Time-Varying Covariates in Software Implementations

Why Covariate Timing Is Part of the Model

A renal-function covariate is not fully specified by its name and effect equation. If estimated or measured renal function changes during follow-up, the implementation must also say when the model uses each value and what value applies between measurements. Two tools can contain the same clearance equation and the same covariate data yet produce different predictions because one holds values constant between records while another interpolates them.

The distinction matters whenever the covariate affects a parameter used to predict concentrations. A change in renal function can alter clearance; clearance in turn affects the rate at which drug is removed. A time-varying input therefore changes the model's predictions over time, not just a static subject characteristic. The earlier material on covariate effects and ODE translation provides the general parameter and state-variable context; here the focus is on how time-indexed covariate values are represented in software.

Neither rule is automatically correct for every renal-function variable. A record-based step function can be suitable when a value is treated as unchanged until a new measurement becomes available. Linear interpolation can be a defensible approximation when the underlying quantity is assumed to change gradually. The choice is a modeling assumption, not a cosmetic software option. It should reflect the measurement process and be applied consistently to all individuals.

Define the Time Rule Before Writing Tool Syntax

Let \(R(t)\) denote renal function at time \(t\), measured in mL/min, and let clearance be related to it by a power model. For illustration, suppose typical clearance at \(R_{\mathrm{ref}}=90\ \mathrm{mL/min}\) is \(5.00\ \mathrm{L/h}\), and the renal-function exponent is \(0.75\). The structural relationship is

The ratio is dimensionless because both renal-function values use mL/min. This example treats the coefficient and exponent as fixed solely to isolate the timing question; it does not imply that these are clinically established values. In a population model, the equation could include between-individual effects and estimated parameters, but the time-handling issue remains.

Suppose \(R(0)=90\ \mathrm{mL/min}\) and \(R(12)=60\ \mathrm{mL/min}\). Under last observation carried forward (LOCF), \(R(t)=90\) for \(0\leq t<12\), then the value changes to 60 at 12. Under linear interpolation,

The expression only defines the interval between the two observations. A separate rule is needed before the first covariate record and after the last one; silently extending a rule beyond observed times can amount to an unexamined extrapolation assumption.

Nonlinear Effects Make the Interpolation Target Matter

A second choice is whether to interpolate the measured covariate and then apply its effect, or to calculate the parameter at observed covariate times and interpolate that parameter. These approaches are generally unequal for a nonlinear effect, though equality can also occur for particular endpoints or interpolation weights.

For linear interpolation of \(R\), define \(w=t/12\). The model first forms \(R(t)=(1-w)R(0)+wR(12)\), then evaluates \(g(R(t))\), where \(g(R)=5.00(R/90)^{0.75}\). Interpolating the endpoint clearances instead gives \((1-w)g(R(0))+wg(R(12))\). These expressions need not match because applying a nonlinear function and taking a weighted average are different operations.

At 12 hours, the clearance calculated from the second renal-function value is approximately \(5.00(60/90)^{0.75}=3.689\ \mathrm{L/h}\). Linear interpolation of clearance at the midpoint, using unrounded endpoint values, gives

This differs from \(4.361\ \mathrm{L/h}\), the clearance obtained by interpolating renal function to \(75\ \mathrm{mL/min}\) and then applying the power model. A modeler should therefore identify what is being interpolated, not just select a setting called “linear.”

Represent the Same Rule in Four Tools

Across tools, separate three tasks: prepare time-stamped covariate values, declare or make those values available to the model, and specify the rule between records. In every case, inspect the data actually passed to the engine. A covariate entered only on laboratory-measurement rows may not be available at other event records in the way intended; preprocessing may need to carry values to relevant records or the tool must be configured to use the time-varying series.

NONMEM. Include renal function as an input variable and reference it in the parameter calculation, for example in a clearance expression corresponding to \(CL=TVCL(R/90)^{\theta_R}\). For a record-based implementation, supply the intended renal-function value on records at which it must be available, and ensure that the model recomputes the relevant parameter at the update. Do not assume that two measurements create a continuous line between them: the data records alone do not state that assumption. If continuous interpolation is intended, make the time-dependent rule explicit and verify how it is evaluated by the selected model structure and estimation setup.

Monolix. Make the renal-function column available as a covariate, classify it as time-varying, and use it in the structural parameter relationship. Then inspect the covariate configuration and its interpolation setting. A continuous covariate may be treated differently from a categorical one, and the setting determines what value the model uses between observed times. The equation may look identical to the NONMEM equation, but that alone does not establish equivalent timing behavior.

nlmixr2. Put renal function in the longitudinal data and refer to its column name in the model expression for clearance. The fitting workflow must also use a defined covariate interpolation behavior; the available control and defaults can depend on the estimation interface and software version. For reproducibility, record the selected behavior explicitly rather than relying on a default. Check that covariate times are correctly ordered within each individual and that values are available across the records needed for the likelihood calculation.

mrgsolve. Provide the time-varying covariate in the event data and use the corresponding data variable when calculating clearance in the model. For a step update, include a record at each time when the value changes, with the new value assigned from that point forward according to the event-data semantics. Do not treat data records as a continuous trajectory unless interpolation is explicitly coded or otherwise configured. When a simulation requires continuously changing renal function, implement that trajectory deliberately and verify it at interior times.

Verification, Failure Modes, and Limitations

A reliable translation tests the covariate behavior independently of parameter estimation. Use a small hypothetical dataset with known values and evaluate the covariate and derived clearance at three times: just before an update, exactly at the update, and between two measurements. At time 6 in the example, LOCF predicts 5.00 L/h, whereas linear interpolation of renal function predicts about 4.361 L/h. These checks reveal whether an update is delayed, applied at the wrong record, or interpolated unexpectedly.

Also check the following before comparing model predictions across software:

• Units and scale. Renal function in the data and the reference value in the equation must use the same units. A value in a different unit system changes the ratio unless converted.
• Record alignment. Confirm that every relevant observation and event sees the intended covariate value, especially when measurements and dosing records occur at different times.
• Update timing. Establish whether the new value applies at the record time itself or only after it, and check duplicate records at the same time.
• Boundaries. Decide what the model uses before the first covariate measurement and after the last. Avoid unintentional extrapolation.
• Transformation order. Determine whether interpolation applies to renal function or to a derived parameter such as clearance.
• Version and defaults. Record software versions and explicit settings; a default is not a portable model specification.

A useful cross-tool diagnostic is to hold all parameters and random effects fixed and compare the implied clearance trajectory before comparing concentrations. If clearances disagree, the source may be covariate timing, data handling, or the transformation order; investigating those first is more informative than changing estimation settings. Once the clearance trajectories match, concentration predictions can be compared under the same doses, events, and ODE model.

Record-based updates are transparent and often align naturally with measured covariate histories, but they imply abrupt changes and may be a poor representation of a continuously evolving quantity. Interpolation gives intermediate values but assumes a trajectory that was not directly observed; linear interpolation can be particularly questionable across long gaps or abrupt clinical changes. A static baseline covariate is simpler, but it discards within-subject changes. The suitable choice depends on the scientific question and measurement schedule, and sensitivity analysis can show whether conclusions depend on it.