A screening tool might identify many records correctly in a test set, yet the share of its positive flags that are correct can differ in another population. The frequency of the condition being screened matters as well as the tool's true-positive and false-positive rates. A flag is a reason to check evidence, not a final eligibility decision.

NIST defines positive predictive value using correct positives divided by all positive results. Its examples concern detection and medical testing. This article uses the mathematical definition in an invented administrative exercise, not NIST health data, a real admission model or a medical test.

Hold the tool rates fixed

Imagine two fictional groups of 1,000 records. In both, the tool flags 80% of records that truly meet a narrowly defined condition and also flags 10% of records that do not meet it. Assume the rates are known exactly for each whole group and the true status has been verified independently. These assumptions are deliberately stronger than most real-world evidence.

Count Group A: 100 meet condition Group B: 10 meet condition
Correctly flagged 80 8
Do not meet condition 900 990
Incorrectly flagged 90 99
All flags 170 107
Correct share of flags 80 / 170 ≈ 47.1% 8 / 107 ≈ 7.5%

The same stipulated true-positive and false-positive rates produce different shares of correct flags because the underlying composition differs. In Group B, the lower number of true cases means false flags make up more of the flagged set. Those percentages describe the fictional groups; they are not probabilities for a particular student's application.

Preserve the two denominators

The 80% rate uses all records that truly meet the condition as its denominator. The 47.1% and 7.5% figures use all flagged records as their denominators. They answer different questions. Calling both 'accuracy' without defining the denominator conceals why they differ.

A missing true case also matters. Group A has 20 qualifying records not flagged; Group B has two. Reviewing only flagged records cannot establish how many qualifying records were missed. To evaluate both types of error, obtain an independently checked reference set that includes flagged and unflagged records.

Avoid importing the fictional rates into real decisions

In practice, the true status may be expensive to establish and a sample may be incomplete. Performance can vary by campus, record language, application cycle or document type. The assumed rates may not transfer to a different population. Record how the reference labels were created, how the sample was selected and whether the evaluation includes uncertainty.

For a financial-aid or admission workflow, a flag should prompt a check against the relevant current rules and evidence. It cannot establish an individual's eligibility or predict a decision. Avoid using a historical population average to override a person's actual documents or an institution's formal assessment.

Make the review path visible

For each flagged record, identify the evidence supporting the condition, the unresolved question and the person or process responsible for verification. For unflagged records, establish an appropriate method to detect missed cases. Protect personal information and follow the institution's access rules; the arithmetic does not justify collecting unnecessary sensitive data.

A useful report includes the size of the reviewed population, the number meeting the condition, the numbers correctly and incorrectly flagged, and the reference method. Comparing the same tool rates across two fictional base rates shows why a positive flag's reliability cannot be inferred from the true-positive rate alone.