Initial Values for Residual Error Models

Why Residual-Error Starting Values Need Their Own Check

A pharmacokinetic prediction can be plausible while its residual-error starting values are not. If the residual error is initialized too small, the likelihood can treat ordinary assay variation as strong evidence against the model. If it is initialized too large, meaningful concentration differences may initially carry little weight. Either problem can make optimization less stable or encourage the fit to compensate for an unsuitable error structure.

The purpose of an initial value is not to declare the final assay error known. It is to give estimation a reasonable scale, in the units and parameterization of the model being fit. A useful first check compares observed concentrations with predictions from a preliminary structural model. Those predictions need not be final, but they should be calculated consistently for the records used in the check.

Residual-Error Models and Their Parameters

Let \(Y_{ij}\) denote an observed concentration for subject \(i\) at observation \(j\), and let \(F_{ij}\) be the corresponding model prediction. Define the prediction residual as \(r_{ij}=Y_{ij}-F_{ij}\). The residual model describes the variation of \(Y_{ij}\) around \(F_{ij}\), conditional on that prediction. The examples below use untransformed concentrations.

Under these assumptions, the conditional variances are \(\sigma_{\mathrm{add}}^2\), \(F_{ij}^2\sigma_{\mathrm{prop}}^2\), and \(\sigma_{\mathrm{add}}^2+F_{ij}^2\sigma_{\mathrm{prop}}^2\), respectively. If the errors are also normal, these variance expressions define normal observation models conditional on \(F_{ij}\). Normality is an additional assumption, not a consequence of the variance equations.

Some software uses a different combined-error parameterization: for example, it may define the conditional standard deviation as a sum, rather than define the variance as a sum of independent components. Other implementations use variance parameters where this tutorial uses standard deviations. Check the model equation and units in the estimation code before transferring a numerical starting value. A value for a variance cannot be substituted directly for a value for its standard deviation.

Estimating a Starting Scale From Residuals

For a proposed prediction \(F_{ij}\), a residual gives a signed discrepancy in concentration units. The squared residual gives a rough contribution to residual variance. Starting values can be obtained by summarizing these discrepancies in a way that matches the model: roughly constant absolute scatter suggests an additive scale; roughly constant relative scatter suggests a proportional scale. These summaries are only preliminary because residuals also reflect prediction error, sampling, and possible model misspecification.

For an additive model, the corresponding rough scale is the root mean square of the residuals in concentration units. The calculation assumes residuals are centered around zero. A persistent nonzero residual mean is a warning about prediction bias, not a reason to inflate the error scale until the bias is hidden.

Deriving a Combined-Error Starting Value

In the combined model, the additive and proportional components are independent and have mean zero. Expanding the variance of their sum shows how residual spread should change with the prediction:

The last equality uses both independence, which makes the covariance zero, and the specified component variances. Consequently, if residual variance can be summarized at different prediction levels, it should be approximately linear in \(F^2\): the intercept corresponds to \(\sigma_{\mathrm{add}}^2\), and the slope corresponds to \(\sigma_{\mathrm{prop}}^2\). This is a diagnostic and a source of rough starting values, not a guarantee that an ordinary regression of individual squared residuals will estimate those quantities reliably. Squared residuals are noisy, and a small sample can produce an unstable or even negative fitted intercept or slope. A negative component variance is not a valid starting value; reconsider the summaries, the data, or the model rather than interpreting it as a negative error scale.

Assay Noise Near the Quantification Limit

A purely proportional model predicts that residual standard deviation shrinks toward zero as \(F\) approaches zero. Real assays may retain an absolute noise floor, so relative error can become large at low concentrations even when the absolute error remains modest. An additive component can represent such a floor over a suitable range; a combined model allows the absolute component to dominate at low predictions and the proportional component to matter more as predictions rise.

For instance, with \(\sigma_{\mathrm{add}}=0.30\) mg/L and \(\sigma_{\mathrm{prop}}=0.10\), the combined standard deviation at \(F=1\) mg/L is \(\sqrt{0.30^2+(0.10\times1)^2}=0.316\) mg/L. At \(F=10\) mg/L it is \(\sqrt{0.30^2+(0.10\times10)^2}=1.044\) mg/L. The relative size of the additive contribution is much greater at the lower prediction. This illustrates why a relative-error summary calculated near zero can be misleading.

Use assay validation or quality-control information at relevant concentration levels when available. It can help distinguish known analytical imprecision from discrepancies caused by the structural model or other unmodeled variability. Observations below a quantification limit are not ordinary exact concentration measurements. Replacing them by the limit or by a fixed fraction of it and treating that substituted value as exact can distort residual summaries; handle such records consistently with the model and data-analysis plan.

Diagnostics, Alternatives, and Limits

Before using a residual summary as a starting value, check whether its assumptions are at least plausible. Residuals should not show a strong systematic shift from zero across prediction, time, or other relevant design features. A trend can signal a poor preliminary prediction, not merely an incorrect residual scale. Compare absolute residuals and relative residuals against predictions: roughly constant absolute spread supports an additive component, while roughly constant relative spread supports a proportional component. A spread that changes with prediction may support a combined model, but does not by itself prove it is the right model.

Where there are enough observations, summarize residual spread across prediction ranges rather than relying on a few individual squared residuals. Check for extreme observations, negative predictions, data-unit mismatches, and inconsistent handling of records near the quantification limit. Also check that the residual definition matches the prediction used: mixing prediction types or including records with substantially different model conditions can obscure the error pattern. Residuals within a subject may be related, so treating every record as independent evidence for a precise scale can overstate how informative a small data set is.

The main practical alternatives are to start from assay precision data, use a simple residual summary from model predictions, or choose a plausible scale from prior model-development experience and assess whether estimation is sensitive to it. Assay data can be more directly informative about analytical noise, but may not describe every source of unexplained observation-level variation in the pharmacometric data. Prediction residuals reflect both assay noise and any shortcomings in the predictions. A plausible informed value avoids an unstable calculation but should be checked rather than assumed correct. These approaches inform initialization; none makes a misspecified error model adequate.

Residual-based calculations are approximate because the preliminary predictions are themselves uncertain and can be wrong. They do not establish the final parameter estimates, prove normality, or identify the cause of unexplained variation. Their value is practical: they give the optimizer an appropriate order of magnitude and reveal obvious mismatches before a more complex estimation step. If residual spread changes sharply, the mean pattern is systematic, or candidate scales depend heavily on a few records, revisit the model and data rather than trusting one numerical summary.