Random Effects on Initial Conditions

Why the Initial Condition Can Vary Between Subjects

A pharmacokinetic trajectory depends not only on the kinetic law and the stochastic forcing, but also on where the system starts. Subjects may differ in their amount in a compartment at the beginning of observation because of prior exposure, incomplete washout, variable loading, or uncertainty in the amount actually present. Treating every subject as if the same known initial amount were present can therefore attribute starting-level differences to the wrong part of the model.

A random initial condition is an individual-level quantity drawn once for a subject and then propagated by that subject’s SDE. It is not a new random draw at each time. This differs from a random kinetic parameter, such as a subject-specific elimination rate, which changes the law governing the trajectory over time. Nested integration over random effects and latent paths provides the general marginal-likelihood framework; the focus here is the specific role of a random starting state.

For example, consider the linear SDE

where \(k>0\) and \(\sigma\geq 0\) are population parameters, \(W_i(t)\) is standard Brownian motion, and \(F_0\) is the population distribution of the initial state. Unless stated otherwise, assume \(X_i(0)\) is independent of \(W_i\), and \(E[X_i(0)^2]<\infty\). The solution is interpreted on \(t\geq0\). This additive-noise model is useful for deriving propagation and variance calculations, although its state is not constrained to remain nonnegative.

How Initial-State Variability Propagates

The integrating-factor solution separates the part inherited from the initial state from the part accumulated through process noise. That separation makes the source of variability explicit.

Proof. Apply Itô’s product rule to \(e^{kt}X_i(t)\). Since \(e^{kt}\) has finite variation, its quadratic covariation with \(X_i(t)\) is zero, and \(d(e^{kt}X_i(t))=\sigma e^{kt}\,dW_i(t)\). Integrating from 0 to \(t\), then multiplying by \(e^{-kt}\), gives the stated solution. The stochastic integral has mean zero, so taking expectations gives the mean formula.

For the variance, write the solution as the sum of \(e^{-kt}X_i(0)\) and the stochastic-integral term. The two terms are independent because \(X_i(0)\) is independent of the Brownian path. Itô’s isometry gives the integral’s variance:

The initial-state term has variance \(e^{-2kt}\operatorname{Var}(X_i(0))\). Adding the two variances proves the result. \(\square\)

Thus, initial-state variance is attenuated by \(e^{-2kt}\), while process-noise variance accumulates over time. The factors differ: the initial-state mean is multiplied by \(e^{-kt}\), and its variance by \(e^{-2kt}\). For \(s\leq t\), the same solution also yields

The first term is the shared initial-state contribution to the covariance. It decreases as both observation times move away from zero.

Random Initial States Are Not Random Kinetics

A useful way to distinguish these effects is to ask which feature of the trajectory changes. With a common fixed \(k\), different initial states change the amplitude, while the deterministic decay rate remains the same. With a common initial state and different \(k_i\), subjects can have different decay rates. Process noise adds further within-subject variation and does not, by itself, represent persistent between-subject differences.

Estimation must account for observation error. For instance, if \(Y_{ij}=X_i(t_{ij})+\epsilon_{ij}\), with independent mean-zero errors of variance \(\sigma_e^2\), then at time zero \(\operatorname{Var}(Y_{i0})=\operatorname{Var}(X_i(0))+\sigma_e^2\). Observed variability at the first sample is not automatically all between-subject initial-state variability. With later observations, the marginal distribution also includes propagated process noise. A likelihood must integrate over the unobserved \(X_i(0)\), rather than treating its individual realized value as known.

Variance Decomposition, Diagnostics, and Limits

The separation of latent variability from residual variability is an instance of the law of total variance. It remains useful when the initial distribution or kinetic model is more complicated than the linear example.

Proof. Write \(X-E[X]=(X-E[X\mid Z])+(E[X\mid Z]-E[X])\). The expected cross-product of these terms is zero: conditioning on \(Z\), the second term is fixed and \(E[X-E[X\mid Z]\mid Z]=0\). Squaring and taking expectations therefore gives

The first term is \(E[\operatorname{Var}(X\mid Z)]\), and the second is \(\operatorname{Var}(E[X\mid Z])\), proving the identity. \(\square\)

For the SDE, conditioning on \(X_i(0)\) separates conditional process variation from variation in the conditional mean. In particular, the law gives the same initial-state and process-noise contributions as the explicit solution. If \(k_i\) is random or associated with \(X_i(0)\), the simple fixed-\(k\) formulas cannot generally be applied by substituting a population-average rate. Instead, condition on the random effects, compute conditional moments under the model, and then average over their joint distribution.

Several patterns should prompt a model check. An estimated initial-state variance at its zero boundary may mean little evidence for between-subject starting variation, but it can also reflect weak early sampling or confounding with measurement error. A large estimated variance, implausible starting concentrations, or strong dependence between fitted initial states and fitted kinetic parameters may indicate an overly flexible model or inadequate sampling. Compare predictions near the start and later in time, and check whether the model reproduces both between-subject spread and within-subject evolution.

The main alternatives answer different questions. A common fixed initial state is appropriate when the start is known and genuinely shared. A random initial state represents uncertainty or heterogeneity in that starting amount. A random kinetic parameter represents heterogeneity in the subsequent dynamics. These effects may be included together only when the design can support their separation; adding both automatically can weaken identifiability.

Finally, the additive-noise linear SDE is not a universally suitable amount model: Gaussian noise can assign positive probability to negative amounts. If nonnegativity is essential, use a state-space formulation whose support and dynamics match the application. The propagation calculations here explain the role of random initial conditions, but do not remove the need to justify the chosen SDE, initial-state distribution, and observation model.