Diagnosing Multimodal and Boundary Individual Posteriors

Why Posterior Shape Matters

An individual PK posterior is not always a single, roughly symmetric peak. Sparse samples, nonlinear exposure relationships, and trade-offs among individual parameters can produce distinct regions of high posterior probability. A posterior may also put substantial probability close to a parameter’s allowed minimum or maximum. Both features matter because a point estimate can conceal uncertainty about which parameter values remain plausible, while a forecast can change sharply across those values.

The concern is practical: if two plausible parameter regions imply very different exposure, a forecast based only on one summary parameter may poorly represent the information in the data. This is relevant to the exposure forecasts and dose decisions developed in earlier material on predicting AUC and Cmax from individual posteriors and posterior probabilities of meeting an exposure target. The task here is to diagnose the posterior shape and trace its consequences, not to replace the PK model or the clinical decision rule.

Definition: Let \(y_i\) denote an individual’s observed data and \(\theta_i\) that individual’s PK parameters. The individual posterior \(p(\theta_i\mid y_i)\) is multimodal when its density has multiple distinct regions of local maximum separated by regions of lower density. It is boundary-concentrated when substantial posterior probability lies near a boundary of the allowed parameter space, such as a lower bound on clearance. A local maximum at a boundary is not necessarily a separate interior mode.

From Posterior Shape to Forecast Uncertainty

A posterior forecast for a future quantity \(Z\), such as AUC or Cmax under a proposed dose, averages over parameter uncertainty and future observation uncertainty. Its distribution is obtained by integrating the conditional forecast distribution over the individual posterior:

p(z\mid y_i)=\int p(z\mid \theta_i,y_i)\,p(\theta_i\mid y_i)\,d\theta_i.

If the posterior has distinct modes, this integral combines forecasts from each mode. The modes need not have similar consequences: different parameter combinations can fit the observed concentrations comparably well yet imply very different future exposures. Replacing the posterior with a single point can discard this between-mode uncertainty.

The law of total variance makes the distinction precise. If \(Z\) is a future forecast quantity, its posterior predictive variance is the average conditional variance plus the variance of the conditional means:

\operatorname{Var}(Z\mid y_i) = E\!\left[\operatorname{Var}(Z\mid\theta_i,y_i)\mid y_i\right] + \operatorname{Var}\!\left(E[Z\mid\theta_i,y_i]\mid y_i\right).

The first term reflects future variability conditional on parameters; the second reflects uncertainty about the parameters themselves. With separated modes, the second term can be large even when each mode is individually narrow. A forecast conditional on one mode omits that uncertainty.

Worked Example: Two Clearance Modes, Two AUC Forecasts

Suppose a hypothetical sparse concentration record gives a posterior approximated by two equally weighted, narrow regions for clearance: \(CL_i=2\) L/h and \(CL_i=8\) L/h. For a hypothetical 400 mg dose under linear PK, assume the AUC over the relevant dosing interval is \(AUC= D/CL_i\), with the dose and clearance units giving AUC in mg·h/L. This relation assumes dose-proportional exposure and that clearance is the relevant determinant of AUC.

At the first posterior region:

AUC_1=\frac{400\ \mathrm{mg}}{2\ \mathrm{L/h}}=200\ \mathrm{mg\cdot h/L}.

At the second region:

AUC_2=\frac{400\ \mathrm{mg}}{8\ \mathrm{L/h}}=50\ \mathrm{mg\cdot h/L}.

With posterior probability 0.5 on each region, the posterior mean clearance is \(0.5(2)+0.5(8)=5\) L/h. Plugging this mean into the exposure equation gives \(400/5=80\) mg·h/L. But the posterior mean AUC is \(0.5(200)+0.5(50)=125\) mg·h/L. The discrepancy arises because exposure is nonlinear in clearance: the average of \(D/CL_i\) is not generally \(D\) divided by the average clearance.

Under this simplified two-point approximation, the AUC forecast has probability 0.5 at 50 and 0.5 at 200 mg·h/L. Thus, for a hypothetical target range of 80–120 mg·h/L, the probability of being in range is 0. The plug-in forecast of 80 mg·h/L would misleadingly place the prediction at the lower edge of that range. Continuous posterior regions and conditional predictive variability would smooth the point masses, but would not justify replacing the full forecast distribution with the plug-in value.

Diagnosing Multimodality and Boundary Concentration

A posterior plot should be examined on scales that reveal both marginal shape and dependence among parameters. A one-dimensional marginal may appear unimodal while a joint posterior contains a curved ridge or separated parameter combinations. Conversely, a histogram with multiple bumps can result from too few samples, poor mixing, or an inappropriate plotting scale rather than genuine posterior modes.

For multimodality, inspect trace plots and marginal and joint plots, and check whether independently initialized chains visit the same regions. Record how often chains move between apparent modes; a chain that remains in one region cannot by itself establish the relative posterior mass of the modes. Simulation-based calibration, covered in the earlier material on simulation-based calibration for Bayesian forecasting, assesses computational calibration across simulated data sets, but it does not make a poorly explored fit trustworthy. Posterior predictive checks and residual diagnostics, covered earlier, help assess fit to the observed data; adequate fit alone does not demonstrate that all posterior modes have been explored.

For a boundary, first state the parameter’s allowed range and examine posterior probability near its limits. A pile-up near a lower bound can reflect information in the data, a constraint imposed by the model, prior influence, or inadequate exploration. The pattern should be considered on the parameter scale used in the model and, where useful, on a scientifically meaningful transformed scale. A density’s mode is not invariant to reparameterization: changing variables changes the density through the transformation’s Jacobian. Therefore, a boundary mode or a reported MAP value must be interpreted on its stated parameter scale.

Check posterior exploration. Compare chains, starting points, and mode occupancy; do not infer mode weights from a chain that never switches regions.
Check joint structure. Plot scientifically important parameter pairs, especially where one parameter can compensate for another in predicting concentrations.
Check the support. Mark the allowed bounds and quantify posterior mass in neighborhoods of them, rather than relying only on a reported point estimate.
Check consequences. Calculate forecasts and target probabilities across the distinct regions or near-boundary values, not only at a pooled summary.

Worked Example: A Clearance Posterior Near Its Lower Bound

Consider a hypothetical model in which clearance is constrained to be at least 0.2 L/h. For transparent arithmetic, suppose posterior summaries assign probability 0.6 to a narrow region represented by \(CL_i=0.2\) L/h and probability 0.4 to a region represented by \(CL_i=1.2\) L/h. This two-region approximation represents concentration near the allowed lower bound; it does not assert that a continuous posterior has a point mass exactly at the bound.

For a hypothetical 120 mg dose, assume the same linear-PK AUC relation, \(AUC=D/CL_i\). The two representative forecasts are:

AUC_1=\frac{120}{0.2}=600\ \mathrm{mg\cdot h/L}, \qquad AUC_2=\frac{120}{1.2}=100\ \mathrm{mg\cdot h/L}.

The posterior mean clearance is \(0.6(0.2)+0.4(1.2)=0.6\) L/h, so the plug-in AUC is \(120/0.6=200\) mg·h/L. In contrast, the posterior mean of the two forecast values is \(0.6(600)+0.4(100)=400\) mg·h/L. For a hypothetical target range of 100–250 mg·h/L, the two-region forecast assigns probability 0.4 to being in range, while the plug-in forecast of 200 mg·h/L is in range. The difference matters for a decision based on the probability of meeting a target.

The posterior’s near-boundary region drives a high-exposure forecast because AUC increases as clearance decreases. This example does not establish that the lower-bound values are biologically correct; it shows why the boundary, its posterior mass, and the exposure consequences should be reported and investigated.

Worked Forecast Consequences and Model Checks

Posterior summaries are useful only when their limitations are clear. A posterior mean summarizes the center under squared-error loss, and a posterior mode identifies a density maximum on a specified scale. Neither is a substitute for the full posterior when a decision depends on a nonlinear forecast or a target probability. A single Gaussian approximation may be convenient, but it can merge separated modes into a region with little posterior density and can obscure skewness or boundary concentration. Posterior predictive integration preserves these features if the posterior computation has explored them adequately.

Worked Example: A Mean Forecast Inside a Target but No Modal Forecast Inside

Suppose a hypothetical individual posterior has two narrow clearance regions, \(CL_i=1\) L/h with probability 0.3 and \(CL_i=5\) L/h with probability 0.7. For a hypothetical 100 mg dose, assume linear PK and \(AUC=D/CL_i\), as above. Then the two AUC forecasts are \(100/1=100\) and \(100/5=20\) mg·h/L.

The posterior mean AUC is \(0.3(100)+0.7(20)=44\) mg·h/L. The posterior mean clearance is \(0.3(1)+0.7(5)=3.8\) L/h, whose plug-in AUC is \(100/3.8=26.32\) mg·h/L. These are different summaries of different quantities; neither is the probability of meeting a target.

For a hypothetical target range of 30–60 mg·h/L, neither modal AUC value lies in range, so the two-region posterior probability of being in range is 0. Yet the posterior mean AUC, 44 mg·h/L, lies in the target range. A statement that “the forecast is in range” based solely on that mean would therefore give the wrong description of this posterior. The relevant probability is obtained by counting the posterior predictive mass in the target range, including any conditional forecast variability if it is part of the model.

The assumed PK relation and the stated two-region posterior make the arithmetic explicit. In an applied fit, inspect whether these regions are supported by well-explored posterior draws and whether each predicts the observed concentrations adequately. A mathematically computed target probability is not reliable if the posterior sampler has missed a mode.

Comparisons, Limitations, and Common Pitfalls

Approach	What it summarizes	Risk with multimodal or boundary posteriors
Posterior mean or mode	A single parameter value on a stated scale	Can hide separated regions; a mean can lie between modes, and a mode can depend on parameterization
Single Gaussian approximation	A center and covariance for one approximately elliptical region	Can smooth away modes, skewness, or truncation at a boundary
Posterior predictive integration	Forecast distribution averaged over posterior parameter uncertainty	Reflects the posterior shape only if the posterior computation explores it adequately
Mode-specific forecasts	Forecast consequences conditional on each distinct region	Do not by themselves provide mode probabilities or a complete predictive distribution

A common mistake is to treat a converged-looking trace plot or a plausible point estimate as proof that the posterior is unimodal. A chain may appear stable because it is trapped in one mode, and a mean may appear precise while averaging across regions with sharply different clinical implications. The correction is to check exploration across chains and modes, inspect the joint posterior and boundaries, and report forecast distributions or target probabilities that retain the relevant uncertainty.

Boundary concentration also has limitations as a diagnostic. A posterior near a constraint does not automatically indicate a faulty model: data can support values close to a legitimate physical or clinical limit. But strong sensitivity to the bound, prior, or parameter scale is a reason to qualify interpretation. Likewise, a multimodal posterior is not automatically evidence of a biological subgroup; it may arise from parameter non-identifiability or the geometry of the likelihood. Distinguishing these explanations requires model and data scrutiny, not a point estimate alone.

Key takeaway: Diagnose posterior shape before reducing it to a point. When distinct modes or a boundary materially change exposure, evaluate the posterior predictive distribution and the clinical target probability across those regions, and verify that posterior computation has explored them.

Check Your Understanding

Use the stated assumptions in each question. The solutions follow each problem.

A posterior has two equally weighted clearance regions at 2 and 6 L/h. Under a 120 mg dose and \(AUC=D/CL_i\), what are the two AUC forecasts and their posterior mean? Solution: They are \(120/2=60\) and \(120/6=20\) mg·h/L. Their posterior mean is \(0.5(60)+0.5(20)=40\) mg·h/L.
Why can a posterior mean clearance give a different AUC forecast from the posterior mean AUC? Solution: Because \(D/CL_i\) is nonlinear in clearance, so in general \(E[D/CL_i\mid y_i]\ne D/E[CL_i\mid y_i]\). The first averages exposure over the posterior; the second forecasts at the mean clearance.
One chain remains in a low-clearance region and another remains in a high-clearance region. Is their pooled sample fraction sufficient to establish the posterior probability of each mode? Solution: Not by itself. Without adequate exploration and movement between regions, the chains may not estimate relative mode mass. Check initialization, chain behavior, and whether the posterior computation explores both regions.
A clearance estimate lies close to its allowed lower bound. Name two checks before interpreting the resulting high-exposure forecast. Solution: Confirm the parameter support and quantify posterior mass near the bound; then assess sensitivity to the bound or prior and verify that posterior sampling has adequately explored the region. Also inspect how forecasts change across plausible values near the boundary.