Prior Sensitivity Analysis for Variance Components

Why Examine Prior Sensitivity?

In a Bayesian nonlinear mixed-effects model, a prior on the random-effect covariance structure can influence inference, particularly when the data provide limited information about one or more variance components or correlations. Prior sensitivity analysis asks a focused question: if scientifically defensible prior choices were changed, would the substantive posterior conclusions change? It is not a search for the prior that produces the preferred answer. It is a transparent check on how much the analysis depends on assumptions that are not supplied by the likelihood.

For subject \(i\), write the individual parameters as \(\phi_i=f(\theta,b_i)\), where \(\theta\) denotes population parameters and \(b_i\) is a vector of random effects. In the usual multivariate normal random-effects model, \(b_i\mid\Omega\sim N(0,\Omega)\). A convenient decomposition is \(\Omega=D R D\), where \(D=\operatorname{diag}(\omega_1,\ldots,\omega_q)\) contains random-effect standard deviations and \(R\) is a correlation matrix. The prior sensitivity question here concerns alternative plausible priors on the \(\omega_k\) and on \(R\), while keeping the data, likelihood, and other model choices fixed.

Specify and Calibrate Alternative Priors

A prior should be interpreted in the units and parameterization of the model. If the random effect is on log clearance, for example, its standard deviation is on the log scale, not in clearance units. A prior that looks numerically small or large cannot be judged without checking that scale. For a standard deviation \(\omega>0\), common possibilities include a half-normal distribution, a half-\(t\) distribution, or a prior on \(\log\omega\). These choices differ in how much probability they place near zero and in their upper tails. None is automatically suitable in every application.

For correlations, one common choice is an LKJ prior on the correlation matrix \(R\), with density proportional to \(|R|^{\eta-1}\), where \(\eta>0\). In dimension two, this induces a density for the single correlation \(\rho\) proportional to \((1-\rho^2)^{\eta-1}\) on \((-1,1)\). When \(\eta=1\), the correlation is uniform; larger \(\eta\) places relatively more weight near zero. This prior choice is distinct from a prior on the standard deviations: changing one does not require changing the other.

Priors for covariance parameters should be stated clearly, including their parameterization and scale. Proper priors generally help ensure a proper posterior, but a proper prior is not necessary for a proper posterior. An improper prior can yield a proper posterior; whenever an improper prior is used, verify that the resulting posterior is proper. Proper priors also make prior predictive implications easier to examine and can support meaningful comparisons between prior specifications.

Repeat the Inference and Compare Like With Like

A fair sensitivity comparison changes the prior specification of interest and leaves the likelihood and data unchanged. Use the same random-effects structure, residual model, covariate definitions, and observation model in each fit. If several prior features are changed simultaneously, a posterior difference may be hard to attribute; when practical, compare one change at a time as well as any justified combined alternatives. The alternatives should be selected before inspecting which produces the most appealing estimate.

For prior specification \(j\), posterior inference follows \(p_j(\theta,\Omega\mid y)\propto p(y\mid\theta,\Omega)p_j(\theta,\Omega)\), with all other parameters included as needed. A useful justification for comparing specifications is the identity

where \(p_j(y)\) and \(p_k(y)\) are the marginal likelihoods under the respective priors. For fixed data and likelihood, the prior density ratio is the source of differences in posterior weighting; the marginal-likelihood ratio is a constant in the parameters that normalizes each posterior. The identity does not require calculating marginal likelihoods to do sensitivity analysis. It clarifies why the same likelihood can yield different posterior distributions under different priors.

Compare quantities relevant to the scientific question: posterior medians or means, credible intervals, probabilities of decision-relevant events, and predictions. Report the priors, parameter scales, and posterior summaries together. Differences should be interpreted against both Monte Carlo error and a meaningful scientific scale. A small numerical difference relative to sampling error is not persuasive evidence of sensitivity; conversely, a material shift that exceeds Monte Carlo error should not be dismissed simply because both fits converged.

Predictive Consequences and a Short Derivation

A change in a variance-component posterior can matter even when the population typical value barely changes, because it changes the distribution of individual parameters. For a lognormal random effect \(b\sim N(0,\omega^2)\), the individual-to-typical parameter multiplier is \(\exp(b)\). Its conditional mean and variance, given \(\omega\), are

These expressions follow from the normal moment-generating function: for \(b\sim N(0,\omega^2)\), \(E[\exp(tb)]=\exp(t^2\omega^2/2)\). Taking \(t=1\) gives the mean; taking \(t=2\) gives the second moment \(\exp(2\omega^2)\). Subtracting the squared mean yields \(\exp(2\omega^2)-\exp(\omega^2)=\exp(\omega^2)\{\exp(\omega^2)-1\}\).

More generally, if \(Y_{\mathrm{new}}\) is a future observation and \(\psi\) denotes all model parameters, the posterior predictive distribution is \(p(y_{\mathrm{new}}\mid y)=\int p(y_{\mathrm{new}}\mid\psi)p(\psi\mid y)\,d\psi\). Its variance separates into expected conditional variance and variance of the conditional mean:

A prior sensitivity comparison can therefore assess predicted concentrations or individual parameter values by generating predictions from each fit, using the same target population and dosing scenario. Compare quantities such as predictive medians, intervals, or probabilities of exceeding a clinically relevant threshold. A change in \(\omega\) need not translate uniformly into a change in every prediction; it depends on the model’s nonlinear mapping from individual parameters to observations.

Diagnostics, Alternatives, and Limitations

Before attributing differences to priors, establish that each fit samples its posterior adequately. Check trace plots, rank-normalized \(\hat R\), effective sample sizes, Monte Carlo standard errors, and HMC warnings such as divergent transitions. Poor mixing or unresolved divergences can make summaries unreliable; rerunning with better sampling is not itself a prior sensitivity analysis. Compare prior predictive implications as well: priors that routinely imply implausible individual variability or concentration profiles may be rejected on scientific grounds before fitting.

Common signs that a comparison needs closer examination include posterior intervals that shift substantially between reasonable priors, posterior mass near zero that changes materially, correlations whose sign or magnitude changes, and predictions that differ enough to alter a scientific decision. These patterns are evidence that conclusions are prior-dependent under the tested specifications, not proof of a particular cause. If alternative priors yield similar summaries and predictions, report that robustness with the range of priors examined rather than claiming universal insensitivity.

Prior sensitivity is not the same as choosing a prior by fit quality. Frequentist likelihood-based estimates are an alternative inferential framework, but they do not answer the same posterior-probability questions. Within Bayesian analysis, one can also compare parameterizations or jointly structured priors, but such changes should be described because they may alter the prior induced on \(\Omega\). Sensitivity analysis has limits: a finite set of priors cannot represent every plausible belief, and agreement across a narrow or implausible set offers little reassurance. It also cannot repair a misspecified likelihood or establish that the model predicts well.