Summarizing a Posterior Distribution: Point Estimates, Credible Intervals, and HDI vs. Equal-Tailed Intervals

A Bayesian analysis rarely stops at producing a posterior distribution — the full, continuous curve over a parameter’s possible values after combining a prior with observed data via Bayes’ theorem. A posterior is not something anyone reports directly in a results section; it has to be reduced to a handful of numbers a reader can actually use. Which numbers, and how they are computed, is not a cosmetic choice. For a symmetric, bell-shaped posterior almost any reasonable summary agrees. For a skewed posterior — and a large share of the parameters researchers actually estimate (rates, variances, ratios, anything bounded near zero) produce skewed posteriors — the choice of summary changes the number a reader sees, sometimes by a meaningful amount.

Point summaries: mean, median, and mode

Three single-number summaries are in routine use, and they answer three different questions:

• Posterior mean — the expected value of the parameter under the posterior, E[θ | data]. It is the point estimate that minimizes expected squared error, and it is the most commonly reported summary by default in most Bayesian software output.
• Posterior median — the 50th percentile of the posterior; it minimizes expected absolute error and, unlike the mean, is not pulled toward a long tail.
• Posterior mode (the maximum a posteriori, or MAP, estimate) — the single most probable value, the peak of the posterior density.

For a symmetric, unimodal posterior — the shape a large-sample or well-behaved Gaussian-approximate posterior often takes — the mean, median, and mode coincide, or come close enough that the choice barely matters. That convenient overlap is exactly what breaks down for a skewed posterior, where the three summaries can diverge enough to tell noticeably different stories about “the” estimated value.

Why a skewed posterior pulls the summaries apart

As an illustration, not a specific study — consider a Beta(3, 40) posterior, the kind of shape that arises when estimating a small proportion (an adverse-event rate, a defect rate, a rare-response rate) from a handful of observed occurrences. Its mode — the peak of the density, computed exactly as (α−1)/(α+β−2) for a Beta distribution — sits at 2/41 ≈ 0.049. Its mean, computed exactly as α/(α+β), sits further out at 3/43 ≈ 0.070. The median has no closed form for a Beta distribution, but for a right-skewed unimodal posterior like this one it typically falls between the mode and the mean, not at either endpoint. Report “the estimate” as the mode and a reader sees roughly 4.9%; report it as the mean and they see roughly 7.0% — a real difference in a low-base-rate context, not rounding noise. Neither number is wrong; they are answers to different questions (“most probable single value” vs. “expected value”), and a methods section that doesn’t say which one it used is leaving that choice invisible to the reader.

Credible intervals vs. confidence intervals

A point summary alone discards the posterior’s spread, so Bayesian results are reported with an interval — a credible interval — alongside the point estimate. A 95% credible interval supports a direct probability statement: given the model and the data, there is a 95% probability the parameter lies inside that interval. That is a genuinely different claim from a frequentist confidence interval, whose 95% figure describes the long-run behavior of the interval-construction procedure across repeated sampling, not a probability statement about the parameter itself given this one dataset. The two intervals are also frequently, if confusingly, close in value for simple models with weak priors and large samples — which is exactly why it matters to label the interval correctly rather than call a Bayesian interval a “confidence interval” out of habit. See CASRAI’s frequentist vs. Bayesian comparison for how this distinction plays out in trial design specifically.

Two ways to build a credible interval: equal-tailed vs. HDI

“A 95% credible interval” is not one fixed calculation — there are two standard constructions, and they are not the same interval unless the posterior happens to be symmetric.

Equal-tailed (percentile) intervals

The equal-tailed interval takes the 2.5th and 97.5th percentiles of the posterior for a 95% interval, leaving exactly 2.5% of the probability mass in each tail. It is simple to compute from posterior draws (just two quantiles) and has a useful mathematical property: it is invariant under any monotonic reparameterization — if you transform the parameter (log it, exponentiate it, take its reciprocal), the equal-tailed interval on the new scale is just the old endpoints transformed the same way. It is also always a single, easy-to-report range.

Highest density interval (HDI)

The HDI is instead defined as the narrowest interval containing 95% of the posterior mass. For a unimodal posterior this means every point inside the interval has higher posterior density than every point outside it — the density at both endpoints is equal, and it is the tightest region satisfying that condition. Two consequences follow directly: an HDI is, for a given probability level, never wider than the equal-tailed interval on the same posterior, and for a genuinely multimodal posterior the HDI can be disjoint — two or more separate intervals rather than one continuous range, something an equal-tailed interval can never represent. The trade-off is that the HDI is not invariant under nonlinear reparameterization: computing the HDI, then transforming it, does not generally match computing the HDI directly on the transformed scale.

Why the choice matters specifically for a skewed posterior

For a symmetric unimodal posterior, the equal-tailed interval and the HDI are nearly identical, and reporting either is effectively a stylistic choice. For a skewed posterior — the Beta(3, 40) example above included — the two constructions genuinely disagree. The HDI shifts toward the mode, where the density is highest, and is narrower; the equal-tailed interval, because it insists on exactly 2.5% probability in each tail regardless of density, extends further into the long tail on the skewed side while cutting off some higher-density territory near the short side. A reader comparing an HDI from one paper to an equal-tailed interval from another, without either paper stating which was used, is comparing two different quantities that happen to share the label “95% credible interval.” That is the specific, avoidable failure mode this choice creates.

What to report in a methods or results section

• State which point summary you are reporting — mean, median, or mode — explicitly, not just “the estimate.” For a skewed posterior, reporting the median alongside the mean gives a reader a quick read on how much the mean is being pulled by the tail.
• Name the interval type explicitly: “95% HDI” or “95% equal-tailed credible interval,” never an unqualified “95% CI” for a Bayesian result — that abbreviation reads as a frequentist confidence interval by convention.
• If you used software defaults, check what they actually compute rather than assuming: Bayesian analysis packages do not all default to the same interval construction, and the difference is exactly the one described above.
• Briefly state the prior and model, since the posterior — and every summary of it — is conditional on both.
• For a visibly skewed posterior, consider reporting the interval alongside a simple description of the shape (or a plot) rather than a point estimate and interval alone; the asymmetry itself is information a symmetric-looking “estimate ± margin” write-up would hide.

Common mistakes

• Calling a credible interval a confidence interval. They answer different questions and the terms are not interchangeable, even when the numeric bounds happen to be similar.
• Assuming the interval is symmetric around the point estimate. Neither an HDI nor an equal-tailed interval on a skewed posterior is generally symmetric around the mean, median, or mode — don’t report it as “estimate ± margin” unless the posterior is genuinely close to symmetric.
• Switching interval type mid-analysis without saying so. Comparing an HDI from one model or software run to an equal-tailed interval from another treats two different constructions as directly comparable when they are not.
• Reporting only the mean for a strongly skewed posterior. The mean can sit meaningfully away from both the mode and the bulk of the probability mass; pairing it with the median (or reporting the median instead) avoids overstating precision around a single number.

Frequently asked questions

Is a credible interval the same as a confidence interval?

No. A 95% credible interval supports a direct probability statement about the parameter given the observed data and the model: there is a 95% probability the true value lies in that range. A 95% confidence interval instead describes the long-run coverage rate of the procedure used to construct it across repeated sampling — it does not assign a probability to the parameter itself. They can be numerically close in simple, weak-prior, large-sample cases, but the interpretation is not interchangeable.

Should I report the posterior mean, median, or mode?

For a roughly symmetric posterior it rarely matters which you report. For a skewed posterior, state explicitly which one you’re using, since they can diverge meaningfully — many methodologists favor the median for skewed posteriors because it isn’t pulled by the tail the way the mean is, but the mean remains the more common software default. What matters most is naming the summary, not defaulting silently to whichever your software prints first.

What is the HDI in Bayesian statistics?

HDI stands for highest density interval: the narrowest interval that contains a given proportion (typically 95%) of the posterior probability mass, such that every value inside it has higher posterior density than every value outside it. It differs from an equal-tailed interval, which instead fixes the probability left in each tail (2.5% on each side for a 95% interval) regardless of density.

Why isn’t a 95% credible interval symmetric around the posterior mean?

Because the posterior itself is asymmetric. Both standard interval constructions — the HDI and the equal-tailed interval — are built directly from the shape of the posterior, and neither one imposes symmetry. An interval that is forced to look symmetric around a point estimate (“estimate ± margin”) on a skewed posterior is misrepresenting that posterior’s actual shape.

Related CASRAI resources

• Confidence Interval — the frequentist counterpart’s operational definition, useful as a direct contrast to the credible interval described above.
• Markov Chain Monte Carlo (MCMC) — how the posterior draws used to compute these summaries are actually generated in practice.
• Bayesian Adaptive Design — a clinical-trial context where posterior summaries directly drive interim decisions.
• Frequentist vs. Bayesian Statistics in Clinical Trials — a fuller comparison of the two inferential frameworks.
• Effect Size — the frequentist-side companion concept to a posterior point summary.
• Research Methods hub — CASRAI’s full cluster on study design, sampling, and quantitative and qualitative analysis.