Why a Power Model Becomes a Straight Line
A power model is often used to describe how a pharmacokinetic parameter changes with body weight. For positive weights, the relationship is linear on the original scale when the exponent is 0 or 1; it is generally curved for other exponent values. Taking logarithms transforms that curve into a straight-line relationship, making the exponent visible as the line’s slope. The interpretation is useful only when the model form, axes, and units are clear.
Write the weight effect relative to a reference weight \(WT_{\mathrm{ref}}\):
Here, \(\theta(WT)\) is the parameter value at body weight \(WT\), \(\theta_{\mathrm{ref}}\) is its value at the reference weight, and \(\beta\) is the power exponent. For example, \(\theta\) might be clearance in L/h, and \(WT\) might be measured in kg. The ratio \(WT/WT_{\mathrm{ref}}\) is dimensionless, so the model compares weights on a relative scale.
Using the natural logarithm, take logs of both sides of the power model:
This has the straight-line form \(y=a+\beta x\), with \(x=\log(WT/WT_{\mathrm{ref}})\) and \(y=\log(\theta/\theta_{\mathrm{unit}})\). The slope is \(\beta\). At the reference weight, \(WT=WT_{\mathrm{ref}}\), so \(x=\log(1)=0\). The intercept is therefore the log of the parameter value at that reference weight, expressed in the chosen parameter unit.
The log transformation requires positive weights and positive parameter values. It also preserves the power-model relationship only when the parameter’s systematic dependence on weight is adequately described by one exponent over the range being considered. If logarithms to base 10 are used on both axes instead, the slope remains \(\beta\); the intercept is expressed using that log base.
Worked Examples of Slope and Intercept
Worked Example: Deriving the Exponent from Two Points
Suppose a hypothetical clearance follows a power relationship, and two plotted points are \((WT,\theta)=(25\text{ kg},2.0\text{ L/h})\) and \((100\text{ kg},4.5\text{ L/h})\). Find the exponent implied by these points. For this calculation, use the ratio form, so the logarithms are of dimensionless ratios.
Divide the parameter values and the weights. The reference weight cancels when comparing two points:
Substitute the ratios and calculate:
Thus the line joining these two points on log-log axes has slope approximately 0.585. This is a two-point exponent, not a guarantee that all observations follow one power model. With more than two points, a fitted line generally will not pass through every point.
Worked Example: Reading a Fitted Line from Plotted Data
Consider hypothetical clearance values measured at three body weights: 25 kg with clearance 2.0 L/h, 50 kg with clearance 3.0 L/h, and 100 kg with clearance 4.5 L/h. Plot \(x=\log(WT/50\text{ kg})\) against \(y=\log(\theta/(1\text{ L/h}))\). The reference weight is 50 kg.
| Weight (kg) | Clearance (L/h) | \(x=\log(WT/50)\) | \(y=\log(\theta/(1\text{ L/h}))\) |
|---|---|---|---|
| 25 | 2.0 | -0.6931 | 0.6931 |
| 50 | 3.0 | 0.0000 | 1.0986 |
| 100 | 4.5 | 0.6931 | 1.5041 |
For a straight-line fit \(y=a+\beta x\), the least-squares slope is the sum of the cross-products of centered \(x\) and \(y\), divided by the sum of squared centered \(x\). In this example, the three \(x\) values are symmetric around zero, and the two endpoint \(y\) values are symmetric around \(\log(3)\). Thus \(\bar{x}=0\) and \(\bar{y}=\log(3)\). The slope calculation reduces to:
The fitted intercept is \(\hat{a}=\bar{y}-\hat{\beta}\bar{x}=\log(3)=1.0986\). Returning to the original scale, the intercept corresponds to a clearance of \(\exp(1.0986)=3.000\) L/h at 50 kg. The estimated power exponent is about 0.585. The data have been chosen so that the points lie exactly on this fitted line, apart from displayed rounding.
This calculation illustrates the axes’ interpretation: moving one unit to the right on the horizontal axis means multiplying weight by \(e\), not adding one kg. The fitted line predicts a corresponding multiplication of clearance by \(e^{\hat{\beta}}\).
Worked Example: Translating the Exponent into a Weight Change
Suppose a hypothetical parameter is 3.00 units at 40.0 kg and follows a power model with \(\beta=0.60\). Find the predicted parameter at 44.0 kg and interpret the effect of a 10% increase in weight.
The weight ratio is \(44.0/40.0=1.10\). The model predicts a multiplier of:
The predicted relative change is therefore \((1.0589-1)\times100\%=5.89\%\), approximately. The predicted parameter value is calculated using the unrounded multiplier:
The exponent is not the percentage change for every weight change. Rather, it describes how a proportional change in weight maps to a proportional change in the parameter under this model. In differential terms, \(d\log(\theta)/d\log(WT)=\beta\), so a small 1% increase in weight corresponds approximately to a 0.60% increase in the parameter when \(\beta=0.60\). For finite changes, use the power-model multiplier rather than treating that approximation as exact.
What the Slope and Intercept Do—and Do Not—Mean
The exponent is a dimensionless elasticity: it is the slope of log parameter against log weight and describes relative rather than absolute change. If \(\beta=1\), doubling weight multiplies the parameter by 2. If \(\beta=0.60\), doubling weight multiplies it by \(2^{0.60}\), approximately 1.516. If \(\beta=0\), the model predicts no weight-related change. A negative exponent describes a decrease in the parameter as weight increases.
The intercept needs more care. Its numerical value changes when the reference weight changes, because the point where the horizontal coordinate equals zero changes. Its value also depends on the parameter unit used inside the logarithm. The exponent does not change under these choices, provided the same log base is used for both axes and the data are expressed consistently. Reporting “the intercept is 1.10” without naming the reference weight and parameter unit is therefore incomplete.
A compact derivation explains why the exponent can be found from any pair of points on an exact power curve. For two positive weights \(WT_1\) and \(WT_2\), divide their model equations. The reference parameter cancels, leaving a ratio of weights raised to the exponent. Taking logs and rearranging gives:
This also shows why a log-log slope is not the same as the slope of a parameter-versus-weight plot. The latter is an absolute change in parameter per unit weight and generally varies with weight for a power curve. The log-log slope is a relative scaling exponent.
Assumptions, Plot Checks, and Alternatives
Interpreting a plotted slope as a population weight effect requires more than drawing a straight line through positive values. The power relationship should be plausible over the observed weight range, the plotted parameter values should be comparable estimates on the same scale, and other systematic effects should not be mistaken for weight dependence. In pediatric data, age and maturation can vary with weight; earlier material on distinguishing maturation from size effects and age-by-size interactions is relevant when considering whether a weight-only line is scientifically interpretable.
If a line is estimated by ordinary least squares after transformation, its statistical interpretation also depends on the error model. For example, that method treats the transformed response as the quantity with additive errors for fitting. A multiplicative error model on the original parameter scale can instead imply a different distributional description after logging. The chosen fitting procedure should match the model’s error assumptions; a visually straight plot alone does not establish them.
Useful checks include confirming that all plotted values are positive, labeling the log base and units, inspecting whether the trend is reasonably linear across the observed range, and checking whether one point has disproportionate influence on the fitted slope. A plot that bends systematically may indicate that one exponent does not summarize the whole range. These checks flag concerns; they do not by themselves identify the cause or justify a more complex model.
A power model is one option among several. A linear model on the original scale describes constant absolute change per unit weight, not constant proportional scaling. A power model describes a constant log-log slope. A more flexible nonlinear model can represent curvature, but it introduces additional parameters and requires adequate data to estimate them. Ratio-based scaling is another way to express a power model relative to a reference weight; it is not a different exponent model. The choice should follow the scientific question and observed support, not the convenience of a straight-line plot.
The fitted exponent is an empirical summary of the modeled relationship, not automatically a causal effect or a universal biological constant. Its uncertainty, the adequacy of the functional form, and the range of weights represented all matter. Extrapolating the line far beyond the observed weights assumes the same power relationship continues to hold there, an assumption that the plotted data cannot verify.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- A log-log plot of a positive parameter against weight has slope 0.75. What parameter multiplier does the model predict when weight doubles?
- A power model uses reference weight 60 kg and predicts 5.0 L/h there. What is the vertical intercept when the axes are \(\log(WT/60\text{ kg})\) and \(\log(\theta/(1\text{ L/h}))\)?
- Two points on a hypothetical power curve are (30 kg, 2.0 units) and (60 kg, 3.2 units). Calculate the exponent implied by those points.
- Why is a log-log slope of 0.5 not an increase of 0.5 parameter units per kilogram?
- Name two conditions or checks needed before interpreting a fitted log-log slope as a weight exponent.
Solutions. (1) The multiplier is \(2^{0.75}=\exp(0.75\log 2)=1.682\), so the model predicts about a 68.2% increase. (2) At the reference weight, the horizontal coordinate is zero; the intercept is \(\log(5.0)=1.609\), with the stated parameter unit. (3) \(\beta=\log(3.2/2.0)/\log(60/30)=\log(1.6)/\log(2)=0.6781\). (4) The slope relates log changes, so it describes proportional scaling; absolute change per kilogram varies with the weight and parameter values. (5) Examples include positive weights and parameter values, a reasonably power-shaped relationship over the observed range, comparable parameter estimates, and attention to possible age or maturation effects that co-vary with weight.