Prior Probability¶

Represent uncertainty about a parameter or hypothesis before the focal evidence is incorporated by assigning it a probability distribution that will be combined with a likelihood.

Version: v2 · 2026-09-06 · History
Domain-specific #: 2535
Origin domain: mathematics
Subdomain: bayesian statistics
Aliases: Prior distribution, Prior

Core Idea¶

A prior probability distribution assigns probability to possible values of an unknown quantity before the focal data are incorporated. In a Bayesian model it is combined with the likelihood through Bayes' rule to produce a posterior distribution. 'Before' is inferential rather than necessarily chronological: the prior may encode previous studies, expert information, structural constraints, or a deliberately weak baseline, but it is the distribution not yet conditioned on the dataset currently under analysis.^[1]

The prior is part of the model, not a detachable opinion sticker. Its support can rule parameter values in or out, its concentration determines how strongly data must speak to move the posterior, and its dependence structure can couple quantities. Proper, improper, weakly informative, informative, reference, hierarchical, and conjugate priors make different commitments. Those commitments must be judged relative to the likelihood and inferential target because a numerically diffuse distribution need not be weak on the scale that matters.

Structural Signature¶

Unknown quantity. A parameter, latent state, model, or hypothesis is assigned uncertainty.
Pre-focal information state. The distribution excludes the dataset currently entering the likelihood.
Support. The prior declares which values or hypotheses receive nonzero probability.
Shape and concentration. Relative weights encode location, scale, dependence, and tail commitments.
Likelihood interface. Prior and data model must be defined on compatible quantities.
Posterior update. Multiplication by likelihood and normalization produce revised weights.
Sensitivity surface. Alternative reasonable priors expose how much conclusions depend on the initial commitment.

What It Is Not¶

Not the posterior. It precedes conditioning on the focal evidence.
Not the likelihood. The likelihood scores observed data under parameter values; it is not a probability distribution over those values by itself.
Not necessarily subjective guesswork. Priors can encode regularization, symmetry, earlier data, or population structure.
Not automatically uninformative when flat. Flatness depends on parameterization and can induce strong transformed-scale commitments.
Not Bayesian updating as a whole. It is one input to the updating operation.

Scope of Application¶

Prior distributions travel literally across Bayesian models wherever uncertainty is assigned before focal conditioning. Their precise family and interpretation remain tied to the modeled quantity and sampling process.

Parameter estimation. Regularizing and representing plausible values before observing the current sample.
Hierarchical modeling. Sharing information through population-level distributions.
Model comparison. Assigning prior mass to competing hypotheses and parameter spaces.
Bayesian decision analysis. Supplying uncertainty used with utilities to compare actions.
Machine learning. Inducing shrinkage, sparsity, smoothness, or other structured preferences.
Sequential analysis. Reusing one stage's posterior as the next stage's prior under a coherent data partition.

Clarity¶

State the random quantity, parameterization, conditioning information, support, distribution family, hyperparameters, and whether the prior is proper. Explain the substantive meaning of its scale. Distinguish data used to construct the prior from focal data used in the likelihood, and report prior predictive behavior and sensitivity rather than relying on labels such as objective or weak.

Manages Complexity¶

A prior compresses background constraints and earlier information into a distribution that enters the same calculus as new data. It regularizes underidentified models and makes uncertainty propagation coherent. The compression can conceal strong assumptions, double-use data, or assign negligible mass near the truth; prior predictive checks and sensitivity analysis reopen what the single distribution suppresses.

Abstract Reasoning¶

Define the unknown quantity and its parameterization.
List information available before the focal evidence and constraints on support.
Choose a distribution family and hyperparameters that express those commitments.
Inspect implied quantities through prior predictive simulation.
Combine the prior with the likelihood to obtain the posterior.
Compare plausible alternative priors and locate conclusion-sensitive regions.
Revise the model transparently if prior predictions are indefensible.

Knowledge Transfer¶

The literal instrument remains Bayesian. The transferable parent is Distributional Assumption: uncertain quantities are assigned a family and parameters before conclusions are drawn. Outside probabilistic modeling, 'prior' can mean any earlier belief, but that usage lacks normalized weights, support, likelihood interface, and posterior calculus.

A prior cannot be judged from its marginal plot alone. Its effect is relational: the prior, likelihood, parameterization, and inferential target combine to determine the posterior. A distribution that looks broad on a raw coefficient can be highly concentrated on a probability, odds ratio, or predicted outcome after a nonlinear transformation. Conversely, a modest-looking variance can allow extreme implications in a hierarchical or dynamical model. Prior review should therefore simulate or derive the quantities the model implies before observing the focal data.

Support is the hardest boundary. Bayes' rule cannot assign posterior mass to a value that received exactly zero prior mass, regardless of how compatible the data appear. Structural exclusions such as positive scale, ordered parameters, or physically bounded probabilities may be justified; accidental truncation can make the model unable to learn. A clear audit separates impossible values, values neglected for numerical convenience, and values merely regarded as unlikely. Sensitivity checks should widen or alter uncertain support commitments rather than changing only a cosmetic label such as weak.

The likelihood determines how rapidly evidence can overcome concentration. With little information or partial identification, the posterior may remain prior-sensitive even in a numerically large dataset. With a highly informative likelihood, several reasonable priors can converge to similar posterior conclusions. Neither situation proves that priors are subjective noise or irrelevant decoration. The correct report shows posterior movement from prior to likelihood-informed result and identifies which conclusions are stable across defensible alternatives.

Hierarchical priors distribute assumptions across levels. Group parameters may share a population distribution whose hyperparameters are themselves uncertain, creating partial pooling rather than a fixed penalty pasted onto each estimate. This can improve estimation while making dependence explicit. The hierarchy also creates new choices about exchangeability, scale, and tail behavior. Calling a model data driven because hyperparameters are estimated does not remove the prior; it moves assumptions upward and changes how information is shared.

Improper priors require a separate validity check. A nonintegrable prior can sometimes produce a proper posterior, but it is not itself a probability distribution and may make model comparison or marginal likelihood undefined. Propriety must be proved for the actual likelihood and parameterization. A formula that worked in one simple model cannot be transferred automatically to a mixture, missing-data model, or hierarchy where unidentifiability changes the integral.

Prior predictive checking provides the most portable diagnostic. Generate parameter values from the prior, then generate hypothetical observations through the likelihood. If the resulting data are impossible in scale, frequency, or structure for the application, the joint model needs revision before seeing the focal sample. This check does not tune the model toward the observed answer; it tests whether declared pre-data assumptions imply a coherent observable world.

These mechanics distinguish the node from Bayesian Updating. Updating is the operation that combines a prior with a likelihood; the prior is one structured input and can be studied before the operation occurs. Distributional Assumption is therefore the strict parent: every prior assigns a law to an uncertain quantity under a model. The domain residual includes pre-focal conditioning status, support effects, posterior sensitivity, predictive implication, and hierarchical pooling. Those roles make Prior Probability autonomous without portraying it as an ungrounded personal belief.

Examples¶

Canonical¶

For a binomial success probability p, a Beta(α,β) prior combines with x successes in n trials to give a Beta(α+x,β+n−x) posterior. The prior contributes support on [0,1], concentration, and an interpretable amount of shrinkage; the conjugate form simplifies calculation without making the choice substantively neutral.^[1]

Mapped back: unknown p → pre-data Beta weights → binomial likelihood → parameter-compatible multiplication → posterior Beta update.

Applied / In Practice¶

A multilevel clinical model can assign site effects a common population prior whose scale is itself estimated. Sparse sites then borrow strength from the population while data-rich sites move farther from it. Analysts compare prior predictive outcome ranges and repeat the fit under wider and narrower scale priors to see whether a treatment conclusion is data-driven or prior-sensitive.

Mapped back: hierarchical quantity → structured prior → likelihood → partial pooling → predictive and sensitivity audit.

Structural Tensions¶

Background knowledge vs. data dominance. A prior can stabilize or overwhelm. Diagnostic: How much likelihood evidence is needed to move the posterior?
Convenience vs. meaning. Conjugacy simplifies algebra but may dictate an implausible shape. Diagnostic: Does the prior predictive distribution make sense?
Weakness vs. parameterization. Flatness can disappear under transformation. Diagnostic: On which scale is informativeness being judged?
Reuse vs. double counting. Earlier evidence can inform a prior or leak from the focal dataset. Diagnostic: Are the information sources disjoint or modeled jointly?
Support discipline vs. misspecification. Excluding impossible values helps, but excluding the truth is fatal. Diagnostic: Which conclusions are forced by zero prior mass?

Structural–Framed Character¶

Distribution-before-conditioning is structural, while parameterization, modeling purpose, domain information, and judgments about plausible scales frame the actual prior. The construct is formal and domain-specific rather than culturally framed.

Structural Core vs. Domain Accent¶

The skeleton is uncertain quantity → pre-evidence weighted possibilities → evidence-conditioned revision. Probability normalization, parameter space, likelihood, and posterior calculus are constitutive. Removing them yields a general assumption or expectation, while the accepted prime Distributional Assumption carries the broader pattern.

Distributional Assumption is the strict parent because every prior commits an uncertain quantity to a probability distribution before inference. Bayesian Updating is the operation that consumes the prior; it is related but not a taxonomic parent of the input object.

The prospective workspace queue contains one strict upward edge to prime:distributional_assumption. No live DAG mutation is authorized.

Relationships to Other Abstractions¶

Current abstraction Prior Probability Domain-specific

Parents (1) — more general patterns this builds on

Prior Probability is a kind of Distributional Assumption Prime

Distributional Assumption is the strict parent because every prior commits an uncertain quantity to a probability distribution before inference.

Hierarchy paths (7) — routes to 5 parentless roots

Prior Probability → Distributional Assumption → Assumption → Epistemic Mode Of A Proposition

Neighborhood in Abstraction Space¶

Prior Probability sits in a sparse region of the domain-specific corpus (89^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Posterior probability. The distribution after focal evidence is incorporated.
Likelihood. A function of parameters induced by observed data, not a prior distribution over parameters.
Prior predictive distribution. The implied distribution of future or replicated observations after integrating parameters over the prior.
Regularizer. Often mathematically related to a prior but not automatically a full probabilistic commitment.
Base rate. A population frequency that may inform a prior but is not synonymous with every prior.

References¶

[1] Andrew Gelman et al., Bayesian Data Analysis, 3^rd ed. (CRC Press, 2013), chapters 1–5. registry ↩a ↩b