Bayesian Statistics¶

Bayesian statistics occupies a peculiar epistemic position: simultaneously mathematically rigorous and philosophically contentious. At its core is a simple formula - Bayes' theorem - that is uncontroversially correct as mathematics but whose interpretation as a reasoning framework remains debated. Frequentist statisticians object to treating probability as "degree of belief" rather than long-run frequency. Practitioners note computational challenges and sensitivity to prior assumptions. Yet despite these limitations, Bayesian thinking has colonized domains far beyond statistics: medical diagnosis, machine learning, courtroom evidence evaluation, and everyday decision-making.

Why extract reasoning tools from a contested framework? Because Bayesian statistics systematically corrects a specific class of cognitive errors: our chronic failure to update beliefs proportionally to evidence. Humans anchor too strongly on initial beliefs, ignore base rates, misinterpret conditional probabilities, and fail to distinguish evidence strength from conclusion certainty. These aren't statistical errors - they're reasoning failures that appear whenever we think under uncertainty.

The extraction principle is straightforward: separate the mental operations from the mathematical formalism. Bayes' theorem itself is just arithmetic, but the habit of proportional belief updating transfers to contexts where you'll never calculate exact posterior probabilities. The discipline of making priors explicit works even when those priors are qualitative. The distinction between likelihood and posterior clarifies thinking whether or not you're doing statistics.

These tools survive even when Bayesian epistemology fails. You don't need to believe probability represents subjective belief to find value in asking "how much should this evidence shift my confidence?" You don't need to endorse Bayesian confirmation theory to benefit from separating prior plausibility from evidential strength. The tools outlast the theory because they formalize intuitions about rational evidence-handling that predate probability theory entirely.

Foundational Tools - Evidence and Belief¶

Tools for systematic belief revision based on new information

Structural Tools - Modeling Uncertainty¶

Tools for representing and decomposing complex uncertain situations

Dynamic Tools - Learning and Adaptation¶

Tools for reasoning about changing evidence and evolving uncertainty

Applied Tools - Decision and Action¶

Tools for using probabilistic beliefs to make better decisions

Quick Reference¶

Decision Type -> Tool Mapping¶

When you need to...	Use these tools
Evaluate a surprising claim	Prior-Likelihood-Posterior Decomposition, Base Rate Anchoring
Update beliefs on new evidence	Proportional Belief Updating, Sequential Evidence Accumulation
Assess evidence strength independent of conclusions	Likelihood Ratio Thinking
Consider all explanations systematically	Hypothesis Space Partitioning
Handle multiple pieces of evidence	Conditional Independence Exploitation, Sequential Evidence Accumulation
Evaluate how much priors matter to your conclusion	Prior Sensitivity Analysis
Check if your model explains reality	Posterior Predictive Checking
Improve your probability estimates over time	Evidence Strength Calibration
Respond to genuinely shocking evidence	Model Expansion Under Surprise
Make a decision under uncertainty	Expected Value Decomposition, Probabilistic Scenario Planning
Decide whether to gather more information	Information Value Calculation
Learn from past decisions	Bayesian Regret Minimization

Usage Notes¶

Domain of applicability: These tools work best when:

You can meaningfully assign probabilities, even qualitatively.
Evidence arrives incrementally and you need to update beliefs.
Multiple hypotheses are genuinely in play, not just theoretical possibilities.
Calibration feedback is available - you can check if your probability estimates were right. They work less well when: probabilities are fundamentally unknowable (Knightian uncertainty), evidence is so entangled you can't decompose it, or you need single-shot decisions with no opportunity to calibrate.

Limitations: Bayesian tools cannot tell you what to value (utilities are inputs, not outputs). They don't resolve genuine value uncertainty or moral disagreement. They assume your hypothesis space contains the truth - but sometimes all your models are wrong in ways you can't imagine. They work best with repeated decisions where calibration is possible, less well for one-off choices. They require explicit probability assignment, which is cognitively costly and sometimes misleading (false precision). They don't automatically handle computational intractability - real Bayesian inference in complex models is often impossible; you need approximations.

Composition: Tools combine naturally. Base Rate Anchoring provides the prior for Prior-Likelihood-Posterior Decomposition. Likelihood Ratio Thinking quantifies evidence strength for Proportional Belief Updating. Hypothesis Space Partitioning creates the framework for Sequential Evidence Accumulation. Expected Value Decomposition uses posterior probabilities from Bayesian updating. Information Value Calculation tells you when to gather evidence that would enable more Proportional Updating. The tools form a coherent system rather than isolated techniques.

Common combinations: - Base Rate + Likelihood Ratio + Proportional Update = Complete evidence evaluation - Hypothesis Partitioning + Sequential Accumulation + Posterior Predictive Check = Robust modeling - Expected Value + Information Value = Optimal decision timing - Prior Sensitivity + Posterior Predictive + Model Expansion = Model validation

Integration with other domains: Bayesian tools complement causal reasoning (Pearl's do-calculus), decision theory (utility functions), scientific method (hypothesis testing as Bayesian updating), and systems thinking (feedback loops as evidence accumulation processes). They provide the probabilistic foundation for reasoning under uncertainty that other frameworks build on. Game theory becomes Bayesian when you model opponent beliefs. Economics becomes behavioral when you study how humans violate Bayesian rationality. Machine learning is applied Bayesian inference at scale.

Watch out for: Over-quantification - not everything needs explicit probabilities. Spurious precision - claiming "72.3% confidence" when you mean "probably." Ignoring computational costs - Bayesian inference can be intractable; sometimes simple heuristics beat principled calculation. Miscalibration - humans are notoriously bad at probability estimation without training and feedback. Model misspecification - all these tools assume your model is roughly right; when it's fundamentally wrong, they entrench error rather than correct it.