Propagating Uncertainty in Allometric Scaling

Why Uncertainty Changes Across Body Weights

An allometric model gives clearance at a particular body weight by scaling a reference clearance according to a weight ratio raised to an exponent. Even when the reference clearance and exponent have been estimated from the same data, neither is known without uncertainty. The effect of that uncertainty depends on the weight at which a prediction is made: the exponent has no effect at the reference weight, but its influence grows as the prediction weight moves farther from it on the log scale.

This tutorial isolates uncertainty in the size-scaling parameters. It does not include maturation effects, between-subject variability, residual error, or uncertainty in the measured weight. Those are distinct sources of variation. Here, weight and reference weight are treated as fixed, and the goal is to understand how uncertainty in the estimated reference clearance and allometric exponent propagates to a model-based clearance prediction.

Definition: For an individual with body weight $W$, the allometric clearance prediction is $$ CL(W)=CL_{\mathrm{ref}}\left(\frac{W}{W_{\mathrm{ref}}}\right)^\theta, $$ where $CL_{\mathrm{ref}}$ is clearance at the reference weight $W_{\mathrm{ref}}$, and $\theta$ is the allometric exponent. In this tutorial, $W$ and $W_{\mathrm{ref}}$ are fixed values; uncertainty in $CL_{\mathrm{ref}}$ and $\theta$ is the parameter uncertainty being propagated.

The reference weight is a modeling anchor. At $W=W_{\mathrm{ref}}$, the ratio is one, so $CL(W)=CL_{\mathrm{ref}}$ regardless of $\theta$. At other weights, both the distance from the reference weight and uncertainty in $\theta$ matter. On the log scale, the prediction is linear in the exponent:

\log CL(W)=\log CL_{\mathrm{ref}}+\theta x, \qquad x=\log\left(\frac{W}{W_{\mathrm{ref}}}\right).

This form makes the key dependency visible. The exponent's contribution is multiplied by $x$: it is zero at the reference weight, negative below it, and positive above it. Thus, increasing $\theta$ decreases predicted clearance below the reference weight but increases it above the reference weight.

First Examine the Effect of the Exponent

A useful first step is a sensitivity calculation: hold the reference clearance fixed and compare predictions across plausible exponent values. This is not a confidence interval unless the selected values and method have a formal interval interpretation. It shows how much the prediction changes under the stated alternatives.

Worked Example: Comparing Exponent Values Across Weights

Suppose a hypothetical model uses $CL_{\mathrm{ref}}=5.00$ L/h at $W_{\mathrm{ref}}=70$ kg. Compare the predictions at 10 kg and 120 kg for $\theta=0.65$, $0.75$, and $0.85$. These values are illustrative alternatives, not estimates from a real study.

At 10 kg, the weight ratio is $10/70=1/7$. For $\theta=0.65$:

CL(10)=5.00\left(\frac{10}{70}\right)^{0.65} =5.00\exp\{0.65\log(10/70)\} \approx 1.411\ \text{L/h}.

Repeating the calculation gives $5.00\exp\{0.75\log(10/70)\}\approx1.162$ L/h and $5.00\exp\{0.85\log(10/70)\}\approx0.956$ L/h. The prediction falls as $\theta$ increases because $10/70<1$.

At 120 kg, the ratio is $120/70=12/7$. The three predictions are $5.00\exp\{0.65\log(120/70)\}\approx7.098$ L/h, $5.00\exp\{0.75\log(120/70)\}\approx7.491$ L/h, and $5.00\exp\{0.85\log(120/70)\}\approx7.906$ L/h. Here the prediction rises with $\theta$, because $120/70>1$. The same exponent range therefore produces a substantial spread at both weights, with the direction of change determined by whether the prediction weight is below or above the reference.

This comparison highlights why reporting an exponent estimate alone is not enough to describe prediction uncertainty. An exponent's uncertainty has little effect on predictions near the reference weight and increasingly affects predictions farther away. That pattern is a mathematical property of the model, not evidence that the exponent itself changes with weight.

Propagating Joint Parameter Uncertainty

For an approximate interval, use uncertainty in both the reference parameter and exponent, including their covariance. Let $a=\log CL_{\mathrm{ref}}$, and suppose the estimates of $a$ and $\theta$ have estimated variances and covariance. Since $\log CL(W)=a+\theta x$, the variance of the log prediction is

\operatorname{Var}\{\log CL(W)\} = \operatorname{Var}(a) +x^2\operatorname{Var}(\theta) +2x\operatorname{Cov}(a,\theta).

The covariance term matters because the two parameters may be estimated together. For example, a higher fitted reference clearance may be paired with a lower fitted exponent while maintaining a similar fit over the observed weights. Ignoring this dependence can overstate or understate the prediction uncertainty, depending on the sign of $x$ and the covariance.

A first-order delta-method calculation is appropriate when the parameter estimates are approximately jointly normal and uncertainty is moderate enough that a local approximation is informative. If the estimated variance of the log prediction is $V$, an approximate 95% interval on the log scale is the log prediction plus or minus $1.96\sqrt{V}$. Exponentiating the endpoints gives an interval on the clearance scale. This is an approximate confidence interval for the model-based prediction, conditional on the specified model and fixed weight, not a range for individual observed clearances.

Justification: For $g(a,\theta)=a+\theta x$, the gradient is $(1,x)$. The first-order variance approximation is the gradient multiplied by the parameter covariance matrix and its transpose: $$ (1,x) \begin{pmatrix} \operatorname{Var}(a) & \operatorname{Cov}(a,\theta)\\ \operatorname{Cov}(a,\theta) & \operatorname{Var}(\theta) \end{pmatrix} \begin{pmatrix} 1\\x \end{pmatrix} = \operatorname{Var}(a)+x^2\operatorname{Var}(\theta)+2x\operatorname{Cov}(a,\theta). $$ Because the log prediction is exactly linear in $a$ and $\theta$, this variance calculation is exact for that linear combination when the parameter covariance is known. The approximate normal interval relies additionally on the distribution of the estimated parameters being adequately represented by a joint normal approximation.

Worked Example: A Joint-Parameter Approximation

Consider a hypothetical fit with $CL_{\mathrm{ref}}=5.00$ L/h, $\theta=0.75$, standard error of $a=\log CL_{\mathrm{ref}}$ equal to 0.080, and standard error of $\theta$ equal to 0.10. Suppose their estimated correlation is $-0.30$. Then $\operatorname{Var}(a)=0.080^2=0.00640$, $\operatorname{Var}(\theta)=0.10^2=0.0100$, and $\operatorname{Cov}(a,\theta)=(-0.30)(0.080)(0.10)=-0.00240$.

At 10 kg, $x=\log(10/70)\approx-1.94591$. The estimated log-prediction variance is

V=0.00640+(-1.94591)^2(0.0100) +2(-1.94591)(-0.00240) \approx0.05361.

Thus, $\sqrt V\approx0.2315$. The point prediction is $5.00(10/70)^{0.75}\approx1.162$ L/h. The approximate interval is

\exp\{\log(1.162)-1.96(0.2315)\} \ \text{to}\ \exp\{\log(1.162)+1.96(0.2315)\} \approx 0.738\ \text{to}\ 1.829\ \text{L/h}.

At 120 kg, $x=\log(120/70)\approx0.53900$, and $V=0.00640+(0.53900)^2(0.0100)+2(0.53900)(-0.00240)\approx0.006718$. The standard deviation on the log scale is about 0.08197. With point prediction $7.491$ L/h, the corresponding approximate interval is $\exp\{\log(7.491)\pm1.96(0.08197)\}$, or about 6.379 to 8.797 L/h.

The intervals are asymmetric around the point prediction on the original clearance scale because they are symmetric on the log scale. The covariance contribution is positive at 10 kg and negative at 120 kg: in the formula, $x$ changes sign while the covariance remains negative.

Scenario Ranges Are Not Confidence Intervals

Sometimes only plausible parameter ranges are available, or the purpose is to show consequences of selected assumptions rather than quantify sampling uncertainty. In that case, evaluate the model at parameter combinations that produce the minimum and maximum prediction. For a weight below the reference, a larger exponent gives a smaller prediction; above the reference, a larger exponent gives a larger prediction. Since clearance is directly proportional to $CL_{\mathrm{ref}}$, the smallest reference clearance always produces the smaller prediction when the exponent is held fixed.

Worked Example: Combining Plausible Parameter Values

Use the same hypothetical reference weight of 70 kg, but now suppose the plausible values are $CL_{\mathrm{ref}}=4.50$ to $5.50$ L/h and $\theta=0.65$ to $0.85$. These separate ranges define a scenario envelope only if any combination of their endpoints is being treated as plausible; they do not, by themselves, establish a confidence level.

At 10 kg, the smallest prediction uses the smaller reference clearance and larger exponent: $4.50(10/70)^{0.85}\approx0.861$ L/h. The largest uses the larger reference clearance and smaller exponent: $5.50(10/70)^{0.65}\approx1.553$ L/h. The resulting scenario range is approximately 0.861 to 1.553 L/h.

At 120 kg, the exponent's direction reverses. The smallest prediction is $4.50(120/70)^{0.65}\approx6.388$ L/h, and the largest is $5.50(120/70)^{0.85}\approx8.696$ L/h. Thus, the scenario range is approximately 6.388 to 8.696 L/h. These endpoints are evaluations of specified parameter combinations, not statements that a prediction has a particular probability of lying in the range.

Diagnostics, Alternatives, and Limitations

Before using a propagation calculation, check that the parameter uncertainty summary matches the model fit. The variances and covariance should come from the same fitted model and parameterization. If the fit reports uncertainty in $CL_{\mathrm{ref}}$ rather than $\log CL_{\mathrm{ref}}$, do not insert that standard error directly as the standard error of $a$; the scale differs. Verify the covariance scale as well. The covariance matrix should be positive semidefinite, and its implied variances should not be negative.

A delta-method interval may be unreliable if the exponent is poorly identified, its sampling distribution is strongly skewed, or parameter uncertainty is large enough that a local approximation gives a poor description. Warning signs include very wide or unstable intervals, pronounced changes under small alterations to the fitting assumptions, and a markedly non-normal or asymmetric parameter likelihood. A normal interval on the log scale will remain positive after exponentiation, but positivity alone does not make the approximation adequate.

When the joint parameter distribution is irregular or uncertainty is substantial, a practical alternative is to draw parameter pairs from a suitable joint approximation or from bootstrap refits, calculate $CL(W)$ for every draw, and summarize the resulting predictions. Parameter pairs must preserve their dependence; independently drawing each parameter can give the wrong prediction spread. Profile-likelihood or bootstrap-based intervals can also be more informative when a quadratic, normal approximation is doubtful. A simple scenario range is easier to communicate but answers a different question: it explores selected values rather than quantifying sampling uncertainty.

All these methods condition on the model form. They do not account automatically for uncertainty about whether a power model is appropriate, uncertainty in the reference weight, error in measured weights, or additional variability between individuals. If weight itself is uncertain, it is another input whose uncertainty may need to be propagated. For example, on the log scale the derivative with respect to $W$ is $\theta/W$, so a first-order approximation would add a weight-uncertainty contribution and any relevant covariances. That extension should not be confused with uncertainty in the fitted allometric parameters alone.

Common pitfall: It is incorrect to assume that uncertainty in $\theta$ changes predicted clearance in the same direction at every weight. Since the exponent multiplies $\log(W/W_{\mathrm{ref}})$, its effect reverses at the reference weight. Check whether $W/W_{\mathrm{ref}}$ is below or above one before selecting parameter combinations or interpreting an exponent sensitivity.

The main practical lesson is to report uncertainty at the weight where the prediction is needed. The uncertainty in a reference clearance alone does not describe the full prediction uncertainty away from the reference weight. Conversely, a broad scenario envelope should not be labeled a confidence interval, and a confidence interval based on a normal approximation should not be presented as accounting for model misspecification or individual variability.

Check Your Understanding

Answer each question, then compare with the brief solutions.

For $W=35$ kg and $W_{\mathrm{ref}}=70$ kg, does increasing $\theta$ increase or decrease the predicted clearance? Why?
At the reference weight, what is the effect of exponent uncertainty on the allometric prediction?
In the log-prediction variance formula, what information is lost if $\operatorname{Cov}(a,\theta)$ is set to zero without justification?
At a weight above the reference, which exponent endpoint gives the smaller prediction when $CL_{\mathrm{ref}}$ is fixed?

Solutions. (1) It decreases: $\log(35/70)=\log(0.5)<0$, so increasing $\theta$ makes $\theta\log(35/70)$ more negative. (2) None: the weight ratio is one, so $CL(W_{\mathrm{ref}})=CL_{\mathrm{ref}}$ for every exponent. (3) The dependence between the fitted parameters is omitted; the covariance contribution $2x\operatorname{Cov}(a,\theta)$ may increase or decrease prediction variance. (4) The smaller exponent gives the smaller prediction, because $W/W_{\mathrm{ref}}>1$ and raising a number above one to a larger power increases it.