Building EBE-Based Individual Concentration Predictions

From Individual Parameter Estimates to Concentration Profiles

An empirical Bayes estimate (EBE) is useful because it can be inserted into a structural model to produce an individual concentration-time profile. The resulting curve is not a direct measurement of a person's complete concentration history. It is a model-based prediction, conditional on the fitted population parameters, that person's covariates, and their EBE.

The key comparison is between two profiles. The covariate-specific typical profile uses the typical parameters predicted for the person's covariates, with all individual random effects set to zero. The EBE-based profile uses the same covariate-conditioned parameters but includes the person's estimated random effects. Because EBEs can be shrunk toward zero, an EBE-based profile may be closer to the typical profile than a profile based on more individual information would be.

This construction assumes that the fitted structural model is appropriate for the dosing history and time scale being used; the covariate model specifies the typical parameters correctly; and the random effects have the distribution and covariance structure assumed in the fitted model. The prediction also presumes that the relevant dose, covariates, and model parameter estimates are supplied consistently. An EBE-based curve is conditional on those inputs: it does not by itself show parameter uncertainty, uncertainty in the EBE, or future residual variability.

Plugging an EBE into a Structural Model

Consider a one-compartment model after an intravenous bolus dose \(D\), with first-order elimination. Let \(V_i\) be volume of distribution and \(CL_i\) clearance for individual \(i\). Then the predicted concentration at elapsed time \(t\) is

Suppose clearance has a log-normal random effect and a covariate-conditioned typical value \(CL_{\mathrm{typ},i}\):

The typical profile uses \(\eta_{CL,i}=0\), so its clearance is \(CL_{\mathrm{typ},i}\). The EBE-based profile uses \(\hat{\eta}_{CL,i}\), giving \(\widehat{CL}_i=CL_{\mathrm{typ},i}\exp(\hat{\eta}_{CL,i})\). The exponential matters: an EBE on the log scale is not a clearance difference in the original units. For multiple random effects, insert the full EBE vector into the parameter relationships used by the model; do not omit estimated effects or change the scale on which they are defined.

Why Shrinkage Pulls the Profile Toward the Typical Curve

For a random effect with prior mean zero, shrinkage means that an EBE is closer to zero than the individual information alone would suggest. Setting the random effect to zero restores the covariate-specific typical parameter. Consequently, shrinkage often makes EBE-based predictions closer to the typical profile than predictions made using a less-shrunk individual estimate.

The simple calculation in the example illustrates the mechanism. If a scalar random effect has prior \(\eta\sim N(0,\omega^2)\), and an approximate individual estimate \(z\) has sampling distribution \(z\mid\eta\sim N(\eta,s_z^2)\), the conditional density is proportional to

Completing the square in \(\eta\), or differentiating the exponent, yields the conditional mean and mode

The multiplier lies between zero and one when both variances are positive. As the individual estimate becomes less precise (larger \(s_z^2\)), the multiplier decreases and the conditional estimate moves toward zero. This is a Gaussian illustration, not a universal closed-form formula for nonlinear mixed-effects models. In a nonlinear model the EBE is obtained from the fitted conditional distribution, and its relation to an unshrunk estimate need not follow this exact scalar expression.

The resulting prediction need not be a linear interpolation between two concentration values. The random effect may enter a parameter exponentially, and the structural model may then transform that parameter again. For the clearance model above, a negative EBE lowers clearance, but the concentration change depends on both the time and the other parameters. At \(t=0\), for example, \(C(0)=D/V\), so a clearance EBE alone does not change the predicted initial concentration.

Covariates Define Which Typical Profile Is the Reference

The typical profile is not necessarily a single population-wide curve. If typical clearance depends on a covariate, each individual's covariate values determine the zero-random-effect reference. For example, suppose

where \(WT_i\) is body weight. A person weighing 56 kg has \(CL_{\mathrm{typ},i}=6.000(56/70)^{0.75}\approx5.075\) L/h. The typical profile for that person uses approximately \(5.075\) L/h, not the reference clearance of \(6.000\) L/h. An EBE-based prediction for this person incorporates the same weight-based value and then applies the estimated random effect, for example \(CL_i=5.075\exp(\hat{\eta}_{CL,i})\).

Interpretation, Alternatives, and Failure Checks

An EBE plug-in curve is one useful summary, but it is not the only possible prediction. The typical profile answers what the model predicts at that person's covariates when random effects are zero. The EBE-based plug-in profile answers what the structural model predicts after inserting the person's estimated random effects. A conditional mean prediction instead averages the structural prediction over the conditional distribution of random effects. Because \(f\) is generally nonlinear, \(f(t;g(x_i,\beta,E[\eta_i\mid y_i]))\) need not equal \(E[f(t;g(x_i,\beta,\eta_i))\mid y_i]\). A full predictive distribution can additionally account for uncertainty and residual variability. These summaries answer different questions and should not be treated as interchangeable.

Before interpreting an EBE-based profile, check the model parameterization and calculation. Confirm that the EBE is on the scale expected by the parameter relationship; use the person's covariates in the typical parameter function; convert units consistently; and use elapsed time since the modeled dose, not an incompatible clock time. For the bolus model, \(CL/V\) must have units of inverse time and \(D/V\) units of concentration. A negative post-dose exponent should give a declining profile when clearance is positive.

A prediction that looks implausible can arise from a coding or unit error, but it can also indicate a poor structural model, unsuitable covariate relationship, or weak individual information. Shrinkage is not evidence that an individual truly has exactly typical parameters: it indicates that the conditional estimate is pulled toward the random-effect distribution's center. Nor does a close EBE-based fit to the observations establish that the population model is correct. Compare predictions with the intended use and observed data, and retain the distinction between a conditional point prediction and uncertainty around it.

• Wrong: “The EBE is negative, so concentration must be lower.” Correction: The effect depends on which parameter the EBE modifies. A negative log-clearance effect lowers clearance and, after a bolus, raises concentrations relative to the typical profile for \(t>0\).
• Wrong: “The EBE-based curve is the person's observed concentration curve.” Correction: It is a structural-model prediction conditional on estimated parameters; observations also include residual variation and measurement error.
• Wrong: “Use the population reference parameters for every typical curve.” Correction: First calculate the typical parameters at the person's covariate values; the zero-random-effect profile is covariate-specific.