Why Look for Trends After Scaling?
A size-scaling rule is intended to account for a systematic relationship between body weight and a pharmacokinetic parameter. After applying that rule, a useful diagnostic question is whether the scaled parameter—or the model’s prediction errors—still changes systematically with weight. A remaining pattern can signal that the chosen exponent, weight descriptor, or model form does not adequately describe the data. It is a clue to investigate, not by itself proof of which model should replace it.
For example, a clearance model may use a power relationship between clearance and body weight. If the exponent is appropriate and other relevant features are adequately represented, clearance expressed at a common reference weight should not show a systematic trend with actual weight. Similarly, concentration residuals should not show a persistent weight-related bias once the model has accounted for the relevant predictors. These expectations are conditional on the model and data being suitable for the diagnostic.
What a Trend Implies About the Exponent
A simple derivation clarifies how an exponent mismatch can create a trend. Suppose the underlying relationship for a parameter is \(P_i=P_{\mathrm{ref}}(W_i/W_{\mathrm{ref}})^\gamma\), where \(\gamma\) is the exponent that describes the relationship in this simplified setting. The analyst scales using \(\theta\), which need not equal \(\gamma\). Substitution into the scaling formula gives:
Taking logarithms makes the implication explicit:
Under this simplified power model, the slope of log scaled parameter against log relative weight is \(\gamma-\theta\). A positive slope means the scaled parameter tends to increase with weight, consistent with an exponent that is too small relative to the relationship in the data; a negative slope is consistent with an exponent that is too large. This interpretation assumes the power relationship is adequate, the parameter values are informative, and other systematic influences are not responsible for the pattern. In a population analysis, uncertainty and other covariates complicate this simple slope interpretation.
Worked Example: Interpreting a Trend in Scaled Clearance
Suppose a hypothetical analysis uses a clearance exponent \(\theta=0.75\) and reference weight \(W_{\mathrm{ref}}=20\) kg. Three individual clearance estimates are 1.82, 3.40, and 6.34 L/h at weights of 10, 20, and 40 kg, respectively. Calculate the scaled clearances and assess the pattern.
For 10 kg, \(P_{\mathrm{scaled}}=1.82(20/10)^{0.75}=1.82(1.6818)=3.061\) L/h. For 20 kg, \(P_{\mathrm{scaled}}=3.40(20/20)^{0.75}=3.40\) L/h. For 40 kg, \(P_{\mathrm{scaled}}=6.34(20/40)^{0.75}=6.34(0.5946)=3.770\) L/h.
The scaled estimates increase from 3.061 to 3.770 L/h across the weight range rather than remaining roughly level. A descriptive endpoint slope on the log scale is \( \log(3.770/3.061)/\log(40/10)=0.150 \). This pattern is consistent with the data’s parameter-weight relationship being steeper than the exponent used, under the simplified power-model interpretation. It does not establish that changing the exponent is the correct remedy: the estimates are hypothetical individual estimates, and the pattern could also reflect other model inadequacies or estimation noise.
Parameter Plots and Residual Plots Answer Different Questions
A plot of scaled individual parameter estimates against weight asks whether the estimated parameter, after scaling, still varies with weight. For example, one might plot \(CL_{i,\mathrm{scaled}}\) against \(W_i\). A plot of model residuals against weight asks whether observed concentrations are systematically above or below their predictions at different weights. These diagnostics are related, but they are not interchangeable.
An individual parameter estimate is inferred from that person’s concentration data and the model. Its precision depends on how much information those data contain. Sparse sampling, limited concentration variation, and shrinkage of individual estimates toward the population prediction can blur or distort an apparent association with weight. Therefore, a trend in individual estimates is a prompt to inspect the model and data, not a direct estimate of a population covariate effect.
Residuals depend on the precise residual definition. For illustration, define a raw prediction residual as \(e_{ij}=C_{ij,\mathrm{obs}}-C_{ij,\mathrm{pred}}\), where \(C_{ij,\mathrm{obs}}\) is an observed concentration and \(C_{ij,\mathrm{pred}}\) is its model prediction. With this convention, positive residuals indicate observations above predictions, and negative residuals indicate observations below predictions. Residuals may also be standardized; their expected scale and interpretation then depend on the standardization and error model. In either case, compare like with like—for example, observations at relevantly similar sampling times and conditions—so that a time-related pattern is not mistaken for a weight effect.
Worked Example: Reading Residual Direction Across Weight
In a hypothetical diagnostic, standardized residuals are defined with the observed-minus-predicted sign convention. At comparable sampling times, three lower-weight participants have residuals \(+0.12\), \(+0.18\), and \(+0.15\); three higher-weight participants have residuals \(-0.16\), \(-0.21\), and \(-0.17\). Calculate the group means and interpret the direction.
For the lower-weight group, the mean is \((0.12+0.18+0.15)/3=0.150\). For the higher-weight group, it is \((-0.16-0.21-0.17)/3=-0.180\). Thus, observations tend to be above predictions at lower weights and below predictions at higher weights. If this pattern persists across the data rather than arising from a few points, it is consistent with the model predicting too little at lower weights and too much at higher weights. That is a pattern worth checking after scaling.
These six residuals do not establish a statistically reliable trend or identify its cause. The group means are descriptive, and a proper diagnosis also checks the full residual distribution, sampling-time patterns, and whether the pattern is influenced by a small number of individuals. The conclusion is a direction for investigation, not a causal claim.
Distinguishing a Slope from a More Complex Pattern
A roughly straight trend on a plot of log scaled parameter against log relative weight is compatible with an exponent mismatch under a power relationship. Curvature is different: it suggests that a single constant exponent may not describe the entire weight range, or that other features of the model or data are contributing. A scatterplot with a narrow weight range may not distinguish these possibilities well.
Worked Example: A Pattern That Is Not a Single Straight Trend
Consider hypothetical scaled parameter estimates of 2.8, 3.4, 3.4, and 4.4 L/h at weights of 5, 10, 20, and 40 kg. Compare the change in log scaled parameter over each doubling of weight.
From 5 to 10 kg, the log change is \(\log(3.4/2.8)=0.194\). From 10 to 20 kg, it is \(\log(3.4/3.4)=0\). From 20 to 40 kg, it is \(\log(4.4/3.4)=0.258\). The changes are not approximately constant across successive doublings. This pattern is not well summarized as one clear, constant residual slope on the log scale.
The values might reflect curvature, imprecise estimates, or some other feature of the data or model; four estimates alone cannot decide among these explanations. The practical next step is to inspect the broader pattern and the information supporting each estimate, rather than immediately selecting a more complicated weight function.
How to Investigate a Suspected Trend
A disciplined check separates the visual observation from the explanation. First confirm the scaling formula, reference weight, units, and parameter definition. A unit error or inconsistent scaling implementation can create a pattern that has nothing to do with the biological relationship. Then inspect the values over the full weight range, noting whether the pattern is a slope, curvature, a few influential points, or different behavior in distinct regions.
State whether the vertical axis is a scaled individual parameter, a raw residual, or a standardized residual. Confirm the residual sign convention and the parameter units.
Look for adequate observations across the weight range and consider whether sampling times, age, or other represented covariates differ systematically with weight. A diagnostic cannot separate effects that the data do not distinguish.
Describe its direction, shape, and extent. Check whether it is robust to individual points and whether residuals also have patterns against time or other relevant predictors.
A persistent straight trend may motivate checking the exponent; curvature may motivate checking the power-form assumption; a poor descriptor may motivate considering an alternative size measure. These are hypotheses to evaluate, not conclusions from the plot alone.
After a justified model change, examine whether the pattern is reduced without creating new systematic patterns or undermining parameter precision. A flatter plot alone does not guarantee a better-supported model.
Alternative Explanations, Comparisons, and Limitations
Several approaches can help assess a suspected trend, but they answer different questions. A visual plot reveals shape and influential points. A descriptive slope summarizes a roughly linear pattern on a chosen scale. Comparing model fits after changing an exponent or descriptor assesses the proposed model change, subject to the assumptions and criteria of that comparison. None can compensate for poor data support or a model that omits relevant structure.
| Approach | What it can show | Main limitation |
|---|---|---|
| Plot scaled parameters against weight | Whether estimated parameters show a remaining size pattern and whether it appears linear or curved | Individual estimates may be noisy or affected by shrinkage and unequal information |
| Plot residuals against weight | Whether concentration predictions show weight-related bias under the fitted model | Residual patterns can be affected by time, error structure, or other omitted predictors |
| Summarize a log-scale slope | Direction and approximate magnitude of a linear trend on the log scale | A single slope can hide curvature and does not establish the mechanism |
| Refit with a different exponent, descriptor, or form | Whether the proposed model change improves the description of the available data | Improvement may be unstable, overfit, or hard to interpret when predictors are confounded |
Weight may be associated with age or other features of the population. If such features also affect the parameter, a residual weight pattern does not isolate a pure size effect. Earlier material on age-weight collinearity and identifiability is relevant here: when the data provide little independent variation, plots and refits may not separate competing explanations. Likewise, using an alternative descriptor only helps if it is scientifically defensible and sufficiently measured; replacing weight by a different quantity does not automatically solve confounding.
A common mistake is to reason, “The scaled parameter still rises with weight, so the exponent must be increased until the plot is flat.” The correction is to treat the plot as a diagnostic, verify how the estimates and residuals were obtained, and consider whether a simple power model and the selected descriptor are appropriate. A chosen exponent should be supported by the analysis and the available data, not selected solely to erase one visual trend.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- If the log scaled parameter rises with log weight under a power model, what does the slope represent in terms of the underlying and used exponents?
- With residuals defined as observed minus predicted, what does a negative residual indicate?
- Why should a trend in individual parameter estimates be interpreted cautiously?
- After scaling, a plot shows a curved pattern rather than a steady slope. What is one implication to investigate?
Solutions. (1) Under the simplified model, the slope is \(\gamma-\theta\), the difference between the underlying relationship exponent and the exponent used for scaling. (2) The observed concentration is below its prediction. (3) Individual estimates can be noisy and can have unequal information or shrinkage; their pattern is not automatically a population effect. (4) A single constant power exponent may not describe the full range, although noise or other model and data features must also be considered.