Why These Estimates Need Their Own Explanation
A population pharmacokinetic report is incomplete if it gives typical parameter estimates but leaves the variability model difficult to interpret. Readers need to know which parameters vary between individuals, the scale on which that variation was estimated, whether random effects were correlated, and how unexplained observation-level error was represented. A numerical estimate alone does not answer those questions.
Reporting should separate three things: the model specification, the estimated values, and the interpretation justified by those values. For example, an estimated positive correlation between two random effects describes their fitted association under the model. It does not establish a biological mechanism or show that one parameter causes the other to change. Likewise, a variance estimate at zero is a boundary result in the fitted model; it is not, by itself, proof that the population has no variation in that parameter.
State the Random-Effects Model and Its Scale
A common model for a positive individual pharmacokinetic parameter is an exponential random-effects model. For parameter \(j\) in individual \(i\), it can be written as:
Here, \(\theta_j\) is the typical population value and \(\eta_{ij}\) is the individual random effect on the log scale. A common assumption is that the vector of random effects for an individual has a multivariate normal distribution with mean zero and covariance matrix \(\Omega\). The diagonal elements of \(\Omega\) are the random-effects variances, written \(\omega_j^2\); its off-diagonal elements are covariances. The corresponding correlations are often easier to interpret and are obtained by standardizing the covariance:
A report should state whether a table gives variance, standard deviation, covariance, or correlation. These are different quantities. For an exponential model, \(\eta\) and its standard deviation are dimensionless because they operate on the log scale. A standard deviation of 0.4 is not a 0.4-unit difference in clearance. When a coefficient of variation is reported for the implied lognormal parameter distribution, its relationship to the log-scale variance is:
This expression assumes the exponential random-effects model and a normally distributed random effect. It is not the same as multiplying the log-scale standard deviation by 100, although that can be a rough approximation when the variance is small. If a report includes a CV, it should identify it as the implied between-individual CV on the parameter scale and, ideally, also give the estimated log-scale variance or standard deviation.
Worked Example: Turning a Random-Effects Estimate Into a Clear Report
Suppose a hypothetical final model uses exponential random effects for clearance and volume. Its estimated log-scale variances are \(\omega^2_{CL}=0.16\) and \(\omega^2_V=0.09\), and the estimated correlation between the two random effects is 0.35. Calculate the log-scale standard deviations, implied CVs, and covariance.
Standard deviations. Taking square roots gives \(\omega_{CL}=\sqrt{0.16}=0.400\) and \(\omega_V=\sqrt{0.09}=0.300\). These are standard deviations of dimensionless log-scale random effects.
Implied CVs. For clearance, \(100\sqrt{\exp(0.16)-1}=100\sqrt{0.1735}=41.65\%\). For volume, \(100\sqrt{\exp(0.09)-1}=100\sqrt{0.09417}=30.69\%\). These are the CVs implied by the specified lognormal model, not direct sample CVs calculated from observed individual parameter estimates.
Covariance. Rearranging the correlation formula gives \(\operatorname{Cov}(\eta_{CL},\eta_V)=0.35(0.400)(0.300)=0.0420\). The covariance is on the log-random-effect scale.
A transparent report could state: “The estimated log-scale variances were 0.160 for clearance and 0.0900 for volume, corresponding to log-scale standard deviations of 0.400 and 0.300 and model-implied parameter CVs of 41.7% and 30.7%, respectively. The estimated correlation between clearance and volume random effects was 0.350.” This wording identifies the quantities and their interpretation without implying a biological cause.
Report Correlations Without Overinterpreting Them
A correlation summarizes the fitted direction and strength of linear association between two random effects on the scale used by the model. In the exponential model above, this is association between log-scale deviations from the typical values. A positive correlation means individuals with a higher-than-typical value of one random effect tend, under the fitted distribution, to have a higher-than-typical value of the other. A negative correlation indicates the opposite tendency.
A correlation table should identify the random effects, the estimate, and the scale or parameterization. If space permits, reporting the covariance matrix as well can help readers reconstruct the assumed joint distribution. State whether the matrix entries are estimates or fixed assumptions. Do not describe a fitted correlation as evidence that the pharmacokinetic parameters are causally linked, nor treat a large correlation as inherently important without considering its uncertainty and the model context.
A correlation estimate can be difficult to interpret if either associated variance is near zero. Since correlation divides covariance by the product of standard deviations, it is not defined when a variance is zero. When a variance is very small, a small change in covariance can produce a large change in correlation. Reports should flag this instability rather than give a near-boundary correlation as if it were a precise, stand-alone result.
Specify the Residual-Error Model
Random effects describe between-individual variation in model parameters. Residual error describes the difference between an observation and the model prediction after accounting for the modeled structure. The report should give the actual residual equation, define its error terms, and state the units of any additive component. A label such as “combined error” is not enough, because implementations can differ in how the terms enter the equation.
One common combined additive-proportional model for a concentration observation is:
In this specification, the proportional error term is dimensionless and the additive error term has the same units as concentration. A commonly assumed structure is that both terms have mean zero, are independent of each other and of the individual random effects, and have variances \(\sigma_{\mathrm{prop}}^2\) and \(\sigma_{\mathrm{add}}^2\), respectively. If errors are also assumed independent across observations, say so when relevant to the analysis. The report must match the fitted model: for example, do not imply both terms were estimated if one was fixed, or imply independence if the model allowed dependence.
Under these assumptions, conditional on the prediction, the residual variance is the sum of the two component variances:
This follows because the variance of a sum of independent terms is the sum of their variances, and multiplying a random variable by a constant multiplies its variance by the constant squared. The equation also shows why the components have different roles: the proportional component grows with the prediction, while the additive component remains on the observation scale.
Worked Example: Interpreting Combined Residual Error
For a hypothetical concentration model, suppose the estimated proportional-error standard deviation is 0.20 and the additive-error standard deviation is 0.50 mg/L. At a prediction of 10 mg/L, calculate the conditional residual standard deviation under the combined model. Assume the two error terms are independent, have mean zero, and have the stated variances.
The conditional variance is \(10^2(0.20^2)+0.50^2=100(0.0400)+0.250=4.25\;(\mathrm{mg/L})^2\). The conditional standard deviation is therefore \(\sqrt{4.25}=2.062\) mg/L.
The proportional component contributes a standard deviation of \(10(0.20)=2.00\) mg/L at this prediction; the additive component contributes 0.50 mg/L. These component standard deviations should not be added directly. Their variances add because the terms are assumed independent. If one observation is 12.1 mg/L when the prediction is 10 mg/L, its residual is 2.1 mg/L, but that observed residual does not reveal how much arose from each error term.
A report should state both the equation and the estimates, for example: “Residual error was modeled as \(C_{\mathrm{obs}}=C_{\mathrm{pred}}(1+\epsilon_{\mathrm{prop}})+\epsilon_{\mathrm{add}}\), with estimated standard deviations 0.200 (dimensionless) and 0.500 mg/L, respectively.” This avoids the ambiguity of reporting “proportional error 0.20” without saying whether 0.20 is a standard deviation or variance.
How to Present the Results
A well-designed results table can make the estimates easy to review without replacing the model description. Include only quantities that were estimated or explicitly fixed, and label them accordingly. For example, a table can separate random-effects estimates from residual-error estimates:
| Component | Estimate | Scale or units | Reporting note |
|---|---|---|---|
| Clearance random-effect variance | 0.160 | Log scale; dimensionless | State the exponential parameterization |
| Clearance random-effect CV | 41.7% | Parameter scale | Model-implied from the lognormal variance |
| Correlation: clearance and volume random effects | 0.350 | Dimensionless | Identify both random effects |
| Proportional residual SD | 0.200 | Dimensionless | Distinguish SD from variance |
| Additive residual SD | 0.500 | mg/L | Units match the concentration observation |
The values in this illustrative table are hypothetical. In an actual report, use the estimates from the fitted model and make clear whether a quantity is an SD, variance, covariance, correlation, or transformed summary. If a variance is reported in one table and its CV in another, make the relationship explicit so readers do not mistake them for unrelated estimates.
A concise accompanying paragraph should identify the random effects included in the final model, summarize any estimated correlations, state the residual equation and parameter estimates, and point out estimates at or near constraints. If intervals or other measures of uncertainty are available, indicate the uncertainty method and present them consistently; the point estimate alone should not be made to carry more inferential weight than it can support.
Boundary Estimates and Warning Signs
A variance is constrained to be nonnegative. An estimate of zero is therefore at the boundary of the allowable parameter space, not an ordinary interior estimate. It means that the fitting procedure returned no positive estimate of that variance under the specified model and data. It does not establish that the true population variance is exactly zero. Estimation near a boundary can also make usual symmetric approximations and standard errors unreliable, so report the boundary status and avoid presenting a routine symmetric interval as automatically appropriate.
Other signs that deserve explicit reporting include a correlation at or very near \(-1\) or \(1\), a variance that is extremely small, or a reported correlation paired with a variance effectively equal to zero. These can indicate weak information about the corresponding components, a nearly singular covariance matrix, or sensitivity of the estimates to modeling choices. They do not by themselves diagnose the cause. Describe the numerical result and its limitation, and use the relevant model diagnostics and sensitivity analyses to assess its consequences.
Worked Example: Reporting a Variance at the Boundary
Consider a hypothetical model in which the estimated log-scale variance for a peripheral-volume random effect is exactly 0.000, with the estimation procedure constraining variances to be nonnegative. The estimated clearance variance is 0.040, and the estimated correlation between these effects is not reported because the peripheral-volume variance is zero.
The appropriate report is: “The peripheral-volume random-effect variance was estimated at the lower boundary, 0.000. A correlation involving this effect was not estimated or interpreted because its variance was zero.” This accurately describes the fit and explains why a correlation is absent.
It would be unsupported to write, “There is no between-individual variability in peripheral volume.” The estimate is conditional on the model, data, and estimation procedure; it does not prove the population value is zero. Nor should the report insert a correlation of zero: the correlation formula has a zero denominator when one standard deviation is zero, so the quantity is undefined, not estimated as zero. If the displayed value was rounded to 0.000 rather than exactly constrained to zero, state the unrounded estimate or describe it as near the boundary.
Common Pitfalls and Appropriate Comparisons
The main alternatives in reporting are not competing statistical tests but different choices of scale, error structure, and presentation. A log-scale variance is compact and aligns directly with the exponential random-effects model; a model-implied CV is often easier to interpret on the parameter scale. Reporting both can be useful, provided the conversion and assumptions are stated. For residual error, an additive model, a proportional model, and a combined model encode different patterns of conditional variability. The equation and units let readers see which structure was fitted; a generic phrase such as “measurement error was included” does not.
Avoid these common errors:
- Calling a variance a CV. A log-scale variance of 0.16 is not a 16% CV. Under the exponential-normal model its implied CV is about 41.7%.
- Omitting units or scale. An additive residual SD for concentration must be reported in concentration units; a proportional SD is dimensionless.
- Adding residual standard deviations. Under independent additive and proportional terms, add their variances and then take the square root.
- Treating correlation as causation. A random-effects correlation summarizes fitted association on the model scale; it does not identify a biological mechanism.
- Turning a boundary estimate into a population fact. Zero estimated variance is a constrained fit result, not proof of exactly zero true variability.
- Reporting an unstable correlation without context. When a variance is near zero, the associated correlation may be poorly determined or undefined at zero.
These distinctions matter because readers may use reported variability and residual error to judge how the model represents individual differences and observations. Clear reporting makes those assumptions reviewable and helps prevent numerical summaries from being mistaken for stronger biological conclusions.
Check Your Understanding
Answer each question, then compare your work with the solutions below.
- A log-scale random-effects variance is 0.25 in an exponential model. What is its log-scale standard deviation and its model-implied CV?
- A proportional residual-error SD is 0.15 and an additive residual-error SD is 0.40 mg/L. Under the independent combined-error model, what is the residual SD at a prediction of 8 mg/L?
- Why is a correlation not interpretable when one of the associated random-effects variances is exactly zero?
- Write one appropriately qualified sentence describing an estimated variance of zero.
Solutions. (1) The log-scale SD is \(\sqrt{0.25}=0.500\). The implied CV is \(100\sqrt{\exp(0.25)-1}=100\sqrt{0.2840}=53.29\%\). (2) The variance is \(8^2(0.15^2)+0.40^2=64(0.0225)+0.160=1.600\;(\mathrm{mg/L})^2\), so the SD is \(\sqrt{1.600}=1.265\) mg/L. (3) Correlation divides the covariance by the product of the two SDs; a zero variance makes one SD zero and the denominator zero, so the correlation is undefined. (4) For example: “The random-effects variance was estimated at the lower boundary of zero under the fitted model; this estimate does not establish that the true population variance is exactly zero.”