Crossover Concentration In Combined Residual Error

Why a Crossover Concentration Is Useful

A combined residual error model allows absolute error to matter at low predicted concentrations and concentration-dependent error to matter at higher concentrations. The crossover concentration gives a direct way to locate where those two contributions are equal. It is a property of the fitted error-model parameters, not a concentration at which the structural pharmacokinetic model changes.

The distinction is useful when interpreting a model over its observed concentration range. If most predictions are well below the crossover, the additive contribution may dominate. If most are well above it, the proportional contribution may dominate. A crossover far outside the observed range can still be calculated, but it may have little practical support from the data.

Model and Definition

Let \(C\) denote a nonnegative model-predicted concentration, and let \(Y\) denote the corresponding observed concentration. Define the residual as \(Y-C\). In the combined error model considered here, the conditional residual mean is zero and the conditional variance is the sum of an additive variance contribution and a proportional variance contribution:

Here \(\sigma_{\mathrm{add}}\) is the additive residual standard deviation, with the same units as concentration. The proportional residual standard deviation \(\sigma_{\mathrm{prop}}\) is dimensionless when the same concentration units are used for \(C\) and \(Y\). The model is often paired with conditionally normal residuals, but normality is not needed to define or calculate the crossover: the crossover follows from the variance contributions alone.

Because \(\sigma_{\mathrm{prop}}\) is dimensionless, this ratio has concentration units. The definition compares variance contributions, not a concentration value with a residual standard deviation, and not the total residual standard deviation with either component. At the crossover, each component contributes half of the total variance.

Deriving and Interpreting the Crossover

For positive \(\sigma_{\mathrm{add}}\) and \(\sigma_{\mathrm{prop}}\), equality of the two variance contributions gives:

The nonnegative solution is selected because concentration is nonnegative. To interpret the result at any \(C>0\), compare the proportional variance contribution with the additive one:

Thus, below the crossover, the additive contribution is larger; at the crossover, the contributions are equal; and above it, the proportional contribution is larger. The ratio grows quadratically with concentration. For example, at twice the crossover, proportional variance is four times additive variance—not merely twice as large.

When both components are positive, their fractions of total variance can also be written in terms of the crossover:

These fractions sum to one when the total variance is positive. The crossover-based expressions require both parameters to be positive. Boundary cases must be handled directly: if \(\sigma_{\mathrm{prop}}=0<\sigma_{\mathrm{add}}\), the additive fraction is 1 at every concentration. If \(\sigma_{\mathrm{add}}=0<\sigma_{\mathrm{prop}}\), the proportional fraction is 1 for \(C>0\); at \(C=0\), total variance is zero and the fractions are undefined. If both parameters are zero, total variance is zero at all concentrations and there is no positive variance to partition.

Practical Calculations and Units

The crossover is meaningful only in relation to the units and concentration range of the model. If the concentration unit changes from mg/L to µg/L, the numerical values of \(C\) and \(\sigma_{\mathrm{add}}\) both change by the same factor, while \(\sigma_{\mathrm{prop}}\) does not. The numerical crossover changes accordingly, but it represents the same physical concentration. Before comparing a crossover with data, make sure the additive standard deviation and predicted concentrations use the same units.

A crossover can also show whether one component is practically relevant in the concentration range of interest. If predictions span 5–25 ng/mL in the example above, all are below 80 ng/mL and the additive component is larger throughout that range. This does not prove that a proportional component is unnecessary: model fit, uncertainty, and the data’s information about the parameters also matter. Conversely, a crossover within the observed range signals that both components may contribute substantially at different parts of that range.

Diagnostics, Alternatives, and Limitations

A crossover is an interpretation of fitted parameters, not a diagnostic test by itself. Assess whether the combined model is plausible by examining residual behavior across predicted concentration. Under a suitable model, the residual spread should be reasonably consistent with the variance the model assigns at each prediction. Systematic widening or narrowing that is not captured by \(\sqrt{\sigma_{\mathrm{add}}^2+(\sigma_{\mathrm{prop}}C)^2}\), or persistent patterns in residuals, can signal misspecification. Such patterns may also arise from the structural model or other features of the data, so the crossover alone cannot identify the cause.

The main alternatives make different assumptions about how residual variance changes with concentration:

These models are not interchangeable descriptions. A crossover cannot establish that the combined model is the best choice; that requires evaluating model fit and the behavior of residuals, among other considerations. It is also sensitive to the fitted parameters. If either parameter is estimated imprecisely, the ratio can be imprecise, especially when \(\sigma_{\mathrm{prop}}\) is close to zero. When the estimated proportional component is zero, the finite crossover formula does not apply.

A common mistake is to add the two standard deviations and set \(\sigma_{\mathrm{add}}+\sigma_{\mathrm{prop}}C\) equal to itself at the crossover, or to call that sum the combined standard deviation. In this model, the variances add, so the combined standard deviation is \(\sqrt{\sigma_{\mathrm{add}}^2+(\sigma_{\mathrm{prop}}C)^2}\). Another mistake is to interpret \(C^{\mathrm{cross}}\) as the concentration at which residual variance is minimized. It is instead the point of equal contributions; total variance continues to increase with \(C\) when the proportional component is positive.