Reconstructing a Parameter Equation From a Published Model

From a Model Description to an Individual-Parameter Equation

A published NLME model may describe a parameter in prose, a compact equation, or a mixture of both. Reconstructing the equation means identifying exactly how the population typical value, covariates, and individual random effect combine to define a subject-specific parameter. It is not enough to recognize a familiar model form: a misplaced reference value, an incorrectly coded indicator, or a random effect on the wrong scale changes the model.

The earlier material on typical values, covariate coding, and hierarchical parameter equations establishes the component ideas. Here the emphasis is on a reliable translation process: preserve the scale and reference conventions stated in the description, write the equation explicitly, then check its meaning and numerical consequences.

Definition: A typical parameter value, denoted here by \(\mathrm{TV}\theta\), is the population-level value specified by a parameter model at its stated covariate reference values and with the individual random effect set to zero. An individual parameter, denoted by \(\theta_i\), is the parameter value assigned to individual \(i\) by that model after inserting the individual’s covariates and random effect. The equation must state how those components combine.

For example, a multiplicative model with log-linear covariate effects can be written as

\theta_i = \mathrm{TV}\theta \exp\left( \beta_1 x_{1i}+\cdots+\beta_q x_{qi}+\eta_i \right).

Here \(x_{ki}\) is the coded value of covariate \(k\) for individual \(i\), \(\beta_k\) is its coefficient on the log scale, and \(\eta_i\) is the individual random effect in the exponent. When all covariates are at their reference values, each \(x_{ki}=0\); with \(\eta_i=0\), the equation gives \(\theta_i=\mathrm{TV}\theta\). If the description specifies a distribution for \(\eta_i\), retain it as specified. Do not infer a distribution merely from seeing an eta symbol.

Not every model is multiplicative. An additive parameter-scale model, for instance, may be

\theta_i = \mathrm{TV}\theta + \beta x_i + \eta_i.

In that equation, \(\beta\) and \(\eta_i\) are on the same scale as \(\theta_i\). The written description determines which form is appropriate; the two forms are not interchangeable.

A Translation and Checking Procedure

Before calculating an individual value, separate the description into its ingredients. Record the parameter and its units, the typical value, the covariate reference values and coding, the scale of each covariate effect, and the placement of the random effect. Then write the equation and substitute values only after checking that each term has the units and meaning required by that equation.

Locate the typical value.
Write down the parameter name, its numerical typical value, and its units. Check what covariate values define that value.

Decode covariates.
For each continuous covariate, record its center and scaling. For each categorical covariate, record the reference category and the indicator assigned to each other category.

Identify the effect scale.
Words such as “multiplied by,” “ratio,” or “percent higher” often describe a multiplicative effect. A stated increment in units of the parameter may indicate an additive effect. Use the equation or precise model description, not wording alone, to resolve ambiguity.

Place the random effect and verify.
Determine whether the random effect appears inside an exponential or is added directly. Set covariates to their reference values and \(\eta_i=0\); the equation should return the typical value.

A useful justification for the common exponential form comes from writing the model on the log scale. If the covariate and random-effect terms combine additively in \(\log(\theta_i/\mathrm{TV}\theta)\), then exponentiating both sides gives the multiplicative parameter equation:

\log\left(\frac{\theta_i}{\mathrm{TV}\theta}\right) = \sum_{k=1}^{q}\beta_k x_{ki}+\eta_i \quad\Longrightarrow\quad \theta_i = \mathrm{TV}\theta \exp\left(\sum_{k=1}^{q}\beta_k x_{ki}+\eta_i\right).

This derivation also explains why the terms in the exponent are dimensionless: the ratio inside the logarithm is dimensionless, so its log and the terms equated to it are dimensionless. For an additive model, by contrast, the covariate and random-effect terms must have the units of \(\theta_i\).

Worked Example: A Centered Continuous Covariate

Suppose a hypothetical published description states: “Typical clearance is 4.8 L/h at a body weight of 70 kg. Clearance is related to weight by a log-linear effect with coefficient 0.25 per 70-kg difference unit, and the subject random effect is exponential.” For one individual, weight is 56 kg and \(\eta_{CL,i}=0.08\). Reconstruct the equation and calculate clearance.

The centered and scaled weight covariate is \((WT_i-70)/70\), so it equals zero at the 70-kg reference value. The description puts the random effect in the exponent, giving

CL_i = 4.8\ \mathrm{L/h} \exp\left( 0.25\frac{WT_i-70\ \mathrm{kg}}{70\ \mathrm{kg}}+\eta_{CL,i} \right).

For \(WT_i=56\ \mathrm{kg}\), the covariate term is \(0.25(56-70)/70=0.25(-0.2)=-0.05\). Adding the stated random effect gives an exponent of \(-0.05+0.08=0.03\). Therefore,

CL_i = 4.8\exp(0.03)\ \mathrm{L/h} = 4.8(1.0304545)\ \mathrm{L/h} = 4.94618\ \mathrm{L/h} \approx 4.946\ \mathrm{L/h}.

The equation returns 4.8 L/h at 70 kg when \(\eta_{CL,i}=0\), as required. The individual’s weight lowers the exponent by 0.05, while the specified positive random effect raises it by 0.08; their net effect is a small increase relative to the typical value.

Categorical Covariates and Multipliers

A categorical covariate needs an explicit indicator definition. Suppose females are the reference category and \(M_i=1\) for a male individual and \(M_i=0\) for a female individual. If the description states that the typical parameter is multiplied by 1.30 for males, the covariate term can be written as \(1.30^{M_i}\). This equals 1 for the reference group and 1.30 for the other group. Equivalently, the term is \(\exp\{\log(1.30)M_i\}\). State the indicator coding so the equation cannot be mistaken for a different reference group.

Worked Example: Reconstructing a Categorical Effect

In a hypothetical model, typical volume is 35 L for females. The published description says that the volume multiplier for males is 1.30 and that the individual random effect acts exponentially. A male has \(\eta_{V,i}=-0.12\). Let \(M_i=1\) for males and \(M_i=0\) for females.

The reconstructed equation is

V_i = 35\ \mathrm{L}\times 1.30^{M_i}\exp(\eta_{V,i}).

For this male, \(M_i=1\), so the covariate multiplier is 1.30. Also, \(\exp(-0.12)=0.8869204\). Substitution gives

V_i = 35(1.30)\exp(-0.12)\ \mathrm{L} = 45.5(0.8869204)\ \mathrm{L} = 40.35488\ \mathrm{L} \approx 40.35\ \mathrm{L}.

The negative random effect reduces the covariate-adjusted value of 45.5 L. It does not reverse the male multiplier or make the parameter negative. Setting \(M_i=0\) and \(\eta_{V,i}=0\) returns 35 L, confirming that females are the reference group.

Additive Equations and Scale Checks

Some descriptions specify absolute changes rather than ratios. In those cases, adding a coefficient or random effect inside an exponential would change the meaning. An additive equation also makes units a direct check: if the parameter is measured in inverse hours, every additive term must be measured in inverse hours.

Worked Example: An Additive Parameter-Scale Model

Suppose a hypothetical description gives typical absorption rate \(1.20\ \mathrm{h}^{-1}\) at age 60 years. Each five-year increase above age 60 adds \(0.08\ \mathrm{h}^{-1}\), and an additive subject deviation is included. For an individual aged 70 years, let \(\eta_{ka,i}=-0.10\ \mathrm{h}^{-1}\).

The age increment is scaled in five-year units, and the complete equation is

k_{a,i} = 1.20\ \mathrm{h}^{-1} + 0.08\ \mathrm{h}^{-1}\frac{AGE_i-60\ \mathrm{years}}{5\ \mathrm{years}} + \eta_{ka,i}.

At age 70, the scaled age difference is \((70-60)/5=2\). Thus the covariate contribution is \(0.08(2)=0.16\ \mathrm{h}^{-1}\), and

k_{a,i} = 1.20+0.16-0.10 = 1.26\ \mathrm{h}^{-1}.

All terms have units of inverse hours, and the result is positive. At age 60 with \(\eta_{ka,i}=0\), the equation returns the typical value. This arithmetic follows the stated additive model; replacing the additions with exponentials would instead describe a different model.

Common Reconstruction Errors and Diagnostics

The most effective checks are internal checks of the equation against the description. Confirm that the reference individual gives the stated typical value; every indicator has the stated coding; the covariate scale matches the coefficient’s definition; and the random effect appears on the stated scale. For a multiplicative equation, exponent terms should be dimensionless. For an additive equation, the terms added to the parameter should carry its units.

Wrong reference category. Coding males as 1 is not a harmless choice if the written multiplier was defined for females relative to males. State the reference and indicator values next to the equation.
Wrong center or scaling. A coefficient per 10 kg cannot be multiplied by an unscaled 10-kg difference and also divided by 10. Use one representation consistently.
Random effect on the wrong scale. \(\mathrm{TV}\theta\exp(\eta_i)\) is multiplicative; \(\mathrm{TV}\theta+\eta_i\) is additive. They give different individual values and require different units for \(\eta_i\).
Confusing a typical value with an arithmetic mean. The equation’s reference value at zero random effect is not automatically an arithmetic population mean. Its interpretation depends on the model form and random-effect distribution.
Forcing an interpretation when the description is incomplete. If “10% per 10 kg” does not say whether the percentage is compounded over successive increments or used as a linear approximation, the equation may be ambiguous. Report the ambiguity rather than silently choosing a form.

A reconstructed equation is only as complete as the supplied model description. It does not establish that the covariate relationship is scientifically appropriate, that the parameter is estimable, or that the fitted model describes data adequately. Those questions require model specification and evaluation beyond transcription. Nor should a residual-error term be inserted into an individual-parameter equation unless the model explicitly places it there: residual error describes observations, whereas this equation defines a parameter.

The main alternative forms—multiplicative log-linear effects and additive parameter-scale effects—are useful for different modeling assumptions. Multiplicative forms preserve positivity for positive \(\mathrm{TV}\theta\) and express covariate changes as ratios. Additive forms express absolute changes, but may produce implausible or nonpositive values outside the covariate range. Neither form should be selected merely because it is easier to calculate; follow the model as published and check whether its stated interpretation agrees with the equation.

Key takeaway: A faithful reconstruction identifies the typical value and its reference conditions, defines every covariate code and scale, places the random effect on the specified scale, and passes the reference-value, unit, and numerical checks.

Check Your Understanding

For each item, identify the equation or calculate the individual parameter. Solutions are included below the questions.

A hypothetical clearance model is \(CL_i=6\exp\{0.4(WT_i-70)/10+\eta_i\}\ \mathrm{L/h}\). Calculate \(CL_i\) for \(WT_i=65\ \mathrm{kg}\) and \(\eta_i=0.05\).
A hypothetical model has typical rate \(1.8\ \mathrm{h}^{-1}\) for the reference category. An indicator \(I_i\) is 1 for renal impairment and 0 otherwise; impairment multiplies the rate by 0.75, and the random effect is exponential. Find the rate for an impaired individual with \(\eta_i=0.10\).
In an additive model \(V_i=40+0.5(AGE_i-50)+\eta_i\), \(V_i\) is in litres and age is in years. What units must the coefficient 0.5 and \(\eta_i\) have? What value results at age 56 with \(\eta_i=-1\ \mathrm{L}\)?
In a multiplicative model, the equation is \(\theta_i=\mathrm{TV}\theta\exp(\beta x_i+\eta_i)\). What does the equation give when the covariate is at its reference value and \(\eta_i=0\)?

Solutions. (1) The exponent is \(0.4(65-70)/10+0.05=-0.20+0.05=-0.15\). Thus \(CL_i=6\exp(-0.15)=6(0.860708)=5.164\ \mathrm{L/h}\). (2) The equation is \(k_i=1.8(0.75)^{I_i}\exp(\eta_i)\ \mathrm{h}^{-1}\). For impairment, \(I_i=1\), so \(k_i=1.8(0.75)\exp(0.10)=1.35(1.10517)=1.492\ \mathrm{h}^{-1}\). (3) Since the terms are added on the parameter scale, 0.5 has units \(\mathrm{L/year}\), and \(\eta_i\) has units L. The result is \(40+0.5(56-50)-1=42\ \mathrm{L}\). (4) At the reference covariate value \(x_i=0\) and \(\eta_i=0\), \(\theta_i=\mathrm{TV}\theta\exp(0)=\mathrm{TV}\theta\).