Can a posterior become the prior for another update?

Yes, when the subsequent evidence is handled with an appropriate conditional likelihood and is not double-counted.

Is a posterior probability always a single number?

It can be a probability for a particular event, while a posterior distribution describes uncertainty across a full parameter space.

Does strong evidence erase every prior difference?

It can reduce sensitivity in many well-behaved settings, but this is not an unconditional guarantee.

AI Glossary / Statistics and probability / Intermediate

Posterior probability

Definition

Posterior probability is probability conditional on specified evidence after updating a prior probability through a model. Bayes' theorem combines the prior with the likelihood of the observed evidence and normalizes the result. For a hypothesis H and evidence E with positive probability, P(H|E) = P(E|H)P(H)/P(E). The word posterior means after the evidence in this updating sequence. It does not mean that the resulting number is an unconditional fact independent of modeling assumptions or the information available.

What it means

A posterior probability answers a question such as how plausible a hypothesis is given the evidence and model. It differs from the probability of seeing the evidence if the hypothesis were true. This reversal of conditioning is central: an alert can be common when a fault exists but still have a modest probability of indicating a real fault if faults are rare and false alerts occur. The prior captures information before the current evidence, and the likelihood describes how the evidence relates to the hypotheses under consideration.

How it works

Define the competing hypotheses or parameter space and specify prior probabilities. Describe the observed evidence through likelihoods under those possibilities. Multiply each prior by its corresponding likelihood, then divide by the total across the possibilities to obtain probabilities summing to one. With continuous parameters, posterior densities require integration rather than a simple finite sum, and probabilities concern regions of parameter space. Examine how conclusions depend on prior and likelihood choices. When updating with several pieces of evidence, model their dependence rather than multiplying likelihoods as if they were independent without justification.

Three practical examples

Equipment monitoring: update the probability of a fault after a sensor alert using a model of detection and false alerts.
Quality control: revise uncertainty about a defect rate after observing inspected items.
AI classification: interpret model-based class probabilities conditional on input features, while checking whether the fitted probabilities are well calibrated for the intended population.

Think of it this way

Think of revising a set of weighted possibilities after receiving a clue. A clue matters according to how much more likely it is under one possibility than another, not simply whether it is compatible with a favorite explanation. The initial weights and the reliability of the clue both affect the updated weights.

Worked example

Consider an original artificial monitoring example. Before an alert, a fault has probability 0.10. An alert occurs with probability 0.80 when a fault exists and 0.20 when it does not. The joint probability of a fault and alert is 0.10 times 0.80 = 0.08. The overall alert probability is 0.08 + 0.90 times 0.20 = 0.26. The posterior fault probability after an alert is therefore 0.08/0.26 = 4/13, approximately 0.3077. In a hypothetical group of 1,000 cases with these expected proportions, 80 true-fault alerts and 180 no-fault alerts explain the same ratio. The likelihood ratio is 0.80/0.20 = 4, multiplying prior odds 1/9 into posterior odds 4/9. If the prior fault probability instead were 0.50 with the same alert behavior, the posterior would be 0.80. The evidence does not determine the posterior without its prior context.

Illustrative case study

In a fictional case, Northline Maintenance deploys an alert model across two equipment groups with different baseline fault frequencies. Managers initially apply the same interpretation to every alert because the sensor specification is identical. Analysts show how different priors affect the posterior probability and assess whether the alert likelihoods also differ between groups. The team documents the relevant population and validates the resulting probabilities against later outcomes before changing inspection priorities.

Common misconceptions and mistakes

Posterior probability is not a p-value. A p-value is calculated under a specified null model and is not, by itself, the posterior probability that the null is true. A high likelihood for the observed evidence does not automatically imply a high posterior without considering alternatives and priors. For a continuous parameter, the probability at an exact point is often zero even where the posterior density is large. Reusing the same evidence as if it were new can double-count information. Model-based probabilities can also be misleading when likelihoods are misspecified or the deployment population differs from the one assumed.

Business use

Businesses can use posterior probabilities to combine prior knowledge and new observations in a transparent decision process. State which event or parameter statement the probability concerns and what evidence was included. Decisions also require consequences: the action justified by a 30% fault probability depends on inspection cost and the impact of missing a fault. Record assumptions and examine sensitivity rather than presenting a model output as a uniquely correct number. When probabilities guide repeated operational decisions, assess calibration and update the model when relevant conditions change.

Check your understanding

An alert detects 80% of faults. Does that mean 80% of alerts indicate faults?

Answer: No. Those probabilities condition on different events. The latter also depends on fault prevalence and false-alert behavior.

Related terms

Learn next

Learn to convert probabilities to odds, then compare posterior probabilities with likelihoods and p-values using examples that explicitly state the conditioning events.

Sources

Frank Harrell, Statistical Glossary, posterior probability: https://hbiostat.org/glossary/ An introduction to Sequential Monte Carlo for Bayesian inference and model comparison, Bayes formula: https://pmc.ncbi.nlm.nih.gov/articles/PMC11946982/

Previous: Posterior predictive p-value Next: Posterior sampling