Evaluate with stratified demographic analysis

What it is

Stratified evaluation (subgroup evaluation) means defining the demographic dimensions before evaluation begins, splitting the test set along them, and reporting metrics per stratum — including intersections (age × language, gender × device tier). It differs from "slicing later by whatever fields happen to exist" in timing and design: dimensions are chosen at design time, every stratum gets a minimum sample size, and intersections are treated as first-class strata rather than by-products. The credibility boundary of the evaluation is set by design, not by luck.

Why it happens

Only predefined strata get measured reliably. Post-hoc slicing depends on data availability, and available fields are themselves biased (language, device, and region fields are missing most for exactly the marginal users), so the slices systematically steer around the weakest populations. Intersections are doubly weak: older users and low-tier phones are each visible in univariate reports, but their overlap — "older users on low-tier phones" — gets averaged away in both, and that is precisely where voice and interface performance collapse. Pre-stratification also changes collection behavior: guaranteeing per-stratum samples forces active cell filling, which front-loads the judgment "where might this fail" into data production.

Studying it

The method is the evaluation design itself: fix the stratification dimensions (derived from the feature's mechanism — voice features by accent and age, camera features by lighting and skin tone), the minimum per-stratum sample size (power analysis against expected effect sizes), and the intersection handling (full crossing or only pre-registered key intersections). The literature reports "worst-group error" as the headline: the minimum across strata, forcing evaluation to face the weakest cell. Methodological cautions: stratification uses sensitive attributes, so evaluation data needs governance and consent; proxy strategies for missing attributes must be documented with misclassification rates; the number of dimensions is bounded by total sample — registering every intersection yields empty cells and multiple-comparison noise, so key intersections are pre-specified, not enumerated.

Where it stops holding

Stratified evaluation locates differences; it does not adjudicate how large a difference is unfair — that requires normative judgment against the decision context (automatic visa refusal and recommendation ranking tolerate different gaps). Dimensions designed for one market must be redefined for another (ethnicity categories inapplicable in most countries; language, ethnicity-as-community, and region may matter more). Feasibility is sample-bound: products with small user bases cannot support deep crossings and should fall back to "small-sample deep evaluation of the single highest-risk stratum" rather than dropping stratification.

Applying it

• Fix three fields in the evaluation plan template: dimensions and rationale, minimum per-stratum sample size, pre-registered intersections; a plan missing any field does not pass.
• Lock the report format to the full stratified table with confidence intervals per cell; mark empty cells "not evaluated" and never merge thin cells into neighbors.
• Headline the worst-stratum metric and set acceptance thresholds there; keep the aggregate as a reference line only.
• Verify: map post-launch complaints or failures to their strata — if they land in strata that were already weak in the table, the evaluation had predictive power; if they land in strata that evaluated well, a dimension was missed and the derivation must be redone.

Related

• Same group: P4.03.1 The training distribution decides who bears the failures · P4.03.2 Aggregate metrics mask subgroup differences
• Adjacent: P4.09 Algorithmic fairness and disparate impact · J1 Accessibility criteria and conformance levels
• Search terms: stratified evaluation · subgroup robustness · worst-group error