Causal Inference

The central challenge is that we can never observe the counterfactual—what would have happened to the same unit in the absence of the treatment. A customer either receives a discounted price or they don't; we cannot see both realities simultaneously. All causal inference methodologies are designed to construct a credible proxy for this unobserved counterfactual using control groups, temporal baselines, or statistical matching. Without a valid counterfactual, any observed lift in conversion following a price cut is merely a correlation, potentially confounded by seasonality or a concurrent marketing campaign.

A quasi-experimental technique that estimates the causal effect of a treatment by comparing the change in outcomes over time between a treatment group and a control group. The key assumption is parallel trends: in the absence of the intervention, the difference between the two groups would have remained constant.

• Application: A retailer rolls out a dynamic pricing algorithm in Region A but not Region B. DiD compares the pre-post change in revenue in Region A against the pre-post change in Region B.
• Strength: Controls for both time-invariant unobserved confounders and common temporal shocks affecting both groups.
• Weakness: Fails if the parallel trends assumption is violated, such as when a local competitor launches a promotion only in Region A during the test period.

A method that attempts to mimic randomization by pairing each treated unit with an untreated unit that has a similar probability of receiving the treatment, based on observed covariates. The propensity score is the conditional probability of assignment to treatment given a vector of observed characteristics.

• Process: First, a logistic regression model estimates the propensity score for each unit. Then, treated units are matched to untreated units with near-identical scores using algorithms like nearest-neighbor or caliper matching.
• Pricing Use Case: Matching customers who received a personalized discount with statistically identical customers who did not, based on demographics, browsing history, and past purchase frequency.
• Critical Limitation: PSM can only balance on observed covariates. Any unobserved confounder—like a customer's unmeasured price sensitivity—will still bias the estimate.

An approach used when the treatment assignment is confounded by unobserved variables. An instrument is a variable that influences the treatment but has no direct effect on the outcome, other than through the treatment. This isolates exogenous variation in the treatment.

• Classic Example: Using a randomized encouragement design where some customers are randomly nudged to visit a sale page. The nudge is the instrument; it affects exposure to the price (treatment) but does not directly cause a purchase.
• Two-Stage Least Squares (2SLS): The standard estimation method. The first stage predicts treatment using the instrument; the second stage predicts the outcome using the predicted treatment values.
• Relevance Condition: The instrument must be strongly correlated with the treatment. A weak instrument produces biased and inconsistent estimates.

A formal graphical language for encoding causal assumptions about a system. Nodes represent variables; directed edges represent direct causal relationships. DAGs are essential for determining which variables must be controlled for and which must not be controlled for to identify a causal effect.

• Backdoor Criterion: A set of rules for identifying a sufficient set of covariates to condition on to block all spurious, non-causal paths between treatment and outcome.
• Collider Bias: A critical pitfall where conditioning on a common effect of both the treatment and the outcome opens a non-causal path, inducing a spurious correlation.
• Practical Use: Before running any pricing experiment, a DAG should be drawn to map confounders like competitor actions, inventory levels, and day-of-week effects to justify the chosen identification strategy.

While standard causal inference estimates the Average Treatment Effect (ATE) across a population, uplift modeling focuses on the Conditional Average Treatment Effect (CATE)—the heterogeneous impact for specific sub-groups. The goal is to target the intervention only at the persuadables.

• Persuadables: Customers who will convert only if given a discount.
• Sure Things: Customers who will convert regardless. Discounting them is wasted margin.
• Lost Causes: Customers who will not convert even with a discount.
• Sleeping Dogs: Customers who would convert if left alone but are alienated by the intervention.
• Meta-Learners: Algorithms like T-Learners, S-Learners, and X-Learners use machine learning models to estimate CATEs from observational or experimental data, directly optimizing for incremental profit rather than just lift.

A quasi-experimental technique that estimates the causal effect of a price change by comparing the pre- and post-treatment difference in outcomes between a treatment group (exposed to the price change) and a control group (not exposed).

• Assumes parallel trends: in the absence of treatment, both groups would have followed the same trajectory
• Commonly used when A/B testing is infeasible, such as region-wide pricing rollouts
• Requires careful validation of the parallel trends assumption using pre-period data visualization

A statistical method that constructs a synthetic control group by matching treated units (e.g., customers who received a discount) with untreated units that have similar propensity scores—the estimated probability of receiving treatment based on observed covariates.

• Reduces selection bias in observational pricing data
• Matching algorithms include nearest neighbor, caliper, and kernel methods
• Requires sufficient covariate overlap (common support) between groups to produce valid estimates

An econometric technique that identifies causal effects when the treatment variable is correlated with unobserved confounders. An instrument is a variable that affects the treatment (e.g., price) but has no direct effect on the outcome (e.g., demand) except through the treatment.

• Classic example: using supply-side cost shocks as an instrument for price to estimate demand elasticity
• Must satisfy the exclusion restriction—the instrument affects outcomes only through the treatment
• Two-stage least squares (2SLS) is the standard estimation method

A design that exploits a known cutoff or threshold in the treatment assignment mechanism. Units just above and just below the threshold are assumed to be nearly identical, allowing causal estimation by comparing outcomes around the discontinuity.

• Sharp RDD: treatment is deterministically assigned at the cutoff
• Fuzzy RDD: the probability of treatment jumps at the cutoff but is not deterministic
• Applied in pricing when discounts trigger at specific inventory levels or loyalty point thresholds

A predictive modeling approach that directly estimates the incremental impact of a pricing action on an individual unit. Unlike traditional response models that predict purchase probability, uplift models predict the causal effect of the treatment.

• Segments customers into four groups: Persuadables, Sure Things, Lost Causes, and Sleeping Dogs
• Meta-learners (S-Learner, T-Learner, X-Learner) are common implementation architectures
• Critical for targeting discounts only to customers who will be incrementally persuaded to convert

The gold standard for causal inference. Units are randomly assigned to treatment and control groups, ensuring that any systematic difference in outcomes can be attributed solely to the treatment.

• Eliminates both observed and unobserved confounding through randomization
• Requires careful statistical power analysis to determine adequate sample size
• In pricing, often implemented as user-level A/B tests with holdout groups receiving the baseline price