Why the Reporting Scale Matters
A parameter estimate is not fully interpretable without its scale, units, and reference condition. A fixed-effect coefficient in a model may describe a change in a logarithm, while a clinician needs to know the corresponding ratio of typical clearances or exposures. Likewise, a random-effect standard deviation on a log scale is not a percentage difference on the original scale. Reporting both the fitted quantity and its clinically interpretable translation makes the result understandable without changing what the model says.
This tutorial focuses on point estimates and their interpretation. It does not replace reporting the model equation and parameterization: the relationship between an estimate and its interpretation depends on those details. A reader should be able to identify what quantity was estimated, its units, the conditions held fixed, and the mathematical transformation used to obtain the reported interpretation.
Fixed Effects: From Coefficients to Ratios
Consider a positive parameter \(P\), such as clearance in L/h, modeled as a function of a continuous covariate \(x\):
Here, \(P_{\mathrm{ref}}\) is the typical value at the reference covariate value \(x_{\mathrm{ref}}\), and \(\beta_x\) is the fixed-effect slope. If \(x\) is measured in years, then \(\beta_x\) has units of inverse years. If \(x\) is a dimensionless indicator, such as an indicator coded 0 or 1, its coefficient is dimensionless. In either case, the product \(\beta_x(x_i-x_{\mathrm{ref}})\) is dimensionless, as a log contrast must be.
Subtracting the log value at the reference condition gives the log contrast. Exponentiating both sides gives the ratio of typical values:
The coefficient alone is not generally the ratio for a clinically meaningful comparison. The comparison requires a covariate difference. For a one-unit increase in \(x\), the ratio is \(\exp(\beta_x)\); for a change of \(d\) units, it is \(\exp(\beta_xd)\). The corresponding percentage change is \(100[\exp(\beta_xd)-1]\%\). A negative percentage indicates a decrease relative to the stated reference.
Worked Example: A Continuous Covariate With Units
Suppose a hypothetical population model for typical clearance is \(\log(CL_i)=\log(CL_{40})+0.0150(\mathrm{age}_i-40)\), with age in years and \(CL_{40}=5.00\) L/h. Interpret the estimate for a 10-year age increase and for a 60-year-old relative to the reference age.
Check the units. The age difference is in years, so the slope \(0.0150\) has units year\(^{-1}\). The product \(0.0150\ \mathrm{year}^{-1}\times 10\ \mathrm{years}=0.150\) is dimensionless.
10-year comparison. The ratio is \(\exp(0.0150\times10)=\exp(0.150)=1.1618\). Thus, under this model and holding other covariates at their reference values, typical clearance is estimated to be about 16.18% higher per 10-year increase in age.
Age 60 versus age 40. The age difference is \(60-40=20\) years, so the ratio is \(\exp(0.0150\times20)=\exp(0.300)=1.3499\). The model therefore implies typical clearance at age 60 is about 34.99% higher than at age 40, with other modeled conditions held fixed. Using the stated \(CL_{40}\), the typical value at age 60 is \(5.00\ \mathrm{L/h}\times1.3499=6.750\ \mathrm{L/h}\).
The slope is not itself a 1.5% increase per year: the exact one-year percentage change is \(100[\exp(0.0150)-1]\%=1.511\%\). The approximation of 1.5% is close for this small coefficient, but the exponential calculation gives the model-implied change.
Worked Example: A Categorical Covariate Effect
Suppose a hypothetical model uses males as the reference category and estimates a female coefficient of \(-0.2231\) for log clearance. The typical male clearance at the same values of other covariates is 5.00 L/h. Interpret the coefficient.
For a binary indicator changing from 0 (male) to 1 (female), the log contrast is \(-0.2231\), which is dimensionless. The female-to-male typical clearance ratio is \(\exp(-0.2231)=0.8000\). The estimated percentage difference is \(100(0.8000-1)=-20.00\%\).
The model therefore estimates typical clearance for females to be 20.00% lower than for males at the same values of the other modeled covariates. Applying the ratio to the reference value gives \(5.00\ \mathrm{L/h}\times0.8000=4.00\ \mathrm{L/h}\). This is a model-based comparison, not evidence by itself that sex causes a difference in clearance.
Variability: Translating Lognormal Random Effects
A common population model for a positive individual parameter is \(P_i=P_{\mathrm{typ}}\exp(\eta_i)\), where \(P_{\mathrm{typ}}\) is the typical value at specified covariate conditions and \(\eta_i\sim N(0,\omega^2)\). This model assumes normally distributed, mean-zero random effects on the log scale with variance \(\omega^2\); the resulting individual parameters are lognormally distributed. Under this convention, \(P_{\mathrm{typ}}\) is the median and geometric typical value of the distribution, not its arithmetic mean. The log-scale standard deviation is \(\omega\), and \(\omega^2\) is the variance. Neither is a percentage coefficient of variation on the original scale.
The coefficient of variation (CV) is the standard deviation divided by the arithmetic mean. For a lognormal parameter with log-scale variance \(\omega^2\), the original-scale CV is:
This conversion depends on the lognormal assumption. It describes relative variability on the original scale; it is not the same as saying that every individual lies within a fixed percentage of the typical value. The arithmetic mean is \(P_{\mathrm{typ}}\exp(\omega^2/2)\), so it exceeds the median when \(\omega^2>0\).
Why the Lognormal CV Formula Holds
Let \(P=P_{\mathrm{typ}}\exp(\eta)\), with \(\eta\sim N(0,\omega^2)\). The normal moment-generating result gives \(E[\exp(t\eta)]=\exp(t^2\omega^2/2)\). Substituting \(t=1\) gives \(E[P]=P_{\mathrm{typ}}\exp(\omega^2/2)\). Substituting \(t=2\) gives \(E[P^2]=P_{\mathrm{typ}}^2\exp(2\omega^2)\). Thus:
Taking the square root and dividing by \(E[P]=P_{\mathrm{typ}}\exp(\omega^2/2)\) cancels the positive scale terms, leaving \(CV=\sqrt{\exp(\omega^2)-1}\). This derivation also shows why the CV is dimensionless and does not depend on the units or magnitude of \(P_{\mathrm{typ}}\).
Worked Example: Converting Exponential Variability to a CV
Suppose a hypothetical clearance model reports \(\omega^2=0.0900\) for the random effect on log clearance and a typical clearance of 8.00 L/h. Report the log-scale variability and the corresponding original-scale CV.
The log-scale standard deviation is \(\omega=\sqrt{0.0900}=0.300\), which is dimensionless because it describes variation in log clearance. The original-scale CV is:
The estimate corresponds to a 30.69% CV in individual clearance under the stated lognormal model. It does not mean that each individual’s clearance differs from 8.00 L/h by exactly 30.69%, nor does it make 8.00 L/h the arithmetic mean. The implied arithmetic mean is \(8.00\exp(0.0900/2)=8.00\exp(0.0450)\approx8.368\) L/h; the typical value 8.00 L/h is the median.
How to Present Estimates Clearly
A clear report keeps the estimation scale visible while giving the transformed interpretation needed for clinical reading. For each estimate, state the model quantity, estimate, units, reference condition, and transformation. Avoid presenting a percentage or ratio without saying what is being compared. For variability, identify whether the reported quantity is \(\omega\), \(\omega^2\), or the transformed CV, and name the distributional assumption behind the transformation.
| Estimate | Report on the estimation scale | Clinically interpretable companion |
|---|---|---|
| Continuous-covariate slope on log parameter | Coefficient and reciprocal covariate units; reference value | Ratio and percentage change for a stated covariate increment |
| Indicator coefficient on log parameter | Coefficient and category coding/reference | Category-to-reference ratio and percentage difference |
| Lognormal random effect | \(\omega^2\), or \(\omega\), stating which is estimated | CV on the original scale, with the lognormal assumption |
The alternatives in the table are complementary, not competing descriptions. Reporting only a log coefficient may be mathematically complete but difficult to interpret; reporting only a ratio can hide the parameterization and the reference. A compact results table can include both. Keep a parameter’s physical units (for example, L/h) distinct from a coefficient’s units (for example, year\(^{-1}\)) and from a dimensionless ratio.
Diagnostics and Common Reporting Failures
Before finalizing an interpretation, check that the reported transformation follows the actual model equation. A log-linear covariate effect implies a multiplicative ratio; it does not imply a constant additive change in the original parameter. Confirm that the ratio compares the intended covariate values and that all other covariates are held at the stated reference conditions. For a continuous covariate, check the coefficient’s units; for a categorical covariate, check the coding and reference category.
- Confusing a log coefficient with a percentage. A coefficient of 0.10 is not exactly a 10% change. For a one-unit contrast, the model-implied change is \(100[\exp(0.10)-1]\%=10.52\%\).
- Omitting the covariate increment. A slope with units year\(^{-1}\) is not a ratio by itself. Multiply by a stated age difference before exponentiating.
- Calling \(\omega\) the original-scale CV. \(\omega\) is a log-scale standard deviation. For lognormal variability, compute \(CV=\sqrt{\exp(\omega^2)-1}\).
- Using the wrong reference. A ratio such as 0.80 is incomplete unless the numerator and denominator groups or conditions are named.
- Calling the typical value the arithmetic mean. In the stated lognormal random-effects model, \(P_{\mathrm{typ}}\) is the median; the arithmetic mean is \(P_{\mathrm{typ}}\exp(\omega^2/2)\).
These checks establish interpretability, not precision or model adequacy. A transformed point estimate does not remove uncertainty, and a clinically readable scale does not validate the model assumptions. Report the model definition and the assumptions needed for the transformation alongside the result.
Check Your Understanding
For each question, work through the scale and reference condition before interpreting the result. Solutions follow the questions.
- A log-linear model has a slope of 0.0200 per year for a positive parameter. What ratio and percentage change does the model imply for a 5-year increase?
- A binary covariate has a log-scale coefficient of 0.1823 for group B relative to group A. Interpret the estimate as a ratio and percentage difference.
- A lognormal random effect has \(\omega^2=0.0400\). Calculate \(\omega\) and the original-scale CV.
- For the model in Question 3, is the typical value the arithmetic mean? If not, give the arithmetic mean as a multiple of the typical value.
Solutions. (1) The log contrast is \(0.0200\ \mathrm{year}^{-1}\times5\ \mathrm{years}=0.100\). The ratio is \(\exp(0.100)=1.1052\), so the modeled parameter is 10.52% higher after a 5-year increase, with other conditions held fixed. (2) The group B-to-A ratio is \(\exp(0.1823)\approx1.2000\), corresponding to an estimated 20.00% higher typical value in group B than group A. (3) \(\omega=\sqrt{0.0400}=0.200\). The CV is \(\sqrt{\exp(0.0400)-1}\approx\sqrt{1.0408-1}\approx0.2020\), or about 20.20%. (4) No. The typical value is the median under the lognormal model; the arithmetic mean is \(P_{\mathrm{typ}}\exp(0.0400/2)=P_{\mathrm{typ}}\exp(0.0200)\approx1.0202P_{\mathrm{typ}}\).