Power Variance Functions For Residual Error

Why Let Residual Variability Depend on Concentration?

A residual error model describes the differences between observed concentrations and the concentrations predicted by the structural model. Treating those differences as having the same spread at every concentration can be reasonable, but it is not always appropriate. For example, an assay or prediction process may have greater absolute variability at higher concentrations. A useful way to represent a gradual change in spread is to let residual standard deviation follow a power of the predicted concentration.

The power model is a single-parameter family: its exponent determines how quickly residual variability changes as predicted concentration changes. At one value of the exponent it gives constant absolute error; at another it gives proportional error. Intermediate values describe intermediate scaling. This tutorial defines that family and shows how to interpret it without confusing variance with standard deviation.

Definition and Parameterization

For observation $k$, let $C_{\mathrm{obs},k}$ be the observed concentration and $C_{\mathrm{pred},k}$ the concentration predicted by the structural model. Define the residual as $\epsilon_k=C_{\mathrm{obs},k}-C_{\mathrm{pred},k}$. The power variance function below models the residual conditional on its predicted concentration.

Definition: A power variance function with reference concentration $C_{\mathrm{ref}}>0$, scale parameter $\sigma>0$, and exponent $p$ is $$ \operatorname{Var}(\epsilon_k\mid C_{\mathrm{pred},k})= \sigma^2\left(\frac{C_{\mathrm{pred},k}}{C_{\mathrm{ref}}}\right)^{2p}. $$ Equivalently, the conditional residual standard deviation is $$ \operatorname{SD}(\epsilon_k\mid C_{\mathrm{pred},k})= \sigma\left(\frac{C_{\mathrm{pred},k}}{C_{\mathrm{ref}}}\right)^p. $$ Here, $C_{\mathrm{pred},k}>0$, $C_{\mathrm{ref}}$ is a fixed concentration in the same units, $\sigma$ has concentration units, and $p$ is dimensionless. In particular, $\sigma$ is the residual standard deviation when $C_{\mathrm{pred},k}=C_{\mathrm{ref}}$.

Writing the function in terms of a ratio makes the exponent act on a dimensionless quantity. It also gives $\sigma$ a direct interpretation at a chosen reference concentration. The choice of $C_{\mathrm{ref}}$ is a scaling convention, not an extra fitted exponent: changing it changes the numerical value of $\sigma$ but does not change the model’s power relationship.

The factor $2p$ appears in the variance because variance is the square of standard deviation. This distinction matters: $p=1$ means the standard deviation is proportional to predicted concentration, so the variance is proportional to its square. It would be incorrect to call the variance itself proportional to concentration in that case.

How the Exponent Changes the Error Pattern

The limiting cases give the power model its practical interpretation. Setting $p=0$ removes concentration from the function, yielding the same residual spread across predictions. Setting $p=1$ makes residual standard deviation increase in direct proportion to predicted concentration. Values between zero and one describe increasing absolute spread at a less-than-proportional rate. An exponent greater than one describes a faster-than-proportional increase; a negative exponent describes decreasing spread as predictions rise, provided the predictions are positive.

Exponent	Residual standard deviation	Interpretation
$p=0$	$\sigma$	Constant absolute error
$0<p<1$	Increases as a sublinear power of prediction	Increasing, but less-than-proportional absolute error
$p=1$	$\sigma(C_{\mathrm{pred}}/C_{\mathrm{ref}})$	Proportional error; constant coefficient of variation

The last row follows because the residual coefficient of variation, defined here as conditional residual standard deviation divided by predicted concentration, is $\sigma C_{\mathrm{pred}}^{p}/(C_{\mathrm{ref}}^p C_{\mathrm{pred}})$. When $p=1$, this reduces to $\sigma/C_{\mathrm{ref}}$, which does not depend on concentration. When $p=0$, absolute standard deviation stays fixed while this relative spread decreases as concentration increases.

Justification: Let $r=C_{\mathrm{pred}}/C_{\mathrm{ref}}$, which is positive. If residual standard deviation is proportional to $r^p$, write it as $\sigma r^p$, where $\sigma$ is the standard deviation at $r=1$. Squaring both sides gives $\operatorname{Var}(\epsilon\mid C_{\mathrm{pred}})=\sigma^2r^{2p}$. Thus the variance exponent is twice the standard-deviation exponent. For $p=1$, dividing the standard deviation by $C_{\mathrm{pred}}=rC_{\mathrm{ref}}$ gives $\sigma/C_{\mathrm{ref}}$, a constant coefficient of variation.

Worked Calculations

Worked Example: An Intermediate Power

Suppose hypothetical concentration predictions are measured in mg/L, $C_{\mathrm{ref}}=10$ mg/L, $\sigma=0.80$ mg/L, and $p=0.5$. Calculate the conditional residual standard deviation and variance at predictions of 40 mg/L and 2.5 mg/L.

At $C_{\mathrm{pred}}=40$ mg/L, the concentration ratio is $40/10=4$. Therefore,

\operatorname{SD}(\epsilon\mid 40)=0.80\sqrt{4}=1.60\ \text{mg/L}, \qquad \operatorname{Var}(\epsilon\mid 40)=(1.60)^2=2.56\ (\text{mg/L})^2.

At $C_{\mathrm{pred}}=2.5$ mg/L, the ratio is $2.5/10=0.25$, so

\operatorname{SD}(\epsilon\mid 2.5)=0.80\sqrt{0.25}=0.400\ \text{mg/L}, \qquad \operatorname{Var}(\epsilon\mid 2.5)=(0.400)^2=0.160\ (\text{mg/L})^2.

The predicted concentration at 40 mg/L is 16 times the prediction at 2.5 mg/L, but the residual standard deviation is only four times as large. That is consistent with $p=0.5$: the standard deviation scales as the square root of the concentration ratio.

Worked Example: Comparing the Constant and Proportional Cases

Consider hypothetical predictions of 5, 10, and 20 mg/L, with $C_{\mathrm{ref}}=10$ mg/L and $\sigma=0.60$ mg/L. Calculate the standard deviations for $p=0$ and $p=1$.

For $p=0$, each concentration ratio raised to zero is 1. Thus the standard deviation is $0.60$ mg/L at all three predictions, and the variance is $0.60^2=0.360\;(\text{mg/L})^2$ at each one.

For $p=1$, the standard deviations are

\begin{aligned} C_{\mathrm{pred}}=5 &: \quad 0.60(5/10)=0.300\ \text{mg/L},\\ C_{\mathrm{pred}}=10 &: \quad 0.60(10/10)=0.600\ \text{mg/L},\\ C_{\mathrm{pred}}=20 &: \quad 0.60(20/10)=1.20\ \text{mg/L}. \end{aligned}

The corresponding variances are $0.0900$, $0.360$, and $1.44\;(\text{mg/L})^2$. In the proportional case, doubling predicted concentration doubles residual standard deviation and multiplies residual variance by four. These calculations show why the exponent must be interpreted on the standard-deviation scale before drawing conclusions about variance.

Worked Example: Interpreting a Change in Residual Spread

Suppose a hypothetical exploratory analysis suggests that residual standard deviation is about 0.30 mg/L near a prediction of 5 mg/L and about 1.20 mg/L near a prediction of 20 mg/L. Use these two approximate spreads to find the power exponent they imply, then calculate $\sigma$ for $C_{\mathrm{ref}}=10$ mg/L.

Dividing the standard deviations cancels $\sigma$ and the reference concentration:

\frac{1.20}{0.30}= \left(\frac{20}{5}\right)^p, \qquad 4=4^p, \qquad p=\frac{\log(4)}{\log(4)}=1.

To find the standard deviation at the reference concentration, use the 5 mg/L point:

0.30=\sigma(5/10)^1, \qquad \sigma=\frac{0.30}{0.5}=0.600\ \text{mg/L}.

The implied power relationship is proportional error: it predicts $0.600(20/10)=1.20$ mg/L at 20 mg/L, matching the other approximate spread. This two-point calculation illustrates interpretation; it does not establish that $p=1$ is supported by a data set. Estimation and evaluation require residual information across a range of predictions.

Assumptions, Estimation, and Diagnostics

A power variance function concerns residual spread conditional on the predictions. Its interpretation is most useful when the structural model adequately describes the conditional mean: otherwise, systematic mean-model errors can be mistaken for changing residual variance. The model also assumes that one smooth power relationship is a reasonable description over the concentration range being analyzed, rather than a relationship that changes sharply between ranges.

The variance function alone does not require normally distributed residuals. If the model is fitted using a Gaussian residual likelihood, that likelihood additionally assumes that, conditional on the predictions, residuals have mean zero and a normal distribution with the specified variance. Conditional independence of residuals is also commonly assumed unless dependence is explicitly modeled. Those distributional and independence assumptions should not be attributed to the power function itself.

A practical diagnostic is to examine residuals against predicted concentrations. A fan-shaped increase in raw residual spread may indicate that constant variance is inadequate. After fitting a power function, standardized residuals are formed by dividing each residual by its modeled standard deviation:

r_k^{*}= \frac{\epsilon_k} {\sigma(C_{\mathrm{pred},k}/C_{\mathrm{ref}})^p}.

If the variance model is suitable, the spread of these standardized residuals should be broadly similar across the prediction range. A remaining funnel, a U-shaped pattern, or a sudden change in spread suggests that a single power relationship may not capture the residual behavior. A trend in the residuals’ location, rather than just their spread, points instead toward a possible mean-model problem. Diagnostics should be interpreted with the number and distribution of observations in mind: apparent patterns at a concentration range with very few observations are uncertain.

The power model is not automatically preferable just because it is flexible. The constant-error model is simpler and may be adequate when spread is roughly stable. The proportional-error case is a useful special case when relative spread is roughly constant. The general power model allows the data to indicate an intermediate or other monotone scaling, but it uses an estimated exponent and can be unstable if predictions cover only a narrow range. Model comparisons should consider whether the added flexibility improves fit in a way consistent with residual diagnostics, not just whether a more complex model can be fitted.

Common Mistakes and Limitations

Confusing the variance exponent with the standard-deviation exponent. Under this parameterization, $p=1$ means standard deviation is proportional to concentration, while variance scales with concentration squared.
Calling $p=0$ proportional error. It is constant absolute residual standard deviation. Proportional error corresponds to $p=1$.
Ignoring the reference concentration. The value of $\sigma$ is the residual standard deviation at $C_{\mathrm{ref}}$, not necessarily at zero or at every prediction.
Reading a fitted exponent as proof of an assay mechanism. A power relationship describes an observed residual pattern; by itself it does not identify the physical source of that pattern.
Extrapolating beyond the observed range. A single power law can imply very small or very large spread far outside the predictions represented in the data, so extrapolated behavior deserves caution.

The formula is defined here for positive predictions. At zero, negative exponents are undefined, and the zero-power case needs a separate interpretation rather than treating $0^0$ as an ordinary value. Predictions must also be meaningful on the concentration scale used to define the model. These restrictions matter especially when the fitted structural model can generate predictions at or below zero.

Key takeaway: In the power variance function, $\sigma$ sets residual standard deviation at a reference concentration, while $p$ controls how that standard deviation changes with prediction. $p=0$ gives constant absolute error; $p=1$ gives proportional standard deviation and constant relative spread.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

For $C_{\mathrm{ref}}=8$ mg/L, $\sigma=0.50$ mg/L, and $p=0$, what are the residual standard deviation and variance at a prediction of 32 mg/L?
With $p=1$, by what factor does residual variance change when predicted concentration is multiplied by 3?
If standard deviation scales as the square root of predicted concentration, what is $p$, and how does standard deviation change when concentration is multiplied by 9?
Why should a residual fan pattern not automatically be attributed to the variance function?

Solutions. (1) Since $p=0$, the standard deviation is $0.50$ mg/L at every positive prediction, and the variance is $0.50^2=0.250\;(\text{mg/L})^2$. (2) Variance scales with concentration to the power $2p=2$, so multiplying concentration by 3 multiplies variance by $3^2=9$. (3) A square-root standard-deviation relationship has $p=0.5$. Multiplying concentration by 9 multiplies standard deviation by $9^{0.5}=3$. (4) A misspecified mean model can create systematic residual patterns that look like changing spread. Check whether residuals are centered appropriately across predictions as well as whether their spread changes.

Exponent	Residual standard deviation	Interpretation
\(p=0\)	\(\sigma\)	Constant absolute error
\(0<p<1\)	Increases as a sublinear power of prediction	Increasing, but less-than-proportional absolute error
\(p=1\)	\(\sigma(C_{\mathrm{pred}}/C_{\mathrm{ref}})\)	Proportional error; constant coefficient of variation