Presenting Final Model Equations

Why Present the Equations Together?

A final population pharmacokinetic model is more than a list of parameter estimates. Its equations show how the dose and dosing history produce a predicted concentration, how covariates modify typical parameters, how individuals vary around those typical values, and how observed concentrations differ from predictions. Reporting these pieces explicitly lets a reader reconstruct what the model means rather than infer it from software syntax or prose.

A useful presentation separates the model into four layers: the structural model, the covariate model, the random-effects model, and the residual-error model. The layers must fit together: each symbol needs a definition, each parameter needs a unit or a statement that it is dimensionless, and each covariate effect needs an unambiguous reference value.

Definition: The structural model describes the expected concentration-time behavior for specified individual parameters and dosing. The covariate model relates typical parameter values to measured characteristics. The random-effects model describes between-subject deviations from those typical values. The residual-error model describes remaining differences between observed concentrations and model predictions.

A Compact Example Model

Consider a hypothetical one-compartment model for an intravenous bolus dose. Let \(A_i(t)\) be the amount of drug in the compartment for subject \(i\), and let \(C_i(t)\) be the corresponding predicted concentration. With linear elimination, the structural equations are

\frac{dA_i(t)}{dt}=-CL_i C_i(t), \qquad C_i(t)=\frac{A_i(t)}{V_i}.

Here \(CL_i\) is subject \(i\)'s clearance, measured in L/h, and \(V_i\) is the volume of distribution, measured in L. Since \(C_i=A_i/V_i\), the elimination rate \(CL_iC_i\) has units \((\mathrm{L/h})(\mathrm{mg/L})=\mathrm{mg/h}\), matching the units of change in amount per time. For a single bolus of dose \(D_i\) at time zero, the initial condition is \(A_i(0)=D_i\).

The complete reporting task is to show what determines \(CL_i\) and \(V_i\), how concentration observations relate to \(C_i(t)\), and what the symbols and units mean. The differential equation alone does not describe covariate or between-subject effects.

Worked Example: Writing the Structural Prediction

For a hypothetical subject given a 100 mg intravenous bolus, suppose \(CL_i=5.0\) L/h and \(V_i=50\) L. Derive the concentration prediction and calculate it at 8 h.

From \(C_i=A_i/V_i\), the amount equation becomes \(dA_i/dt=-(CL_i/V_i)A_i\). Separating variables and integrating from \(0\) to \(t\) gives \(\ln[A_i(t)/A_i(0)]=-(CL_i/V_i)t\). Therefore \(A_i(t)=D_i\exp[-(CL_i/V_i)t]\), and

C_i(t)=\frac{D_i}{V_i}\exp\left(-\frac{CL_i}{V_i}t\right).

The initial concentration is \(100\ \mathrm{mg}/50\ \mathrm{L}=2.00\) mg/L, and the elimination rate constant is \(5.0/50=0.100\) h\(^{-1}\). Thus

C_i(8)=\frac{100\ \mathrm{mg}}{50\ \mathrm{L}}\exp[-(0.100\ \mathrm{h}^{-1})(8\ \mathrm{h})] =2.00\exp(-0.800)=0.899\ \mathrm{mg/L}.

The exponent is dimensionless because h\(^{-1}\) is multiplied by h. This solution assumes a single instantaneous bolus, one well-mixed compartment, and linear elimination; it is not a general solution for every dosing regimen or structural model.

Covariates and Reference Values

A covariate equation should state both the functional form and the value against which the effect is expressed. For example, let \(WT_i\) be body weight in kg and \(eGFR_i\) be estimated glomerular filtration rate in mL/min/1.73 m\(^2\). A hypothetical typical-clearance model is

CL_{\mathrm{typ},i} =\theta_{CL} \left(\frac{WT_i}{70\ \mathrm{kg}}\right)^{0.75} \left(\frac{eGFR_i}{90\ \mathrm{mL/min/1.73\,m^2}}\right)^{\theta_{GFR}}.

In this equation, \(\theta_{CL}\) is typical clearance at the stated reference values, in L/h, and \(\theta_{GFR}\) is a dimensionless exponent. The ratios are dimensionless, so the result retains the units of \(\theta_{CL}\). The reference values are part of the model definition: changing them changes the interpretation of the typical parameter, even if a reparameterized model can produce the same predictions.

A report should not merely say that clearance was “adjusted for weight and renal function.” It should give the actual equation, name the covariates and their measurement units, and identify the reference values. If a covariate is centered or transformed, say how. For a categorical covariate, identify the reference category and the factor applied to other categories.

Worked Example: Interpreting a Covariate Equation

Suppose the hypothetical typical-clearance equation above has \(\theta_{CL}=5.0\) L/h and \(\theta_{GFR}=0.50\). Calculate typical clearance for a subject weighing 100 kg with \(eGFR=60\) mL/min/1.73 m\(^2\), before applying any subject-specific random effect.

Substitution gives

CL_{\mathrm{typ},i} =5.0\ \mathrm{L/h} \left(\frac{100}{70}\right)^{0.75} \left(\frac{60}{90}\right)^{0.50}.

The weight ratio is \(100/70=1.429\), and \(1.429^{0.75}\approx1.306\). The renal-function ratio is \(60/90=0.6667\), and \(0.6667^{0.50}\approx0.8165\). Therefore \(CL_{\mathrm{typ},i}=5.0(1.306)(0.8165)=5.33\) L/h, to three significant figures. This is the covariate-adjusted typical value; it is not yet the individual's clearance after between-subject variation is included.

At the reference values of 70 kg and 90 mL/min/1.73 m\(^2\), both ratios equal one, so \(CL_{\mathrm{typ}}=\theta_{CL}=5.0\) L/h. This makes the reference interpretation directly verifiable from the equation.

Adding Between-Subject Variation and Residual Error

For positive pharmacokinetic parameters, a common specification is an exponential random effect. For clearance and volume, the hypothetical model is

CL_i=CL_{\mathrm{typ},i}\exp(\eta_{CL,i}), \qquad V_i=\theta_V\left(\frac{WT_i}{70\ \mathrm{kg}}\right)\exp(\eta_{V,i}).

Here \(\theta_V\) is typical volume at 70 kg, in L. The individual parameters \(CL_i\) and \(V_i\) are positive because exponentials are positive. The random effects \(\eta_{CL,i}\) and \(\eta_{V,i}\) are dimensionless. A complete report states their joint distribution and covariance structure, for example

\begin{pmatrix}\eta_{CL,i}\\\eta_{V,i}\end{pmatrix} \sim N\left[ \begin{pmatrix}0\\0\end{pmatrix}, \begin{pmatrix}\omega^2_{CL}&\omega_{CL,V}\\ \omega_{CL,V}&\omega^2_V\end{pmatrix} \right].

The diagonal entries are variances of the random effects and the off-diagonal entry is their covariance. The equation states the assumed distribution, not that every individual effect is zero. If the covariance is omitted because the effects are assumed independent, that assumption should be stated rather than left unclear.

For observation \(j\) from subject \(i\), let \(C_{\mathrm{pred},ij}\) be the structural prediction at that observation's time and dosing history. One possible combined residual-error model is

C_{\mathrm{obs},ij} =C_{\mathrm{pred},ij}(1+\epsilon_{p,ij})+\epsilon_{a,ij},

where \(\epsilon_{p,ij}\) is a dimensionless proportional error and \(\epsilon_{a,ij}\) is an additive error with concentration units, such as mg/L. A report should state the assumed distributions and independence structure; for instance, both terms may have mean zero, variances \(\sigma_p^2\) and \(\sigma_a^2\), and be mutually independent and independent of the subject random effects. Under those assumptions, the conditional mean is \(C_{\mathrm{pred},ij}\) and the conditional variance is

\operatorname{Var}(C_{\mathrm{obs},ij}\mid C_{\mathrm{pred},ij}) =C_{\mathrm{pred},ij}^{\,2}\sigma_p^2+\sigma_a^2.

Thus \(\sigma_p\) is dimensionless, while \(\sigma_a\) has the same units as concentration. The error equation defines the observation model; it does not alter the structural concentration prediction.

Worked Example: Combining the Model Layers

Continue with the hypothetical bolus model. Use \(\theta_{CL}=5.0\) L/h, \(\theta_{GFR}=0.50\), \(\theta_V=50\) L, \(WT_i=100\) kg, and \(eGFR_i=60\) mL/min/1.73 m\(^2\). Suppose \(\eta_{CL,i}=0.10\), \(\eta_{V,i}=-0.05\), and the dose is 100 mg. Calculate the predicted concentration at 8 h, then show one possible observation under the residual model.

From the covariate calculation, \(CL_{\mathrm{typ},i}=5.33\) L/h. Applying the random effect gives \(CL_i=5.33\exp(0.10)=5.89\) L/h. The volume is \(V_i=50(100/70)\exp(-0.05)=67.9\) L. These individual parameter values, rather than the typical values, enter the structural prediction.

The elimination rate constant is \(CL_i/V_i=5.89/67.9=0.0867\) h\(^{-1}\). Therefore

C_{\mathrm{pred},i}(8) =\frac{100\ \mathrm{mg}}{67.9\ \mathrm{L}} \exp[-(0.0867\ \mathrm{h}^{-1})(8\ \mathrm{h})] \approx0.736\ \mathrm{mg/L}.

For illustration, let \(\sigma_p=0.20\), \(\sigma_a=0.050\) mg/L, and let one possible pair of residual values be \(\epsilon_{p,ij}=0.050\) and \(\epsilon_{a,ij}=0.0272\) mg/L. Then \(C_{\mathrm{obs},ij}=0.736(1+0.050)+0.0272=0.800\) mg/L. These selected residual values illustrate the equation; they are not parameter estimates or evidence about any real study.

Justification, Checks, and Common Failure Modes

The exponential random-effects form has a useful interpretation. Taking logarithms gives \(\log(CL_i/CL_{\mathrm{typ},i})=\eta_{CL,i}\). A zero random effect yields the typical value; a positive effect multiplies it upward, and a negative effect multiplies it downward. This multiplicative structure preserves positivity and expresses subject deviations on a relative rather than additive scale.

Before presenting final equations, check that the notation is consistent from line to line, the structural prediction depends on the individual parameters, and all units reconcile. The exponent in an exponential must be dimensionless; a covariate ratio raised to a power should have compatible numerator and denominator units; and additive residual error must use the concentration unit of the observation. Also verify that the written equation matches the fitted model, including any fixed effects, transformations, covariance assumptions, and error components.

Several presentation failures are especially consequential:

Giving only a verbal summary. “Clearance depends on weight” does not disclose the exponent, reference weight, or functional form.
Leaving reference values implicit. A coefficient cannot be interpreted correctly if a reader does not know which covariate values define the typical parameter.
Confusing typical and individual parameters. Covariate-adjusted typical clearance is not the same as an individual's clearance when random effects are included.
Omitting units or labels. A number without its parameter name and units is difficult to interpret and can invite dimensional errors.
Using an error equation without defining its terms. State whether the error is proportional, additive, or combined, and provide the assumptions for the error terms.

A useful diagnostic is to substitute the stated reference values and zero random effects. The equations should then return the declared typical parameters. If they do not, either the equation or the parameter description is inconsistent. Another check is to set covariates to their references and confirm that their ratios equal one. These algebraic checks can catch reporting mistakes even when the fitted model itself is unchanged.

Choosing the Right Level of Detail

The equations should be complete enough to reproduce the model's mathematical specification, but they need not reproduce software code. A differential equation may be clearest for a general dosing history, while a closed-form concentration equation can be useful for a simple bolus model. If both are shown, explain their relationship and the conditions under which the closed form applies. For more complex structural models, define each compartment or process needed to interpret the reported equations rather than implying that a one-compartment formula applies universally.

A concise equation set is not a substitute for explaining assumptions. State the dosing and structural assumptions relevant to the equations, the covariate reference values and transformations, the random-effect distribution and covariance structure, and the residual-error form. In contrast, avoid burying these definitions in dense prose or relying on readers to infer them from parameter labels. The goal is a model that can be understood and checked directly from the report.

A common mistake is to interpret \(\theta_{CL}\) as clearance for every subject because it appears in the equation. In this example, \(\theta_{CL}\) is clearance only at the covariate reference values and with the random effect set to zero. The correction is to trace the full sequence: covariates determine \(CL_{\mathrm{typ},i}\), the random effect determines \(CL_i\), and \(CL_i\) enters the structural prediction. Similarly, residual error is added only when connecting the prediction to an observation.

Key takeaway: Present the structural, covariate, random-effects, and residual-error equations as a connected model. Define every symbol, provide parameter units and covariate reference values, and make clear which equations describe typical values, individual values, predictions, and observations.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

In a covariate equation, what is the interpretation of the typical clearance parameter when covariates are expressed as ratios?
Why must the exponent in \(\exp[-(CL/V)t]\) be dimensionless?
If a proportional residual term is multiplied by a predicted concentration, what are the units of the proportional error and the additive error?
For a subject at all covariate reference values with random effects equal to zero, what should the covariate and individual-parameter equations return?
Why is the covariate-adjusted typical parameter not necessarily the individual's parameter?

Solutions. (1) It is the typical parameter value at the stated covariate references, before individual random effects are applied. (2) An exponential's argument must be dimensionless; \(CL/V\) has units h\(^{-1}\), which cancel with time in h. (3) The proportional error is dimensionless; the additive error has concentration units. (4) Each covariate ratio is one, and an exponential random-effect multiplier with zero effect is one, so the individual parameter equals the typical parameter at the references. (5) The individual value also includes the subject-specific random effect, which can move it above or below the typical value.