Is a prior the same as a prior probability?

A prior distribution describes uncertainty across a whole parameter space; a prior probability can refer to one event or hypothesis within that space.

Must priors be highly informative?

No. They can range from concentrated to weakly informative, with different implications.

Can priors be checked before fitting?

Yes. Prior predictive simulation examines the observations the prior and likelihood jointly imply.

AI Glossary / Statistics and probability / Intermediate

Prior distribution

Definition

A prior distribution represents uncertainty about unknown parameters or other model quantities before incorporating the particular data used in the current Bayesian update. It is combined with the likelihood to produce a posterior distribution. The prior can express earlier evidence, substantive knowledge, plausible ranges or regularizing assumptions. Before means before the current update, not necessarily before any information has ever been collected. Its meaning depends on the parameterization, scale and conditioning information of the model.

What it means

A prior is more than a single initial guess. It describes which values are plausible and how much uncertainty surrounds them. Two priors can have the same mean but very different concentration, leading to different posterior influence. A broad prior is not automatically neutral or safe: it can place substantial probability on unrealistic parameter combinations or predictions. Conversely, a carefully chosen weakly informative prior can rule out implausible extremes without claiming precise knowledge of the parameter.

How it works

Define the model quantities and their scales before selecting distributions. Use relevant prior evidence and domain constraints, documenting how they enter the specification. Inspect the predictions produced by drawing parameters from the prior and then generating observations from the likelihood. This prior predictive check helps reveal implications that are hard to see from parameter ranges alone. Fit the model and examine sensitivity to reasonable alternative priors. Ensure the posterior is mathematically well defined; an improper prior does not integrate to one and requires particular care because not every likelihood makes the resulting posterior proper.

Three practical examples

Quality modeling: represent uncertainty about a defect probability using a distribution supported between zero and one.
Regression: use priors that regularize implausibly large coefficients on a clearly stated predictor scale.
Forecasting: combine earlier information about a demand rate with newly observed counts, while distinguishing historical relevance from an assumption that old conditions remain unchanged.

Think of it this way

Think of a map of plausible values before examining the latest evidence. Some regions may be more plausible than others, and the width of the map expresses uncertainty. New evidence changes the map, but the starting map must still make sense in the units and setting where the model operates.

Worked example

Consider an original Bernoulli model for an unknown success probability theta. A Beta(2,3) prior has mean 2/(2+3) = 0.4 and variance 2 times 3 divided by [5^2 times 6] = 0.04. Observing six successes and four failures gives a Beta(8,7) posterior, with mean 8/15, approximately 0.5333. A uniform Beta(1,1) prior with the same data gives posterior mean 7/12, approximately 0.5833. A more concentrated Beta(20,30) prior also has mean 0.4, but yields posterior mean 26/60, approximately 0.4333. Thus identical prior means do not imply identical updating: concentration matters. The beta parameters act like prior contributions in this conjugate calculation, but they should not be described as literal observed past counts unless that interpretation is justified by the prior construction. These stipulated priors illustrate sensitivity, not a recommendation for a real application.

Illustrative case study

In a fictional case, Alder Forecasting models a small number of service events. Analysts initially choose very broad priors for several positive rates. Prior predictive simulations generate counts far beyond any feasible system capacity. The team revises the priors using documented operational scales, then checks how posterior forecasts change under other plausible choices. The report explains both the information used and the remaining uncertainty instead of labeling one prior objectively uninformative without examining its consequences.

Common misconceptions and mistakes

A prior distribution is not necessarily a subjective opinion unsupported by evidence, nor is it automatically an objective fact. Uniformity depends on parameterization: a distribution uniform on a probability scale is not uniform on its log-odds scale. Large datasets can reduce prior sensitivity in many settings, but weak identification and sparse subgroups can leave important dependence on priors. A posterior from an earlier analysis can sometimes become a new prior, provided the same evidence is not counted twice. Priors on separate parameters can also imply surprising joint behavior, so checking each marginal distribution alone may be insufficient.

Business use

Businesses should treat prior specification as a documented modeling decision. Explain its practical implications through plausible outcomes, not only mathematical parameter names. Separate historical data used to build a prior from the current likelihood data, especially when datasets overlap. Assess whether assumptions transfer across products, locations or time periods. Sensitivity analysis can reveal whether the evidence supports the decision robustly or whether the result is driven by uncertain prior choices. Good communication makes the role of the prior visible rather than hiding it behind software defaults.

Check your understanding

Beta(2,3) and Beta(20,30) have the same mean. Must they lead to the same posterior after identical data?

Answer: No. Their concentrations differ, so their relative influence on the update differs.

Related terms

Prior probability
Likelihood
Bayesian inference
Conjugate prior
Regularization
posterior distribution
prior predictive check

Learn next

Learn how parameterization affects prior meaning, then practice prior predictive simulation and sensitivity analysis before interpreting posterior conclusions.

Sensitivity analysis

Sources

Stan User's Guide, Posterior and Prior Predictive Checks: https://mc-stan.org/docs/2_36/stan-users-guide/posterior-predictive-checks.html Frank Harrell, Glossary of Statistical Terms, prior and posterior concepts: https://hbiostat.org/glossary/

Previous: Prior belief Next: Prior-independent mechanism