EconML

EconML is a Python library for causal inference designed to estimate heterogeneous treatment effects using a combination of machine learning and econometrics. It serves as a toolkit for calculating conditional average treatment effects to determine how specific interventions impact individuals or subgroups.

The project provides a framework for double machine learning and orthogonal machine learning to isolate causal signals from high-dimensional confounders. It includes specialized implementations for causal forests and instrumental variable learners, allowing for the recovery of causal relationships even in the presence of unobserved confounding.

The library covers a broad range of capabilities including causal model validation through refutation tests and calibration curves, the construction of personalized treatment policies, and the analysis of dynamic treatment regimes. It also supports statistical inference for uncertainty quantification and the interpretation of effect heterogeneity using tree-based models and Shapley values.

The project is primarily implemented and demonstrated through Jupyter Notebooks.

Features

Causal Inference Tools - Provides a comprehensive suite of tools for identifying and measuring cause-effect relationships in observational data.
Heterogeneous Treatment Effect Estimators - Calculates the causal effect of a treatment for individual units using algorithms optimized for unbalanced datasets.
Causal Machine Learning Integrations - Uses machine learning to remove bias from treatment estimates by isolating the causal signal from high-dimensional confounders.
Causal Forests - Implements causal forests to calculate treatment effects using double machine learning and random forest similarity metrics.
Doubly Robust Estimators - Ships doubly robust estimators that merge propensity and outcome models to ensure valid causal estimates.
Instrumental Variable Estimators - Calculates heterogeneous treatment effects using a doubly robust instrumental variable approach.
Causal Estimate Refuters - Ships causal estimate refuters to validate robustness by introducing random confounders or placebo variables.
Causal Policy Learning - Predicts the optimal recommended treatment for users using doubly robust policy learning methods.
Double Machine Learning - Removes bias from treatment effect estimates by using machine learning models to isolate the causal signal.
Average Treatment Effect Calculation - Calculates the average treatment effect as the mean difference in outcomes between treatment points across a population.
Heterogeneity Driver Identification - Trains shallow decision trees to isolate the feature cutoffs that maximize differences in treatment response.
Orthogonalization Methods - Implements orthogonal machine learning to isolate causal signals by removing the influence of confounding variables.
Treatment Effect Predictors - Calculates point estimates, confidence intervals, and covariance matrices for local parameters at specific data points.
Instrumental Variable Orthogonalization - Provides instrumental variable orthogonalization to recover causal effects when unobserved confounding is present.
Causal Forests - Implements causal forests to estimate non-parametric treatment effects using honest random forest architectures.
Causal Model Interpretability - Analyzes complex treatment effect models using tree-based interpreters to uncover the characteristics driving outcomes.
Causal Assumption Validators - Provides causal assumption validators to assess how unobserved confounding affects the reliability of estimates.
Causal Inference Frameworks - Provides a structured framework for modeling the causal impact of sequential treatments over time.
Causal Effect Estimators - Provides interfaces to estimate and then interpret the magnitudes of causal effects from observational data.
Doubly Robust Forests - Estimates effects for categorical treatments by solving local regressions based on doubly robust moments and forests.
Imperfect Compliance Estimators - Uses instrumental variables to isolate random variation in treatment assignment for non-compliant experiment subjects.
Marginal Effect Calculators - Computes the derivative of the treatment effect around a specific point to determine the immediate impact.
Population Effect Summarizers - Provides capabilities to calculate mean point estimates, standard errors, and confidence intervals for population treatment effects.
Cross-Fitting Sample Splitting - Implements cross-fitting sample splitting to ensure nuisance models do not overfit, providing valid causal estimates.
Treatment Policy Learning - Provides offline doubly robust policy learning to predict optimal recommended treatments for given features.
Confidence Interval Calculators - Calculates statistical ranges to determine the significance of predicted causal effects.
Individual Prediction Intervals - Calculates standard errors and confidence intervals for individual treatment effect predictions.
Treatment Effect Uncertainty Quantification - Computes uncertainty intervals for treatment effects using bootstrapping, OLS, or debiased lasso methods.
Dynamic Treatment Regime Analysis - Evaluates the causal impact of sequences of treatments over time using dynamic double machine learning.
Feature Contribution Analysis - Quantifies how individual features contribute to the heterogeneity of estimated causal parameters.
Federated Causal Learners - Enables treatment effect estimation across distributed datasets without sharing raw data between parties.
Federated Treatment Aggregators - Combines multiple pre-fit causal estimators to estimate heterogeneous treatment effects across distributed data sources.
Intent-to-Treat Estimators - Calculates heterogeneous treatment effects for A/B test settings using the DRIV algorithm to handle observational data.
Local Parameter Estimators - Implements local parameter estimation by solving linear moment equations to identify treatment effects in specific subpopulations.
Marginal Treatment Effect Analysis - Determines the marginal impact of a treatment on an outcome for both constant and heterogeneous effects.
Multi-Treatment Attribution - Estimates the causal impact of several discrete outreach efforts while filtering out confounding correlations.
Personalized Treatment Policies - Determines the optimal intervention for specific individuals by predicting counterfactual outcomes to maximize a target metric.
Treatment Effect Visualizations - Display the accuracy of estimated treatment effects using calibration plots, QINI curves, and TOC curves.
Treatment Policy Construction - Allows for the construction of personalized treatment policies by segmenting samples based on costs and effect signs.
Treatment Policy Optimization - Determines the best treatment assignments for individual units using doubly-robust corrections to maximize counterfactual outcomes.
Uplift Models - Validates treatment effect models using uplift curves and specific validation metrics.
Feature Expansion - Provides methods for generating polynomial or interaction features to capture non-linear heterogeneity in treatment effects.
Debiased Lasso - Provides debiased lasso regression to estimate linear coefficients with valid standard errors in high-dimensional settings.
Local Moment Trees - Solves local moment equation problems by partitioning data into regions where parameter estimates are calculated.
Model Performance Evaluators - Provides model performance evaluators to calculate R-square analogues for treatment effect models.
Nuisance Model Selection - Provides tools to optimize the choice of nuisance models to improve the accuracy of treatment effect estimation.
Honest Forests - Implements honest random forests to prevent overfitting in causal effect estimation.
Honest Random Forests - Fits subsampled honest random forest regressors using separate data samples for tree structure and leaf estimation.
Sequential Treatment Estimation - Estimates heterogeneous treatment effects for interventions offered sequentially via adaptive policies.
Causal Counterfactual Simulators - Implements modules to compute alternative outcomes by intervening on causal models and simulating hypothetical scenarios.
CATE Model Validation - Provides tools to evaluate the accuracy of conditional average treatment effect models through calibration metrics.
Temporal Causal Analysis - Estimates how treatment effects evolve over time within panel data settings with repeated observations.
Uplift Modeling - Implements uplift modeling to measure treatment effectiveness using QINI coefficients and uplift curves.
Debiased Lasso Estimators - Computes linear coefficients with L1 regularization and applies a debiasing correction to provide standard errors.
Orthogonal Forests - Estimates non-parametric treatment effects while controlling for high-dimensional confounders using random forests and cross-fitting.
Bootstrap Confidence Estimation - Computes confidence intervals for causal estimators using repeated resampling of observational data.
Prediction Calibrations - Implements prediction calibrations to compare predicted group average treatment effects against observed values.
Causal Graph Creators - Integrates causal graphs with estimators and refutation tests to validate causal inference results.
Causal Inference - Heterogeneous Treatment Effects Estimation by Microsoft.

py-why/dowhy

8,175View on GitHub

DoWhy is an open-source Python library for causal inference that structures the entire analysis into a sequential four-step framework: modeling, identification, estimation, and refutation. It treats causal assumptions as explicit, first-class citizens, represented as directed acyclic graphs that can be automatically validated against observed data. The library distinguishes itself by cleanly separating the causal identification problem from statistical estimation, allowing any compatible estimator to be used for a given target estimand. It includes automated refutation testing that validates

uber/causalml

5,875View on GitHub

CausalML is a machine learning library for causal inference, providing tools to estimate treatment effects and causal impacts using experimental and observational data. It functions as a framework for uplift modeling and the estimation of heterogeneous treatment effects to distinguish causation from correlation. The library focuses on identifying how different user segments respond to specific interventions. This includes calculating the incremental gain of target metrics to optimize marketing campaigns, targeting high-response customer segments, and personalizing user engagement through the

ObservedObserver/visual-insights

4,653View on GitHub

Visual Insights is an automated exploratory data analysis platform and causal inference tool designed to discover patterns and cause-and-effect relationships within datasets. It functions as an interactive data visualization library using a grammar-of-graphics approach to generate multi-dimensional charts and dashboards. The project distinguishes itself through a natural language interface that translates plain-text questions into data answers and visualizations via a language model. It provides a specialized framework for causal discovery and inference, allowing users to identify variable li

achael/eht-imaging

5,313View on GitHub

This project is a suite of software for radio interferometry imaging, specialized in the processing, analysis, and reconstruction of Very Long Baseline Interferometry (VLBI) observations. It provides tools for reconstructing images from interferometry data using regularized maximum likelihood methods and managing the end-to-end data processing pipeline from raw visibilities to final images. The software distinguishes itself with a dedicated interstellar scattering simulator that models thin-screen scattering effects and applies scattering kernels to radio images. It also features a radio imag

py-whyEconML

Features