Simpson's Paradox

The hallmark of the paradox: a statistical association (e.g., a positive correlation) observed in multiple separate groups reverses direction (becomes negative) or disappears entirely when the groups are combined into a single aggregate dataset. This is not a mathematical error but a structural feature of the data.

The reversal is driven by a lurking variable that influences both the independent and dependent variables. This confounder has an unequal distribution across the aggregated groups, creating a spurious association in the combined view. In supply chains, this is often a variable like order size or supplier lead time class.

The paradox manifests when the subgroups being aggregated have substantially different sample sizes. The larger group's internal trend is diluted, while its weight dominates the aggregate calculation, masking the true, consistent relationship found within each individual subgroup.

Simpson's Paradox starkly illustrates the danger of inferring causation from raw correlation. A naive analysis of aggregate data yields a statistically valid but causally false conclusion. Resolving it requires a causal graph (DAG) to identify the confounder and compute the true causal effect using techniques like stratification or the Backdoor Criterion.

A classic operational scenario: Supplier A has a higher on-time delivery rate than Supplier B for both small and large orders individually. However, when all orders are aggregated, Supplier B appears to have the higher overall rate. The confounder is order size: Supplier A handles mostly large, complex orders, while Supplier B handles mostly small, simple ones.

The primary defense against the paradox is data stratification. By disaggregating data by the confounding variable (e.g., analyzing delivery performance separately for each order size class), the true, consistent subgroup trends are revealed. This prevents a misleading aggregate metric from driving a flawed operational decision, such as firing a superior supplier.

An extraneous variable that influences both the treatment and the outcome, creating a spurious association. In Simpson's Paradox, a lurking confounder—such as order size or shipment urgency—is the hidden driver that reverses the trend when data is aggregated. Identifying and controlling for confounders is the primary statistical remedy for the paradox.

A graphical rule for identifying a sufficient set of covariates to condition on to block all spurious paths between a treatment and outcome. Applying the backdoor adjustment to a Directed Acyclic Graph (DAG) allows analysts to systematically determine which variables must be stratified to prevent Simpson's Paradox from distorting causal estimates.

A visual representation of causal assumptions where nodes represent variables and directed edges represent direct causal relationships. Constructing a causal DAG before analysis forces explicit articulation of the data-generating process, making it immediately apparent when a variable like 'warehouse location' acts as a collider or confounder that could trigger a reversal.

A distortion that occurs when conditioning on a variable that is a common effect of both the treatment and the outcome. Unlike confounding, collider stratification can induce a spurious association where none exists. In supply chains, filtering analysis to only 'expedited orders' can create a collider that falsely suggests a reliable supplier is underperforming.

The variation in the causal effect of an intervention across different subgroups. Simpson's Paradox is a dramatic manifestation of treatment effect heterogeneity where subgroup effects are consistent in direction but the aggregate reverses. Estimating conditional average treatment effects (CATE) for each stratum prevents misleading aggregate conclusions.

A formal framework defining causal relationships using structural equations that represent the data-generating mechanism. An SCM provides a rigorous mathematical language to decompose observed associations into causal and spurious components, offering a complete resolution to Simpson's Paradox by explicitly modeling the unobserved confounding structure.