Saturable Emax Functions for Covariate Effects

Why Use a Saturating Covariate Effect?

A covariate may influence a model parameter strongly over part of its range but have progressively less additional influence at higher values. A straight-line relationship would keep predicting a constant change per covariate unit, while a saturating function allows the effect to level off. In pharmacometrics, this can be a useful way to represent diminishing changes in a parameter across a specified covariate range.

The function describes a relationship in the model; by itself it does not establish that changing the covariate would cause the parameter to change. As with other covariate forms, the scale of the modeled parameter should be selected deliberately, as discussed in the earlier material on choosing the parameter scale. The Emax relationship is defined on that selected scale.

Definition: Let $x_i$ be a continuous covariate for individual $i$, and let $x_{\mathrm{ref}}$ be its reference value. For covariate values $x_i \ge x_{\mathrm{ref}}$, define $z_i=x_i-x_{\mathrm{ref}}$. A saturable Emax covariate relationship on a chosen parameter scale is $$ \psi_i=\psi_{\mathrm{ref}}+E_{\max}\frac{z_i}{EC_{50}+z_i}. $$ Here, $\psi_{\mathrm{ref}}$ is the modeled parameter at the reference covariate value; $E_{\max}$ is the signed maximum change from that baseline on the $\psi$ scale; and $EC_{50}>0$ is the covariate increment that produces half of that maximum change. The units of $EC_{50}$ are the same as the units of $x_i$.

The subscript “50” refers to half of the maximum covariate effect, not necessarily to a concentration or to a dose. If the covariate is weight in kilograms, for example, $EC_{50}$ is measured in kilograms. The reference value matters: $\psi_{\mathrm{ref}}$ is the predicted value at $x_{\mathrm{ref}}$, not automatically the value at covariate zero.

Shape, Limits, and Interpretation

For $z_i=0$, the fraction is zero, so the prediction equals the baseline. When $z_i=EC_{50}$, the fraction is $1/2$, and the prediction has moved halfway from baseline toward its limiting value. As $z_i$ grows without bound, the fraction approaches one. Thus, for a positive $E_{\max}$, the curve increases toward $\psi_{\mathrm{ref}}+E_{\max}$; for a negative $E_{\max}$, it decreases toward that same expression, which is then below baseline.

The limiting value is approached rather than reached at any finite covariate increment. It is therefore more precise to call $E_{\max}$ the maximum or asymptotic change, rather than to say that the model predicts the full change at some finite $x_i$.

\frac{z}{EC_{50}+z}=\frac{1}{2}\quad\text{when }z=EC_{50}, \qquad \lim_{z\to\infty}\frac{z}{EC_{50}+z}=1.

The diminishing slope also follows directly from differentiation. Holding the parameters fixed and differentiating with respect to the covariate increment gives

\frac{d\psi}{dz} =E_{\max}\frac{EC_{50}}{(EC_{50}+z)^2}.

For $EC_{50}>0$ and $z\ge0$, this derivative has the sign of $E_{\max}$, so the relationship is monotone in the direction specified by that parameter. Its absolute value declines as $z$ increases, which is the mathematical reason the effect flattens. This shape is an assumption of the chosen function; it should be checked against the scientific question and the covariate range in the data.

Worked Example: Predicting Clearance on the Additive Scale

Suppose a hypothetical model uses weight, measured in kilograms, to predict clearance, $CL$, in liters per hour. The reference weight is 40 kg, and the proposed relationship is $CL_i=4.5+6z_i/(40+z_i)$, where $z_i=x_i-40$ kg. Interpret the terms and predict clearance at 60 kg and 120 kg.

The baseline is $4.5$ L/h at 40 kg. The maximum change is $+6$ L/h, so the limiting clearance is $4.5+6=10.5$ L/h. The half-effect increment is $EC_{50}=40$ kg: at 80 kg, which is 40 kg above the reference, the predicted change is 3 L/h.

At 60 kg, $z=60-40=20$ kg. Substitution gives $CL=4.5+6(20)/(40+20)=4.5+120/60=6.5$ L/h.

At 120 kg, $z=120-40=80$ kg. Thus $CL=4.5+6(80)/(40+80)=4.5+480/120=8.5$ L/h. The prediction is higher than at 60 kg, but the curve is already moving toward its 10.5 L/h limit.

Choosing the Scale of the Model Parameter

The additive form is convenient when changes in the parameter itself are meaningful and all predictions remain in a scientifically plausible range. It can, however, predict an invalid value for a parameter that must be positive if the baseline and effect sizes are poorly chosen or extrapolated beyond the data.

For a strictly positive parameter, one alternative is to model its logarithm:

\log(\theta_i)=\log(\theta_{\mathrm{ref}}) +E_{\max}\frac{z_i}{EC_{50}+z_i}.

In this form, the baseline on the original scale is $\theta_{\mathrm{ref}}$, and the limiting value is $\theta_{\mathrm{ref}}\exp(E_{\max})$. Thus, $\exp(E_{\max})$ is the limiting multiplicative change relative to baseline; $E_{\max}$ itself is a change on the log scale. At the half-effect covariate increment, the predicted value is $\theta_{\mathrm{ref}}\exp(E_{\max}/2)$. The choice between scales changes the interpretation and predictions, so the scale must be stated rather than left implicit.

Worked Example: Predicting a Positive Parameter on the Log Scale

Consider a hypothetical model for clearance with reference value $CL_{\mathrm{ref}}=5$ L/h at a fat-free mass of 50 kg. Let $z=x-50$ kg, $EC_{50}=25$ kg, and $E_{\max}=\log(1.5)$. Calculate the predictions at 75 kg and 100 kg.

At 75 kg, $z=25$ kg, so $z/(EC_{50}+z)=25/(25+25)=0.5$. Therefore, $\log(CL)=\log(5)+0.5\log(1.5)$, and $CL=5\exp(0.5\log(1.5))=5\sqrt{1.5}\approx6.124$ L/h. This is the halfway point on the log scale, not the arithmetic midpoint between 5 and 7.5 L/h.

At 100 kg, $z=50$ kg and the fraction is $50/(25+50)=2/3$. The prediction is $CL=5\exp((2/3)\log(1.5))=5(1.5)^{2/3}\approx6.552$ L/h. The limiting value as weight increases without bound is $5\exp(\log(1.5))=7.5$ L/h.

Assumptions, Domain, and Data Support

The standard interpretation above relies on several conditions. The covariate is measured on a meaningful continuous scale; the reference value is stated; $EC_{50}$ is positive; and the modeled covariate values are within the domain $x_i\ge x_{\mathrm{ref}}$. Within that domain, the formula has no zero denominator and its direction is controlled by the sign of $E_{\max}$. The model also assumes a single monotone saturating trend on the selected parameter scale.

That domain restriction is consequential. If observations lie on both sides of $x_{\mathrm{ref}}$, then simply extending the formula can create a singularity at $z=-EC_{50}$, and the usual baseline-to-asymptote interpretation no longer applies across the full covariate range. Choose a reference at or below the modeled range when scientifically justified, or consider a functional form designed for the observed range. Do not silently apply the nonnegative-increment interpretation to negative $z$.

The data must also contain enough information to distinguish the curve’s features. If all observed covariate increments are much smaller than $EC_{50}$, then $z/(EC_{50}+z)$ is approximately $z/EC_{50}$; the observations may reveal a nearly linear trend but provide little information about the asymptote. If all increments are far larger than $EC_{50}$, the fraction is near one, and the data may estimate the limiting level but reveal little about the baseline or the bend. Reliable estimation of all three features generally requires useful covariate coverage around the transition as well as observations near the reference region.

Worked Example: A Decreasing Saturable Effect

Suppose a hypothetical absorption-rate parameter $k_a$, in inverse hours, is modeled additively as $k_{a,i}=1.2-0.8z_i/(10+z_i)$. Here $z_i$ is a nonnegative biomarker increment in its stated units from the reference value. Find the prediction at increments of 10 and 30 units, and interpret the sign and limit.

At $z=10$, the fraction is $10/(10+10)=0.5$. The prediction is $k_a=1.2-0.8(0.5)=1.2-0.4=0.8$ h$^{-1}$. This is halfway from the baseline of 1.2 h$^{-1}$ toward the limiting value.

At $z=30$, the fraction is $30/(10+30)=0.75$. Thus $k_a=1.2-0.8(0.75)=1.2-0.6=0.6$ h$^{-1}$. As $z$ increases without bound, $k_a$ approaches $1.2-0.8=0.4$ h$^{-1}$. The negative maximum change encodes a decreasing relationship; it does not mean that the half-effect covariate is negative.

The predictions remain positive over the modeled domain in this example. That check matters because an additive model for a positive parameter could otherwise produce impossible negative values at large covariate increments.

Diagnostics and Alternatives

A fitted saturable relationship should be evaluated as a model assumption, not accepted merely because its curve looks familiar. Useful checks include plotting observed covariate values against individual or population parameter estimates with the fitted curve, examining residual or prediction diagnostics across covariate ranges, and checking whether the estimated baseline, $E_{\max}$, and $EC_{50}$ are plausible and stable. In an NLME model, parameter uncertainty and the distribution of covariates across individuals also matter: a visually smooth fitted curve does not guarantee that its asymptote is well supported.

Warning signs: estimates of $EC_{50}$ or $E_{\max}$ change substantially with starting values; uncertainty is very large; the estimated bend lies outside the observed covariate range; predictions become implausible; or residual patterns remain systematic across the covariate.
Check the shape: a saturating function is monotone on its intended domain. If the observed relationship clearly reverses direction, a single Emax curve is not an adequate description.
Check the range: display the observed covariate values along with the fitted curve. Avoid interpreting an estimated asymptote as data-established when observations do not approach the plateau.

A linear function is simpler and may be preferable when the observed range shows no meaningful curvature. An exponential covariate function has a different shape and does not generally approach a finite additive limit. Splines can accommodate more flexible patterns, including curvature not specified in advance, but require choices about their structure and may be harder to summarize with a single half-effect covariate. These alternatives answer different modeling needs; the Emax form is not automatically superior because it has a biologically familiar label.

Common Mistakes and Practical Interpretation

Calling $EC_{50}$ the covariate value at half effect. It is the increment from the reference, so the corresponding covariate value is $x_{\mathrm{ref}}+EC_{50}$.
Ignoring the scale. In an additive model, $E_{\max}$ has the units of the parameter. In a log-scale model, it is dimensionless and the limiting ratio is $\exp(E_{\max})$.
Treating the limit as an observed prediction. The full maximum change is approached as the covariate increment grows without bound; it is not reached at a finite increment.
Using the curve outside its intended domain. The simple interpretation assumes nonnegative increments from the reference. Values below it require reconsidering the functional form and its domain.
Overinterpreting a fitted plateau. If the data do not cover the transition or high-covariate region, the maximum change may be weakly identified and largely extrapolative.

Key takeaway: A saturable Emax relationship specifies a baseline, a signed limiting change, and a covariate increment producing half that change. Interpret all three on the chosen parameter scale, state the reference and domain, and assess whether the data support the curve’s transition and plateau.

Check Your Understanding

Answer each question, then compare with the solutions below.

A model has $x_{\mathrm{ref}}=20$ mg/L and $EC_{50}=15$ mg/L. At what covariate value does the modeled effect reach half its maximum change?
For $\psi=3+8z/(12+z)$, what are the baseline, maximum change, limiting value, and prediction at $z=12$?
On a log scale, a parameter has baseline 4 units and $E_{\max}=\log(2)$. What is its limiting value, and what is its prediction at the half-effect increment?
Why may data containing only covariate increments much smaller than $EC_{50}$ provide weak information about the asymptote?
What domain issue arises if the simple Emax formula is interpreted for covariate increments below zero?

Solutions. (1) The half-effect covariate value is $x_{\mathrm{ref}}+EC_{50}=20+15=35$ mg/L. (2) The baseline is 3 on the $\psi$ scale; $E_{\max}=8$; the limiting value is $3+8=11$; and at $z=12$, $\psi=3+8(12)/(12+12)=3+4=7$. (3) The limit is $4\exp(\log 2)=8$ units. At half effect, the prediction is $4\exp(\log(2)/2)=4\sqrt{2}\approx5.657$ units. (4) For small $z$, $z/(EC_{50}+z)$ is approximately $z/EC_{50}$, so the observed trend can look nearly linear and may not reveal the plateau. (5) Negative increments can invalidate the usual baseline-to-asymptote interpretation and can make the denominator zero at $z=-EC_{50}$; the function should not be extended across such values without a justified domain and interpretation.