Applying Dynamic Covariates to Individual Parameters

Why Evaluate a Covariate at Each Time?

An individual’s weight or renal function may change during a study. If a pharmacokinetic parameter depends on one of these measurements, using only a baseline value means the model does not update that parameter as the covariate changes. A dynamic covariate model instead evaluates the covariate function at the relevant time. The individual random effect remains the same: it represents a subject-specific deviation, not a new deviation drawn at every observation.

This distinction matters for prediction. A subject may have persistently higher-than-typical clearance because of their individual effect, while that subject’s clearance also changes as weight or renal function changes. The two contributions have different roles and should not be conflated.

Definition: A dynamic covariate model for an individual parameter specifies that parameter as a function of the individual’s covariate value at time $t$, together with a subject-specific random effect that is held fixed over the modeled period. For a positive parameter $P$, a common form is $$ P_i(t)=P_{\mathrm{typ}}\,g(x_i(t);\theta)\exp(\eta_{P,i}), $$ where $P_{\mathrm{typ}}$ is the typical parameter value, $x_i(t)$ is the covariate value for individual $i$ at time $t$, $g$ is a specified covariate function, $\theta$ is its parameter or parameter vector, and $\eta_{P,i}$ is the individual random effect for $P$.

In the usual NLME formulation, the random effects are modeled as population-level random variables across individuals. For an individual prediction, however, the subject’s $\eta_{P,i}$ is one fixed value across the times being evaluated. The covariate value can change from time to time; the individual random effect does not change unless the model explicitly gives it a time-varying structure.

The function $g$ must be specified before using the model. For a power covariate effect, $g(x_i(t);\theta)=(x_i(t)/x_{\mathrm{ref}})^\theta$, where $x_{\mathrm{ref}}$ is a reference value. A covariate history can be represented as stepwise or interpolated; the parameter calculation uses the value supplied by that representation at the time of evaluation. This tutorial concerns applying that value to the parameter, not choosing a history representation.

How the Dynamic Parameter Is Evaluated

For a weight effect on clearance, one possible model is

CL_i(t)=CL_{\mathrm{typ}}\exp(\eta_{CL,i}) \left(\frac{WT_i(t)}{WT_{\mathrm{ref}}}\right)^{\theta_{WT}}.

Here, $CL_{\mathrm{typ}}$ is typical clearance at the reference weight for an individual with $\eta_{CL,i}=0$. The exponent $\theta_{WT}$ controls the weight relationship. Holding the individual effect fixed, the ratio of the parameter at two times is

\frac{CL_i(t_2)}{CL_i(t_1)} = \left(\frac{WT_i(t_2)}{WT_i(t_1)}\right)^{\theta_{WT}}.

The typical clearance and the individual effect cancel. This ratio isolates the model-implied change associated with weight; it does not imply that weight caused a change in a real patient’s clearance.

Worked Applications

Worked Example: Updating Clearance as Weight Changes

Hypothetical model. Let $CL_{\mathrm{typ}}=5.0$ L/h, $WT_{\mathrm{ref}}=60$ kg, and $\theta_{WT}=0.75$. For one individual, let $\eta_{CL,i}=0.0953$, and suppose the represented weight history gives 60 kg at $t_1$, 66 kg at $t_2$, and 72 kg at $t_3$. Assume the random effect remains fixed at 0.0953 at all three times.

First, the individual multiplier is $\exp(0.0953)\approx1.1000$. At 60 kg, $$ CL_i(t_1)=5.0\exp(0.0953)(60/60)^{0.75} \approx5.50\ \mathrm{L/h}. $$ At 66 kg, the weight multiplier is $(66/60)^{0.75}=1.1^{0.75}\approx1.0741$, so $$ CL_i(t_2)=5.0\exp(0.0953)(1.1)^{0.75} \approx5.91\ \mathrm{L/h}. $$ At 72 kg, $1.2^{0.75}\approx1.14653$, which gives $$ CL_i(t_3)=5.0\exp(0.0953)(1.2)^{0.75} \approx6.31\ \mathrm{L/h}. $$

Using unrounded predictions, the ratio of clearance at 72 kg to clearance at 60 kg is $1.2^{0.75}\approx1.147$. Thus, the model predicts about 14.7% higher clearance at 72 kg than at 60 kg for this individual, with the individual effect unchanged. The calculation does not support changing $\eta_{CL,i}$ between those times.

Worked Example: Updating Clearance With Renal Function

Hypothetical model. Suppose clearance has a nonrenal fraction of 0.25 and a renal fraction of 0.75 at a reference eGFR of 90 mL/min/1.73 m². Use the model

CL_i(t)=CL_{\mathrm{typ}}\exp(\eta_{CL,i}) \left[0.25+0.75\left(\frac{eGFR_i(t)}{90}\right)\right].

Let $CL_{\mathrm{typ}}=6.0$ L/h and $\eta_{CL,i}=\ln(1.05)$, so $\exp(\eta_{CL,i})=1.05$. For this hypothetical individual, eGFR is 90 mL/min/1.73 m² at $t_1$, 60 at $t_2$, and 45 at $t_3$. The same individual effect is used at every time.

At $t_1$, the bracketed multiplier is $0.25+0.75(90/90)=1.00$, giving $6.0(1.05)(1.00)=6.30$ L/h. At $t_2$, it is $0.25+0.75(60/90)=0.75$, giving $6.0(1.05)(0.75)=4.725$ L/h. At $t_3$, it is $0.25+0.75(45/90)=0.625$, giving $6.0(1.05)(0.625)=3.9375$ L/h, or 3.938 L/h to four significant figures.

The modeled clearance at eGFR 60 is 0.75 times its value at the reference eGFR, and at eGFR 45 it is 0.625 times that reference value. These ratios follow from the specified hypothetical linear renal contribution. They should not be treated as a universal renal-clearance relationship.

Worked Example: Applying Separate Weight Effects to Clearance and Volume

Hypothetical model. Let clearance and volume both depend on weight, but with different exponents: $$ CL_i(t)=5.0\exp(\eta_{CL,i}) \left(\frac{WT_i(t)}{60}\right)^{0.75}\ \mathrm{L/h}, \qquad V_i(t)=40\exp(\eta_{V,i}) \left(\frac{WT_i(t)}{60}\right)^{1.0}\ \mathrm{L}. $$ Suppose $\exp(\eta_{CL,i})=1.10$, $\exp(\eta_{V,i})=1.05$, and weight changes from 60 to 72 kg. Both individual effects are held fixed.

At 60 kg, $CL_i=5.0(1.10)=5.50$ L/h and $V_i=40(1.05)=42.0$ L. At 72 kg, the clearance weight factor is $1.2^{0.75}\approx1.14653$, so $CL_i\approx5.50(1.14653)=6.306$ L/h, or 6.31 L/h. The volume weight factor is $1.2$, so $V_i=42.0(1.2)=50.4$ L.

The derived rate constant $k=CL_i/V_i$ is approximately $5.50/42.0=0.1310\ \mathrm{h}^{-1}$ at 60 kg and $6.306/50.4=0.1251\ \mathrm{h}^{-1}$ at 72 kg. Its ratio is $1.2^{0.75}/1.2=1.2^{-0.25}\approx0.9554$. Although both parameters increase, they do not increase by the same proportion, so their ratio decreases. This conclusion follows from the two covariate functions, not from a change in either random effect.

Why the Random Effect Cancels in a Time Ratio

For a positive parameter with the multiplicative model $P_i(t)=P_{\mathrm{typ}}g(x_i(t);\theta)\exp(\eta_{P,i})$, evaluate it at two times using the same individual effect:

Write both parameter values.
$P_i(t_1)=P_{\mathrm{typ}}g(x_i(t_1);\theta)\exp(\eta_{P,i})$ and $P_i(t_2)=P_{\mathrm{typ}}g(x_i(t_2);\theta)\exp(\eta_{P,i})$.

Form their ratio.
Divide the second expression by the first. The shared $P_{\mathrm{typ}}$ and $\exp(\eta_{P,i})$ cancel, leaving $P_i(t_2)/P_i(t_1)=g(x_i(t_2);\theta)/g(x_i(t_1);\theta)$.

Interpret what remains.
The within-individual change is determined by the change in the modeled covariate function. The cancellation requires the same individual effect and the same model parameters at both times.

This is an algebraic property of the specified model, not evidence that the covariate relationship is correct. If the effect is additive rather than multiplicative, or if the random effect is explicitly modeled as time-varying, this particular ratio result need not hold.

Assumptions, Diagnostics, and Alternatives

A dynamic covariate calculation is meaningful only under the assumptions built into its model. In particular, the covariate value at the evaluation time must be represented appropriately; the chosen function must be defined for the values used; and the individual random effect must be intended to remain constant over the interval. For the examples above, the power model assumes a positive weight and a fixed exponent, while the renal example assumes a linear eGFR contribution and a fixed renal fraction. Those are modeling choices, not automatic physiological facts.

Useful checks include plotting each individual’s covariate history alongside the resulting parameter curve, checking that reference values produce the intended reference multiplier, and verifying that calculations at two times reproduce the model’s implied ratio. Investigate sudden parameter jumps that arise from an unintended covariate value, implausible predictions at extreme values, or systematic residual patterns across covariate levels. A smooth-looking curve alone does not establish that the covariate effect is supported by the data.

A baseline-covariate model uses one value, often baseline weight or eGFR, throughout the modeled period. It is simpler and may be appropriate when changes are negligible or when only baseline information is available, but it cannot represent parameter updates associated with later covariate values. A dynamic model can use more information, but it relies on the covariate history and the assumed function being adequate. A direct time effect is another possible model component, but it represents a different relationship: change with time itself rather than change through the measured covariate. These approaches answer different modeling questions and should not be interchanged without justification.

Dynamic covariate models also have limits. Sparse or noisy covariate measurements can make the history uncertain; values far outside the observed range require extrapolation; and an association encoded in a covariate function is not, by itself, a causal effect. If the true individual parameter changes for reasons not captured by the covariate function, holding $\eta_{P,i}$ fixed does not make that change disappear; it means the model may fail to represent it.

Common Mistake

A common error is to recalculate or redraw the individual random effect whenever the covariate changes. For example, treating $\eta_{CL,i}$ as one value at 60 kg and a newly sampled value at 72 kg confounds the covariate effect with unexplained between-individual variability. In the standard model used here, the same subject-specific effect multiplies the parameter at both times; only $g(x_i(t);\theta)$ is reevaluated. Another error is to compare rounded predictions and report a ratio that disagrees with the model’s exact covariate ratio. Carry enough digits through the calculation, then round the reported values.

Key takeaway: Evaluate the covariate function using the individual’s value at each time, but keep that individual’s random effect fixed. For a multiplicative parameter model, the within-individual ratio across times is the ratio of the covariate functions.

Check Your Understanding

Answer each question, then check the solutions below.

In a model $P_i(t)=P_{\mathrm{typ}}(X_i(t)/X_{\mathrm{ref}})^{0.5}\exp(\eta_{P,i})$, what changes between times if the covariate changes but the individual effect is fixed?
For a power weight effect with exponent 0.75, find the parameter ratio when weight changes from 50 kg to 80 kg.
In the hypothetical renal model above, what is the bracketed multiplier at eGFR 30 mL/min/1.73 m²?
Why is a baseline-covariate model unable to represent a parameter update due to a later measured covariate value?
Name one diagnostic that could reveal a problem with a dynamic covariate model.

Solutions. (1) The covariate function changes; $P_{\mathrm{typ}}$ and $\eta_{P,i}$ stay fixed. (2) The ratio is $(80/50)^{0.75}=1.6^{0.75}\approx1.423$. (3) The multiplier is $0.25+0.75(30/90)=0.500$. (4) It uses the same selected covariate value throughout, so it does not reevaluate the function at the later value. (5) For example, plot predicted individual parameter curves against covariate histories and check for implausible jumps or systematic residual patterns across covariate levels.