Why the Scale of an Effect Matters
A covariate effect can be reported as a ratio, a percentage change, an absolute change in parameter units, or a standardized contrast. These forms can describe the same underlying model, yet they answer different questions. A 1 L/h change in clearance cannot be compared directly with a 1 L change in volume of distribution, and a 20% change in one parameter does not necessarily have the same clinical consequences as a 20% change in another.
The first task is therefore not to choose the scale that makes an effect appear largest. It is to state the contrast being made and the scale used, then interpret that result within the parameter’s units, reference value, and covariate range. The earlier material on presenting covariate effects with clinical context motivates this reporting discipline; here the focus is on how different scales affect the comparison itself.
A ratio is dimensionless, while an absolute change retains the parameter’s units. When \(P_0>0\), a ratio greater than 1 indicates an increase and a ratio between 0 and 1 indicates a decrease. A ratio of 1.20 corresponds to a 20% increase, not a 120% increase. Conversely, a ratio of 0.80 corresponds to a 20% decrease.
What a Standardized Contrast Does—and Does Not—Mean
“Standardized contrast” can refer to more than one operation. In a continuous-covariate model, one useful convention is to report the effect over a specified covariate increment, such as one population standard deviation of the covariate. This makes the size of the covariate change explicit. It does not, by itself, standardize the parameter change.
A different standardized contrast expresses the absolute parameter change in units of a reference standard deviation of that parameter. For clarity, define \(s_{P,0}\) as the standard deviation of \(P\) in a specified reference population, measured on the original parameter scale. The standardized parameter contrast is \(\Delta P/s_{P,0}\). This answers how large the change is relative to that reference variability. It is meaningful only when the reference population and the source of \(s_{P,0}\) are stated. For example, a between-individual standard deviation and a measurement-error standard deviation are not interchangeable.
Deriving the Scale Conversions
The conversions follow directly from the definitions. Since \(R=P_1/P_0\), multiplying both sides by \(P_0\) gives \(P_1=RP_0\). Subtracting \(P_0\) gives the absolute change, and dividing that change by \(P_0\) gives the relative change:
Thus, a ratio and a percentage change contain the same relative-effect information, expressed differently. To obtain an absolute change, the reference parameter value is also needed. For example, the same ratio of 1.20 implies a change of 1 unit when \(P_0=5\), but 10 units when \(P_0=50\).
For a power model \(P(x)=P_{\mathrm{ref}}(x/x_{\mathrm{ref}})^\theta\), where \(x_{\mathrm{ref}}>0\), the ratio for a covariate contrast from \(x_0\) to \(x_1\) is
The reference parameter \(P_{\mathrm{ref}}\) cancels in this ratio, but the covariate values do not. Consequently, an exponent alone does not specify the effect over every possible contrast. For a one-standard-deviation increase from a particular \(x_0\), the ratio is \(((x_0+s_x)/x_0)^\theta\), provided \(x_0+s_x\) is within the range where the model is appropriate.
Worked Example: Converting a Clearance Ratio to a Percentage and Absolute Change
Suppose a hypothetical power model for typical clearance is \(CL(x)=5.00(x/70)^{0.75}\) L/h, where \(x\) is body weight in kg. Compare 84 kg with 70 kg.
Ratio. The weight ratio is \(84/70=1.20\). Therefore,
The 84-kg value is about 1.147 times the 70-kg value. The percentage change is \(100(1.1465\ldots-1)\%\approx 14.65\%\), an increase.
Absolute change. At 70 kg, \(CL_0=5.00\) L/h. At 84 kg, \(CL_1=5.00(1.1465\ldots)\approx 5.733\) L/h. Thus \(\Delta CL=5.733-5.00=0.733\) L/h, to the shown precision. The ratio and percentage summarize relative change; the absolute change gives the difference in clearance units.
This calculation assumes the stated power model applies across 70–84 kg and that the comparison holds other modeled conditions fixed. It is a model-based contrast, not evidence that changing an individual’s weight would cause the calculated change.
Comparing Parameters on Multiple Scales
Ratios and percentages are often convenient when parameters differ in units or typical magnitudes. Absolute changes remain important when the amount of change in the parameter’s own units is relevant. A standardized parameter contrast adds a comparison with reference variability, but depends on how that variability was defined. No one scale removes the need to specify the contrast and the reference.
| Scale | Calculation | Question it answers | Main qualification |
|---|---|---|---|
| Ratio | \(P_1/P_0\) | How many times the reference value is the comparison value? | Requires a clearly specified reference and contrast. |
| Percentage change | \(100(P_1/P_0-1)\%\) | What is the relative increase or decrease? | Can obscure differences in parameter units and baseline magnitude. |
| Absolute change | \(P_1-P_0\) | How much does the parameter change in its own units? | Changes in different units are not directly comparable. |
| Parameter-standardized contrast | \((P_1-P_0)/s_{P,0}\) | How large is the change relative to a specified reference SD? | Depends on the chosen SD and reference population. |
| One-SD covariate contrast | Compare \(x_0\) with \(x_0+s_x\) | What effect corresponds to a specified covariate increment? | Depends on the covariate distribution and starting value. |
Worked Example: The Same Percentage Scale, Different Parameter Changes
In a hypothetical categorical covariate comparison, the estimated clearance ratio is 0.82 and the estimated volume ratio is 1.12. Suppose the reference values are \(CL_0=4.80\) L/h and \(V_0=35.0\) L. Suppose also that the specified reference-population standard deviations are \(s_{CL,0}=1.44\) L/h and \(s_{V,0}=10.5\) L.
Clearance. The percentage change is \(100(0.82-1)\%=-18.0\%\). The comparison clearance is \(0.82(4.80)=3.936\) L/h, so \(\Delta CL=3.936-4.80=-0.864\) L/h. The parameter-standardized contrast is \(-0.864/1.44=-0.600\).
Volume. The percentage change is \(100(1.12-1)\%=12.0\%\). The comparison volume is \(1.12(35.0)=39.2\) L, so \(\Delta V=39.2-35.0=4.20\) L. The parameter-standardized contrast is \(4.20/10.5=0.400\).
These are different descriptions of the modeled effects: clearance decreases by 18.0% and volume increases by 12.0%; their absolute changes are \(-0.864\) L/h and \(4.20\) L, respectively. Relative to the supplied reference SDs, their standardized contrasts are \(-0.600\) and \(0.400\). The standardized values are not universal properties of the covariate effects: changing the reference population or SDs would change them. The calculation assumes the ratios, reference values, and SDs refer to compatible populations and parameter scales.
Worked Example: Comparing Effects Over One-SD Covariate Increments
Consider two hypothetical power relationships. For clearance, let \(CL(x)=CL_{\mathrm{ref}}(x/70)^{0.75}\), with body weight \(x\) in kg and a covariate SD of 14 kg. For a second parameter \(Q\), let \(Q(z)=Q_{\mathrm{ref}}(z/60)^{0.50}\), with covariate \(z\) in its stated units, a reference value of 60, and a covariate SD of 15 units. Compare each reference value with an increase of one stated SD.
Clearance contrast. The comparison is 70 to 84 kg. Its ratio is
Thus, the modeled clearance increase over this one-SD increment is about \(14.65\%\).
\(Q\) contrast. The comparison is 60 to 75 covariate units. Its ratio is
Thus, the modeled \(Q\) increase over this one-SD increment is about \(100(1.1180-1)\%=11.80\%\). On the specified one-SD covariate contrasts, the relative effect is larger for clearance (about 14.65% versus 11.80%). The exponents alone would not establish that comparison: the starting values and covariate increments enter the ratios. The comparison also assumes both model relationships are suitable over their respective contrasts; it does not establish that the parameters have equivalent clinical consequences.
Choosing and Reporting a Scale
A useful comparison begins by stating what is held fixed and what changes. For a categorical covariate, name the reference and comparison groups. For a continuous covariate, name both values or the increment, including its units. Then identify whether the reported quantity is a ratio, a percentage, an absolute difference, or a standardized contrast. If a standardized parameter contrast is used, state the source and scale of its reference SD.
For a power model, different starting values can produce different ratios for the same exponent over an equal additive covariate increment. For instance, an increase of 10 units is a larger proportional increment when the starting value is 20 than when it is 100. Likewise, a one-SD increment depends on the covariate distribution in the population used to calculate \(s_x\). A standardized covariate contrast is therefore tied to a population, just as a parameter-standardized contrast is tied to a chosen parameter SD.
Even a well-specified parameter effect does not alone determine exposure or treatment consequences. Those depend on how the parameter enters the PK model and on the clinical setting. For example, a clearance ratio can be translated into an exposure implication only under the relevant dosing and model assumptions; the percentage change in clearance is not itself a percentage change in every exposure measure.
Diagnostics and Common Pitfalls
Before comparing reported effects, check that the contrasts and scales actually match. Two percentages are not comparable if one is calculated for a small covariate increment and the other for a large one. Two standardized contrasts are not comparable if their denominators represent different populations or different kinds of variability. A ratio should also be checked against the stated reference: reversing the groups changes the ratio to its reciprocal, not simply to the same number with a minus sign.
- Comparing raw differences across units. A difference of 2 L and a difference of 2 L/h are not the same magnitude. Use units and a clearly chosen comparison scale.
- Calling a ratio a percentage change. A ratio of 0.82 is an 18% decrease, not an 82% decrease. Calculate \(100(R-1)\%\).
- Comparing exponents as if they were effects over any contrast. In a power model, the ratio also depends on \(x_1/x_0\). State the covariate values used.
- Leaving a standardized contrast undefined. Say whether the covariate increment is one SD or whether the parameter change is divided by a parameter SD, and state the relevant reference.
- Equating scale with clinical importance. A large standardized contrast or percentage does not by itself establish an exposure or treatment consequence.
If a computed ratio does not equal \(P_1/P_0\), or if the percentage does not equal \(100(R-1)\%\), revisit the group order, reference value, units, and arithmetic. For power-model calculations, retain sufficient precision in logarithms and exponentiation until the final rounding; rounded intermediate values can change the reported result.
Check Your Understanding
For each question, identify the scale and reference before interpreting the number. Solutions follow the questions.
- A parameter ratio is 1.30. What is the percentage change?
- A parameter changes from 8.0 to 6.8 mg/L. Calculate its ratio and percentage change.
- A power model has exponent 0.50. Calculate the ratio for a covariate change from 25 to 36 units.
- An absolute parameter change is \(-2.0\) units and the specified reference parameter SD is 5.0 units. What is the parameter-standardized contrast?
- Why is it not sufficient to compare two power-model exponents to decide which covariate effect is larger?
Solutions. (1) \(100(1.30-1)\%=30.0\%\), an increase. (2) \(R=6.8/8.0=0.850\), so the percentage change is \(100(0.850-1)\%=-15.0\%\), a decrease. (3) \(R=(36/25)^{0.50}=\sqrt{1.44}=1.20\), a 20.0% increase over that contrast. (4) \(-2.0/5.0=-0.400\) reference SDs; this interpretation depends on the stated SD and reference population. (5) The ratio over a contrast also depends on the covariate values, since \(R=(x_1/x_0)^\theta\). The exponents alone do not specify effects over different covariate ranges or increments.