Additive Linear Effects on the Parameter Scale

What an Additive Effect Says

An additive linear covariate effect is useful when a one-unit increase in a covariate is expected to correspond to the same absolute change in a model parameter across the covariate range being studied. For example, a slope of 0.04 L/h per kg for clearance means that each additional kilogram is associated with 0.04 L/h higher typical clearance under the specified model. It does not mean a constant percentage increase.

The parameter-scale choice matters: an additive relationship predicts changes in the units of the parameter itself. In contrast, a relationship on a transformed scale may imply a different pattern of absolute changes. The introductory material on centering continuous covariates explains why choosing a meaningful reference value can make the intercept easier to interpret.

Definition: Let $x_i$ be a continuous covariate for individual $i$, $x_{\mathrm{ref}}$ a chosen reference value, and $P_{\mathrm{typ}}(x_i)$ the typical value of a model parameter at that covariate value. An additive linear covariate model is $$ P_{\mathrm{typ}}(x_i)=\theta_P+\beta_P(x_i-x_{\mathrm{ref}}). $$ Here, $\theta_P$ is the typical parameter at the reference value, and $\beta_P$ is the absolute change in the typical parameter per one-unit increase in $x$.

The model is linear in the covariate, with an intercept at the reference value. The slope’s units are the units of $P$ divided by the units of $x$. For clearance measured in L/h and weight measured in kg, $\beta_{\mathrm{CL}}$ has units of L/h per kg. A positive slope predicts increasing typical clearance as weight increases; a negative slope predicts decreasing typical clearance. The slope describes the modelled relationship, not proof that the covariate causes a change.

Embedding the Covariate Relationship in an NLME Model

The covariate relationship specifies how the typical parameter varies across individuals. An NLME model may also allow individual parameters to vary around that typical value. One common specification for a positive parameter is

P_i=P_{\mathrm{typ}}(x_i)\exp(\eta_i), \qquad \eta_i\sim N(0,\omega_P^2).

In this specification, $\eta_i$ is an individual random effect and $\omega_P^2$ its variance. When $\eta_i=0$, the individual parameter equals the typical value at that individual’s covariate value. The exponential random-effect term keeps $P_i$ positive if $P_{\mathrm{typ}}(x_i)>0$. With mean-zero normal $\eta_i$, $P_{\mathrm{typ}}$ is the median of the individual parameter distribution at fixed $x_i$, not its arithmetic mean.

The additive effect in this tutorial is the relationship between $x_i$ and $P_{\mathrm{typ}}(x_i)$. It is not a claim that the random effects must also enter additively. Other random-effect structures may be used when justified, but the covariate function and random-effect structure should be stated separately so their meanings are not confused.

The model assumes that the straight-line relationship is a reasonable description over the covariate range of interest, that the reference value and units are correctly specified, and that the parameter predictions satisfy the model’s constraints. Estimation also requires enough covariate variation and information in the data to distinguish a slope from a constant parameter. As with other NLME specifications, the structural and residual-error models must be adequate for the data; a covariate slope cannot repair misspecification elsewhere.

Interpreting and Reparameterizing the Slope

The central implication follows by subtracting predictions at two covariate values. For any $x_1$ and $x_2$,

\begin{aligned} P_{\mathrm{typ}}(x_2)-P_{\mathrm{typ}}(x_1) &=\{\theta_P+\beta_P(x_2-x_{\mathrm{ref}})\} -\{\theta_P+\beta_P(x_1-x_{\mathrm{ref}})\}\\ &=\beta_P(x_2-x_1). \end{aligned}

Thus, a 10-unit increase in $x$ implies a change of $10\beta_P$ in the typical parameter. The reference value cancels from this difference: centering changes the intercept’s interpretation, not the fitted straight line or its slope. This also explains the key assumption embedded in the model: the absolute change is constant per covariate unit, including at the edges of the range.

The same line can be written without centering as $P_{\mathrm{typ}}(x)=\alpha_P+\beta_Px$. The two intercepts are related by $\theta_P=\alpha_P+\beta_Px_{\mathrm{ref}}$. The centered form makes $\theta_P$ a prediction at the chosen reference, while the uncentered intercept predicts the parameter at $x=0$, which may be outside the data range or scientifically unhelpful.

Worked Example: Clearance Versus Body Weight

Suppose a hypothetical NLME analysis uses body weight from 50 to 90 kg and estimates the following typical clearance relationship, with reference weight 70 kg:

\mathrm{CL}_{\mathrm{typ}}(x) =4.20\ \mathrm{L/h} +(0.0350\ \mathrm{L/h/kg})(x-70\ \mathrm{kg}).

At 50 kg, the centered weight is $50-70=-20$ kg. The prediction is $4.20+0.0350(-20)=4.20-0.700=3.50$ L/h.

At 90 kg, the centered weight is $90-70=20$ kg. The prediction is $4.20+0.0350(20)=4.20+0.700=4.90$ L/h.

The difference is $4.90-3.50=1.40$ L/h over a 40 kg difference in weight, or $1.40/40=0.0350$ L/h per kg, matching the fitted slope. The interpretation is that each additional kilogram is associated with 0.0350 L/h higher typical clearance according to this model. These are typical predictions; an individual’s prediction under the random-effect specification may differ.

Worked Example: Using the Reference Value and Slope Units

In another hypothetical model, typical volume is 30.0 L at a reference age of 40 years, and the additive slope is $-0.120$ L/year. Calculate typical volume at ages 25 and 55 years, assuming these ages are within the observed range.

The model is $V_{\mathrm{typ}}(x)=30.0\ \mathrm{L}+(-0.120\ \mathrm{L/year})(x-40\ \mathrm{years})$. At age 25, the difference from reference is $-15$ years, so $V_{\mathrm{typ}}=30.0+(-0.120)(-15)=30.0+1.80=31.8$ L.

At age 55, the difference is 15 years, so $V_{\mathrm{typ}}=30.0+(-0.120)(15)=30.0-1.80=28.2$ L.

The predictions are symmetric around 30.0 L at age 40 because the ages are equally spaced around the reference. The negative slope means a one-year increase in age corresponds to a 0.120 L decrease in typical volume under this model. The units check: L/year multiplied by years gives L, which can be added to the reference volume.

Parameter Constraints and the Observed Range

A straight line can continue decreasing or increasing indefinitely, even when the parameter cannot. For a parameter that must be positive, the additive prediction must remain positive over the covariate values for which the model will be used. With a negative slope, the lowest prediction occurs at the largest covariate value in the range; with a positive slope, it occurs at the smallest value. Check the endpoints and any relevant prediction range rather than assuming positivity from a positive reference estimate.

Worked Example: Checking Positivity Before Prediction

Consider a hypothetical typical clearance model over creatinine clearance values from 20 to 100 mL/min: $\mathrm{CL}_{\mathrm{typ}}(x)=1.80\ \mathrm{L/h}+(0.0250\ \mathrm{L/h\ per\ mL/min})(x-60\ \mathrm{mL/min})$. The slope is positive, so the minimum prediction on this range is at 20 mL/min: $1.80+(0.0250)(20-60)=1.80-1.00=0.800$ L/h. The maximum is at 100 mL/min: $1.80+(0.0250)(100-60)=1.80+1.00=2.80$ L/h.

Both endpoint predictions are positive, so the line remains positive throughout this range. However, the same formula evaluated at 0 mL/min gives $1.80+(0.0250)(0-60)=1.80-1.50=0.300$ L/h, and sufficiently low extrapolated values would eventually produce nonpositive clearance. That behavior is a warning about using the line outside its supported range, not evidence that the fitted relationship is suitable there.

A positive prediction at the observed endpoints is a necessary practical check for a positive parameter over that interval, because a straight line between two positive endpoint values remains positive. It does not establish that the functional form is scientifically adequate, that the endpoints are well supported by data, or that predictions beyond them are valid.

Diagnostics, Alternatives, and Limitations

Assess the additive relationship using both model-based checks and the data’s covariate distribution. In particular, examine whether residual patterns across the covariate suggest systematic curvature or persistent underprediction and overprediction. Check whether the covariate range is broad enough to identify a slope, whether there are sparse regions or influential observations, and whether the estimated relationship remains plausible at both ends. A statistically estimable slope may still imply a clinically negligible change, while a small subgroup or narrow covariate range may make the slope unstable.

A model with no covariate effect is the direct simpler alternative: it sets the slope to zero, so the typical parameter is constant across $x$. The additive model is appropriate when a constant absolute change is scientifically and empirically defensible. A multiplicative relationship instead implies that equal covariate increments correspond to a constant ratio, not a constant absolute difference. More flexible functions can represent curvature, but they require enough data to support their additional structure and must be evaluated without relying on unsupported extrapolation.

The main limitations of the additive form follow from its assumptions. A constant slope may be unrealistic across a broad range; predictions may violate parameter constraints; and predictions far beyond the data are especially fragile. Centering makes the reference prediction interpretable but does not fix curvature, confounding, sparse data, or model misspecification. Report the covariate units, reference value, parameter units, and range used for interpretation so the numerical slope cannot be mistaken for a unit-free effect.

Common Mistakes

Calling the slope a percentage change. The coefficient is an absolute change in parameter units per covariate unit. A percentage interpretation does not follow from this additive form.
Ignoring units. A slope of 0.035 is incomplete without specifying, for example, L/h per kg.
Interpreting the intercept at zero covariate value. In the centered model, $\theta_P$ is the typical parameter at $x_{\mathrm{ref}}$, not at $x=0$.
Treating a typical prediction as every individual’s parameter. The covariate function gives the typical value; individual random effects can produce individual parameters that differ from it.
Assuming a good fit within the data guarantees safe extrapolation. Check the covariate support and parameter constraints wherever predictions will be used.

Key takeaway: In an additive linear covariate model, the slope is a constant absolute change in the typical parameter per covariate unit. State its units and reference value, calculate predictions only over a defensible range, and check that those predictions remain scientifically plausible.

Check Your Understanding

Answer each question, then compare with the solutions below.

A parameter is modelled as $P_{\mathrm{typ}}=12.0+0.50(x-4)$, where $P$ is measured in mg/L and $x$ in hours. What is the slope’s interpretation and unit?
For that model, calculate the typical parameter at $x=6$ hours.
If a model predicts a positive typical value at both endpoints of an interval and is linear between them, can its prediction be nonpositive inside the interval? Explain.
What changes, and what stays the same, if the reference value is changed from 4 to 6 hours while preserving the same fitted straight line?

Solutions. (1) The typical parameter increases by 0.50 mg/L per one-hour increase in $x$; the slope unit is mg/L per hour. (2) $12.0+0.50(6-4)=12.0+1.00=13.0$ mg/L. (3) No. Every value on a straight line between two positive endpoint predictions is a weighted average of those predictions and is therefore positive. (4) The slope and predictions stay the same; the intercept becomes the typical parameter at 6 hours, namely $12.0+0.50(6-4)=13.0$ mg/L.