Average Treatment Effect

ATE is defined as the expected difference between the potential outcome under treatment (Y(1)) and the potential outcome under control (Y(0)):

[ATE = E[Y(1) - Y(0)]]

• Potential Outcomes Framework: Each unit has two latent outcomes; only one is observed.
• Population Averaging: The expectation is taken over the entire population of interest.
• Linear Contrast: It is a simple difference in means, making it highly interpretable for business stakeholders.

To estimate ATE from observational data, three core assumptions must hold:

• Unconfoundedness (Ignorability): All variables that affect both treatment assignment and the outcome are observed and controlled for.
• Positivity (Overlap): Every unit has a non-zero probability of receiving either treatment or control, ensuring comparable groups.
• Stable Unit Treatment Value Assumption (SUTVA): The treatment applied to one unit does not affect the outcome of another unit, and there is only one version of the treatment.

ATE is a global average, but it can be decomposed into subgroup-specific effects:

• Average Treatment Effect on the Treated (ATT): (E[Y(1) - Y(0) | T=1]). The effect specifically for units that actually received the intervention.
• Average Treatment Effect on the Untreated (ATU): (E[Y(1) - Y(0) | T=0]). The hypothetical effect if the control group had been treated.
• Selection Bias: The difference between ATT and ATE reveals the presence of systematic selection into treatment.

Multiple statistical approaches exist to estimate ATE while adjusting for confounding:

• Outcome Regression (G-computation): Model the outcome as a function of treatment and covariates, then average predicted differences.
• Inverse Probability of Treatment Weighting (IPTW): Create a pseudo-population by weighting each unit by the inverse of its propensity score, breaking the link between confounders and treatment.
• Doubly Robust Methods: Combine outcome regression and propensity score weighting to provide two chances for correct specification; the estimator is consistent if either model is correct.

Applying ATE to disruption analysis requires caution due to systemic complexities:

• Interference Violations: A port closure (treatment) in one region affects shipping times globally, violating SUTVA.
• Dynamic Treatments: Inventory policies change continuously over time, making static binary treatment definitions insufficient.
• Heterogeneity Masking: A zero ATE may hide significant positive effects for one product category and negative effects for another, leading to misguided operational decisions.

ATE is the aggregate summary, but modern causal inference focuses on its decomposition:

• Conditional Average Treatment Effect (CATE): (E[Y(1) - Y(0) | X=x]). The treatment effect for a specific subgroup defined by features (X).
• Causal Forests: A machine learning method that recursively partitions data to discover where treatment effects are strongest.
• Uplift Modeling: Directly models the difference in response probability, targeting interventions only at persuadable units to maximize ROI.

The variation in the causal impact of an intervention across different subgroups or individuals. While ATE provides a single population average, HTE reveals that a disruption mitigation strategy might be highly effective for one supplier tier but negligible for another. Estimating HTE is critical for precision targeting of resources.

• Conditional ATE (CATE): The ATE for a specific subgroup defined by covariates.
• Uplift Modeling: A technique to directly model the incremental impact at the individual unit level.

The process of estimating what would have happened to an outcome if a specific treatment had been different. The ATE fundamentally relies on the missing counterfactual—we observe a factory's output after a disruption but must estimate what its output would have been without the disruption.

• Rubin Causal Model: Formalizes this as the difference between potential outcomes.
• Synthetic Control Method: Constructs a weighted combination of untreated units to serve as the counterfactual for a single treated unit.

An extraneous variable that influences both the treatment assignment and the outcome, creating a spurious association. Uncontrolled confounding is the primary threat to a valid ATE estimate. For example, a supplier's size might affect both their likelihood of being audited (treatment) and their on-time delivery rate (outcome).

• Backdoor Criterion: A graphical rule for identifying a sufficient set of covariates to block confounding paths.
• Propensity Score Matching: Balances treated and control groups on observed confounders.

A visual representation of causal assumptions where nodes are variables and directed edges represent direct causal relationships. DAGs are the blueprint for identifying the ATE, as they encode the data-generating process and reveal which variables must be controlled for to isolate the causal effect.

• Do-Calculus: A set of rules by Judea Pearl for transforming interventional distributions into observational ones using the DAG.
• Collider Bias: A distortion introduced by incorrectly conditioning on a common effect of treatment and outcome.

A variable that affects the treatment but has no direct effect on the outcome except through the treatment. IVs are a powerful tool for estimating the ATE when unobserved confounding is present, such as using a supplier's distance from a weather event as an instrument for shipment delays.

• Relevance Condition: The instrument must be strongly correlated with the treatment.
• Exclusion Restriction: The instrument must affect the outcome only through the treatment.

A method that uses flexible machine learning models to control for high-dimensional confounders while providing valid statistical inference for the ATE. It works by orthogonalizing the treatment and outcome with respect to controls, removing regularization bias.

• Neyman Orthogonality: The key property that ensures small errors in nuisance parameter estimation do not bias the treatment effect estimate.
• Cross-Fitting: Splits the data to avoid overfitting bias when estimating nuisance functions.