeneity.mbt, multi_arm.mbt, continuous.mbt, panel.mbt, survival.mbt, and mediation.mbt cover extended designs.
  • uncertainty.mbt, bootstrap_advanced.mbt, cross_validation.mbt, and simulation.mbt provide inference and reproducible evaluation.
  • cmd/main contains a runnable benchmark application.

#Benchmark and reproducibility

The simulation API is deterministic for a fixed SyntheticConfig seed. run_synthetic_benchmark reports:

  • sample size and configured ground-truth ATE;
  • estimated IPW ATE and absolute error;
  • standard error and effective sample size;
  • propensity-model AUC.

This makes the benchmark suitable for regression testing, examples, and comparing implementation changes without relying on external data.

#Tests and quality checks

The test suite covers normal paths, empty and singleton inputs, dimension mismatches, non-finite values, extreme probabilities, repeated survival times, numerical boundaries, integration paths, and deterministic simulation.

moon fmt --check
moon info
moon check --deny-warn
moon test --deny-warn
moon build --target wasm-gc
moon build --target native

#CI

GitHub Actions installs the latest MoonBit stable toolchain from the official installer and runs:

  • format and warning-as-error checks;
  • wasm-gc and native builds;
  • public-interface regeneration and clean-diff verification;
  • wasm-gc tests with coverage summary;
  • native type checks and tests.

#License

Apache License 2.0. See LICENSE.

#Project links

///|
test "README example" {
  inspect(mean([1.0, 3.0, 5.0]), content="3")
}

⚠️ Package is too large to be displayed

This package contains over 1061 items.

Source Files

Powered by MoonBit

Site sourceReport issuePackagesBuild queueSkillsStatistics

© 2026 mooncakes.io