Definition
Stratified sampling divides a population into nonoverlapping groups and draws a probability sample within each group. An overall mean or rate combines group estimates using their population shares. Sampling equal counts from unequal groups does not make those groups equally common in the population.
Why It Matters
- A small, operationally important return category can receive too few inspections in an undifferentiated sample.
- The Commerce Intelligence OS framework needs both meaningful subgroup evidence and a correctly weighted batch estimate.
How It Works
- Freeze the audit population and define strata from known attributes, such as warehouse or disposition route. Every eligible unit must belong to exactly one stratum.
- Choose sample counts and randomly select units within each stratum. Preserve group sizes, inclusion probabilities, and missing inspection records.
- Combine stratum rates using population shares. Calculate uncertainty using the actual stratified design, including finite population effects where appropriate.
- Report subgroup findings beside the weighted total. Keep nonresponse and inspection errors visible rather than assuming sampling weights correct every defect.
Ecommerce Example
Context: Illustrative example: a fixed batch contains 900 units in route A and 100 in route B. A random audit of 50 from each route finds issue rates of 2% and 20%.
Recommended move: Weight the rates by 90% and 10% to estimate 3.8% for the batch. The pooled sample rate of 11% overrepresents route B.
Why it matters: These hypothetical values show why deliberate oversampling needs design-aware estimation; they do not report an iKawn audit result.
iKawn Framework
Partition
The iKawn framework defines return groups before selecting the sample.
Inspect
Give relevant routes explicit probability-based audit coverage.
Weight
Reconnect sample findings to the actual return population.
Act
Route subgroup issues to owners while preserving the overall context.
Concise Summary
Stratified audits improve control over group coverage. Use population weights and design-aware uncertainty when combining the findings.