Posterior Intervals for Variance Components Near Zero

Why Intervals Near Zero Need Care

A variance component in a generalized linear mixed model describes variation among groups or other clusters. Write \(\tau^2\) for one such variance and \(\tau\) for its standard deviation. Since \(\tau^2 \geq 0\), a posterior distribution concentrated near zero is constrained by a boundary. It may also be strongly right-skewed: many plausible values can be small, while a smaller set of larger values extends the upper tail.

In that setting, a single estimate can hide important features of the posterior. The posterior mean may be pulled upward by the tail, while a symmetric interval can extend below the allowable value zero. Credible intervals make uncertainty visible, but their construction matters: equal-tailed and highest-posterior-density intervals answer related, not identical, questions.

The intervals here are conditional on the specified likelihood, prior, and observed data. For a continuous prior on \(\tau^2\), the posterior is also continuous, so the posterior probability of the exact value \(\tau^2=0\) is zero. An interval that includes zero does not, by itself, assign positive posterior probability to an exact zero variance.

Two Posterior Interval Constructions

For \(\tau^2\), let \(F\) denote its posterior cumulative distribution function. The equal-tailed interval is

Its lower and upper tails each have posterior probability \(\alpha/2\). This construction is defined by posterior quantiles and remains available for skewed distributions. If the posterior is continuous and has support on \([0,\infty)\), its lower endpoint cannot be below zero.

An HPD set is defined by a density threshold: it includes values where the posterior density is at least that threshold, with the threshold chosen so that the set has posterior probability \(1-\alpha\). If the density is unimodal and an HPD interval has both endpoints in the interior of the parameter space, the endpoint densities are typically equal, with higher density inside. If the interval includes a boundary, the boundary endpoint need not have the same density. More generally, HPD sets comprise values above a density threshold, subject to the required probability. In a multimodal posterior, an HPD set may be disconnected rather than a single interval.

A Boundary Case: A Decreasing Posterior Density

A simple posterior illustrates why the boundary changes the HPD construction. Suppose the posterior distribution of \(\tau^2\) is exponential with rate \(\lambda\), so its density and cumulative distribution function for \(v \geq 0\) are

The density decreases from its maximum at \(v=0\). Thus, the 95% HPD set starts at zero: the highest-density values are those nearest the boundary. To find its upper endpoint \(U\), set the posterior probability from zero to \(U\) equal to 0.95:

The equal-tailed interval has different endpoints. Solving \(F(L\mid y)=0.025\) and \(F(U\mid y)=0.975\) gives \(L=-\log(0.975)/\lambda\) and \(U=-\log(0.025)/\lambda\). Both intervals contain 95% posterior probability, but they allocate probability differently. In particular, the HPD interval includes the boundary and does not require equal density at its two endpoints.

Interpreting Posterior Intervals on Different Scales

Variance and standard deviation describe the same component on different scales. Because the square-root function is increasing on \([0,\infty)\), transforming the endpoints of a posterior interval for \(\tau^2\) gives the corresponding interval for \(\tau\), provided the transformation is applied to the posterior distribution consistently. In particular, posterior quantiles transform as

This relationship does not mean that summaries such as means can be transformed in the same way: in general, \(E(\tau\mid y)\) is not \(\sqrt{E(\tau^2\mid y)}\). Nor should a standard deviation interval be reported with variance units. The scale and units should always be named.

Choosing and Reporting an Interval

Equal-tailed intervals are often straightforward to calculate from posterior draws: estimate the two specified quantiles. They are especially useful when the goal is to report how much posterior probability lies in each tail. An HPD interval emphasizes the region of highest posterior density and is often shorter than an equal-tailed interval for a unimodal posterior. However, HPD intervals depend on the parameter scale: a nonlinear transformation can change which values have the highest density. Therefore, calculate an HPD interval on the scale whose density-based interpretation you intend to report; do not assume that transforming an HPD interval preserves its HPD property.

For intervals estimated from MCMC draws, the quantiles are estimates, not exact values. Check that the chains mix adequately for the variance component and that effective sample size is sufficient for the tails being reported. Tail quantiles can be noisy even when the posterior median appears stable. Monte Carlo uncertainty in an endpoint should not be mistaken for posterior uncertainty in the parameter itself.

Diagnostics, Alternatives, and Limitations

A posterior interval is only as reliable as the posterior computation and model behind it. A credible interval close to zero can reflect genuine weak information about a variance component, but it can also be affected by poor exploration of the posterior. Look for chains that move through the relevant range, and examine the stability of both lower and upper quantiles across chains or longer runs. A pile-up of sampled values at exactly zero deserves particular attention: it may arise from a model with a point mass at zero, a constrained computational representation, or an implementation issue. Those cases have different interpretations and should not be conflated with a continuous posterior that merely concentrates near zero.

Frequentist confidence intervals and Bayesian credible intervals use different probability statements and depend on different assumptions. A frequentist interval is defined by the long-run coverage of its procedure under repeated sampling; a Bayesian credible interval summarizes posterior probability conditional on the model, prior, and observed data. Standard symmetric approximations can be awkward for a nonnegative variance close to its boundary. A Bayesian interval avoids reporting negative values when its posterior has nonnegative support, but it does not remove sensitivity to the prior, likelihood, or mixing of the posterior computation.

Report the interval type, credibility level, parameter scale, and units. When the posterior is very skewed, the median and interval often communicate location and uncertainty more clearly than the mean and a symmetric range. If practical interpretation depends on whether the standard deviation is below a scientifically meaningful threshold, report that probability separately and state the threshold and units. Do not describe a near-zero interval as proof of no variation.