Bayes factor
The ratio of the true positive rate to the false positive rate (equivalently the likelihood ratio), which acts as the multiplier by which a prior probability is converted into a posterior probability and serves as the measure of how much evidence confirms a hypothesis.
As defined in 315 | Branden Fitelson on the Logic and Use of Probability
Where it comes up · 1
Most important claims · 8 of 8Search all claims for “Bayes factor” →
- Inductive logic is no different from any other science: it provides a theory (probability calculus plus the Bayes factor and conditional probability), but applying it requires constructing models of particular arguments and contexts, which is partly an art involving statistics, measurement, idealization and approximation, is not knowable a priori, and is often controversial—yet the theory can still be true even though all its models are false.315 | Branden Fitelson on the Logic and Use of Probability
- Good experimental design should be understood as trying to maximize the confirmational power of the evidence generated—measured by the absolute value of the confirmation (Bayes factor) measure—where confirmational power is neutral between positive and negative relevance, so the goal is the largest possible multiplier or divider of prior probabilities, not maximizing the probability that the hypothesis is true.315 | Branden Fitelson on the Logic and Use of Probability
- Even a highly reliable test with a large Bayes factor will not yield a high posterior probability if the prior is extremely low—a biologically male person getting a positive pregnancy test should not conclude he is probably pregnant because his prior probability is essentially zero, demonstrating that the multiplier alone cannot determine the posterior.315 | Branden Fitelson on the Logic and Use of Probability
- Within the inductive logic tradition, adding the criterion that a confirmation measure must generalize entailment—assigning maximal value when evidence entails the hypothesis and minimal value when it refutes it—uniquely narrows the dozens of candidate measures down to the Bayes factor, the very same quantity reported on diagnostic test boxes.315 | Branden Fitelson on the Logic and Use of Probability
- Argument strength is irreducibly two-dimensional, given by an ordered pair: the conditional probability of the conclusion given the premise (which should exceed one-half) and the Bayes factor measuring the evidence's relevance/confirmation; no single number can totally order all arguments, because evidence that is irrelevant—even if the conclusion is probable given it—is not a reason to believe the conclusion.315 | Branden Fitelson on the Logic and Use of Probability
- The base rate fallacy identified by Kahneman and Tversky occurs because people confronting a reliable test for a rare disease conflate two genuinely relevant but distinct dimensions—the low probability of having the disease and the high confirmation (probability increase) provided by the positive result—and tend to answer with a number clustering around the normalized Bayes factor, deferring to confirmation rather than probability.315 | Branden Fitelson on the Logic and Use of Probability
- Diagnostic tests reliably report the true positive and false positive rates—whose ratio is the Bayes factor (likelihood ratio)—because these error rates can be determined objectively in a laboratory without knowing anything about the individual; but the probability that you actually have the disease given a positive result also requires a prior probability, which depends on facts about the specific person and cannot be known by the test maker.315 | Branden Fitelson on the Logic and Use of Probability
- There is an irony in Kahneman and Tversky's research: they criticize humans for failing to compute posterior probabilities, yet their own scientific paper, like all scientific papers and like diagnostic-test makers, only reports likelihood ratios (Bayes factors) and cannot tell readers how probable its hypothesis is, because the posterior depends on each reader's prior information.315 | Branden Fitelson on the Logic and Use of Probability