Bayesian Forecasts With Changing Covariates During Follow-Up

Why Covariate Change Matters in an Individual Forecast

An individual PK forecast is conditional on more than the patient’s estimated parameters and dosing record. If renal function, body size, or another covariate changes during follow-up, and the PK model links that covariate to a parameter, the forecast should reflect the expected covariate path. Otherwise, the forecast can remain mathematically precise while describing a patient whose physiology is no longer represented by the model inputs.

The key distinction is between a patient-specific parameter and a covariate that changes over time. In a common model, an individual’s random effect, such as \(\eta_{CL,i}\), represents a persistent patient-level deviation over the forecast window. A changing covariate modifies the typical parameter value as time passes. The model’s covariate relationship determines how large that modification is; a linear eGFR relationship used below is illustrative, not a universal clinical rule.

Definition: A changing-covariate Bayesian forecast is a prediction of future PK quantities that conditions on the patient’s observed data and dosing history, represents covariates along their expected or uncertain future trajectory, and propagates uncertainty in the individual parameters and, when relevant, in the covariate trajectory.

For example, a clearance model might be written as

CL_i(t)=CL_{\mathrm{ref}}\exp(\eta_{CL,i})f_R(R_i(t))f_W(W_i(t)),

where \(R_i(t)\) and \(W_i(t)\) are renal function and body size at time \(t\), and \(f_R\) and \(f_W\) are the specified covariate functions. A typical allometric body-size term could be \(f_W(W)=(W/W_{\mathrm{ref}})^{0.75}\), but the selected relationship must be justified by the fitted model. The posterior distribution for \(\eta_{CL,i}\) comes from the individual’s available data and the population model; changing \(R_i(t)\) or \(W_i(t)\) does not, by itself, mean that the individual random effect has changed.

From a Covariate Path to a Concentration Forecast

Consider a one-compartment model after an IV bolus, with amount \(A_i(t)\), constant volume \(V_i\), and clearance \(CL_i(t)\) that may vary with time. Between doses, the mass-balance equation is

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

This form assumes linear elimination, fixed volume over the interval, and a clearance trajectory specified by the covariate model. Dividing the differential equation by \(A_i(t)\) and integrating from \(s\) to \(t\) gives

\log\left(\frac{A_i(t)}{A_i(s)}\right) =-\int_s^t\frac{CL_i(u)}{V_i}\,du, \qquad C_i(t)=C_i(s)\exp\left[-\int_s^t\frac{CL_i(u)}{V_i}\,du\right].

Thus the forecast depends on the integrated elimination rate over the interval, not merely the covariate value at its start or end. With multiple bolus doses and linear disposition, each dose contributes its surviving concentration. For doses \(D_j\) given at times \(t_j\), the concentration at a later time \(t\) is

C_i(t)=\sum_{j:t_j\le t}\frac{D_j}{V_i} \exp\left[-\int_{t_j}^{t}\frac{CL_i(u)}{V_i}\,du\right].

This sum presumes that the doses and their times are known, that superposition applies, and that the same individual parameter and covariate path govern each contribution. The earlier material on Forecasting Under Nonlinear Pharmacokinetics covers models where superposition does not apply; in that setting, the covariate path must instead enter the model’s state equations.

A Bayesian forecast averages over uncertainty rather than replacing every unknown by a single estimate. Let \(D_i\) denote the patient’s observed data, and let \(\theta_i\) collect relevant individual parameters. If the future covariate path \(X_i(t)\) is uncertain, the predictive distribution has the form

p(C_{\mathrm{future}}\mid D_i) =\int p(C_{\mathrm{future}}\mid\theta_i,X_i(\cdot),\text{dosing}) \,p(\theta_i,X_i(\cdot)\mid D_i)\,d\theta_i\,dX_i(\cdot).

When the covariate path is treated as known, the integral over \(X_i(\cdot)\) is omitted. In practice, posterior draws of individual parameters and, where appropriate, covariate paths can each be passed through the PK model. The resulting predictions form a distribution for future concentrations. A prediction for the latent concentration and a prediction for a future measured concentration are not identical: the latter also includes the applicable residual and assay error, as described in the earlier material on Propagating Assay and Residual Error Into Forecasts.

Worked Example: A Renal-Function Change After a Bolus

Suppose a hypothetical patient receives an IV bolus that produces a concentration of \(12.0\) mg/L at time zero. Volume is \(30.0\) L. The fitted model uses \(CL_i(t)=CL_{\mathrm{ref}}\{eGFR(t)/90\}\), with eGFR in the model’s specified units. For illustration, consider two equally weighted posterior draws for \(CL_{\mathrm{ref}}\): \(3.00\) and \(4.00\) L/h. eGFR is 90 for the first 2 hours and 60 for the next 4 hours. Find the posterior mean of the latent concentration at hour 6 using these two draws.

The covariate ratios are 1 for 2 hours and \(60/90=2/3\) for 4 hours. For a posterior draw \(CL_{\mathrm{ref}}\), the integrated elimination rate is

\int_0^6\frac{CL_i(u)}{V_i}\,du =\frac{CL_{\mathrm{ref}}}{30.0}\left(2(1)+4\left(\frac{2}{3}\right)\right) =\frac{7}{45}CL_{\mathrm{ref}}.

For the \(3.00\) L/h draw, the integrated rate is \((7/45)(3.00)=0.4667\), so \(C_i(6)=12.0\exp(-0.4667)\approx7.525\) mg/L. For the \(4.00\) L/h draw, it is \((7/45)(4.00)=0.6222\), so \(C_i(6)=12.0\exp(-0.6222)\approx6.441\) mg/L. The two draws have equal weight, so their posterior mean forecast is \((7.525+6.441)/2\approx6.983\) mg/L. This is a mean over the two illustrative posterior draws, not a complete prediction interval.

The assumptions here are a constant volume, linear elimination, the stated eGFR relationship, and the specified piecewise-constant eGFR path. If the covariate relationship were different, or the eGFR transition occurred at another time, the integrated rate and forecast would change.

Multiple Doses and a Covariate Change During Follow-Up

For repeated doses, each dose’s contribution is attenuated according to the covariate-dependent elimination rate over the full time since that dose. The same covariate change can therefore affect dose contributions unequally: an older dose accumulates exposure to the changing clearance over a longer interval.

Worked Example: Forecasting After a Drop in Renal Function

In a hypothetical one-compartment example, \(V_i=20.0\) L and one posterior draw gives \(CL_{\mathrm{ref}}=2.00\) L/h. The model sets clearance equal to \(CL_{\mathrm{ref}}\) when eGFR is 90, and proportional to eGFR. A 100 mg IV bolus was given 8 hours before the forecast origin and another 100 mg bolus 4 hours before it. eGFR was 90 up to the forecast origin, then is expected to remain 45 for the next 2 hours. Calculate the latent concentration 2 hours after the forecast origin.

During the historical 8 or 4 hours at eGFR 90, clearance is \(2.00\) L/h. During the 2 forecast hours at eGFR 45, it is \(2.00(45/90)=1.00\) L/h. Each bolus initially contributes \(100/20.0=5.00\) mg/L. For the older dose, the integrated rate is \((2.00/20.0)(8)+(1.00/20.0)(2)=0.900\). For the more recent dose, it is \((2.00/20.0)(4)+(1.00/20.0)(2)=0.500\). Therefore,

C_i=5.00e^{-0.900}+5.00e^{-0.500} =2.033+3.033 \approx5.066\ \text{mg/L}.

If clearance were incorrectly held at \(2.00\) L/h throughout the forecast period, the calculation would instead be \(5.00e^{-1.00}+5.00e^{-0.600}=1.839+2.744\approx4.583\) mg/L. The difference arises because the forecast uses lower clearance during follow-up. This comparison isolates the effect of the covariate path for one parameter draw; a complete Bayesian forecast would repeat the calculation across the patient’s posterior parameter distribution.

Uncertain Future Covariates and Practical Alternatives

A covariate trajectory may itself be uncertain. For instance, the future eGFR value may depend on clinical events not yet observed. A forecast can represent that uncertainty with scenarios or with a probability model for the future trajectory. Averaging the covariate first and then predicting is generally not equivalent to averaging predictions across covariate paths, because the PK prediction is a nonlinear function of the integrated rate.

Worked Example: Why the Average Covariate Is Not the Average Forecast

Suppose a hypothetical patient has \(C_i(0)=10.0\) mg/L, \(V_i=25.0\) L, and \(CL_{\mathrm{ref}}=2.50\) L/h in a posterior draw. During the next 4 hours, assume two possible eGFR paths, each with posterior predictive probability 0.5: eGFR remains 90, or it remains 45. The model uses \(CL_i(t)=CL_{\mathrm{ref}}\{eGFR(t)/90\}\). Calculate the mean forecast over these paths and compare it with a forecast made at the average eGFR of 67.5.

If eGFR remains 90, the elimination rate is \(2.50/25.0=0.1000\) per hour, giving \(10.0e^{-(0.1000)(4)}\approx6.703\) mg/L. If eGFR remains 45, the rate is \(1.25/25.0=0.05000\) per hour, giving \(10.0e^{-(0.05000)(4)}\approx8.187\) mg/L. The mean over the two equally probable paths is \((6.703+8.187)/2\approx7.445\) mg/L.

The average eGFR is 67.5, which gives a rate of \(\{2.50(67.5/90)\}/25.0=0.07500\) per hour. A plug-in forecast at that average is \(10.0e^{-(0.07500)(4)}\approx7.408\) mg/L, not 7.445 mg/L. The difference illustrates that averaging covariates before a nonlinear prediction does not generally reproduce the average of predictions. The two calculations answer different questions; when paths are uncertain, the desired predictive distribution should reflect the distribution over paths.

The main alternatives are to hold a covariate fixed at its latest value, insert a planned future value, forecast over several plausible scenarios, or model a probability distribution for the future covariate. Holding the value fixed can be adequate when change is negligible over the forecast horizon. A planned value is suitable when there is a credible clinical plan and the forecast is explicitly conditional on that plan. Scenarios are transparent when only a small number of future paths are plausible. A probability model is more complete when path uncertainty can be characterized and is important to the decision.

Assumptions, Diagnostics, and Common Pitfalls

A changing-covariate forecast depends on both the PK model and the covariate trajectory supplied to it. Before interpreting it, check that the covariate definition, units, measurement times, and transformations match those used in model development. Confirm that the trajectory is plausible over the forecast horizon and that the fitted covariate relationship is being used within a clinically defensible range. If a covariate changes the volume as well as clearance, the fixed-volume derivation above is insufficient: amount and concentration must be related using the specified time-varying volume model.

Useful diagnostics include plotting predicted clearance or other covariate-dependent parameters against time, marking observed and assumed covariate values, and checking for implausible jumps. Compare forecasts under reasonable alternative covariate paths to see whether the clinical conclusion is sensitive to path assumptions. Examine patient-level predictive checks and forecast calibration where available; a systematic pattern in future residuals may indicate a poor covariate relationship, erroneous timing, or another structural problem. An apparently smooth concentration curve does not establish that the assumed covariate path is correct.

Freezing a changing covariate without justification. Carrying the last eGFR or body weight forward may be a useful scenario, but it is not automatically an unbiased forecast.
Changing the individual effect whenever the covariate changes. A covariate acts through the model’s specified parameter relationship. It does not imply that the patient’s posterior \(\eta\) must be re-estimated at each time point.
Using a future measurement as if it were already known. If that value is uncertain at forecast time, treating it as fixed understates predictive uncertainty. State whether the forecast is conditional on a planned or assumed value.
Plugging in a mean path and calling it the full Bayesian forecast. A forecast at an average covariate is not generally the same as the average over uncertain paths, as the worked example shows.
Extrapolating a covariate relationship without scrutiny. A relationship estimated over one range need not remain valid far outside it; the resulting forecast may be model-driven extrapolation rather than patient-specific evidence.

Key takeaway: For a changing-covariate forecast, carry the covariate path through the PK model over time, then propagate posterior uncertainty through the resulting prediction. Make clear which parts of the path are observed, assumed, or uncertain.

Check Your Understanding

Answer each question, then compare with the brief solutions below.

With constant volume and time-varying clearance, what quantity determines the fraction of concentration remaining between two times?
A patient’s covariate changes during a forecast, but the individual random effect is held fixed. Is this internally consistent with the model described here? Explain.
For a linear one-compartment model with two bolus doses, why can the older dose be affected differently from the newer dose by a clearance change?
When might a forecast at a fixed future covariate value be appropriate, and what should be stated about its interpretation?

Solutions. (1) The integrated elimination rate, \(\int_s^t CL_i(u)/V_i\,du\); the remaining fraction is its negative exponential. (2) Yes, if the model specifies a persistent individual random effect and the covariate acts through its time-varying parameter relationship. (3) The older dose has been exposed to the clearance trajectory for longer, so its integrated elimination rate can differ. (4) It can be appropriate when the value is planned or is a defensible scenario; the forecast should be identified as conditional on that assumed value rather than presented as accounting for all covariate uncertainty.