Definition:Simpson's paradox
🔄 Simpson's paradox is a statistical phenomenon in which a trend or relationship that appears in several subgroups of data reverses or disappears when the subgroups are combined — a pitfall that can lead insurers, actuaries, and underwriters to draw exactly the wrong conclusion from aggregate portfolio data. In the insurance industry, where data is routinely segmented by line of business, territory, risk class, and policy year, the paradox arises more often than many practitioners realize. A property and casualty carrier might observe that its overall loss ratio worsened year over year, yet find that loss ratios improved within every individual rating territory — because the mix of business shifted toward higher-loss territories, dragging the aggregate figure in the opposite direction of the underlying trend.
⚙️ The paradox occurs whenever a lurking confounding variable — one that is associated with both the grouping structure and the outcome — changes its distribution across the groups being compared. In health insurance, for example, a new claims-management vendor might appear to increase average claim costs across the combined book, even though costs fell for both high-severity and low-severity segments, simply because the vendor was disproportionately assigned the high-severity segment. If analysts or executives rely on the aggregate comparison without stratifying by severity mix, they may wrongly terminate an effective program. Similarly, reinsurers benchmarking cedants' performance must guard against the paradox when comparing combined ratios across portfolios with different compositions of commercial and personal lines business. The remedy lies in thoughtful stratification, proper regression adjustment, or causal-inference techniques that account for confounders rather than ignoring them.
💡 Awareness of Simpson's paradox is especially critical as the insurance industry adopts more data-driven decision-making tools. Automated dashboards and summary statistics that aggregate across heterogeneous sub-portfolios can mask genuine improvements — or conceal genuine deterioration — if the underlying composition is shifting. Regulators evaluating whether a pricing change improved consumer outcomes, or whether a fraud intervention reduced costs, can reach erroneous conclusions without disaggregated analysis. For insurtech companies presenting performance metrics to carrier partners, failure to address the paradox can undermine credibility. The lesson is deceptively simple yet operationally demanding: never trust an aggregate trend until you have examined whether it holds within the meaningful subgroups that compose it, and always consider whether selection effects or shifting business mix could be inverting the story the data appears to tell.
Related concepts: