Why Random-Effect Correlations Need Careful Interpretation
Population PK models often allow an individual's clearance and volume, for example, to vary together. A fitted correlation may represent genuine variation across individuals, but it can also appear when the model has difficulty separating parameters or when its structural equations are inadequate. In the latter case, one random effect may partly compensate for another so the predicted concentration profiles fit the observed data.
The goal is not to eliminate correlations simply because they are large, or to retain them simply because they improve the objective function. Instead, examine what the correlation means under the model, whether the data can support it, and whether it persists when plausible structural alternatives are considered. The earlier material on separating structural misfit from random-effect misfit and reading residual patterns provides useful context for this evaluation.
For positive PK parameters, a common model is \(P_{ij}=P_j^{\mathrm{typ}}\exp(\eta_{ij})\), where \(P_j^{\mathrm{typ}}\) is the typical value and \(\eta_{ij}\) is the individual's random effect. In that model, \(\Omega\) describes variation and association on the log-parameter scale. A positive correlation between \(\eta_{\mathrm{CL}}\) and \(\eta_{\mathrm{V}}\) means that individuals with above-typical log clearance tend, under the model, also to have above-typical log volume. It is not a statement that clearance causes volume to increase, nor is it automatically the correlation between the untransformed parameters.
How Compensation Can Produce a Correlation
A model estimates parameters using the information available in the observed concentration-time profiles. If different combinations of individual parameters produce similar predictions over the observed sampling times, estimates can trade off against one another. The population model may then assign correlated random effects to represent those combinations across individuals. This can improve fit without demonstrating that the underlying biological parameters truly covary.
A simple one-compartment intravenous bolus model illustrates the issue. With dose \(D\), clearance \(CL_i\), and volume \(V_i\), the concentration is \(C_i(t)=(D/V_i)\exp[-(CL_i/V_i)t]\). The terminal decay rate depends on the ratio \(CL_i/V_i\). If observations are concentrated in the terminal phase, they may constrain that ratio more clearly than the separate values of clearance and volume. A tendency for both to increase together can preserve the ratio, producing a positive association between their random effects.
The ratio is not the whole profile: the initial concentration \(D/V_i\) also depends on volume. Early observations, reliable dose information, and a suitable model can help distinguish the parameters. If those sources of information are weak or the structural model is wrong, however, an estimated correlation may become extreme or unstable. Correlation is therefore a clue to investigate, not a diagnosis by itself.
Worked Example: Same Terminal Rate, Different Parameter Values
Consider two hypothetical individuals given a 1,000 mg intravenous bolus. Suppose one has \(CL=5\) L/h and \(V=50\) L, while the other has \(CL=8\) L/h and \(V=80\) L. Compare their terminal rates and predicted concentrations at 12 h under the one-compartment model.
For the first individual, \(CL/V=5/50=0.1000\) h\(^{-1}\), and the initial concentration is \(1{,}000/50=20.00\) mg/L. Thus,
For the second individual, \(CL/V=8/80=0.1000\) h\(^{-1}\), but the initial concentration is \(1{,}000/80=12.50\) mg/L. Therefore,
The terminal rates are identical, although the concentrations differ because the volumes differ. Terminal observations alone would not distinguish these parameter pairs by their decay rate; information about the profile's level is also needed. This example shows a route to parameter trade-off, not evidence that a real population must have correlated clearance and volume.
Estimating and Testing a Covariance
Suppose a model has \(q\) random effects and assumes \(\eta_i\sim N(0,\Omega)\), with individuals independent and the same population covariance structure applying to each individual. A full symmetric \(\Omega\) estimates \(q\) variances and \(q(q-1)/2\) covariances. A diagonal \(\Omega\) sets all covariances to zero, asserting that the random effects are uncorrelated under this distributional model. It does not assert that the corresponding PK parameters are biologically unrelated in every possible sense.
A covariance matrix must be positive semidefinite; a nonsingular covariance matrix must be positive definite. This constraint matters when correlations approach \(1\) or \(-1\): at those limits, two random effects are perfectly linearly related in the modeled distribution, and the matrix is singular if both variances are positive. Estimation software commonly uses a parameterization designed to maintain a valid covariance matrix, but a boundary-like estimate can still indicate that the data do not identify the separate variance directions reliably.
A standard comparison is between a model with a selected covariance free and the otherwise identical model with that covariance fixed to zero. Let \(M_0\) denote the restricted model and \(M_1\) the model with the covariance estimated. If both use the same observations, likelihood, structural model, fixed effects, and residual-error model, then the likelihood-ratio statistic is the difference in their objective functions, where the objective function is \(-2\) times the log likelihood.
Under the null hypothesis that the covariance is zero, and under regular likelihood assumptions, \(\Delta OFV\) is approximately chi-squared with one degree of freedom. This comparison concerns one covariance restriction; testing several restrictions requires accounting for the number and form of the restrictions. The usual approximation can be unreliable in small or weakly informative data sets or when other parameters are on boundaries. In particular, a test of a covariance fixed to zero with positive variances is not the same as testing whether a variance itself is zero.
Worked Example: Testing One Covariance Restriction
In a hypothetical analysis, a model with diagonal random-effect covariance has \(OFV=1248\). The otherwise identical model estimates one covariance between log clearance and log volume and has \(OFV=1240\). Assume both models converged to valid solutions, the variances are positive, and the likelihood-ratio approximation is reasonable.
The statistic is
There is one additional estimated covariance, so \(df=1\). For a chi-squared distribution with one degree of freedom, the upper-tail p-value for 8.000 is approximately 0.0047. Under the null model and the stated assumptions, a statistic at least this large would be unusual. The result is evidence that allowing this covariance improves likelihood-based fit; it does not establish that the estimated association is biological or that the structural model is correct.
The next steps are to inspect the estimated variances and correlation, verify numerical stability, and assess whether the correlated model improves profile diagnostics and prediction. The test alone cannot decide whether a covariance or a structural change is the more credible explanation.
Reading the Estimate Alongside Model Diagnostics
A correlation estimate is most informative when considered with its uncertainty and its sensitivity to model choices. A value near \(1\) or \(-1\) deserves particular attention, but there is no universal cutoff that distinguishes plausible from implausible correlations. Review whether the estimate changes substantially with starting values, parameterization, data subsets, or modest changes to the structural model. Check that the covariance matrix is valid and that the estimated variances are not so small that the correlation is poorly defined in practice.
Also inspect the model's predictions and residual patterns across time, concentration, dose, and relevant covariates. A correlation may absorb systematic profile features that the structural model fails to represent. For instance, if one structural alternative removes a time-dependent residual pattern and the extreme random-effect correlation becomes moderate, that change supports investigating the original structure. It is not proof that the alternative is true: compare plausibility, fit, and predictive performance rather than relying on the correlation's movement alone.
Worked Example: Interpreting a Large Estimated Correlation
Suppose a hypothetical model estimates standard deviations of 0.30 for \(\eta_{\mathrm{CL}}\) and 0.20 for \(\eta_{\mathrm{V}}\), with a correlation of \(0.95\). Calculate the covariance and interpret the estimate.
By definition,
The covariance is 0.0570 on the log-random-effect scale, while 0.95 is the dimensionless correlation. Under the fitted distribution, high log-clearance effects tend to accompany high log-volume effects. The estimate is close to the upper limit, so the analyst should examine its uncertainty, whether both random-effect directions are informed by the data, and whether it remains stable under reasonable model checks. It should not be reported as proof that clearance and volume are biologically linked.
Comparing Covariance Restrictions With Structural Alternatives
A diagonal covariance model is a useful restriction when correlations are not supported or are unstable. It is simpler, can reduce estimation burden, and avoids assigning meaning to a weakly identified association. Its cost is that it may fit less well if the population genuinely has correlated random effects. A full covariance model can capture that association, but it uses more parameters and can conceal shortcomings in the structural description.
A structural alternative changes the model for the concentration-time course, such as the absorption, distribution, or elimination description. Because the covariance model and the structural model answer different questions, one should not treat them as interchangeable fixes. Fit-based comparisons are straightforward by a likelihood-ratio test only when the compared models are nested and the required regularity conditions hold. For non-nested alternatives, compare appropriate information criteria or predictive performance, alongside parameter plausibility and residual diagnostics; do not apply a nested-model likelihood-ratio reference distribution automatically.
A practical evaluation can proceed in stages:
Record the variances, covariances, correlations, uncertainty, and whether the matrix is valid.
Compare the unrestricted and restricted covariance models, keeping other model components and the data unchanged.
Assess residual patterns, prediction, convergence, and sensitivity of the estimate to reasonable modeling choices.
Fit scientifically plausible structural alternatives and compare their diagnostics and predictive performance without assuming that a correlation identifies the correct structure.
Common Pitfall and Limitations
Wrong reasoning: “The estimated clearance-volume correlation is 0.95, so clearance and volume must be strongly associated in the population.” This treats a conditional model estimate as direct biological evidence and ignores possible compensation, weak identification, and structural error.
Correction: State that the fitted model estimates a strong positive association between the log-scale random effects, then assess uncertainty, stability, profile fit, and structural alternatives. The conclusion should remain conditional on the model and available data.
There are limits to what this evaluation can establish. A better likelihood does not by itself validate a random-effect distribution, and a diagonal covariance does not guarantee a correct structural model. Nor does a high correlation uniquely identify a specific structural defect. Different data designs and model structures can produce similar trade-offs. Conclusions should therefore distinguish what the model estimates from what the data can establish, and report the alternatives considered.
Check Your Understanding
Try each question, then compare with the solutions below.
- In a lognormal parameter model, what does a positive correlation between \(\eta_{\mathrm{CL}}\) and \(\eta_{\mathrm{V}}\) mean? Does it establish causation?
- A diagonal model has \(OFV=910\), and a model adding one covariance has \(OFV=906.2\). Under the usual one-degree-of-freedom likelihood-ratio approximation, calculate the statistic and give an approximate p-value.
- Why can terminal-phase observations make clearance and volume difficult to separate in a one-compartment intravenous bolus model?
- Name two checks to make before interpreting a correlation estimate near \(1\) as credible population association.
Solutions. (1) Under the fitted distribution, above-typical log clearance tends to accompany above-typical log volume; this is an association under the model, not evidence of causation. (2) \(\Delta OFV=910-906.2=3.8\), with \(df=1\); the approximate p-value is 0.051. This is not below a 0.05 threshold, and the result should be interpreted with the test assumptions in mind. (3) The terminal rate depends on \(CL/V\), so different clearance-volume pairs can have the same rate; information about the profile level is needed to distinguish them. (4) Examples include checking uncertainty and covariance-matrix validity, testing sensitivity to starting values or reasonable model changes, and examining residual and predictive diagnostics.