Reporting the Final NLME Model Specification

Why the Final Specification Must Be Reconstructible

A final NLME model is more than a list of parameter estimates. Its specification describes how dose and time produce a predicted observation for each individual, how individual parameters vary around typical values, and how observed data vary around predictions. A report is adequate only when another analyst can translate it into the same mathematical model without guessing about units, covariates, random effects, residual error, or dose handling.

This tutorial concerns the model definition itself. It is distinct from reporting the computational settings used to fit that model: a parameter estimate without a defining equation is ambiguous, while a complete equation does not by itself document how estimation was carried out.

Definition: A final NLME model specification is a precise statement of the structural model, parameter definitions and units, covariate relationships, random-effects distribution, residual error model, and dosing and observation conventions used to generate the model predictions. It should distinguish fixed parameters, individual random effects, and residual variability.

A Complete Example of the Specification

Consider a hypothetical one-compartment model for concentration after intravenous bolus doses. For individual $i$, let $A_i(t)$ be the amount of drug in the compartment, in mg, and let $C_i(t)$ be concentration, in mg/L. Between doses, assume linear elimination with individual clearance $CL_i$ and volume $V_i$. The predicted concentration is $f_i(t)=A_i(t)/V_i$.

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

For a dose $D_{ij}$, in mg, given as an instantaneous intravenous bolus to individual $i$ at time $t_{ij}$, the amount immediately after the dose is $A_i(t_{ij}^{+})=A_i(t_{ij}^{-})+D_{ij}$. The superscripts denote the values just after and just before the dose. This jump condition makes the dose event part of the model rather than an unstated implementation detail.

Suppose clearance and volume depend on body weight $WT_i$, measured in kg, and have correlated individual random effects:

CL_i=\theta_{CL}\left(\frac{WT_i}{70\ \mathrm{kg}}\right)^{0.75}\exp(\eta_{CL,i}), \qquad V_i=\theta_V\left(\frac{WT_i}{70\ \mathrm{kg}}\right)^{1.00}\exp(\eta_{V,i}).

Here $\theta_{CL}$ is typical clearance for a 70 kg individual with both random effects equal to zero, and $\theta_V$ is the corresponding typical volume. The exponents 0.75 and 1.00 are fixed covariate exponents in this specification. Each $\eta$ is dimensionless and acts on the log scale. A report must say whether exponents are fixed or estimated; merely writing “weight adjusted” is not enough.

One possible population distribution is a bivariate normal random-effects model. Let $\boldsymbol{\eta}_i=(\eta_{CL,i},\eta_{V,i})^\mathsf{T}$. Then state, for example, that $\boldsymbol{\eta}_i$ are independently distributed across individuals with mean zero and covariance matrix $\Omega$, and give the matrix or its standard deviations and correlation. Conditional on the individual effects and model prediction, a combined additive-proportional residual model for concentration can be written as

Y_{ij}=f_i(t_{ij})+ \sqrt{\sigma_{\mathrm{add}}^2+ \left(\sigma_{\mathrm{prop}}f_i(t_{ij})\right)^2}\,\epsilon_{ij}, \qquad \epsilon_{ij}\sim N(0,1).

In this model, $\sigma_{\mathrm{add}}$ has concentration units and $\sigma_{\mathrm{prop}}$ is dimensionless. The standard-normal residual terms are conditionally independent across observations if that is the intended specification. If the residual model differs by assay, occasion, or observation type, that distinction must also be stated rather than left implicit.

Worked Example: Calculating an Individual’s Parameters

Use the hypothetical model above with $\theta_{CL}=5.00\ \mathrm{L/h}$, $\theta_V=36.0\ \mathrm{L}$, $WT_i=84.0\ \mathrm{kg}$, $\eta_{CL,i}=0.100$, and $\eta_{V,i}=-0.0500$. Calculate the individual clearance and volume.

The weight ratio is $84.0/70.0=1.20$. Thus,

CL_i=5.00(1.20)^{0.75}\exp(0.100) =5.00(1.14653)(1.10517) \approx 6.336\ \mathrm{L/h}.

For volume,

V_i=36.0(1.20)^{1.00}\exp(-0.0500) =36.0(1.20)(0.951229) \approx 41.093\ \mathrm{L}.

The implied elimination rate constant is $CL_i/V_i\approx6.33555/41.0931=0.15416\ \mathrm{h}^{-1}$. For a 400 mg bolus at time zero, assuming $A_i(0^-)=0$, the predicted concentration immediately after dosing is $400\ \mathrm{mg}/41.0931\ \mathrm{L}\approx9.734\ \mathrm{mg/L}$. These calculations rely on the stated one-compartment, instantaneous-mixing, linear-elimination assumptions; they are not valid merely because clearance and volume estimates are available.

What to State About Each Model Component

The equations are the central part of the report, but each needs definitions around it. A useful specification identifies the observation being modeled (for example, plasma concentration), its units, the time origin, and how predictions relate to the recorded observations. For an ODE model, give the state variables, initial conditions, dose inputs, and parameter-dependent rates. For an algebraic model, give the function and define its arguments. State any structural assumptions that affect the form, such as instantaneous mixing, first-order elimination, or linearity.

Covariate relationships require more than naming the covariate. Report its units, any transformations, the reference value used for centering or scaling, and whether the effect is categorical, linear, power, or another functional form. Clarify whether a coefficient is fixed or estimated. A reference value is important for interpretation: in the weight model above, each typical parameter is defined at 70 kg, not at an abstract covariate value of zero.

For random effects, report which parameters vary between individuals, the scale on which each effect enters, their assumed distribution, and their covariance structure. If the covariance matrix is diagonal, say that the random effects are modeled as uncorrelated. If correlations are estimated, report them or provide the full covariance matrix. Do not use “independent random effects” if the model includes nonzero correlations. Also state if random effects are absent from a parameter: omission should not be mistaken for an undocumented assumption.

For residual variability, identify the observation scale and give the exact variance or standard-deviation relationship. The symbols alone are insufficient: define units and say whether the residual terms are assumed conditionally independent and normally distributed. If concentrations are modeled on a log scale, make that explicit and distinguish the model for log concentration from a model for concentration.

Worked Example: Reporting the Random-Effects Structure

Suppose a hypothetical fit uses $\omega_{CL}=0.300$, $\omega_V=0.200$, and correlation $\rho=0.300$, where these are the standard deviations and correlation of the log-scale random effects. Give the covariance matrix and interpret the scales.

The covariance is $\rho\omega_{CL}\omega_V=(0.300)(0.300)(0.200)=0.0180$. Therefore, the report can state:

\boldsymbol{\eta}_i\sim N\left[ \begin{pmatrix}0\\0\end{pmatrix}, \begin{pmatrix} 0.0900 & 0.0180\\ 0.0180 & 0.0400 \end{pmatrix} \right].

The diagonal entries are variances, so they are the squares of the reported standard deviations. The positive correlation means that, under this assumed population distribution, individuals with above-typical log clearance effects tend also to have above-typical log volume effects. It does not mean that clearance causes volume to increase. Because the effects enter exponentially, a one-standard-deviation increase in $\eta_{CL}$ multiplies clearance by $\exp(0.300)\approx1.350$, rather than adding 0.300 L/h.

Justifying the Structural Equations

A short derivation can make the connection between a reported equation and its assumptions transparent. Between bolus doses, the one-compartment amount equation is $dA_i/dt=-(CL_i/V_i)A_i$. For $A_i(0)>0$, separating variables and integrating from time zero gives $\log\{A_i(t)/A_i(0)\}=-(CL_i/V_i)t$; if $A_i(0)=0$, the solution is $A_i(t)=0$, also given by the formula below. Exponentiation yields $A_i(t)=A_i(0)\exp[-(CL_i/V_i)t]$; division by $V_i$ gives the concentration prediction. This solution presumes constant individual clearance and volume over the interval and a single well-mixed compartment.

This derivation also helps detect inconsistencies. The ratio $CL_i/V_i$ must have units of inverse time, so the exponential argument is dimensionless. The amount is in mg and volume in L, so predicted concentration is in mg/L. A report that assigns incompatible units to these quantities has not specified a coherent model.

Worked Example: Making Oral Dosing Assumptions Explicit

Consider a separate hypothetical one-compartment oral model with first-order absorption, no lag time, and linear elimination. A dose enters a gut depot with bioavailability fixed at $F=1$. Let $A_g(t)$ and $A_c(t)$ be depot and central amounts in mg. Suppose $k_a=1.20\ \mathrm{h}^{-1}$, $CL=4.00\ \mathrm{L/h}$, and $V=40.0\ \mathrm{L}$. For a single 100 mg dose at time zero, state the equations and calculate predicted concentration at 2.00 h, taking random effects to be zero.

The dose event is $A_g(0^+)=100\ \mathrm{mg}$, $A_c(0^+)=0$. Between doses, the model is

\frac{dA_g}{dt}=-k_aA_g,\qquad \frac{dA_c}{dt}=k_aA_g-\frac{CL}{V}A_c,\qquad C(t)=\frac{A_c(t)}{V}.

Here $k=CL/V=4.00/40.0=0.100\ \mathrm{h}^{-1}$. With $k_a\ne k$, solving the two equations gives

C(t)=\frac{F D k_a}{V(k_a-k)} \left[\exp(-kt)-\exp(-k_at)\right].

Substituting the hypothetical values at 2.00 h,

C(2.00)=\frac{(1)(100)(1.20)}{(40.0)(1.20-0.100)} \left[\exp(-0.100\times2.00)-\exp(-1.20\times2.00)\right]

$$ =2.72727(0.818731-0.090718)\approx1.985\ \mathrm{mg/L}. $$ The specification also states what is not included: no absorption lag, no nonlinear elimination, and no estimated bioavailability in this example. Repeated doses would require an explicit dose amount and time for each event.

Comparing Specification Choices and Checking for Failure

Different mathematical forms may encode similar scientific ideas, but they are not interchangeable without explanation. A power covariate model centered at a reference value directly defines the typical parameter there; an uncentered model can make the intercept correspond to an implausible covariate value. A diagonal random-effects covariance matrix is simpler than a correlated structure, but it asserts zero correlations. A combined additive-proportional error model allows the residual SD to remain nonzero near zero prediction while increasing at larger predictions; an additive-only or proportional-only model makes a different variance assumption.

Specification choice	What must be reported	Common ambiguity
Centered versus uncentered covariate	Reference value, units, transformation, and coefficient	“Adjusted for weight” without stating the functional form
Diagonal versus correlated random effects	Varying parameters, distribution, variances, and correlations or covariances	Calling effects independent while reporting nonzero covariance
Additive, proportional, or combined residual error	Exact formula, observation scale, and parameter units	Reporting error estimates without saying whether they are SDs or variances
Bolus, infusion, or oral dose input	Event timing, dose units, input duration or absorption, and bioavailability assumptions	Giving the dose amount but not how it enters the model

Specification checks are practical diagnostics. Confirm that units agree across equations, that each covariate has a defined reference, that every random effect is assigned to a parameter, and that the residual model matches the modeled observation scale. Check that dose times, infusion durations, dose units, and time units are consistent. If a reader cannot determine whether a dose is a bolus or an infusion, or whether an exponential effect is on a parameter or on a prediction, the report is incomplete.

The main limitation of a specification is that completeness does not establish scientific adequacy. A fully stated one-compartment model may still be a poor description of the system, and a precise residual equation does not prove that its assumptions hold. The specification makes assumptions inspectable and reproducible; assessment of whether they are supported requires appropriate model evaluation. Keep the final equations distinct from prose about the rationale for selecting them, so that the implemented model is not confused with alternatives considered during development.

Common pitfall: Writing “clearance was allometrically scaled, with interindividual variability and proportional error” sounds informative but leaves the model underdetermined. It does not give the covariate reference, exponent, random-effects distribution or covariance, residual equation, or parameter units. Replace such shorthand with explicit equations and definitions; use prose to clarify assumptions and dose handling.

Define the prediction.
Name the modeled observation, structural states or function, time scale, units, and dose-event rules.

Define individual parameters.
Give covariate equations, reference values, fixed or estimated coefficients, and random-effect placement.

Define variability.
State the random-effects distribution and covariance structure, then specify the residual model on the observation scale.

Audit consistency.
Check units, event conventions, equation dimensions, and whether every shorthand statement has an unambiguous mathematical meaning.

Check Your Understanding

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

In a power model using $WT_i/70$, what does the typical parameter represent at a body weight of 70 kg?
If a model reports a random-effects variance of 0.09 for a log-scale effect, what is its standard deviation?
Why must an oral model report how the dose enters the gut depot, not only the dose amount?
For a combined additive-proportional residual model, what are the units of the additive and proportional error parameters when concentration is in mg/L?
Name two consistency checks for a one-compartment model with clearance in L/h and volume in L.

Solutions. (1) At 70 kg the weight ratio is one, so the covariate multiplier is one; with the random effect set to zero, the parameter equals its typical value. (2) The standard deviation is $\sqrt{0.09}=0.300$. (3) Bolus depot input, infusion, and other input assumptions can generate different concentration-time predictions even for the same dose amount; the dose event and absorption assumptions define the model input. (4) The additive error parameter is in mg/L, while the proportional error parameter is dimensionless because it multiplies a concentration. (5) Check that $CL/V$ has units $(\mathrm{L/h})/\mathrm{L}=\mathrm{h}^{-1}$, and that the exponential argument $(CL/V)t$ is dimensionless when time is in hours. Also check that amount divided by volume has concentration units.