Observational Causal Inference and De‑Confounding Techniques for Industrial Applications

The presentation begins with a brief overview of causal inference, distinguishing causation from correlation and introducing three typical causal graph structures: fork, chain, and collider.

Fork: a variable X causes both Y and Z (e.g., a cold causing fever and runny nose).

Chain: a variable X causes Y, which in turn causes Z (e.g., rain → wet ground → slip).

Collider: two independent variables influence a common effect (e.g., parents' blood types determine a child's blood type).

Selection bias is discussed, covering bias caused by conditioning on colliders and bias from insufficient sample coverage.

The concept of confounding is illustrated with the classic smoking‑lung‑cancer example, where family smoking acts as a hidden confounder affecting both treatment (smoking) and outcome (lung cancer).

Randomized experiments are presented as the ideal way to eliminate confounding, but the article acknowledges many practical scenarios where randomization is infeasible.

The Stable Unit Treatment Value Assumption (SUTVA) is explained through three requirements: random independent treatment assignment, observable independent outcomes, and no interference of one unit’s treatment on another’s potential outcomes.

The do‑operator, introduced by Judea Pearl, is described as a tool that converts observational probabilities into interventional (causal) probabilities by removing edges that represent confounding.

Four major de‑confounding methods are then detailed:

Resampling based on limited RCT data: Train a propensity‑score model, stratify the observational data, and resample within strata to match the treatment‑control ratio of the RCT.

Feature decomposition (representation learning): Separate features into confounders and adjustment variables using orthogonal loss functions (e.g., D2VD, AutoIV) to reduce variance of causal effect estimates.

Double Machine Learning (DML): Apply Neyman orthogonalization to obtain unbiased causal effect estimates by constructing a residualized treatment variable and regressing the outcome on it.

Back‑door and front‑door adjustments: Use known causal graphs to block confounding paths (back‑door) or leverage mediators (front‑door) when the confounder distribution is unknown.

Practical applications are showcased:

In recommendation systems, the recommendation algorithm itself is a confounder; de‑confounding improves click‑through and watch‑time predictions.

Resource allocation problems are modeled with price‑demand and budget‑return curves; de‑confounding helps estimate marginal ROI across cities for optimal budget distribution.

The article concludes by emphasizing the importance of de‑confounding for reliable causal effect estimation in data‑driven decision making.