Bayes' Rule

Bayes' Rule is a framework for changing a belief in the light of evidence without abandoning the starting probabilities that describe the relevant population. Its central discipline is to combine a prior belief, often expressed as a base rate, with the diagnosticity of new information: evidence should raise or lower the probability of a hypothesis in proportion to how much more likely that evidence is under the hypothesis than under its alternatives. The rule therefore turns probability from an impression into a constrained form of reasoning.

Definition and Core Logic

Bayes' Rule begins with a subjective degree of belief but limits what that belief may become. If the probability of rain tomorrow is 40 percent, the probability of no rain must be 60 percent; if a candidate has a 30 percent chance of being elected and an 80 percent chance of reelection conditional on winning, the probability of two consecutive victories is 24 percent. These relationships are not optional features of intuition. They are requirements imposed by the logic of probability.

The rule is named after the eighteenth-century English minister Thomas Bayes, whose work addressed how people should revise beliefs when evidence appears. In practical terms, Bayes' Rule combines a prior probability with a likelihood ratio, the relative diagnosticity of the evidence. A description that is four times more likely for one hypothesis than for its alternatives does not automatically make that hypothesis likely; its effect depends on where the estimate began.

A numerical example makes the dependence on base rates explicit. If 3 percent of graduate students are enrolled in computer science and a description of Tom W is four times more likely for a computer-science student than for a student in another field, the posterior probability that Tom W is a computer scientist is 11 percent, not 400 percent and not a majority probability. If the base rate were 80 percent instead, the same evidence would produce a posterior probability of 94.1 percent. The evidence has the same diagnosticity in both cases, but the conclusion changes because the priors differ.

Base Rates and Diagnostic Evidence

The practical lesson is that base rates matter even when case-specific evidence is available. People often underweight statistical base rates because a vivid description appears to tell a coherent story about the individual. In the Tom W problem, the correct response is to remain close to the prior probabilities of different academic fields, slightly lowering the high probabilities associated with well-populated areas such as humanities and education or social science and social work, while slightly raising the low probabilities of rarer specialties such as library science and computer science. The small amount of evidence is not sufficiently trustworthy to displace the population distribution.

Bayesian reasoning also requires questioning the apparent diagnosticity of evidence. Associative coherence and the tendency to construct a story from whatever information is available can make a description feel more informative than it is. The discipline is therefore summarized by two linked instructions: anchor the judgment on a plausible base rate, and ask how strongly the evidence truly distinguishes the hypothesis from its alternatives. The difficulty is not that either instruction is obscure; it is that applying them deliberately feels unnatural.

The cab problem shows why a base rate cannot be treated as a single undifferentiated kind of fact. In the initial scenario, 85 percent of the city's cabs are Green and 15 percent are Blue, while a witness identifies the vehicle as Blue and has been judged correct about either color 80 percent of the time. The witness's testimony is salient and case-specific, but the large statistical prevalence of Green cabs still matters. A judgment based only on the testimony risks neglecting the population from which the cab came.

A second version changes the meaning of the base-rate information. If Green-cab drivers cause more than five times as many accidents as Blue-cab drivers, the high accident rate is not merely a fact about how many cabs exist. It changes the causal account of how this particular accident may have occurred. The resulting Green-recklessness story conflicts with the witness's Blue identification, and the two inferences approximately cancel; the Bayesian estimate for Green is about 41 percent, leaving the colors roughly equally likely. Statistical base rates are commonly underweighted when specific information is present, whereas causal base rates are more readily incorporated because they bear directly on the individual case.

Odds, Updating, and Prediction

In odds form, Bayes' Rule states that posterior odds equal prior odds multiplied by the likelihood ratio. A medical example illustrates why even frightening evidence may leave a rare hypothesis improbable. If a serious disease occurs in only 1 of every 600 cases sent for testing, and a positive result has a likelihood ratio of 25 to 1, the odds rise from 1 in 600 to 25 in 600, corresponding to a probability of about 4 percent. A positive test is important evidence, but it does not erase the disease's rarity.

The same calculation applied to Tom W starts with prior odds of .03/.97, or approximately .031, for computer science. Multiplying those odds by a likelihood ratio of 4 yields posterior odds of about .124, which convert to a posterior probability of 11 percent. Expressing the calculation in odds makes the procedure transparent: the evidence multiplies the prior rather than replacing it.

The rule also supplies a foundation for reference-class prediction. When estimating the outcome of a large project, a forecaster should first ask what happens to comparable projects before allowing the distinctive features of the current case to dominate. In one project discussion, an informal survey suggested a reference-class base rate of 40 percent failure and seven to ten years for completion. That baseline was already grim, and comparison with similar teams suggested that the project at hand was no better than the class. The inside view had extrapolated from early progress while ignoring unknown unknowns, including illnesses, divorces, bureaucratic delays, coordination crises, and long periods of little or no progress. Bayes' Rule does not demand blind obedience to the base rate; it demands that departures from it be justified by evidence whose diagnosticity is real rather than merely persuasive.

Calibration and the Limits of Intuition

Bayesian discipline is also connected to the measurement of prediction quality. In a twenty-year study, Philip Tetlock collected more than 80,000 forecasts from 284 people who earned their living commenting on political and economic trends. The experts assessed alternative futures, including persistence of the status quo, increases, and decreases in political freedom or economic growth. Their performance was worse than that of an equal-weighting strategy that assigned the same probability to each outcome, and specialists were not significantly better than nonspecialists even in regions they knew best. Greater knowledge produced only a slight improvement on average, while the most knowledgeable and famous forecasters often became more overconfident.

Overconfidence appears in financial forecasting as well. In a survey of 11,600 forecasts by chief financial officers of large corporations, the correlation between predicted and actual short-term Standard & Poor's returns was slightly less than zero. When the executives predicted a decline, the market was slightly more likely than not to rise. Their stated 80 percent confidence intervals produced surprises in 67 percent of cases, rather than the roughly 20 percent expected from properly calibrated intervals. These results show how easily a coherent story built from available information can be mistaken for reliable knowledge.

Calibration asks whether stated probabilities and intervals match observed frequencies. If a judge assigns a value X01 such that only 1 percent of true values should fall below it, and X99 such that only 1 percent should exceed it, the true values should lie between those limits in 98 percent of cases. In practice, actual values in many studies fell outside the stated extreme bounds in about 30 percent of cases. The intervals were systematically too narrow, reflecting certainty not warranted by the available knowledge. Starting from a best estimate and making insufficient adjustments upward and downward is one mechanism through which anchoring produces this overconfidence.

Relation to Judgment and Decision Theory

Bayes' Rule concerns belief revision, but its importance extends into decision theory because beliefs enter choices among uncertain outcomes. Prospect theory treats outcomes as gains and losses relative to a reference point rather than simply as changes in total wealth, and it identifies diminishing sensitivity and loss aversion as features of evaluation. These psychological facts do not replace Bayesian updating: a person must still form beliefs about probabilities before weighing possible outcomes. They do show, however, that disciplined probabilities alone cannot guarantee stable choices if the values assigned to outcomes shift with framing or reference points.

The broader normative tradition likewise distinguishes logical requirements from psychological tendencies. Modern decision theory includes dominance, under which an option at least as good in every respect and better in one should be preferred, and invariance, under which equivalent descriptions should produce the same preference. Human judgments often violate such principles, just as they neglect base rates or exaggerate diagnostic evidence. Bayes' Rule therefore functions as one part of a larger discipline of judgment: it constrains how beliefs should respond to evidence, while calibration, careful reference classes, and tests of consistency reveal when intuitive confidence has outrun what the evidence can support.