How Covariates and Individual Effects Enter a Parameter
A covariate can explain systematic differences between individuals, while an individual random effect represents remaining between-individual variation. In a log-linear parameter model, these two contributions enter the exponent, so they alter a positive parameter multiplicatively rather than by adding fixed amounts on its original scale.
The focus here is the individual parameter \(P_i\) given by \(P_i=\theta\exp(\beta x_i+\eta_i)\). The interpretation depends on distinguishing the parameter’s natural scale from the scale on which the exponent is linear. For example, if \(P_i\) is clearance, its units might be L/h; \(x_i\) could be body weight in kg; \(\beta\) then has units kg\(^{-1}\), while \(\eta_i\) is dimensionless.
The equation specifies how a covariate value and a random effect enter the parameter. The coefficient \(\beta\) describes the covariate’s linear contribution to \(\log(P_i)\), conditional on the individual effect. The value \(\eta_i\) adds an individual-specific displacement on that same log scale. Because exponentiation maps every real exponent to a positive value, the model guarantees \(P_i>0\) whenever \(\theta>0\).
The algebraic form does not, by itself, require a particular probability distribution for \(\eta_i\). A common distributional specification is \(\eta_i\sim N(0,\omega^2)\), with \(\omega^2\) the random-effect variance. When this specification is used, it is an assumption about the distribution of individual effects, not a consequence of the exponential equation. Interpretations below treat \(\eta_i\) as a value held fixed when comparing covariate values.
What the Covariate Coefficient Means
Taking logarithms gives a linear expression in the covariate: \(\log(P_i)=\log(\theta)+\beta x_i+\eta_i\). Thus, for fixed \(\eta_i\), a one-unit increase in \(x_i\) changes the log parameter by \(\beta\). On the original parameter scale, the change depends on both the current parameter and the value of \(\beta\); it is not generally a constant number of parameter units.
The local effect can be found by differentiating with respect to \(x_i\), keeping \(\theta\) and \(\eta_i\) fixed:
This derivative is the change in parameter units per one covariate unit at the specified \(x_i\) and \(\eta_i\). If \(\beta>0\), the parameter increases as the covariate increases; if \(\beta<0\), it decreases. If \(\beta=0\), the covariate has no effect in this model. For nonzero \(\beta\), the derivative’s size changes with \(P_i\), so the same covariate increment need not produce the same absolute parameter change at different starting values.
For a finite covariate change \(\Delta x\), the equation also gives an exact comparison. Keeping the same individual effect in both parameter values, \(P_i(x_i+\Delta x)=P_i(x_i)\exp(\beta\Delta x)\). This follows by dividing the two expressions and cancelling the shared \(\theta\) and \(\exp(\eta_i)\). It makes clear that the covariate acts proportionally on the natural parameter scale, while its contribution remains linear on the log scale.
Worked Example: A Body-Weight Effect on Clearance
Suppose a hypothetical model for clearance in L/h is \(P_i=4\exp(0.015x_i+\eta_i)\), where \(x_i\) is body weight in kg and \(\beta=0.015\) kg\(^{-1}\). Compare the parameter values at 60 kg and 80 kg for the same individual effect, \(\eta_i=0\). The equation is positive in both cases, and the covariate effect is assumed linear in the log parameter over this range.
At 60 kg, the exponent is \(0.015(60)+0=0.90\), so \(P_i=4\exp(0.90)=4(2.45960)=9.838\) L/h.
At 80 kg, the exponent is \(0.015(80)+0=1.20\), so \(P_i=4\exp(1.20)=4(3.32012)=13.28\) L/h.
The exact finite-change calculation is \(4\exp(0.90)\exp(0.015(20))=4\exp(1.20)\approx13.280468\) L/h, consistent with direct substitution. The increase is approximately \(13.280468-9.838412=3.442056\) L/h, or 3.442 L/h. The local slope at 60 kg is \(0.015(9.838412)=0.147576\) L/h per kg; at 80 kg it is \(0.015(13.280468)=0.199207\) L/h per kg. The positive coefficient gives a positive association in this specified model, but the natural-scale slope is not constant over body weight.
What the Individual Random Effect Means
The random effect \(\eta_i\) enters additively in \(\log(P_i)\). Holding the covariate and all fixed model quantities constant, a difference in \(\eta_i\) changes the exponent and therefore changes \(P_i\) on its natural scale. For instance, if two individuals have effects \(\eta_a\) and \(\eta_b\) at the same covariate value, then \(P_a/P_b=\exp(\eta_a-\eta_b)\). This comparison isolates the individual-effect contribution: the covariate term is shared and cancels.
The random effect is not a measurement error in a single concentration observation. It is a component of the individual parameter in the hierarchical model. A positive value raises the parameter relative to the value with the same fixed quantities and a smaller \(\eta_i\); a negative value lowers it. This direction statement follows from the exponential function being increasing. The amount of change in parameter units depends on the parameter’s scale.
Worked Example: Comparing Two Individual Effects
Consider the hypothetical parameter model \(P_i=10\exp(0.02x_i+\eta_i)\), with \(P_i\) in mg/h and \(x_i\) measured in covariate units. For two individuals with the same \(x_i=5\), suppose \(\eta_a=-0.20\) and \(\eta_b=0.30\). The model assumes the same fixed coefficient and scale for both, and the comparison holds the covariate fixed.
For individual \(a\), the exponent is \(0.02(5)-0.20=-0.10\), giving \(P_a=10\exp(-0.10)=10(0.904837)=9.048\) mg/h.
For individual \(b\), the exponent is \(0.02(5)+0.30=0.40\), giving \(P_b=10\exp(0.40)=10(1.491825)=14.918\) mg/h.
The random-effect difference is \(0.30-(-0.20)=0.50\). The ratio implied by that difference is \(\exp(0.50)\), and \(9.048374\exp(0.50)\approx14.918247\) mg/h. The natural-scale difference, using the unrounded values, is \(14.918247-9.048374=5.869873\) mg/h. These calculations show why the random effect is described on the log scale even though its consequence is expressed through the individual parameter.
Covariate and Individual Contributions Together
The two terms operate simultaneously in the exponent. A covariate effect does not replace an individual effect, and the individual effect does not change the coefficient \(\beta\) in this model. Rather, each contributes to the same log parameter: \(\beta x_i\) is the covariate contribution and \(\eta_i\) is the individual's random contribution. Their sum is exponentiated along with the scale \(\theta\).
Worked Example: Calculating a Parameter With Both Terms
Suppose, hypothetically, \(P_i=12\exp(0.03x_i+\eta_i)\) L/h, with \(x_i=4\) covariate units and \(\eta_i=-0.15\). Assume the displayed log-linear relationship is appropriate for the covariate range and that this individual effect applies to the parameter being calculated.
The covariate contribution is \(0.03(4)=0.12\), and adding the individual effect gives the total exponent \(0.12-0.15=-0.03\). Therefore, \(P_i=12\exp(-0.03)=12(0.9704455)=11.6453\) L/h.
To isolate the effect of changing the covariate while holding this same \(\eta_i=-0.15\) fixed, set \(x_i=0\). Then \(P_i=12\exp(-0.15)=12(0.8607080)=10.3285\) L/h. The covariate change from 0 to 4 increases the parameter by \(11.6453-10.3285=1.3168\) L/h, using the displayed rounded parameter values. The unrounded difference is approximately 1.3168 L/h as well. The random effect is unchanged in this comparison, so the change is attributable to the covariate term under the stated model.
Assumptions, Diagnostics, and Alternatives
An interpretation is only as useful as the assumptions supporting it. The model presumes a positive parameter and a log parameter that is linear in the specified covariate, with coefficient \(\beta\). The particular covariate scale and units must be known: changing from kilograms to grams changes the numerical coefficient needed to represent the same relationship. If a normal distribution is assigned to \(\eta_i\), that distribution is an additional assumption; independence of the random effects from the covariate is also commonly assumed in a basic model, but is not guaranteed by the equation itself.
Model assessment should look for evidence that the assumed covariate relationship is inadequate. A systematic pattern in parameter-related residuals or individual-effect estimates across covariate values can indicate a missing curvature, an incorrect covariate scale, or an omitted predictor. Such patterns do not prove a particular alternative, but they warn against interpreting \(\beta\) as a stable linear log-scale effect. It is also important to check whether the covariate has been coded in the units stated and whether the parameter and coefficient units are consistent.
A central alternative is an additive parameter model, such as \(P_i=\theta+\beta x_i+\eta_i\). In that form, the covariate and individual effect add directly in parameter units, so the change associated with a unit of \(x_i\) is constant. But an additive model can predict nonpositive values for a parameter that must be positive. The exponential form instead guarantees positivity and makes effects proportional on the original scale. It imposes a different shape, however: a linear trend in \(\log(P_i)\), not a linear trend in \(P_i\). Neither form should be selected solely because its coefficients appear easier to explain; plausibility, fit, and the parameter’s constraints matter.
| Feature | Log-linear parameter model | Additive parameter model |
|---|---|---|
| Equation | \(P_i=\theta\exp(\beta x_i+\eta_i)\) | \(P_i=\theta+\beta x_i+\eta_i\) |
| Scale of covariate and individual effects | Log-parameter scale | Natural parameter scale |
| Positivity | Guaranteed when \(\theta>0\) | Not guaranteed |
| Covariate’s natural-scale change | Depends on current parameter value | Constant per covariate unit |
A frequent mistake is to read \(\beta\) as an additive change in \(P_i\). For example, saying “a one-unit increase in \(x_i\) adds \(\beta\) L/h” is wrong for the exponential model: \(\beta\) is the change in log parameter per covariate unit and has reciprocal covariate units, not parameter units. The correction is to report the scale explicitly, and, when a natural-scale effect is needed, calculate the parameter at the specified covariate values and compare those values. Another error is to treat \(\eta_i\) as if it were an observed residual attached independently to every concentration; here it modifies the individual parameter.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- For \(P_i=6\exp(0.04x_i+\eta_i)\) mg/h, what are the units of \(\beta\) if \(x_i\) is measured in hours? What scale does \(\eta_i\) act on?
- With \(\theta=5\) L/h, \(\beta=-0.02\) kg\(^{-1}\), \(x_i=10\) kg, and \(\eta_i=0\), calculate \(P_i\). State whether the covariate effect is increasing or decreasing with \(x_i\).
- Why is it incorrect to say that \(\beta\) is the parameter-unit increase for each one-unit increase in \(x_i\)?
- Name one model diagnostic pattern that could make a linear covariate effect on the log parameter questionable.
Solutions. (1) \(\beta\) has units hour\(^{-1}\), so that \(\beta x_i\) is dimensionless. The random effect \(\eta_i\) is dimensionless and acts additively on \(\log(P_i)\). (2) The exponent is \(-0.02(10)+0=-0.20\), so \(P_i=5\exp(-0.20)=5(0.818731)=4.0937\) L/h. Since \(\beta<0\), the model predicts a decreasing parameter as \(x_i\) increases, with \(\eta_i\) fixed. (3) \(\beta\) is the change in log parameter per covariate unit, not a change in parameter units; the natural-scale change depends on the current \(P_i\). (4) A systematic trend or curvature in parameter-related residuals or individual-effect estimates across covariate values suggests that the assumed linear relationship on the log scale may be inadequate.