liveBook · Manning

Posterior odds

As explained in Grokking Bayes

Posterior odds are the ratio of the posterior probabilities of two competing hypotheses or models after incorporating observed data.

Posterior odds compare two hypotheses or models after observed data updates their prior probabilities. They combine prior odds with the Bayes factor, showing how much more one explanation is supported than another. section 4.10

Interpret posterior odds by keeping these distinctions in mind:

  • Unlike a Bayes factor, posterior odds depend on subjective prior preferences for the competing models. section 4.10
  • A small probability for data under one hypothesis does not by itself show that the hypothesis is unlikely; alternative explanations and their priors also matter. section 11.7
  • Posterior probabilities, rather than an isolated data likelihood or p-value, should guide the final decision. section 11.7

A court weighs evidence

Suppose the prior probability that a randomly selected nurse is guilty is 1/100,000, while innocence has probability 0.99999. If the evidence has probability 0.5 under guilt and 0.000001 under innocence, the posterior odds of innocence to guilt are 0.2, or equivalently guilt is five times more likely than innocence. The evidence therefore does not justify treating guilt as certain. section 11.7

p(H0∣D)p(H1∣D)=p(D∣H0)p(H0)p(D∣H1)p(H1)=0.000001×0.999990.5×0.00001=0.2.

The equation makes the update explicit by multiplying each hypothesis's likelihood by its prior probability before taking their ratio. section 11.7

How Grokking Bayes covers it

How deep it goes
Explains the ratio, its dependence on priors, and a numerical court example without developing computational implementation.
Examples
Uses competing Bayesian models and a court case involving evidence about a nurse's guilt or innocence.
What it assumes
Builds from the ideas of hypotheses, models, prior probabilities, likelihoods, and observed data.
What you can do after
Evaluate competing explanations by combining their prior probabilities with how well each explains observed data.
How it connects
Connects posterior odds to Bayes factors, prior probability, likelihood, posterior probability, and hypothesis testing.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Grokking Bayes. The section links point to the parts of the book it draws on.