Centering and Referencing Body-Size Covariates

Why the Reference Weight Matters

A power-function model describes how a typical pharmacokinetic parameter changes with body weight. Its reference weight is the point at which the typical parameter is reported. For example, a typical clearance reported at 70 kg answers a different descriptive question from one reported at 60 kg, even when both values come from the same fitted relationship. The numerical estimate changes because the point of reference changes; the model’s predictions for patients do not have to change.

This distinction is useful when interpreting or comparing parameter estimates. A reference near the weights represented in a data set can make the typical parameter easier to interpret and can avoid reporting it at an unobserved, potentially unhelpful value. But a different reference is not, by itself, evidence of a different size effect or a different fit.

For clearance, one common model is

Here \(W_i\) is the body weight of individual \(i\), \(CL_{\mathrm{ref}}\) is typical clearance at \(W_{\mathrm{ref}}\), and \(\theta\) is the allometric exponent. At \(\eta_{CL,i}=0\), the individual parameter equals the typical parameter predicted by the covariate model. With normally distributed random effects on the log scale, this is the model’s typical, or median, individual value conditional on weight; it is not generally the arithmetic mean across individuals. The same reference-weight logic applies to volume or another parameter modeled with a power function.

The ratio \(W_i/W_{\mathrm{ref}}\) is dimensionless, so the weight and reference weight must use the same units. The model describes a proportional change: if weight is multiplied by a factor \(r\), the typical parameter is multiplied by \(r^\theta\). The exponent may be fixed or estimated, as discussed in the earlier material on fixed versus estimated allometric exponents.

Converting Between Reference Weights

Suppose the same power relationship is written first with reference weight \(A\), and then with reference weight \(B\). The exponent and all individual predictions are to remain the same. Equating the two expressions for the typical parameter at any weight \(W\) gives the conversion.

Here \(P_A\) and \(P_B\) are the typical parameter values at reference weights \(A\) and \(B\). The conversion follows by dividing both sides by \((W/B)^\theta\), or equivalently by evaluating the original curve at \(W=B\). It does not require refitting the model. If the exponent is positive, moving the reference upward increases the typical parameter; for a negative exponent, it decreases it.

On the log scale, the same conversion is an intercept shift:

This form also shows why the typical parameter and exponent can be statistically associated when both are estimated: changing the reference changes the intercept by a quantity involving the exponent. The conversion changes the parameter coordinates, not the underlying curve.

What Changes—and What Does Not

Changing the reference weight changes the numerical value and interpretation of the typical parameter. In the example, 5.20 L/h describes typical clearance at 70 kg, while 4.632 L/h describes typical clearance at 60 kg. Neither value is a universal clearance for all individuals; each is the value at a specified point on the size relationship.

If only the reference changes, the exponent, the predicted typical curve across weight, and the individual predictions remain the same. In the model with random effects, the individual’s \(\eta_{CL,i}\) also remains the same under this pure re-expression. The reference change does not remove between-individual variability or alter the modeled proportional effect of weight.

The conversion can be made for any positive reference weights, but the reference should be chosen for a reason. A value near a clinically meaningful or well-represented body weight often gives the typical parameter a direct interpretation. Choosing a reference far outside the observed range may still be mathematically valid, but the corresponding typical parameter can be an extrapolated summary rather than a parameter anchored near the data.

Centering is often clearest on the log-weight scale. The covariate term can be written as \(\theta[\log(W_i)-\log(W_{\mathrm{ref}})]\), which equals zero at the reference weight. This is equivalent to using \(\theta\log(W_i/W_{\mathrm{ref}})\). It is not generally equivalent to replacing the ratio with a raw difference such as \(W_i-W_{\mathrm{ref}}\); that would define a different model and a different effect scale.

Uncertainty and Estimation

When the exponent is estimated, uncertainty in the converted typical parameter depends on uncertainty in both the original typical parameter and the exponent. In particular, their covariance matters. Let \(a=\log P_A\) and \(c=\log(B/A)\). Then \(\log P_B=a+c\theta\). For estimates of \(a\) and \(\theta\), the approximate variance of the transformed log parameter is

This is the variance formula for a linear combination; here the transformation on the log scale is linear. Omitting the covariance term can overstate or understate uncertainty. If \(\theta\) is fixed rather than estimated, the conversion is a constant rescaling of \(P_A\), so its standard error on the original scale is multiplied by \((B/A)^\theta\). If the exponent is estimated, use the joint parameter uncertainty when reporting precision for the converted estimate.

Checks, Failure Modes, and Alternatives

A useful diagnostic is to calculate a prediction at one or more weights using both parameterizations. Agreement, up to rounding, checks the conversion arithmetic and the direction of the ratio. If the predictions disagree materially, inspect the reference-weight ratio, exponent, and units before concluding that the model changed.

• Ratio reversed. The conversion from \(A\) to \(B\) uses \((B/A)^\theta\), not \((A/B)^\theta\). A quick check is that, for a positive exponent, increasing the reference weight must increase the typical parameter.
• Inconsistent units. A ratio such as \(70\text{ kg}/60\text{ lb}\) is invalid. Convert both weights to the same unit before calculating.
• Exponent changed unintentionally. Re-referencing does not itself change \(\theta\). Changing it alters the weight relationship rather than merely changing where the typical value is reported.
• Raw difference substituted for a ratio. A model using \(W-W_{\mathrm{ref}}\) has a different functional form from one using \((W/W_{\mathrm{ref}})^\theta\).
• Converted estimate mistaken for an individual prediction. \(P_B\) is the typical value at \(B\); an individual prediction may also include that individual’s random effect.

Alternative parameterizations can express the same curve, such as using a log-scale intercept rather than a typical parameter at a stated reference. They are mathematically equivalent when the transformation is applied consistently. Re-fitting after changing the reference can produce small numerical differences because of rounding, optimization tolerances, constraints, or implementation choices. Such differences should be checked; they are not a consequence required by the reference change itself. The comparison is exact when the model, data, exponent treatment, and parameter space are unchanged and the conversion is exact.

A reference choice is therefore a matter of interpretation and parameterization, not a substitute for assessing whether the power-function model is appropriate. This tutorial concerns how to express a specified size relationship. It does not establish which body-size measure or exponent should be selected for a particular drug or population.