Why an Allometric Point Prediction Is Not the Whole Answer
An allometric equation can turn an animal pharmacokinetic (PK) parameter into a human prediction, but a single predicted value hides uncertainty in its inputs. The animal parameter may be imprecisely estimated, the relevant body weights may be uncertain, and the scaling exponent may not be known exactly. These uncertainties matter when a predicted clearance or volume is used to inform a first-in-human decision: the same point prediction can support quite different decisions depending on how wide its plausible range is.
This tutorial derives an approximate interval for that uncertainty. It treats the allometric form as specified and focuses on uncertainty in its inputs. It does not account for uncertainty that the allometric relationship itself is appropriate, or for unmodeled species differences. Those are separate sources of uncertainty, not made to disappear by calculating an interval.
The formula assumes that the animal and human parameters are comparable, that the body weights and parameter use consistent units, and that one exponent \(b\) applies over the species and weight range in question. For example, if \(\theta\) is clearance, the parameter values must refer to the same clearance definition and dose convention. The weight ratio is dimensionless, so the predicted parameter retains the units of \(\theta_a\).
Putting the Equation on the Log Scale
Because the allometric equation is multiplicative and all its inputs are positive, taking logarithms makes the propagation easier to see. Define \(A=\ln(\theta_a)\) and \(L=\ln(W_h/W_a)\). Then the logarithm of the human prediction, \(Z=\ln(\theta_h)\), is
Suppose \(A\), \(b\), and \(L\) are independent, with means \(m_a\), \(m_b\), and \(m_L\), and variances \(s_a^2\), \(s_b^2\), and \(s_L^2\), respectively. Here \(s_a\) is uncertainty on the log scale for the animal parameter, while \(s_b\) and \(s_L\) quantify uncertainty in the exponent and log body-weight ratio. These are assumptions about the uncertainty inputs, not statements that the underlying estimates are always independent or normally distributed.
Under independence, the mean of \(Z\) is \(m_a+m_bm_L\). The variance includes a product term because both \(b\) and \(L\) are uncertain:
The last term is the additional variance from the joint uncertainty in \(bL\). A first-order delta-method calculation omits this product-of-variances term; that can be reasonable when both uncertainties are small, but the fuller expression is straightforward under the independence assumptions above. If the uncertain inputs are dependent, this expression is not sufficient: covariance terms must also be considered.
Deriving an Approximate Interval
If \(Z\) is approximately normal, an approximate 95% interval on the log scale is its mean plus or minus 1.96 standard deviations. Exponentiation returns the interval to the original, positive PK scale. The approximation is most credible when the input uncertainties are moderate and the distribution of \(Z\) is not strongly skewed or dominated by a product of highly uncertain quantities.
The interval is symmetric on the log scale, but not on the original scale. Its center in the multiplicative sense is the prediction \(\exp(m_a+m_bm_L)\); this is not generally the arithmetic mean of all possible predictions. If \(Z\) were exactly normal, the arithmetic mean of \(\theta_h\) would be \(\exp(E[Z]+\operatorname{Var}(Z)/2)\), which is larger than the median when uncertainty is nonzero. For decision-making, state which summary is being reported rather than treating “the prediction” as unambiguous.
For the body-weight ratio, uncertainty can be represented directly through \(L\). If the human and animal log weights have variances \(s_h^2\) and \(s_a^2\), and are independent, then \(s_L^2=s_h^2+s_a^2\). In many applications the weights are treated as fixed, making \(s_L^2=0\). That simplification should be explicit: it does not mean that uncertainty in body weight has been included.
Worked Example: Propagating a Combined Log-Scale Uncertainty
Suppose a hypothetical animal clearance estimate is \(0.800\) L/h, the human-to-animal body-weight ratio is \(3.50\), and the exponent is \(0.75\). Assume the combined uncertainty inputs give an approximate log-scale standard deviation of \(s_Z=0.1140\). This summary may include animal-parameter, exponent, and weight-ratio uncertainty; the interval calculation uses it only after those contributions have been appropriately combined.
The plug-in prediction is $$ 0.800(3.50)^{0.75}=2.04711\ldots\ \text{L/h}, $$ or approximately \(2.047\) L/h. The 95% log-scale margin is \(1.96(0.1140)=0.22344\), approximately \(0.2234\). Thus the interval, using the unrounded prediction and margin, is $$ [2.04711\exp(-0.22344),\ 2.04711\exp(0.22344)] \approx [1.637,\ 2.560]\ \text{L/h}. $$
Under the stated uncertainty model, this is an approximate 95% uncertainty interval for human clearance. It is not a statement that there is a 95% probability that a fixed, true human clearance lies in this particular interval. Its interpretation depends on the model and uncertainty distribution used to construct it.
Worked Example: Combining Three Uncertainty Contributions
Consider a second hypothetical prediction. Let the animal clearance have geometric central estimate \(0.500\) L/h and log-scale standard deviation \(s_a=0.080\). Set the nominal body-weight ratio to \(4.00\), so \(m_L=\ln(4)=1.386294\), with \(s_L=0.100\). Let the exponent have mean \(m_b=0.70\) and standard deviation \(s_b=0.12\). Assume the three inputs are independent and the log-scale normal approximation is adequate.
The mean log prediction is $$ m_Z=\ln(0.500)+(0.70)(1.386294) =-0.693147+0.970406=0.277259. $$ The plug-in prediction is \(\exp(0.277259)\approx1.3195\) L/h. The log-scale variance is $$ s_Z^2=(0.080)^2+(0.70)^2(0.100)^2 +(1.386294)^2(0.12)^2+(0.12)^2(0.100)^2 \approx0.039118. $$ Thus \(s_Z\approx0.1978\), and the 95% margin is \(1.96(0.1978)\approx0.3877\). Exponentiating the endpoints gives $$ [\exp(0.277259-0.3877),\ \exp(0.277259+0.3877)] \approx[0.8955,\ 1.945]\ \text{L/h}. $$
The exponent contribution, \(m_L^2s_b^2\), is the largest variance contribution here. The calculation therefore shows why reporting only uncertainty in the animal clearance could understate uncertainty in the human prediction. It also illustrates that body-weight uncertainty contributes through \(m_b^2s_L^2\), while joint exponent-and-weight uncertainty contributes through \(s_b^2s_L^2\).
Worked Example: Translating a Stated Log-Scale Standard Error
For a third hypothetical clearance prediction, suppose the plug-in value is \(6.014\) L/h and the total log-scale standard error is \(0.1418\), obtained from the chosen input uncertainty model. The margin is $$ 1.96(0.1418)=0.277928\approx0.2779. $$
Using the unrounded log-scale margin and prediction, the interval is approximately $$ [6.014\exp(-0.2779),\ 6.014\exp(0.2779)] \approx[4.555,\ 7.941]\ \text{L/h}. $$
The interval is wider than the point estimate alone suggests, and its upper endpoint is farther from the point estimate in absolute units than its lower endpoint. That is expected for an interval symmetric on the log scale. Reporting extra digits would not make the assumptions or the underlying prediction more certain.
Assumptions, Diagnostics, and Alternatives
Before interpreting an interval, check that the prediction target and inputs match the calculation. The animal parameter must be positive for the log transformation, the weight ratio must use the intended species and representative weights, and the exponent must refer to the same allometric relationship. Then check the uncertainty assumptions: are the log animal estimate, exponent, and log weight ratio reasonably summarized by the stated means and variances? Is independence plausible? A shared source of estimation or data can induce dependence, in which case an independence-based interval may be too narrow or otherwise misplaced.
Several warning signs call for more scrutiny. A very wide interval may reflect genuine uncertainty or an exponent applied over a large weight contrast. A highly asymmetric interval on the original scale is not itself a flaw, but an endpoint near zero or an extreme upper tail can signal that a normal approximation for \(Z\) deserves checking. If the exponent uncertainty is large, the product \(bL\) can be notably non-normal; the approximate interval may then poorly represent the tails. Likewise, an interval that changes substantially under small, plausible changes to the exponent or weight ratio is evidence that the translation is sensitive, not a reason to hide that sensitivity.
A plug-in prediction is simpler but suppresses input uncertainty. A first-order delta-method interval uses local linearization and may be adequate for small uncertainties; the product variance term in the derivation above gives a modest refinement under independence. A range analysis can show outcomes under specified low and high inputs, but its endpoints depend on those chosen bounds and are not probability intervals. More elaborate distribution-based propagation can represent non-normal inputs or dependencies, but it requires defensible input distributions and should not be used to disguise weak evidence.
No interval from this calculation accounts automatically for model misspecification, species-specific biology, differences in study design, or uncertainty about whether a single power law is appropriate. These limitations are especially important when extrapolating far beyond the observed species or parameter range. The interval quantifies only the uncertainties that were put into it.
Name the PK parameter, units, animal and human weights, and the exponent used.
Summarize uncertainty in the log animal parameter, exponent, and log weight ratio; state any independence assumptions.
Compute the log prediction and its variance, including the product term when both exponent and log weight ratio are uncertain.
Assess whether the approximation is credible, exponentiate the endpoints, and state what sources of uncertainty remain outside the calculation.
Check Your Understanding
Answer each question using the assumptions stated in the question. Brief solutions follow.
- If the exponent is treated as known, which exponent-uncertainty and product-uncertainty terms disappear from the log-scale variance?
- Why is an interval symmetric on the log scale generally asymmetric on the original PK scale?
- For \(b=0.8\), \(s_L=0.1\), \(m_L=1.2\), and \(s_b=0.05\), calculate the two variance contributions involving weight-ratio uncertainty and exponent uncertainty.
- What does an independence assumption mean in this calculation, and what is a warning sign that it may not be appropriate?
- Name two sources of uncertainty that the derived allometric interval does not automatically include.
Solutions. (1) If \(s_b=0\), both \(m_L^2s_b^2\) and \(s_b^2s_L^2\) are zero; the term \(m_b^2s_L^2\) remains if the weight ratio is uncertain. (2) Exponentiation is nonlinear: equal distances above and below a log-scale center become multiplicative, rather than additive, distances on the original scale. (3) The weight-ratio term is \(m_b^2s_L^2=(0.8)^2(0.1)^2=0.0064\); the exponent term is \(m_L^2s_b^2=(1.2)^2(0.05)^2=0.0036\). (4) Independence means that uncertainty in one input does not provide information about another; shared estimation data or a common source of error can make that assumption doubtful. (5) Examples include species-specific biology, allometric model misspecification, and differences in study design.