Sandhya Indurkar

Math foundations

Conditional Probability: Given That B Happened

P(A given B) versus P(B given A)

The idea

Unconditional probability asks how common A is in the full population. Conditional probability asks how common A is among cases where B already happened. That restriction changes the denominator.

Screening is the classic trap: a positive test is not the same event as having the condition. P(disease | positive test) and P(positive test | disease) use the same joint counts but answer different questions.

Conditional probability answers: Among cases where B is true, how often is A true?

Example: P(A | B) is not P(B | A)

Conditional probability restricts you to cases where B already happened. Swapping A and B changes the denominator.

A positive test does not mean the same thing as a sick patient.

P(A)

2.0%

Has condition

P(B)

9.0%

Positive test

P(A | B)

20.0%

Given Positive test

P(B | A)

90.0%

Given Has condition

P(Has condition | Positive test) = 20.0%, but P(Positive test | Has condition) = 90.0%. They swap numerator and restrict to different groups.

The math

Definition

P(A | B) = P(A and B) ÷ P(B)

Restrict to outcomes where B happened. What share of that slice also has A? If P(B) is zero, the conditional is undefined.

Chain form

P(A and B) = P(A | B) × P(B)

Joint probability factors into a conditional times the probability of the condition. This is the bridge to dependent chains and Bayes updates.

Do not swap

P(A | B) ≠ P(B | A) in general

High test accuracy on sick patients does not automatically mean a positive test implies a high chance of disease. Base rate and false positives still matter.

A simple application

Before you staff a review queue from model alerts, ask what a flag means in your population. The conditional read is what reviewers experience, not the model's recall on known positives alone.

The habit: name the given event out loud. “Given this alert fired” is a different sample than “given this order is fraud.” The applied posts on base rates, screening, and threshold tradeoffs build on this definition.