STATISTICS GRADUATE LEVEL • BAYESIAN STATISTICS

Posterior Summaries & Credible Intervals — Posterior summaries and credible intervals

Distilling the full posterior distribution into interpretable point estimates and probability intervals for principled Bayesian inference.

Historical Context & Motivation

The idea that probability can express degrees of belief rather than merely long-run frequencies has deep roots in the work of Thomas Bayes and Pierre-Simon Laplace. Once a researcher computes a posterior distribution over an unknown parameter, the natural next question becomes: how do we summarize that distribution in a way that is both concise and faithful to its shape? This challenge motivated the development of posterior summaries—point estimates, interval estimates, and loss-function-based decisions—that have become central tools in modern Bayesian practice.

Frequentist statistics addressed the summarization problem through confidence intervals and maximum likelihood estimators, but these tools carry interpretive constraints: a 95% confidence interval, for instance, does not claim that the parameter lies within it with probability 0.95. Bayesians sought a framework in which intervals would carry exactly that direct probability interpretation. The resulting concept, the credible interval, assigns posterior probability directly to a set of parameter values, providing an intuitive and coherent summary of uncertainty.

1763
Bayes' Theorem Published Posthumously
Richard Price publishes Thomas Bayes' essay, introducing the notion of updating beliefs via the likelihood, laying the groundwork for posterior inference.
1812
Laplace's Théorie Analytique
Laplace formalizes inverse probability and uses posterior means and modes as point estimators, treating probability as a measure of rational belief.
1961
Decision-Theoretic Foundations
James Berger and others systematize the connection between loss functions and optimal Bayesian point estimates, showing that the posterior mean minimizes squared-error loss and the posterior median minimizes absolute-error loss.
1990s
MCMC Revolution
Markov chain Monte Carlo methods (Gibbs sampling, Metropolis-Hastings) make it computationally feasible to summarize complex, high-dimensional posteriors, popularizing credible intervals and HPD regions in applied work.
2010s–Present
Probabilistic Programming & Automated Summaries
Tools such as Stan, PyMC, and Turing.jl compute posterior summaries, credible intervals, and diagnostics automatically, making Bayesian workflow accessible to a broad range of applied disciplines.

The central question this lesson addresses is: given a full posterior distribution π(θ | y), what are the principled ways to collapse it into point summaries (posterior mean, median, mode) and interval summaries (equal-tailed intervals, highest posterior density regions), and how do different choices reflect different loss functions and inferential goals?

Core Principles & Definitions

After observing data y and specifying a prior π(θ), the posterior distribution π(θ | y) encodes all updated information about the parameter θ. Because this distribution is often a complex density over a continuous space, analysts require principled tools to extract actionable summaries. The following core ideas underpin the choice of point estimates and interval summaries in Bayesian inference.

1

Posterior Mean

The expected value E[θ | y] of the posterior distribution. It minimizes squared-error loss and is the most common Bayesian point estimate when the posterior is roughly symmetric.
2

Posterior Median

The value that divides the posterior probability in half. It minimizes absolute-error loss and is more robust than the mean when the posterior is skewed or heavy-tailed.
3

Posterior Mode (MAP)

The maximum a posteriori estimate, the value with the highest posterior density. It minimizes 0-1 loss and corresponds to the peak of the posterior.
4

Equal-Tailed Credible Interval

An interval [q_{α/2}, q_{1−α/2}] containing (1 − α)×100% posterior probability, with equal probability in each tail. Straightforward to compute from posterior quantiles.
5

Highest Posterior Density (HPD) Region

The shortest interval (or set of intervals for multimodal posteriors) containing (1 − α)×100% posterior probability. Every point inside has density at least as high as every point outside.
KEY TAKEAWAY
Think of the posterior distribution as a topographic map of plausibility over parameter space. A point estimate is like dropping a single pin on the map—whether you place it at the highest peak (mode), the center of mass (mean), or the elevation median depends on how you penalize errors. A credible interval is like drawing a contour line that encloses a region where most of the 'probability mass' sits, giving a range of plausible values rather than a single guess.

Visual Explanation — Posterior Distribution & Summaries

The diagram below illustrates a mildly skewed posterior density for a parameter θ, along with the three common point estimates and two types of 95% credible intervals. Observe how the posterior mode sits at the peak, the posterior mean is pulled toward the heavier right tail, and the posterior median lies between them.

The cyan curve is the posterior density π(θ | y). Dashed vertical lines mark the mode (cyan), median (pink), and mean (violet). Below the x-axis, the amber bar shows the equal-tailed 95% credible interval and the emerald bar shows the 95% HPD interval. Note that the HPD interval is shorter because it captures the region of highest density.

The key observation from this diagram is that the three point estimates differ whenever the posterior is asymmetric. For a perfectly symmetric posterior—such as a normal distribution—the mode, median, and mean coincide. As skewness increases the mean is pulled toward the heavier tail because it integrates over the entire density, whereas the mode remains anchored at the peak. Similarly, the equal-tailed interval places identical probability (α/2) in each tail, which can result in a wider interval compared to the HPD region that instead seeks the shortest set with the required coverage.

Mathematical Framework

Each posterior summary can be derived as the optimizer of a specific expected loss under the posterior distribution. Let π(θ | y) denote the posterior density and L(θ, δ) the loss incurred when the true parameter is θ and our summary is δ. The Bayes estimator is the δ that minimizes the posterior expected loss, also called the posterior risk.

POSTERIOR RISK
ρ(δ | y) = ∫ L(θ, δ) π(θ | y) dθ
ρ(δ | y) = posterior expected loss for action δ; L(θ, δ) = loss function; π(θ | y) = posterior density.
SQUARED-ERROR LOSS → POSTERIOR MEAN
L(θ, δ) = (θ − δ)² ⟹ δ* = E[θ | y] = ∫ θ π(θ | y) dθ
Under squared-error loss the Bayes estimator δ* is the posterior mean. This follows by differentiating the posterior risk with respect to δ and setting it to zero.
ABSOLUTE-ERROR LOSS → POSTERIOR MEDIAN
L(θ, δ) = |θ − δ| ⟹ δ* = median(θ | y)
The posterior median minimizes expected absolute deviation. The median is the value m such that P(θ ≤ m | y) = 0.5.
0-1 LOSS → POSTERIOR MODE (MAP)
L(θ, δ) = 1 − 𝟙(|θ − δ| < ε) ⟹ δ* → arg max_θ π(θ | y) as ε → 0
Under 0-1 loss in the continuous case (with a limiting argument), the optimal estimator is the posterior mode, also called the maximum a posteriori (MAP) estimate.

Credible Intervals

A (1 − α) credible interval is any set C ⊆ Θ such that P(θ ∈ C | y) = 1 − α. Two canonical constructions dominate practice.

EQUAL-TAILED INTERVAL
C_{ET} = [q_{α/2}, q_{1−α/2}] where P(θ ≤ q_p | y) = p
The equal-tailed interval places probability α/2 in each tail. The quantiles q_p are obtained from the inverse CDF of the posterior.
HIGHEST POSTERIOR DENSITY (HPD) REGION
C_{HPD} = { θ : π(θ | y) ≥ k_α } where k_α satisfies P(θ ∈ C_{HPD} | y) = 1 − α
The HPD region collects all parameter values whose posterior density exceeds a threshold k_α. For unimodal posteriors this yields a single interval; for multimodal posteriors it may consist of disjoint pieces.

Detailed Comparison — Equal-Tailed vs. HPD Intervals

While both the equal-tailed and HPD intervals achieve the same nominal coverage probability, they can differ substantially in width and interpretation when the posterior is skewed or multimodal. The diagram below contrasts both interval types on a Gamma(3, 1) posterior, a distribution with notable right skewness. The HPD interval is shorter because it capitalizes on the asymmetry, capturing the tall left portion of the density rather than trimming equal probability from both tails.

Panel A shows the equal-tailed interval (amber) with endpoints at the 2.5th and 97.5th percentiles. Panel B shows the HPD interval (emerald) whose endpoints are defined by a horizontal density cutoff kα. For this right-skewed Gamma(3,1) posterior the HPD interval is about 5% shorter while covering the same 95% posterior probability.
Comparison of equal-tailed and HPD credible intervals
FeatureEqual-Tailed IntervalHPD Interval
DefinitionQuantile-based: places α/2 probability in each tail.Density-based: collects all θ with π(θ | y) ≥ k_α.
WidthMay be wider for skewed posteriors.Always the shortest interval (for unimodal posteriors).
Transformation invarianceYes — quantiles are preserved under monotone transformations.No — density values change under reparameterization.
Multimodal posteriorsAlways a single contiguous interval.May consist of disjoint intervals.
Ease of computationVery easy from MCMC samples (sort and take percentiles).Requires density estimation or specialized algorithms.
💡 When are they identical?
For any symmetric, unimodal posterior (e.g., a normal distribution), the equal-tailed interval and the HPD interval coincide exactly. In practice, when the posterior is approximately symmetric—as it often is with large sample sizes thanks to the Bernstein–von Mises theorem—the two intervals are nearly indistinguishable.

Worked Example — Beta-Binomial Model

Suppose we observe y = 7 successes in n = 20 Bernoulli trials and adopt a Beta(2, 2) prior on the success probability θ. Because the Beta prior is conjugate to the Binomial likelihood, the posterior is available in closed form: θ | y ∼ Beta(α + y, β + n − y) = Beta(9, 15). We wish to compute the posterior mean, median, mode, and both the 95% equal-tailed and 95% HPD credible intervals.

Posterior Summaries for θ | y ∼ Beta(9, 15)
1
Step 1 — Identify the PosteriorWith a Beta(2, 2) prior and observing y = 7 successes in n = 20 trials, conjugacy gives θ | y ∼ Beta(2 + 7, 2 + 13) = Beta(9, 15). The posterior is a right-skewed unimodal density on (0, 1).
θ | y ∼ Beta(9, 15)
2
Step 2 — Posterior MeanThe mean of a Beta(a, b) distribution is a / (a + b). Therefore E[θ | y] = 9 / (9 + 15) = 9 / 24 = 0.375. This is the Bayes estimator under squared-error loss.
Posterior mean = 0.375
3
Step 3 — Posterior Mode (MAP)The mode of Beta(a, b) (for a, b > 1) is (a − 1) / (a + b − 2). So the MAP estimate is (9 − 1) / (9 + 15 − 2) = 8 / 22 ≈ 0.364.
Posterior mode ≈ 0.364
4
Step 4 — Posterior MedianThere is no closed-form expression for the median of a Beta distribution; we evaluate numerically. Using the inverse Beta CDF, the median satisfies F(m; 9, 15) = 0.5, giving m ≈ 0.371. As expected the median falls between the mode and the mean for this slightly right-skewed density.
Posterior median ≈ 0.371
5
Step 5 — 95% Equal-Tailed Credible IntervalWe compute the 2.5th and 97.5th percentiles of Beta(9, 15): q₀.₀₂₅ ≈ 0.199 and q₀.₉₇₅ ≈ 0.571. Therefore the equal-tailed 95% credible interval is [0.199, 0.571], with width ≈ 0.372.
ET 95% CI: [0.199, 0.571]
6
Step 6 — 95% HPD IntervalWe seek the shortest interval [l, u] such that P(l ≤ θ ≤ u | y) = 0.95 and π(l | y) = π(u | y). For Beta(9, 15) this can be found numerically; the result is approximately [0.190, 0.563], with width ≈ 0.373. Since the Beta(9, 15) is only mildly skewed, the HPD and equal-tailed intervals are nearly identical. With greater skewness the HPD interval would be noticeably shorter.
HPD 95% CI: [0.190, 0.563]

Credible Intervals vs. Confidence Intervals

One of the most persistent sources of confusion in applied statistics is the distinction between a Bayesian credible interval and a frequentist confidence interval. Although both produce numeric intervals and can even coincide numerically in certain conjugate settings, they answer fundamentally different questions and carry distinct philosophical interpretations. Understanding this difference is essential for reporting results accurately and for choosing the right framework for a given inferential goal.

Bayesian credible intervals vs. frequentist confidence intervals
AspectBayesian Credible IntervalFrequentist Confidence Interval
Probability statementP(θ ∈ C | y) = 1 − α. The parameter has a (1 − α) probability of lying in C given the observed data.Under repeated sampling, (1 − α)×100% of constructed intervals contain the true θ. No probability statement about a specific interval.
θ treated asRandom (has a distribution reflecting uncertainty).Fixed but unknown (randomness comes from the data).
Depends on prior?Yes — the prior influences the posterior and hence the interval.No — the interval is determined solely by the sampling distribution of the estimator.
InterpretationDirect: 'There is a 95% probability θ is in this interval.'Indirect: 'The procedure covers θ 95% of the time.'
Numerical agreementCoincides with confidence interval under flat priors and regular likelihoods (Bernstein–von Mises).Can differ from credible intervals, especially in small samples or with informative priors.
KEY TAKEAWAY
Imagine you've lost your keys somewhere in your apartment. A credible interval says: 'Given everything I know—the layout, where I usually drop them, my memory of today—there's a 95% chance the keys are in the living room.' A confidence interval says: 'If I repeatedly lost my keys and each time searched the same way, 95% of my searched areas would contain them.' The credible interval speaks about this particular instance; the confidence interval speaks about the long-run performance of the search procedure.

Connections to Advanced Bayesian Theory

Posterior summaries and credible intervals serve as the foundation for more advanced Bayesian methods. As models grow in complexity—hierarchical structures, high-dimensional parameter spaces, non-parametric priors—the challenges of summarizing posteriors intensify. This section surveys the connections between the core tools introduced above and their extensions in modern Bayesian practice.

From core posterior summaries to advanced Bayesian methods
Core ConceptAdvanced ExtensionKey Idea
Posterior meanRao-Blackwellized estimatesConditioning MCMC estimates on sufficient statistics reduces variance, yielding sharper posterior mean estimates.
HPD regionLoss-function-based decision regionsGeneralized to arbitrary loss functions and multivariate parameter spaces; HPD becomes a highest-density credible set in ℝᵈ.
Equal-tailed intervalQuantile regression & posterior predictive intervalsThe quantile machinery extends from parameter intervals to prediction intervals over future observations.
MAP estimateVariational inference (VI)VI approximates the posterior with a tractable family; the mode of the variational distribution is a MAP-like summary of the approximate posterior.
Credible interval calibrationPosterior predictive checks & calibration diagnosticsSimulation-based calibration (SBC) verifies that nominal credible interval coverage matches empirical coverage across simulated datasets.

In high-dimensional settings the concept of a credible interval extends to credible regions and marginal credible intervals. A joint 95% HPD region in ℝd is the smallest-volume set with 95% posterior probability—analogous to the one-dimensional HPD interval but often computationally intractable for large d. Practitioners frequently report marginal intervals (the credible interval for each component θj obtained from its marginal posterior), acknowledging that these do not capture dependence between parameters. The Bernstein–von Mises theorem provides an asymptotic bridge: under regularity conditions, the posterior converges to a normal distribution centered at the MLE, so credible and confidence intervals become numerically equivalent as n → ∞.

Practice Problems

PROBLEM 1CONCEPTUAL
Explain why the posterior mean, posterior median, and posterior mode generally give different values when the posterior distribution is skewed. Under what condition do all three coincide?
PROBLEM 2BASIC CALCULATION
Suppose θ | y ∼ Beta(4, 8). Compute the posterior mean, posterior mode, and the 90% equal-tailed credible interval using the Beta distribution properties. (Recall: mean = a/(a+b), mode = (a−1)/(a+b−2) for a,b > 1, and use the Beta inverse CDF for quantiles.)
PROBLEM 3INTERMEDIATE
A researcher collects n = 50 observations from a Normal(μ, σ² = 4) population (σ² known) and obtains a sample mean of x̄ = 12.3. With a Normal(10, 9) prior on μ, derive the posterior distribution and compute the posterior mean, the 95% equal-tailed credible interval, and explain whether the HPD interval would differ.
PROBLEM 4APPLIED
A clinical trial tests a new drug. Out of n = 100 patients, y = 62 respond positively. Using a Beta(1, 1) prior (uniform on [0,1]) on the response probability θ, compute the posterior distribution, the posterior mean, and the 95% HPD interval. A regulatory agency requires that the lower bound of the 95% credible interval exceed 0.50 for approval. Would the drug be approved under this criterion?
PROBLEM 5CRITICAL THINKING
Consider a bimodal posterior arising from a two-component mixture: π(θ | y) = 0.6 × N(2, 0.25) + 0.4 × N(5, 0.25). Discuss qualitatively what happens to (a) the posterior mean as a point estimate, (b) the 95% equal-tailed credible interval, and (c) the 95% HPD region. Which summary is most informative? Could reporting only a single point estimate and a single interval be misleading here?

Summary

The posterior distribution π(θ | y) encapsulates all updated knowledge about a parameter θ after observing data y, but communicating its full shape is often impractical. Point summaries reduce the posterior to a single number: the posterior mean (optimal under squared-error loss), the posterior median (optimal under absolute-error loss), and the posterior mode (MAP) (optimal under 0-1 loss). These estimators coincide for symmetric unimodal posteriors but diverge under skewness or multimodality.

Credible intervals provide interval summaries with a direct probability interpretation: a 95% credible interval contains the parameter with posterior probability 0.95. The equal-tailed interval places equal probability in each tail and is invariant to monotone reparameterization, while the highest posterior density (HPD) region is the shortest credible set and can reveal multimodality by splitting into disjoint pieces. Choosing among these tools depends on the loss function relevant to the problem, the posterior's shape, and the communication goals of the analysis. Together, posterior summaries and credible intervals form the essential toolkit for translating Bayesian inference into actionable conclusions.

Varsity Tutors • Statistics Graduate Level • Posterior Summaries & Credible Intervals