Likelihood Ratios and Belief Updates

Graduate Depth 73 in the knowledge graph I know this Set as goal
Unlocks 19 downstream topics
bayesian likelihood-ratio belief-updating quantitative-reasoning

Core Idea

The likelihood ratio is the engine of Bayesian updating: it measures how much more (or less) likely a piece of evidence is under one hypothesis compared to another. A likelihood ratio of 10 means the evidence is 10 times more likely if the hypothesis is true than if it is false — a strong update. A ratio near 1 means the evidence is equally expected either way — no update warranted. Thinking in likelihood ratios rather than raw probabilities makes Bayesian reasoning more intuitive: instead of juggling joint probabilities, you ask "how much more expected is this evidence under my hypothesis?" and shift your confidence proportionally. In log-odds form, updates become simple addition: log-odds posterior = log-odds prior + log likelihood ratio.

How It's Best Learned

Practice with the classic medical diagnosis example: a test with 99% sensitivity and 5% false positive rate gives a likelihood ratio of 99/5 ≈ 20. For a disease with 1% base rate (prior odds 1:99), a positive test gives posterior odds of 20:99, or about 17% — far from certainty despite a 99% accurate test. Work problems in both probability and log-odds form to build fluency with both representations.

Common Misconceptions

Explainer

From Bayesian thinking in practice, you know the habit of updating beliefs on evidence. From Bayes' theorem, you know the formula: P(H|E) = P(E|H) x P(H) / P(E). Likelihood ratios are the engine that makes this machinery intuitive rather than formulaic. Once you understand them, Bayesian updating becomes something you can do in your head, in real time, for everyday reasoning.

The likelihood ratio for a piece of evidence is simply: how much more likely is this evidence if my hypothesis is true than if it is false? Formally, LR = P(E|H) / P(E|not-H). A likelihood ratio of 10 means the evidence is 10 times more expected under the hypothesis than under the alternative -- a strong update toward H. A ratio of 1 means the evidence is equally likely either way -- no update at all. A ratio of 0.1 means the evidence is 10 times more expected if the hypothesis is false -- a strong update against H. The beauty of this formulation is that it separates the evidence's diagnostic power from your prior beliefs. The likelihood ratio tells you how much to shift; your prior tells you where you started. Together they determine where you end up.

Consider the classic medical diagnosis example. A test has 99% sensitivity (it catches 99% of true cases) and a 5% false positive rate. The likelihood ratio for a positive test is 99/5 = roughly 20. For a disease with a 1% base rate, your prior odds are 1:99. A positive test multiplies those odds by 20, giving posterior odds of 20:99, or about 17% probability of disease. This is far from the 99% that the test's "accuracy" might suggest -- and the likelihood ratio framework makes it immediately clear why. The test is good (LR of 20 is a strong update), but the prior is low (1:99 odds), and the prior has not been overcome. Without likelihood ratios, this result is counterintuitive and frequently miscalculated; with them, it is transparent.

The log-odds representation makes multi-evidence updating even more natural. In log-odds, Bayesian updating becomes simple addition: log-odds(posterior) = log-odds(prior) + log(LR1) + log(LR2) + ... . Each piece of evidence contributes an additive term. Weak evidence (log-LR near 0) barely shifts the sum; strong evidence (large log-LR) moves it substantially. Opposing pieces of evidence cancel. You can see at a glance how many weak pieces of evidence it takes to equal one strong one, and how a strong prior (large negative or positive log-odds) resists movement from moderate evidence. This additive structure is not just computational convenience -- it reflects the genuine way evidential weight accumulates, and it makes the relationship between evidence strength and belief change immediately legible.

What did you take from this?

Topics in reflective domains aren't scored by quiz answers. Read, reflect, and mark when you've thought it through.

Quiz me anyway →

Prerequisite Chain

Counting to 10Counting to 20Understanding ZeroThe Number ZeroCounting to FiveOne-to-One CorrespondenceCombining Small Groups Within 5Addition Within 10Addition Within 20Two-Digit Addition Without RegroupingTwo-Digit Addition with RegroupingAddition Within 100Repeated Addition as MultiplicationMultiplication Facts Within 100Division as Equal SharingDivision as Grouping (Measurement Division)Division: Grouping (Repeated Subtraction) ModelDivision: Fair Sharing ModelDivision as Equal SharingDivision as GroupingBasic Division FactsDivision Facts Within 100Two-Digit by One-Digit DivisionDivision with RemaindersRemainders and Quotients in DivisionDivision Word ProblemsIntroduction to Long DivisionFactors and MultiplesPrime and Composite NumbersEquivalent FractionsRelating Fractions and DecimalsDecimal Place ValueReading and Writing DecimalsComparing and Ordering DecimalsAdding and Subtracting DecimalsMultiplying DecimalsDividing DecimalsDividing FractionsMixed Number ArithmeticOrder of OperationsInteger Order of OperationsVariable ExpressionsCombining Like TermsOne-Step EquationsTwo-Step EquationsSolving Multi-Step EquationsEquations with Variables on Both SidesLiteral EquationsSlope-Intercept FormPoint-Slope FormWriting Linear EquationsParallel and Perpendicular Line SlopesGraphing Linear EquationsPiecewise FunctionsStep FunctionsComposition of FunctionsInverse FunctionsRadical Functions and GraphsRational ExponentsExponential Functions and GraphsLogarithms IntroductionBig-O Notation and Asymptotic AnalysisBreadth-First Search (BFS)Shortest Paths in Unweighted GraphsDijkstra's Shortest Path AlgorithmAlgorithm Analysis and Big-O NotationTuring MachinesTime Complexity and the Class PNondeterministic Turing MachinesNP and Polynomial-Time VerificationProbabilistic Computation and BPPBayesian EpistemologyBayesian Thinking in PracticeLikelihood Ratios and Belief Updates

Longest path: 74 steps · 467 total prerequisite topics

Prerequisites (3)

Leads To (1)