Conditional probability works by changing the probability question after information has arrived. Instead of asking how likely event A is across the entire original probability space, we restrict attention to the part of the world in which event B has occurred and ask how much of that remaining probability also belongs to A. For P(B)>0, P(A|B)=P(A∩B)/P(B). This simple ratio is one of probability’s most important transformations because it turns information into updated uncertainty and provides the foundation for dependence, independence, Bayes’ theorem, diagnostic reasoning, prediction, sequential models and much of modern statistics.
A medical test comes back positive.
A student has already answered the first question correctly.
Rain has started.
A machine has survived its first thousand operating hours.
In every case, the probability question has changed because the information set has changed.
The governing question: now that I know B happened, what probability should I assign to A inside the smaller world consistent with B?
Quick Read
ORIGINAL PROBABILITY SPACE → OBSERVE B → RESTRICT TO B → RENORMALISE PROBABILITY → P(A|B) → COMPARE WITH P(A) → DEPENDENCE / INDEPENDENCE → MULTIPLICATION RULE → TOTAL PROBABILITY → BAYES → DECISION
1. Conditioning Restricts the World
Suppose a class contains 30 students, 18 of whom study French and 12 of whom study Japanese. If we learn that a randomly selected student is in the Japanese group, the original 30-person reference set is no longer the relevant denominator. The probability question now lives inside those 12 students.
2. The Formula Renormalises Probability
For events A and B with P(B)>0:
P(A|B)=P(A∩B)/P(B).
The numerator keeps outcomes satisfying both A and B. The denominator rescales the entire B-world back to probability one.
3. Conditioning Is Not Intersection
P(A∩B) is the probability that both events occur in the original world. P(A|B) is the probability of A after B has become the reference world. The same intersection appears in the numerator, but the denominator changes the meaning.
4. The Multiplication Rule Reverses the Definition
Rearranging gives:
P(A∩B)=P(A|B)P(B).
Equally, P(A∩B)=P(B|A)P(A). This lets joint events be decomposed into a sequence of conditional steps.
5. Probability Trees Are Conditional Factorisations
A tree branch first assigns probability to one event, then conditional probabilities to what happens next. Multiply along a path to obtain the joint probability of that path. Add mutually exclusive paths to obtain broader events.
6. Dependence Means Conditioning Changes Probability
If P(A|B) differs from P(A), learning B changes our probability for A. The events are dependent under the probability model.
7. Independence Means Conditioning Does Not Change Probability
If A and B are independent and P(B)>0, then P(A|B)=P(A). Equivalently, P(A∩B)=P(A)P(B). Independence is therefore a precise statement that one event carries no probability information about the other within the model.
8. Conditional Probability Is Directional Even When Intersection Is Symmetric
A∩B equals B∩A. But P(A|B) generally does not equal P(B|A), because the denominators differ.
This asymmetry is the source of many diagnostic and base-rate mistakes.
9. Sensitivity Is Not Positive Predictive Value
In diagnostic testing, P(test positive|disease) describes sensitivity. A patient usually wants P(disease|test positive). These are reverse conditional probabilities and can differ dramatically when disease prevalence is low.
10. Base Rates Enter Through the Denominator
A highly accurate test can still produce many false positives when applied to a population where the condition is rare. Conditional reasoning must combine test performance with the prevalence of the underlying state.
11. Bayes’ Theorem Is Conditional Probability Turned Around
From the two multiplication identities:
P(A|B)=P(B|A)P(A)/P(B).
Bayes does not create new probability from nowhere. It reverses the conditioning direction using a prior probability and the probability of the observed evidence.
12. The Law of Total Probability Builds the Evidence Probability
If H₁,…,Hₖ partition the possible states:
P(B)=ΣP(B|Hᵢ)P(Hᵢ).
The denominator in Bayes is therefore a weighted average of how compatible the evidence is with every possible state.
13. Conditioning Can Create Dependence
Two variables can be independent marginally and dependent after conditioning on a third variable. Selection into a subgroup can create relationships that were absent in the full population.
14. Conditioning Can Remove Dependence
Ice-cream sales and drownings may be associated overall because both increase with temperature. Conditional on temperature, much of that association can disappear. Conditioning can reveal a common cause.
15. Simpson’s Paradox Is a Conditional-Probability Warning
An association can reverse when data are stratified by a third variable. Marginal and conditional probabilities answer different questions because they average over groups differently.
16. Collider Bias Shows That Not Every Condition Is Helpful
If two variables both influence a third variable, conditioning on that common effect can induce a spurious association between the causes. “Control for everything” is therefore not a safe causal strategy.
17. Selection Bias Is Often Conditioning in Disguise
A dataset containing only admitted students, only hospital patients or only successful companies is already conditioned on a selection event. Relationships inside that selected group need not match relationships in the source population.
18. Conditional Probability Drives Sequential Prediction
Weather tomorrow depends on what is known today. Language models predict the next token conditional on preceding context. Time-series models predict future observations conditional on past information. Sequential uncertainty is conditional uncertainty.
19. The Chain Rule Factorises Joint Probability
For several events or variables:
P(A₁,…,Aₙ)=P(A₁)P(A₂|A₁)P(A₃|A₁,A₂)…
This factorisation is fundamental in graphical models, Bayesian networks and generative modelling.
20. Conditional Distributions Extend Conditioning Beyond Events
Instead of asking P(A|B), we can ask for the entire distribution of Y given X=x. Conditional means, variances and quantiles are then derived from that conditional distribution.
See How Probability Distributions Work.
21. Conditional Expectation Is a Probability-Weighted Centre After Information Arrives
E(Y|X) is not merely an average inside an arbitrary subgroup. It is a random quantity indexed by information, and it is central to prediction, regression, martingales and stochastic processes.
22. The Law of Iterated Expectations Rebuilds the Marginal Mean
E[Y]=E[E(Y|X)].
Average inside information states, then average those conditional expectations over the distribution of the information itself.
23. Conditional Variance Separates Within-State Uncertainty
Var(Y|X) describes remaining variation in Y after X is known. The law of total variance decomposes total variation into expected within-X variance plus variance of the conditional mean across X.
24. Information Can Reduce Uncertainty Without Determining the Outcome
Knowing a student’s prior score may narrow the distribution of the next score without fixing it. Conditional probability lets information sharpen uncertainty rather than forcing false certainty.
25. More Information Does Not Always Improve a Decision
Information can be noisy, costly, delayed or selected. A variable can predict an outcome while being unusable at decision time. Conditional probability needs an information architecture, not just more columns.
26. Conditioning on Future Information Creates Leakage
A predictive model evaluated using variables that would only be known after the prediction time can appear extraordinarily accurate. It has conditioned on information unavailable in deployment.
27. Missingness Can Be a Conditioning Problem
Observed cases are often those satisfying a response or measurement event. If observation probability depends on the variable itself or related variables, the observed conditional distribution can differ from the target distribution.
28. Survival Analysis Conditions on Having Survived
A hazard describes an instantaneous event rate among units that have survived to a given time. The risk set changes as time passes. Survival analysis is full of conditional probability structure.
29. Reliability Uses Conditional Survival
The probability that a component survives another 100 hours can depend on the fact that it has already survived 5,000 hours. Memoryless exponential models are special precisely because elapsed survival does not change the conditional future distribution.
30. Markov Models Restrict Which Past Information Matters
A first-order Markov property says the future is conditionally independent of the more distant past given the present state. This is a modelling assumption about conditional probability, not a universal property of time.
31. Bayesian Networks Encode Conditional Independence
A directed acyclic graph factorises a joint distribution into local conditional distributions. Missing edges represent conditional-independence claims under the graph’s semantics.
32. Classification Models Conditional Class Probabilities
A calibrated classifier aims to estimate quantities such as P(Y=1|X=x). A score of 0.8 should mean that among comparable cases receiving that probability, roughly 80% experience the event under stable conditions.