Parameter Correlation in Individual Bayesian Fits

Why Parameter Correlation Matters for a Forecast

An individual Bayesian fit produces uncertainty about a patient’s pharmacokinetic parameters, not just a single clearance or volume estimate. That uncertainty is generally joint: parameter values that are plausible together may not be plausible in every combination. For example, a patient’s posterior may give relatively high clearance values greater support when paired with relatively high volume values. A forecast that combines marginal parameter ranges independently can then describe parameter combinations poorly supported by the patient’s data and prior.

This matters because a future concentration is a function of several parameters at once. If uncertainty in two parameters moves the forecast in the same direction, their positive posterior correlation can increase forecast uncertainty. If their effects oppose one another, a negative correlation can increase uncertainty instead. The direction and size of the effect depend on the forecast time and dosing context; a correlation cannot be labelled universally “good” or “bad” for precision.

The earlier material on identifiability of clearance and volume from sparse levels establishes why data may constrain combinations of parameters more strongly than each parameter separately. Here the focus is how that dependence appears in the individual posterior and how it changes forecasts.

Posterior Covariance and Correlation

For a parameter vector \(\theta=(CL,V,k_a,F)\), the posterior covariance matrix, \(\Sigma\), contains the variance of each parameter on its diagonal and covariances between pairs off the diagonal. It must be positive semidefinite, but it need not be diagonal. A pairwise correlation describes only a two-parameter summary: it does not fully describe a nonlinear, skewed, or multimodal joint posterior.

Correlations depend on parameterization. A correlation computed for \(CL\) and \(V\) need not equal the correlation for their logarithms. In pharmacokinetic models, parameters such as \(CL\), \(V\), and \(k_a\) are positive, while \(F\) is bounded between zero and one. Posterior summaries may therefore use logarithms for positive parameters and a logit transformation for \(F\). Always identify the scale on which a reported correlation was calculated. A strong correlation on a transformed scale is not, by itself, evidence of a fixed biological relationship between parameters; it describes dependence in the fitted posterior, conditional on the model, prior, and patient data.

How Dependence Enters Forecast Uncertainty

Let \(g(\theta)\) denote a predicted concentration at a specified future time under a specified dosing history. Near a representative parameter value, a first-order Taylor approximation gives \(g(\theta)\approx g(\theta_0)+\nabla g(\theta_0)^\mathsf{T}(\theta-\theta_0)\). Consequently, a local approximation to the variance of the forecast due to parameter uncertainty is \(\operatorname{Var}(g(\theta)\mid\text{data})\approx\nabla g^\mathsf{T}\Sigma\nabla g\). For two parameters \(A\) and \(B\), this includes the term \(2(\partial g/\partial A)(\partial g/\partial B)\operatorname{Cov}(A,B)\), in addition to the separate variance contributions.

The covariance contribution is positive when the parameter covariance and the product of the forecast sensitivities have the same sign; it is negative when their signs differ. Thus a positive parameter correlation can either widen or narrow a particular forecast distribution. This local approximation is useful for understanding the mechanism, but may be inaccurate when the forecast is strongly nonlinear or the posterior is wide or non-normal. Drawing complete parameter vectors from the joint posterior and calculating a forecast for each draw handles such cases more directly.

Correlations Among Clearance, Volume, Absorption, and Bioavailability

The four parameters can influence an oral concentration forecast in distinct but overlapping ways. Clearance affects elimination over time; volume affects the concentration scale and, through \(CL/V\), the elimination rate; \(k_a\) affects the absorption-time profile; and \(F\) scales the amount reaching systemic circulation. Because a concentration depends jointly on them, the posterior can contain dependence among any of these parameters. The data and prior determine the pattern for a particular patient and model.

For illustration, in a one-compartment model with first-order absorption and no lag time, the concentration after a single oral dose is \(C(t)=FDk_a/[V(k_a-k)]\{\exp(-kt)-\exp(-k_at)\}\), where \(k=CL/V\) and \(k_a\ne k\). This formula makes clear that changing one parameter can alter both the scale and shape of the concentration-time curve. The effect of changing \(k_a\), in particular, depends on time: it changes the absorption phase and need not have the same direction or magnitude of effect at every time point.

Using Joint Posterior Draws and Checking the Result

A direct forecasting procedure is to retain the complete parameter vector from every posterior draw. For draw \(m\), calculate the concentration under the intended future dosing history using that draw’s \(CL\), \(V\), \(k_a\), and \(F\), as applicable. The resulting set of concentrations approximates the posterior distribution of the model-predicted concentration. If the target is a future measured concentration rather than the underlying model-predicted concentration, the appropriate observation-error component must also be included. The distinction is between uncertainty about the model’s expected concentration and uncertainty about an individual future observation.

Before interpreting a correlation, check that posterior computation has adequately explored the distribution. Inspect trace plots and marginal and pairwise plots, and assess whether the sampler has mixed adequately and produced enough effective draws for the summaries of interest. A tilted cloud can suggest pairwise dependence, while separate clusters or curved bands warn that one correlation coefficient may be a poor description. These checks matter because an apparent correlation estimated from a poorly explored chain may reflect sampling failure rather than the posterior.

Also check that the reported correlations and forecasts correspond to the intended parameterization, model, and dosing history. A correlation matrix is a summary, not a model-validation test: a plausible-looking matrix does not establish that the pharmacokinetic model fits the patient data. Nor does a strong correlation alone identify which additional data would be most useful; that requires examining the specific design and estimation question.

Common Pitfall and Practical Limits

Wrong reasoning: “Clearance and volume are positively correlated, so uncertainty in both must make every concentration forecast less precise.” This conclusion ignores forecast sensitivities. In the IV bolus example, their sensitivities had the same sign at 4 h but opposite signs at 15 h, so the covariance term changed sign. The correction is to assess the joint uncertainty for the particular forecast, using the covariance and the parameter sensitivities or, preferably for nonlinear cases, complete posterior draws.

Another common mistake is to report two marginal credible intervals and assume they describe the joint posterior. They do not: separate intervals omit which parameter combinations occur together. Likewise, plugging in posterior means or MAP estimates gives a point forecast, not a forecast distribution, and generally does not convey uncertainty. A diagonal covariance approximation or independent resampling can be useful as a deliberately simplified comparison, but should not replace the fitted joint posterior when dependence is material.

Correlation has limits as a summary. It is scale-dependent, captures only linear pairwise association, and can obscure nonlinear dependence or multiple posterior modes. It also depends on the assumed model and prior; it is not a model-free property of the patient. Joint-draw forecasts address dependence more faithfully, but they cannot correct misspecification, unrepresentative data, or inadequate posterior sampling. Forecast uncertainty should therefore be interpreted as conditional on the model, prior, and available patient information.