Conditional probability:
how new information changes the odds

When extra information changes the probability

Conditional probability is the chance that something happens once you know that something else already has. A coin toss gives 50 percent heads regardless of anything else, but the chance that a card is a king rises from 4 in 52 to 4 in 12 once you learn it is a face card. The condition shrinks the world you are drawing from.

That shrinking is the core idea. You stop looking at everyone and look only at the group where the condition holds, then ask what fraction of that group also has the outcome you care about.

The formula and its pieces

The denominator is the probability of the condition, not of the outcome. Once B is known to have occurred, only the part of A that overlaps with B is still possible. Dividing by P(B) rescales that overlap so the probabilities inside B add to 1.

Rearranged, the same relationship gives the multiplication rule: P(A and B) = P(B) × P(A | B). It is how you build probabilities of chains of events, like two draws without replacement.

A useful habit is to say the sentence out loud: among all the people who are B, what fraction are also A? If the answer to that phrasing matches the number you computed, the denominator is right. If you find yourself dividing by the grand total, you have found the joint probability instead.

Worked example from a two-way table

A school has 120 students. 40 play a sport and 12 of those also play a musical instrument. What is the chance a randomly chosen sports player also plays music?

The shortcut matters. When the table is given in counts, divide the count in the overlap by the count in the condition group. The 120 cancels out and you never need to turn anything into a decimal.

Counts make this kind of problem easy to follow because you can see the groups. Writing the numbers into a small two-way table, with sport against music across the rows and columns, shows 12 in the overlap, 28 who play only sport, 18 who play only music and 62 who do neither. The totals check: 12 + 28 + 18 + 62 = 120.

Order matters: P(A | B) is not P(B | A)

Reversing the condition is the most common error in everyday reasoning. Suppose 30 students play music. Then the chance that a musician also plays a sport is 12 ÷ 30 = 40%, which is not the 30% from before. The overlap is the same, yet the group it is measured against changed.

Media reports often blur this difference, for example by quoting how many criminals had a trait as though it were how many people with that trait become criminals. Always ask which group is the denominator.

Chains of events and the multiplication rule

The product form of the rule is what you use when events happen one after another and each step changes the odds of the next. Drawing two cards from a deck without putting the first back is the standard case.

A tree diagram lays the same thing out visually. Each branch carries a conditional probability, and the probability of any complete path is the product of the branches along it. Adding the paths that end in the same outcome gives the total probability of that outcome, which is exactly the denominator used in Bayes' rule.

Flipping the condition with Bayes' rule

A spam filter makes this concrete. Suppose 20% of mail is spam, 60% of spam contains the word free, and 5% of ordinary mail does too.

Checking independence

Two events are independent when learning one tells you nothing about the other. In that case P(A | B) equals P(A). If the school above had 30% of all students playing music and 30% of sport players doing so as well, the activities would be independent. Because the overall music rate is 25% (30 of 120) and the rate among athletes is 30%, they are mildly connected.

Independence is useful but rare in data. A calculator can help you compare P(A | B) with P(A) quickly; if they differ meaningfully, the events are related and any model that treats them as independent will mislead.

Real-world conditioning also hides selection effects. If you only ever see data from people who walked into a clinic, a shop or a website, the group you are conditioning on was already filtered, and the percentages you compute describe that filtered group. Ask who is missing from the table before you generalise the answer to everyone.

Pitfalls and limits

• The condition must have a probability above zero. You cannot condition on something impossible.
• Small groups give unstable percentages. 3 of 10 is also 30%, but with far more uncertainty than 300 of 1,000.
• Conditional probability describes association, not cause. Sport players playing music more often does not mean one produces the other.
• Base rates matter. A test that is right 95% of the time can still be mostly wrong on a positive result if the condition is rare.