Skip to lesson

17 / 18 · Reason beyond certainty

Count before you trust the percentage

Distinguish P(flag | defect) from P(defect | flag) using natural frequencies.

Builds on Support is not entailment

Go to practice ↓

A question to keep in mind

A detector finds 90% of defects. Is a flagged item 90% likely to be defective?

A conditional probability P(A | B) describes the share of A cases within the B group. Reversing A and B changes the denominator. A detector's sensitivity is about flagged items among defects; what a user wants after a flag is about defects among flagged items.

Teaching model: exactly 1,000 items, 10 defective. The detector flags 90% of the defective items and 10% of the 990 nondefective items. It flags 9 real defects and 99 nondefects, for 108 flags total. Only 9/108, about 8.33%, of flags are defects.

The base rate matters because a small false-positive rate applied to a large group can produce many flags. These stipulated counts are not measured performance of a real product. In real use the rates and prevalence are estimates that can shift between populations.

Work through an example

  1. Split the population first: 10 defective + 990 nondefective = 1,000. Apply each rate only to its own group.
  2. True flags: 10 × 0.9 = 9. False flags: 990 × 0.1 = 99. The post-flag denominator is 108, not 10 or 1,000.
  3. Bayes' rule records the same accounting: P(D | F) = P(F | D)P(D) / P(F). Counting groups makes the denominator visible.

Your turn

0 / 3
Practice 1Not checked

A new batch: 1,000 items, 100 defective, 90% sensitivity, 10% false-positive rate. Calculate the flag groups.

Your answer
100 defective900 nondefective
Read solution · does not award completion

True flags / false flags / all flags / percentage: 90 / 90 / 180 / 50%

There are 90 true flags and 90 false flags, 180 flags total. Half of the flags are defective: 50%. The detector rates stayed fixed; the base rate changed.

Practice 2Not checked

To calculate P(defect | flag), which group belongs in the denominator?

Your answer
Read solution · does not award completion

All flagged items

Always name the population after the conditioning bar before doing arithmetic.

Practice 3Not checked

A low-prevalence deployment has many false flags. What is a reasonable response?

Your answer
Read solution · does not award completion

Use a confirmation step and compare error costs and deployment-specific rates.

A probability does not decide an action by itself. Costs, benefits, and further evidence matter.

Bring it back to your own work

Find a percentage in a claim you care about. Write its numerator, denominator, and whether either group was selected before measurement.

Saved only on this browser0 / 6000