Do-Why Library

Separates causal assumptions from data analysis by requiring users to explicitly encode domain knowledge as a Directed Acyclic Graph (DAG). This graph defines the causal relationships between variables—including confounders, mediators, and instruments—before any estimation occurs. The model step forces practitioners to declare their beliefs transparently, making the entire analysis auditable and debatable. Supports multiple model types including Structural Causal Models (SCM) and potential outcomes frameworks.

Applies do-calculus and graphical criteria automatically to determine whether a causal effect can be estimated from observational data. The library inspects the user-provided DAG and applies rules like the backdoor criterion, front-door criterion, and instrumental variable identification to output a mathematical expression for the interventional distribution. This eliminates manual derivation errors and handles complex graphs where the identification path is non-obvious to human analysts.

Provides a unified interface to a wide range of estimation methods, allowing practitioners to switch between techniques without rewriting code. Supported methods include:

• Linear regression and instrumental variable estimators
• Propensity score matching and stratification
• Inverse Probability of Treatment Weighting (IPTW)
• Double Machine Learning for high-dimensional confounders
• EconML integration for heterogeneous treatment effects Each estimator receives the same identified estimand, ensuring consistent inputs across methods.

Systematically stress-tests causal estimates to assess their sensitivity to unobserved confounding and model misspecification. Refutation methods include:

• Random common cause: Adding a simulated confounder to see if the estimate changes
• Placebo treatment: Replacing the treatment with an independent random variable
• Data subset validation: Re-estimating on random subsets to check stability
• Bootstrap refutation: Assessing estimate variance through resampling This step provides a quantitative confidence assessment that goes beyond p-values.

Enables estimation of individual-level what-if scenarios using the three-step abduction-action-prediction process from structural causal models. Given observed outcomes, the library updates the noise distribution of exogenous variables (abduction), sets the treatment to a hypothetical value (action), and computes the resulting outcome (prediction). This supports use cases like identifying which specific shipments would have been delayed under a different routing policy.

Integrates with the broader Python causal inference and machine learning ecosystem through standardized interfaces. Key integrations include:

• EconML: Microsoft's library for heterogeneous treatment effect estimation using causal forests and meta-learners
• Scikit-learn: Compatible with any sklearn-compatible model for nuisance function estimation
• NetworkX: Graph construction and manipulation for DAG specification
• Pyro: Probabilistic programming backend for Bayesian causal models This modularity allows practitioners to leverage state-of-the-art ML models as components within a rigorous causal framework.

The formal backbone of DoWhy's modeling step. An SCM defines causal relationships using structural equations that represent the data-generating mechanism. DoWhy uses SCMs to encode domain knowledge, specifying how each variable is determined by its direct causes and an independent exogenous noise term. This allows the library to mathematically distinguish between observational and interventional distributions.

The visual and computational input to DoWhy's CausalModel. A DAG encodes causal assumptions as a graph where nodes are variables and directed edges represent direct causal effects, with no feedback loops. DoWhy uses the DAG to automatically apply graphical criteria like the backdoor criterion and front-door criterion during the identification step, determining which variables must be controlled to estimate a target causal effect.

A set of three inference rules developed by Judea Pearl that DoWhy leverages internally. Do-calculus transforms interventional probability distributions—expressions containing the do() operator—into standard observational probabilities. This is the mathematical engine that enables DoWhy's identification step to determine if a causal query can be answered from available data, even when simple covariate adjustment is insufficient.

The fourth and most granular step in DoWhy's workflow. Counterfactual analysis estimates what would have happened to a specific unit's outcome if its treatment had been different, given what actually occurred. DoWhy implements this through structural equation models to answer 'what-if' questions at the individual level—for example, attributing a specific supply chain delay to a particular supplier failure rather than to concurrent weather events.

A critical validation step unique to DoWhy's workflow. After estimating a causal effect, DoWhy applies robustness checks to stress-test assumptions:

• Placebo Treatment: Replaces the real treatment with an independent random variable to verify the estimate drops to zero
• Random Common Cause: Adds an unobserved confounder to assess sensitivity
• Data Subset Validation: Re-estimates on random subsets to check stability
• Bootstrap Refutation: Uses resampling to compute confidence intervals

Methods that DoWhy integrates when a pre-specified DAG is unavailable. These algorithms infer causal structures directly from observational data by testing conditional independencies. DoWhy supports constraint-based algorithms like the PC algorithm and score-based methods, allowing users to bootstrap a plausible causal graph from raw data before proceeding through the standard identify-estimate-refute pipeline.