让证据推动信念

#1. Executive Summary

This chapter argues one central thesis: Bayes' theorem is not a school of thought or a statistical taste — it is a theorem, the unique logic of how a coherent mind must revise its beliefs when new evidence arrives. Everything else in the chapter follows from taking that claim seriously and then confronting the fact that human beings do not natively obey it.

The core move of Bayesian thinking is to treat a belief as a number between 0 and 1 — a degree of confidence — and to treat evidence as something that multiplies that number by a measurable factor (the likelihood ratio). This gives us, for the first time, a concrete and checkable answer to the question every thinking person faces daily: how much should this new piece of evidence move me?

Three conclusions organize the report. First, the mathematics is uncontroversial and old (Bayes & Price 1763; Laplace 1774), but its use as a general engine of inference only became practical in the 1990s with computational methods (Gelfand & Smith 1990). Second, humans are systematically non-Bayesian in identifiable ways — we update too little in some settings (conservatism; Edwards 1968) and ignore prior probabilities in others (base-rate neglect; Kahneman & Tversky 1973) — but these failures are partly artifacts of format, and dissolve substantially when information is presented as natural frequencies (Gigerenzer & Hoffrage 1995). Third, the grander claims — that the brain is a Bayesian engine (the "Bayesian brain," free-energy principle) — are best classified as plausible theory shading into unfalsifiable metaphor, and should be held far more loosely than the theorem itself.

The practical payoff is a discipline you can run in your head and in your notebook: state your prior probability, estimate how much more likely the evidence is under your hypothesis than under the alternatives, multiply, and log the result. This is the "Belief Update Protocol" of Section 10, and it connects directly to the superforecaster habit of updating frequently and in small increments established in Chapter One.

#2. Why This Topic Matters

Most cognitive tools are heuristics — useful rules of thumb with a domain of validity. Bayesian updating is different in kind. It is the normative standard itself: under a small set of reasonable assumptions, it is provably the only way to hold and revise degrees of belief without contradicting yourself.

Two independent lines of argument establish this. The Dutch book argument (Ramsey 1926; de Finetti 1937) shows that if your degrees of belief violate the probability axioms, a clever bookmaker can construct a set of bets you consider individually fair that guarantee you lose money whatever happens; coherence — immunity to a sure loss — is mathematically equivalent to obeying the probability rules. Cox's theorem (Cox 1946) approaches from another direction: if you insist that degrees of plausibility be represented by real numbers and combine consistently, the rules you are forced into are the rules of probability. Together these say: rational belief just is probabilistic belief, and rational updating just is conditioning on evidence. [This is well-established as a philosophical foundation; but see Section 11 on Cox's theorem's known gaps.]

This is why Bayes sits underneath so much of modern intellectual life. It is the logic of medical diagnosis (updating disease probability on a test result), of the scientific method (updating theory credence on data), of spam filters, of machine-learning uncertainty estimates, and of the courtroom (weighing evidence for guilt). Learning to think in its terms is not learning one more trick; it is learning the grammar that all careful evidence-handling secretly obeys.

And crucially, it gives you a number. Vague injunctions like "keep an open mind" or "consider the evidence" provide no way to check whether you did it right. Bayes replaces them with a quantity — how much did this evidence move my probability, and should it have? — that you can write down, defend, and later audit.

#3. Foundations

#3.1 The core concepts

• Conditional probability P(A|B): the probability of A given that B is known to be true. This is the atom from which everything is built.
• Prior P(H): your degree of belief in a hypothesis H before seeing the new evidence.
• Likelihood P(E|H): how probable the evidence E would be if the hypothesis H were true. (Note the direction — this is the probability of the data given the theory, not the theory given the data.)
• Posterior P(H|E): your updated degree of belief in H after seeing E. This is the output.
• Base rate: the prior probability of a condition in the relevant population (e.g., disease prevalence). The single most-forgotten number.
• Evidence / model / uncertainty: evidence is what you observe; the model is the assumed structure connecting hypotheses to observations; uncertainty is what remains, and Bayes quantifies it rather than hiding it.

#3.2 The theorem

From the definition of conditional probability, P(H,E) = P(E|H)P(H) = P(H|E)P(E). Rearranging gives Bayes' theorem:

In words: posterior = likelihood × prior, normalized. The denominator P(E) is just the total probability of the evidence across all hypotheses — a bookkeeping constant that makes the posteriors sum to 1.

#3.3 Historical development

The mathematics has a long and international lineage:

The 200-year gap between Bayes (1763) and the theorem's practical dominance is telling: the idea was correct but computationally intractable. The posterior for any realistic problem is an integral that cannot be solved by hand. The Markov Chain Monte Carlo lineage — the Metropolis algorithm (Metropolis et al. 1953), its generalization by Hastings (1970), the Gibbs sampler (Geman & Geman 1984), and the paper that made the wider statistical community realize its Bayesian power (Gelfand & Smith 1990) — is what finally let practitioners "sample from the posterior when you cannot write it down."

#4. Current Scientific Understanding

It is essential to distinguish four claims that are often conflated, because they sit at very different confidence tiers.

(a) Bayes' theorem as mathematics — well-established (in fact, certain). The theorem is a trivial consequence of the definition of conditional probability. No one disputes it. If you accept that P(H,E) = P(E|H)P(H), you have already accepted Bayes.

(b) Bayesian statistics as a methodology — well-established but genuinely contested against frequentism. Here "Bayesian" means treating unknown parameters as having probability distributions and using priors. This is a real methodological choice, and the frequentist–Bayesian debate is a live, legitimate disagreement about the foundations of inference, not a settled matter. The modern consensus, exemplified by the American Statistical Association's 2016 statement on p-values (Wasserstein & Lazar), is that naive frequentist tools (especially p < 0.05) are widely misused — which has increased interest in Bayesian and likelihood-based alternatives, without crowning a single winner.

(c) Bayesian decision theory — well-established as a normative framework. Combining a posterior with a utility function to maximize expected utility is the standard rational-choice model, and it connects Bayes directly to action (Section 6).

(d) The Bayesian brain — plausible theory, not settled fact. The claim that neural computation implements Bayesian inference is supported by suggestive behavioral evidence but faces serious challenges about falsifiability and redescription (Section 6, Section 11).

#5. Interdisciplinary Perspectives

The power of Bayes is that four disciplines converge on the same equation from different angles — and then argue productively about what it means.

The complementarity is elegant: mathematics provides the target, psychology measures the miss, philosophy justifies the target, and AI tries to hit it at scale. The deepest cross-disciplinary tension is the problem of priors (Section 11) and, in cognitive science, the worry that calling a behavior "Bayesian" merely redescribes it rather than explaining it (Jones & Love 2011; Bowers & Davis 2012).

#6. Mental Models

#6.1 The likelihood ratio as "evidence strength" (the odds form)

The single most useful reframing of Bayes for everyday thinking is the odds form. Instead of probabilities, work in odds, and the theorem becomes a simple multiplication:

This is the whole of Bayesian thinking in one line: posterior odds = prior odds × likelihood ratio. The likelihood ratio (LR) is the strength of the evidence. An LR of 10 means the evidence is ten times more expected under your hypothesis than against it — multiply your odds by 10. An LR near 1 means the evidence is worthless no matter how dramatic it feels. This model works everywhere and rarely fails; its only trap is forgetting the denominator (how likely the evidence is under the alternative).

#6.2 Base-rate-first reasoning

Always ask how common is this thing? before reacting to case-specific evidence. This is the direct antidote to base-rate neglect and the natural home of the "outside view" / reference-class forecasting habit from Chapter One.

#6.3 Prior sensitivity ("does my conclusion survive a different prior?")

Rather than agonize over the "true" prior, run the conclusion under a range of reasonable priors. If the posterior is robust, the prior didn't matter and objections to it are moot. If the conclusion flips, you've learned that you are reasoning from assumption, not evidence — a crucial thing to know.

#6.4 Expected value of information

Before gathering more evidence, ask: would any plausible result actually change my decision? If not, collecting it is theater. This is the formal bridge between belief (posterior) and action, and it governs the explore/exploit tradeoff below.

#6.5 The "what would change my mind?" test

Name, in advance, the observation that would lower your confidence — and the likelihood ratio it would carry. A belief for which no such evidence exists is not a Bayesian belief at all; it is held with probability 1, which Bayes forbids updating away from (Section 7).

#6.6 Shrinkage / regression to the mean as implicit Bayesian updating

When you temper an extreme observation toward the average — expecting a stellar quarter to be followed by a merely good one — you are doing Bayes without the algebra. The population mean is your prior; the extreme data point is the likelihood; the shrunk estimate is the posterior. This is why "regression to the mean" and "Bayesian estimation" are the same phenomenon.

#7. Common Misconceptions

"Bayesian = subjective = anything goes." The prior is subjective, but updating is not — given a prior and the evidence, the posterior is forced. And with enough data, priors wash out (see the next misconception for the limit of this). Coherence (the Dutch book) is a hard mathematical constraint, not a matter of taste.

"Priors are just guesses, so the method is arbitrary." Priors are explicit, auditable assumptions — which is a feature, not a bug. Frequentist methods also embed assumptions; Bayes just makes them visible. And prior sensitivity analysis (6.3) turns the "arbitrariness" into a checkable question.

"Bayes' theorem is only for statisticians." The odds form (6.1) is arithmetic a child can do. The theorem is a thinking discipline first and a computational tool second.

"More data always fixes a bad prior." True only if the prior assigns nonzero probability to the truth and the model is correct. A prior of exactly 0 or 1 can never be updated (the likelihood ratio is multiplied against dead weight), and a wrong model can be updated beautifully toward a confident falsehood (Section 11). Data cannot rescue a hypothesis space that excludes reality.

"The Bayesian brain proves humans are rational." This confuses a research program with a verdict. The evidence that brains approximate Bayesian inference in low-level perception is real; the leap to "therefore human judgment is rational" is contradicted by the entire heuristics-and-biases literature (Section 4d, Section 11).

#8. Real-World Applications

The direction of application is always the same: identify the prior (base rate), assess the likelihood ratio of the evidence, and update.

• Medicine. Interpreting a positive screening test requires the base rate (prevalence). Ignoring it is the classic error (Section 9.2). Bayesian methods are also increasingly used in clinical-trial design and reanalysis.
• Business / A/B testing. Bayesian A/B testing reports "probability that B beats A," which is what decision-makers actually want — but it is not magically immune to the "peeking" problem (Section 9.4).
• Investing. Base rates for how often companies with certain characteristics succeed, and awareness of regime uncertainty (the model that worked in one era may not hold), are pure Bayesian reasoning. The 2008 crisis is the cautionary tale (Section 9.3).
• Law. Evidence should be weighed as a likelihood ratio (how much more probable is this evidence under guilt than innocence?), not as a raw probability — the failure to do so is the prosecutor's fallacy (Section 9.1).
• Personal life / news. When you hear a claim, ask: what's the base rate for claims like this being true, and how much does this source's saying it move me? Extraordinary claims (low prior) require extraordinary evidence (high likelihood ratio) to reach a high posterior.
• Using AI. Knowing that a model's stated confidence may be miscalibrated (Section 12) is itself a Bayesian prior about your tool.

#9. Case Studies

#9.1 The prosecutor's fallacy: Sally Clark (inverted Bayesian reasoning)

Sally Clark, an English solicitor, was convicted in 1999 of murdering her two infant sons after both died in infancy. The paediatrician Sir Roy Meadow testified that the chance of two cot deaths (SIDS) in one affluent family was 1 in 73 million, obtained by squaring an estimated single-death rate of about 1 in 8,500.

Two Bayesian errors sank the case. First, squaring assumes the two deaths were independent, ignoring shared genetic and environmental factors — the Royal Statistical Society, in its October 2001 statement, called this "statistically invalid" and noted there were "very strong a priori reasons for supposing that the assumption will be false." Second, and more deeply, the "1 in 73 million" was presented as if it were the probability of Clark's innocence. This is the prosecutor's fallacy: confusing P(evidence | innocent) with P(innocent | evidence). The correct question is the likelihood ratio — how much more likely are two dead infants under murder than under SIDS? Since double murder is also extremely rare, the ratio is nowhere near 73 million to one. Clark's conviction was quashed in 2003; she died in 2007. The case is the definitive demonstration that ignoring base rates and inverting conditional probabilities can destroy a life.

#9.2 The medical-test problem: base-rate neglect among physicians

Casscells, Schoenberger & Grayboys (1978, NEJM 299:999–1001) asked 60 Harvard physicians, house officers, and students: if a disease has prevalence 1/1,000 and a test has a 5% false-positive rate, what is the chance a person with a positive result actually has the disease? The most common answer was 95%. The correct answer is about 2%. The intuition ignores the base rate: among 1,000 people, ~1 has the disease (and tests positive), but ~50 healthy people also test positive (5% of 999). So a positive result is a true positive only about 1 in 51 times ≈ 2%.

This is not merely a lay error; trained doctors made it. But the crucial follow-up: Gigerenzer & Hoffrage (1995, Psychological Review 102:684–704) showed the error is largely an artifact of format. Rephrase the identical problem in natural frequencies ("10 out of every 1,000 women have breast cancer; of these, X test positive; of the remaining 990, Y test positive...") and performance jumps: the percentage of participants giving the correct Bayesian answer roughly tripled — in Gigerenzer & Hoffrage's words, "the percentage of correct (i.e., Bayesian) inferences tripled, specifically, from 16 to 46%." This is the single most hopeful finding in the chapter: the human "bug" is partly a display problem, and it has a fix.

(Caveat: the Casscells study had only 60 participants selected in "hallway encounters," and the effect size in the original is often overstated. The direction of the finding, however, has replicated repeatedly.)

#9.3 The 2008 crisis: beautifully updating a wrong model

The Gaussian copula formula (Li 2000, "On default correlation: a copula function approach," Journal of Fixed Income) gave a single number for the probability that many mortgage-backed securities would default together. It was elegant, tractable, and widely adopted. Its fatal flaw was a model assumption — a low, stable default correlation calibrated on a period during which US national house prices had not fallen. The model could be updated on incoming data, but the hypothesis space itself excluded the possibility of a nationwide housing collapse.

The signature quote is Goldman Sachs CFO David Viniar in August 2007 (as reported in the Financial Times): "We were seeing things that were 25-standard-deviation moves, several days in a row." A 25-sigma event should occur far less than once in the history of the universe; several in a row is not bad luck but proof the model was wrong. This is the deepest Bayesian lesson: you can update flawlessly inside a false model and march confidently off a cliff. The map is not the territory (Section 11).

#9.4 Bayesian A/B testing: a genuine but oversold tool

Technology companies increasingly use Bayesian A/B testing (promoted by firms like VWO) because it answers the decision-relevant question — "what is the probability B is better than A?" — and permits multi-armed-bandit designs (Thompson sampling; Thompson 1933) that shift traffic toward winners while learning. This is the formal home of explore vs. exploit: try things, learn, and increasingly exploit what works, allocating traffic in proportion to the posterior probability each arm is best.

But a widely repeated claim — that Bayesian tests are "immune to peeking" (continuous monitoring with optional stopping) — is false as usually stated. If you stop as soon as the posterior looks good, you inflate error rates whether your statistics are Bayesian or frequentist; only a properly specified loss function or genuine prior discipline controls this. The honest lesson: the framework is better matched to how businesses decide, but it does not repeal the mathematics of optional stopping.

#9.5 Laplace's sex-ratio analysis: the first real posterior inference

Working with Parisian birth records (1745–1770) showing roughly 251,527 boys to 241,945 girls, Laplace (from 1774) computed the posterior probability that the true probability of a male birth exceeds 1/2, and found it "almost infallibly" greater than 1/2. This was arguably the first substantive scientific conclusion drawn from a posterior distribution — Bayes' machinery used not on a toy urn but on real demographic data, decades before the tools to compute it easily existed.

#9.6 A personal-scale example

Consider a belief you likely hold too confidently: "I'm a good judge of character; I can tell when someone is lying." Set a prior — say you believe you're right 80% of the time. Now apply the "what would change my mind" test. The base rate matters enormously: deception is rare in most interactions, so even a genuinely good lie-detector (high likelihood ratio) generates mostly false alarms, exactly as in the medical-test case. And the likelihood ratio is far weaker than intuition suggests: Bond & DePaulo's 2006 meta-analysis (Personality and Social Psychology Review) of 206 studies and 24,483 judges found that "people achieve an average of 54% correct lie-truth judgments, correctly classifying 47% of lies as deceptive and 61% of truths as nondeceptive" — barely above the 50% you'd get by guessing. Running this as an explicit update — prior 80%, confronted with evidence that your "tells" carry a likelihood ratio near 1 — should shrink your confidence sharply. Most people never run the update, because the belief is protected (Section 11, motivated reasoning).

#10. Practical Framework: The Belief Update Protocol

This is the chapter's actionable core. Run it in four steps.

1. 1. STATE THE PRIOR

   Write P(H) as a number BEFORE evidence.
   Ask: what is the base rate?

2. 2. ASSESS THE LIKELIHOOD RATIO

   How much more expected is this evidence
   under H than under the alternatives?
   LR = P(E|H) / P(E|not-H)

3. 3. UPDATE EXPLICITLY

   Posterior odds = Prior odds × LR
   Move as much as the LR dictates — no more, no less

4. 4. LOG IT

   Record prior, evidence, LR, posterior, and
   the pre-committed 'what would change my mind'

5. Did I actually move?

6. Didn't move despite real LR

   Check for CONSERVATISM

   or motivated reasoning

   Moved wildly on weak LR

   Check for base-rate neglect

   or over-reaction to vivid data

Step 1 — State the prior. Force yourself to write a number before the evidence. This defeats hindsight bias (Chapter One) and creates an auditable record.

Step 2 — Estimate the likelihood ratio. This is where the real thinking happens. Do not ask "does this evidence fit my theory?" (almost everything fits). Ask "is this evidence more expected under my theory than under the leading alternative?"

Step 3 — Update. Multiply. If the LR is 3, your odds triple; if it's 1.2, nudge; if it's ~1, do nothing however loud the evidence.

Step 4 — Log it in a decision/belief journal — the natural home, established in Chapter One, of a running update record.

#The two-week exercise

1. Pick three contested beliefs you hold (one personal, one professional, one about the world).
2. Write each as an explicit prior probability, and write down in advance the evidence that would move it and roughly what likelihood ratio it would carry.
3. Over two weeks, collect evidence and update explicitly each time, recording the size of each move.
4. At the end, review: were your moves too small (conservatism), too large (over-reaction), or absent (motivated reasoning)? The pattern of your errors is the lesson.

#Checklist for detecting that you are NOT updating

• You are collecting evidence but your probability never changes → confirmation-seeking / conservatism.
• You can't name any evidence that would change your mind → the belief is held at probability 1 (illegitimate).
• You explain away every disconfirming datum but accept every confirming one → asymmetric likelihood weighting (the mechanism of confirmation bias).
• Your confidence tracks how much you want the belief to be true → motivated reasoning (Kunda 1990).

#11. Criticisms and Limitations

The problem of priors. Where do priors come from? Subjective Bayesians (de Finetti) say they are personal and disciplined only by coherence. Objective Bayesians (Jeffreys, Jaynes) seek "non-informative" or reference priors that encode maximal ignorance. Neither program fully succeeds: there is no unique "ignorance" prior invariant under all reparameterizations. The pragmatic resolution is prior sensitivity analysis — make the prior explicit and show how much (or little) the conclusion depends on it.

Cox's theorem is not airtight. The clean "Bayes is forced by logic" story rests on Cox's theorem, but Halpern (1999, JAIR 10:67–85) produced a counterexample showing the theorem fails on finite domains without additional assumptions, and that even in infinite domains the assumptions are stronger than they look. The foundation is strong but not the unconditional proof it is sometimes sold as.

Model uncertainty — the deepest limit. Bayes updates beliefs within a fixed hypothesis space. If the truth lies outside that space — the wrong model — updating can converge confidently to a falsehood (the 2008 copula, Section 9.3). Gelman and Shalizi (2013, "Philosophy and the practice of Bayesian statistics") argue that in practice good Bayesian statistics is falsificationist: you must check the model against the data (posterior predictive checks) rather than merely update within it. Bayes is a superb updater and a poor model-critic; the human must supply the criticism.

Computational limits in human brains. Exact Bayesian inference is intractable even for computers (hence MCMC). The brain cannot be doing exact Bayes; at best it approximates. This is both a practical limit and a challenge to the strong Bayesian-brain thesis.

The Bayesian brain: plausible theory bordering on unfalsifiable metaphor. Behavioral evidence that humans integrate priors and sensory likelihoods near-optimally in perception and motor control (Körding & Wolpert 2004, Nature 427:244–247) is genuinely impressive — but even this landmark has been challenged: a 2021 reanalysis of the original trial-level data (Duffy, Igan, Pinheiro & Smith) found subjects were less sensitive to more-precise feedback and showed a recency bias, "not more" — i.e., the data do not cleanly exclude non-Bayesian explanations. At the level of high cognition, critics are sharper. Jones & Love (2011, "Bayesian Fundamentalism or Enlightenment?", Behavioral and Brain Sciences) and Bowers & Davis (2012, "Bayesian just-so stories," Psychological Bulletin 138:389–414) argue that because priors, likelihoods, and utility functions can be freely adjusted to fit almost any behavior, many Bayesian models are unfalsifiable and often no better-predicting than simpler non-Bayesian accounts. The Bayesian modelers replied (Griffiths, Chater, Norris & Pouget 2012) that the framework's job is to characterize the problem (computational-level analysis), not to claim neurons literally compute posteriors. The honest verdict: the Bayesian brain is a productive research program and a plausible theory for perception/sensorimotor control; it is not established fact for cognition, and the free-energy principle in particular risks becoming an unfalsifiable metaphor.

Does teaching transfer? Encouragingly, Sedlmeier & Gigerenzer (2001, JEP: General 130:380–400) showed that training people to construct natural-frequency representations (rather than plug into Bayes' formula) raised the rate of correct Bayesian solutions from a median of about 14% at baseline to 93%, a gain that showed no decay over 15 weeks — whereas conventional rule-training decayed. Broader debiasing work (Morewedge et al. 2015, Policy Insights from the Behavioral and Brain Sciences 2:129–140) reported "two longitudinal experiments that found medium to large effects of one-shot debiasing training interventions," with reductions in biases including base-rate neglect persisting up to three months. So the skill is teachable — but whether it reliably transfers to messy real-world judgment (as opposed to structured problems) remains only partly demonstrated.

When precise updating is worse than robust heuristics. Under deep uncertainty (unknown probabilities, unknown model — the Knightian and "wicked" regimes of Chapter One), precise Bayesian updating can be overconfident theater. When one bad outcome is catastrophic and irreversible (ruin), expected-utility maximization over a fragile posterior is dangerous; robust or precautionary heuristics that simply avoid the ruinous branch can dominate. Gigerenzer's ecological-rationality program shows "less is more" — simple heuristics beat complex models — precisely when data are scarce and the environment is unstable.

#12. Future Directions

Uncertainty in AI: epistemic vs. aleatoric. A central modern distinction (Kendall & Gal 2017) is between aleatoric uncertainty (irreducible noise in the world — no amount of data removes it) and epistemic uncertainty (the model's own ignorance — reducible with more data). Bayesian neural networks and their cheaper approximations (MC dropout, Gal & Ghahramani 2016; deep ensembles, Lakshminarayanan et al. 2017) aim to estimate the epistemic part — to make a model know what it doesn't know. This matters enormously for trusting machines: a self-driving car that is confidently wrong is far more dangerous than one that flags "I'm uncertain here, hand over."

LLM calibration. Modern deep networks are often miscalibrated — confidently wrong (Guo et al. 2017, "On calibration of modern neural networks"). For large language models the picture is nuanced: Kadavath et al. (2022, "Language Models (Mostly) Know What They Know") found that large models are well-calibrated on multiple-choice and true/false questions when the format is right, and can be trained to predict the probability that they know an answer. But subsequent work documents a calibration gap: Steyvers et al. (2024, Nature Machine Intelligence) found that "users tend to overestimate the accuracy of LLM responses when provided with default explanations," and that "longer explanations increased user confidence, even when the extra length did not improve answer accuracy" (n = 301 US participants, MMLU questions). The Bayesian frame is exactly the right one for the human–AI relationship: treat the model's stated confidence as evidence with its own (often poor) likelihood ratio, not as truth.

Science reform. Bayesian evidence standards (Bayes factors), pre-registration, and the retreat from p < 0.05 (Wasserstein & Lazar 2016) are reshaping how science handles evidence — pushing toward the explicit, cumulative updating Bayes describes.

The open question. Whether structured Bayesian education can measurably and durably improve real-world judgment — not just test performance — remains the frontier. The superforecaster evidence (below) suggests the answer is a qualified yes, but the mechanism (is it Bayes, or general open-mindedness?) is unresolved.

Connecting to Chapter One — superforecasters as applied Bayesians. The Good Judgment Project provides the best real-world evidence that Bayesian habits improve judgment. Superforecasters update more frequently and in smaller increments than ordinary forecasters. Mellers et al. (2015, Perspectives on Psychological Science 10:267–281) found that where typical forecasters tended to give estimates divisible by 10%, "superforecasters were most likely to make forecasts divisible by 1% and only 1%" (p. 276); Tetlock & Gardner's Superforecasting (2015) reports that "fully one-third of their forecasts used the single percentage point scale," and that the trick is "to update a forecast frequently but in most cases make only small adjustments." Atanasov et al. ("Small steps to accuracy: Incremental belief updaters are better forecasters") quantify the frequency: superforecasters submitted a mean of about 5.1 forecasts per question versus about 2.0 in the core sample, and incremental updaters were more accurate without more apparent effort. This is the Belief Update Protocol in the wild — and the direct link to Chapter One's finding that decision quality lives in process.

#13. Recommended Resources

Beginner

• Annie Duke, Thinking in Bets (2018). The lightest entry — teaches probabilistic belief and "resulting" without equations. (Directly extends Chapter One.)
• Will Kurt, Bayesian Statistics the Fun Way (2019). Genuinely gentle, worked examples, builds the odds-form intuition.
• Nate Silver, The Signal and the Noise (2012). Narrative case studies of prediction done well and badly; motivates why Bayes matters across domains.

Intermediate

• Richard McElreath, Statistical Rethinking (2nd ed., 2020). The best modern on-ramp to real Bayesian data analysis; emphasizes model-building and criticism, not button-pushing. Accompanying lecture videos are excellent.
• Sharon Bertsch McGrayne, The Theory That Would Not Die (2011). The human history of Bayes — Bayes, Laplace, wartime codebreaking, the frequentist wars.

Advanced

• E.T. Jaynes, Probability Theory: The Logic of Science (2003). The passionate objective-Bayesian manifesto; Bayes as extended logic. Opinionated and profound.
• David MacKay, Information Theory, Inference, and Learning Algorithms (2003). Ties Bayes to information theory and machine learning; freely available; superb intuition pumps.
• Andrew Gelman et al., Bayesian Data Analysis (3rd ed.). The professional reference.

Landmark papers (worth reading for the ideas, not just citation): Bayes & Price (1763); Laplace (1774); Cox (1946); Edwards (1968, conservatism); Kahneman & Tversky (1973, base rates); Gigerenzer & Hoffrage (1995, natural frequencies); Gelfand & Smith (1990, the computational revolution); Friston (2005, free energy); Griffiths & Tenenbaum (2006, Bayesian models of cognition); Kendall & Gal (2017, uncertainty in deep learning).

#14. Self-Check

Attempt these from memory before checking anything.

1. Derive Bayes' theorem in odds form from the definition of conditional probability, in your own words. Why is the odds form more useful for everyday reasoning than the standard form?
2. Explain why a 99%-accurate test can still mean a disease is unlikely after a positive result — and name the number everyone forgets.
3. Describe two documented ways human updating deviates from Bayes (name the researchers), and give a personal example of each from your own life.
4. What is the likelihood ratio, and why is a dramatic-feeling piece of evidence with a likelihood ratio near 1 worthless?
5. State the prosecutor's fallacy precisely, using the two conditional probabilities it confuses. How did it operate in the Sally Clark case?
6. Explain the difference between epistemic and aleatoric uncertainty, and why the distinction matters for trusting an AI system.
7. What is the strongest objection to the claim that "the brain is Bayesian"? Why does the free-energy principle attract the charge of unfalsifiability?
8. "More data always fixes a bad prior." Give two conditions under which this is false.

Synthesis (verify your own recall): A strong set of answers will connect the arithmetic to the psychology to the philosophy. The odds form (posterior odds = prior odds × likelihood ratio) should feel like the spine of the whole chapter: base-rate neglect is forgetting the prior-odds term; conservatism is under-multiplying by the likelihood ratio; confirmation bias is misjudging the likelihood ratio asymmetrically. The medical-test and Sally Clark cases are the same error (ignoring the base rate / inverting the conditional) wearing different clothes. The 2008 copula and the Bayesian-brain critiques are both about the limits of updating: you can update perfectly inside a wrong model, and calling behavior "Bayesian" can redescribe rather than explain it. If you can explain the Belief Update Protocol and say honestly whether your own recent updates were too small, too large, or absent, you have the chapter.

#Knowledge Card — Bayesian Thinking

## Knowledge Card — Bayesian Thinking
- Core terms:
  - Prior P(H): degree of belief before evidence (often the base rate)
  - Likelihood P(E|H): how expected the evidence is if the hypothesis holds
  - Posterior P(H|E): updated belief after evidence (the output)
  - Likelihood ratio: P(E|H)/P(E|¬H) — the strength of the evidence
  - Base rate: prior prevalence of a condition in the population
  - Conjugate prior / MCMC: shortcuts and sampling methods for computing posteriors you can't write down
  - Epistemic vs. aleatoric uncertainty: reducible model-ignorance vs. irreducible world-noise
  - Coherence (Dutch book): having beliefs that can't be turned into a guaranteed loss = obeying probability rules
- Core mental models:
  - Odds form: posterior odds = prior odds × likelihood ratio (the whole of Bayes in one line)
  - Base-rate-first: ask "how common is this?" before reacting to case evidence
  - Prior sensitivity: does my conclusion survive a different reasonable prior?
  - "What would change my mind?" — name the disconfirming evidence and its likelihood ratio in advance
- Connections to prior chapters:
  - Chapter One (Deciding Well): incremental, small-step updating IS the superforecaster habit; base-rate neglect is both a bias and a Bayesian failure; the decision journal is the natural home of a belief-update log; deep uncertainty/ruin is where precise updating yields to robust heuristics
  - Cognitive Biases chapter: base-rate neglect, conservatism, and confirmation bias are all deviations from the odds-form update; ecological rationality (natural frequencies) shows biases are partly format artifacts
- Recommended next chapter: Causal Inference & Correlation vs. Causation — Bayes tells you how to update on evidence, but not which correlations license action; causal reasoning is the missing half of good judgment
- One habit to keep: Before updating on any striking piece of news, write your prior as a number and estimate the likelihood ratio — then move only as much as the ratio dictates, and log it.