Review questions

From a decision to interpretable evidence

Begin with the decision, alternatives and authorized decision maker. Turn the request into a question with a clear population, period, observational unit and comparison. A larger total need not imply a higher rate, and a change following an action does not establish causation. When denominator data or causal evidence are unavailable, narrow the conclusion and identify the information still needed.

Specify the metric before calculating: numerator, denominator, eligibility, timing, unit and exclusions. For an overall rate, combine numerators and denominators rather than assigning equal weight to unequal groups without justification. Differences between percentages use percentage points; relative change divides the difference by the starting value. A zero baseline cannot serve as a divisor. Correcting eligibility is not operational improvement.

Ask what a row represents before summing columns or deleting duplicates. An order, its line and its payment movement have different grains. Use the appropriate key, distinguish point-in-time balances from flows over a period, and align aggregation levels before comparing targets with actuals. Summing distinct monetary amounts does not identify distinct transactions; legitimate transactions may have equal values.

Link each possible explanation to an expected pattern and a check that could weaken it. Review data quality and reconciliation before interpretation, and label patterns discovered afterward as exploratory. Test sensitivity to unknown outcomes; the known-case rate need not equal the overall rate. Conclude with observations, unresolved questions and a next step that can reduce uncertainty, preserving the conditions required to reproduce the calculation.