Scaling and Reparameterizing Emax Models for Estimation

Why scale or reparameterize an Emax model?

An Emax curve can describe the same exposure-response relationship using concentration in, for example, nanograms per milliliter or micrograms per liter. The predictions do not depend on the chosen units, but a numerical estimation procedure works with the numerical values supplied to it. If concentrations are around \(10^{-5}\), responses are around \(10^4\), and parameters have very different magnitudes, the calculations can be less well conditioned: small numerical changes in one parameter may have a very different effect from similar changes in another. Scaling puts quantities on more comparable numerical ranges. Reparameterization expresses the same curve using different parameter coordinates.

These operations can make computation more stable, but they do not create information in the data. In particular, changing units cannot resolve weak identifiability, such as having observations only far below EC50. The earlier material on identifiability and parameter confounding addresses that distinct issue.

Definition: Scaling changes the numerical units or origin used to express concentration or response. Reparameterization replaces one or more model parameters with a one-to-one alternative coordinate system over the parameter region of interest. A transformation is prediction-equivalent if it leaves the predicted response at every concentration unchanged after parameters are translated consistently.

Exact scaling of the Emax model

Use the baseline-relative direct Emax model

R(C)=E_0+E_{\max}\frac{C}{EC_{50}+C},

where \(C\) and \(EC_{50}\) have the same concentration units, \(E_0\) and \(E_{\max}\) have response units, and \(EC_{50}>0\). The signed value of \(E_{\max}\) permits either an increasing or a decreasing response. Choose a positive concentration scale \(C_s\), a positive response scale \(R_s\), and a response origin \(R_r\). Define the scaled variables

c=\frac{C}{C_s},\qquad r=\frac{R-R_r}{R_s}.

The transformed parameters are

\beta_0=\frac{E_0-R_r}{R_s},\qquad \beta_{\max}=\frac{E_{\max}}{R_s},\qquad c_{50}=\frac{EC_{50}}{C_s}.

Substituting \(C=C_s c\) and the transformed parameters into the model gives

r(c)=\beta_0+\beta_{\max}\frac{c}{c_{50}+c}.

Thus the original and scaled equations describe exactly the same curve. Recover original-unit parameters using \(E_0=R_r+R_s\beta_0\), \(E_{\max}=R_s\beta_{\max}\), and \(EC_{50}=C_s c_{50}\). The response origin shifts the baseline coordinate; the response scale changes the numerical values of both response parameters. Neither changes the fraction of the maximum effect achieved at a given concentration.

Assumptions: The concentration scale \(C_s\) and response scale \(R_s\) must be positive and fixed for the analysis; \(C\) and \(EC_{50}\) must use the same units; and the response transformation must be applied consistently to observations and predictions. The transformation is algebraically exact. Whether it improves estimation depends on the data, implementation, and numerical method.

Worked Example: Scaling Concentration and Response

Suppose hypothetical data are described by \(E_0=100\) response units, \(E_{\max}=60\) response units, and \(EC_{50}=20\) concentration units. Set \(C_s=10\) concentration units, \(R_r=100\) response units, and \(R_s=20\) response units. The transformed parameters are

\beta_0=\frac{100-100}{20}=0,\qquad \beta_{\max}=\frac{60}{20}=3,\qquad c_{50}=\frac{20}{10}=2.

At \(C=20\), the original prediction is

R(20)=100+60\frac{20}{20+20} =100+30=130.

Here \(c=20/10=2\), so the scaled prediction is

r(2)=0+3\frac{2}{2+2}=1.5.

Converting back gives \(R=R_r+R_s r=100+20(1.5)=130\), exactly the original prediction. At \(C=5\), the original prediction is \(100+60(5/25)=112\); the scaled prediction is \(r(0.5)=3(0.5/2.5)=0.6\), which converts back to \(100+20(0.6)=112\). The agreement at both concentrations illustrates that this is a coordinate change, not a different curve.

Alternative parameter coordinates

Scaling changes units; reparameterization changes the parameter coordinates. Two common alternatives express the concentration dependence using a reciprocal or a logarithm. With scaled concentration \(c=C/C_s\), define \(\kappa=C_s/EC_{50}=1/c_{50}\). Then

\frac{c}{c_{50}+c} =\frac{\kappa c}{1+\kappa c}.

This reciprocal form is one-to-one for positive \(EC_{50}\). The log form uses \(\eta_{50}=\log(c_{50})\), so \(c_{50}=\exp(\eta_{50})\). It represents the same positive values of \(c_{50}\) while making positivity automatic when the parameter is estimated on an unrestricted real scale. Neither alternative is universally better: reciprocal coordinates can be awkward when \(EC_{50}\) is close to zero, and log coordinates can make interpretation less direct because a change in \(\eta_{50}\) is a relative, rather than additive, change in \(c_{50}\).

Worked Example: Reciprocal EC50 Coordinates

Consider a hypothetical curve with \(EC_{50}=0.006\) mg/L and choose \(C_s=0.010\) mg/L. Then \(c_{50}=0.006/0.010=0.6\). At \(C=0.010\) mg/L, \(c=1\), and the fraction of maximum effect is

\frac{c}{c_{50}+c}=\frac{1}{0.6+1} =\frac{1}{1.6}=0.625.

The reciprocal coordinate is \(\kappa=1/0.6=1.667\), giving

\frac{\kappa c}{1+\kappa c} =\frac{1.667(1)}{1+1.667} \approx\frac{1.667}{2.667}=0.625.

For a hypothetical \(E_0=2.3\) response units and \(E_{\max}=1.2\) response units, both coordinates therefore predict \(2.3+1.2(0.625)=3.05\) response units. The reciprocal parameter is dimensionless here because it is defined from scaled concentration; \(1/EC_{50}\) on its own would instead carry inverse-concentration units.

Why scaling can improve numerical conditioning

A numerical procedure often updates parameters using derivatives of predicted responses with respect to those parameters. If these derivatives have very different scales, one parameter change can produce a large prediction change while a similarly sized change in another parameter produces almost none. The following derivative comparison explains why unit scaling can help, without guaranteeing it.

\frac{\partial R}{\partial E_0}=1,\qquad \frac{\partial R}{\partial E_{\max}}=\frac{C}{EC_{50}+C},\qquad \frac{\partial R}{\partial EC_{50}} =-E_{\max}\frac{C}{(EC_{50}+C)^2}.

These derivatives have different units as well as potentially different numerical magnitudes. After scaling, the derivatives are with respect to dimensionless concentration and response coordinates, and are expressed in the corresponding scaled units. A more balanced set of parameter and derivative magnitudes may reduce avoidable numerical difficulty.

Worked Example: Comparing Derivative Scales

Suppose a hypothetical model has \(E_0=2500\), \(E_{\max}=1200\), and \(EC_{50}=0.030\) nM. At \(C=0.030\) nM, the derivatives in original units are

\frac{\partial R}{\partial E_0}=1,\qquad \frac{\partial R}{\partial E_{\max}}=\frac{0.030}{0.060}=0.5,

\frac{\partial R}{\partial EC_{50}} =-1200\frac{0.030}{(0.060)^2} =-10000\ \text{response units per nM}.

Use \(C_s=0.010\) nM, \(R_r=2500\), and \(R_s=1000\). The scaled parameters are \(\beta_0=0\), \(\beta_{\max}=1.2\), and \(c_{50}=3\), and the concentration is \(c=3\). The scaled prediction is \(r=1.2(3/6)=0.6\), corresponding to \(R=2500+1000(0.6)=3100\). In scaled coordinates, the derivatives at this point are

\frac{\partial r}{\partial \beta_0}=1,\qquad \frac{\partial r}{\partial \beta_{\max}}=0.5,\qquad \frac{\partial r}{\partial c_{50}} =-1.2\frac{3}{(3+3)^2}=-0.1.

The derivative values now refer to scaled quantities, so they should not be compared as though they had the same units as the original derivatives. They illustrate how a change of coordinates can make numerical magnitudes more manageable. Whether the complete estimation problem is better conditioned depends on all observations and parameters, not just this one concentration.

For the sigmoid Emax model, the same concentration and response transformations apply, with the Hill coefficient unchanged:

r(c)=\beta_0+\beta_{\max}\frac{c^\gamma}{c_{50}^{\gamma}+c^\gamma}.

The ratio inside the fraction is unchanged by consistent concentration scaling. Consequently, the curve shape parameter \(\gamma\) does not need a unit conversion.

Checking equivalence and choosing a representation

A reliable comparison uses the same hypothetical parameter set and evaluates predictions at the same concentrations in both representations. For an Emax model, verify that the transformed predictions convert back to the original predictions, and that all transformed quantities have sensible units. A discrepancy usually indicates an inconsistent scale factor, a response origin applied to only one side of the equation, or an incorrect conversion of \(EC_{50}\).

The common representations have different practical features:

Representation	Useful feature	Point to check
Original \(EC_{50}\), response, and concentration units	Directly interpretable in the reported units	Numerical magnitudes may be very different
Scaled \(c_{50}\), \(\beta_0\), and \(\beta_{\max}\)	Can place variables and parameters on more comparable scales	Report the scale factors so parameters can be translated back
Reciprocal \(\kappa=1/c_{50}\)	May be convenient when the concentration-response equation is written in terms of concentration multiplied by a rate-like coordinate	Large reciprocal values can be inconvenient when \(c_{50}\) is small
Log coordinate \(\eta_{50}=\log(c_{50})\)	Enforces \(c_{50}>0\) through \(c_{50}=\exp(\eta_{50})\)	Interpret effects on the log scale and convert back before reporting

If response scaling divides every residual by the same positive constant, an unweighted sum of squared residuals is multiplied by a constant factor \(1/R_s^2\). Its minimizing predictions are therefore unchanged. If the response origin is also shifted consistently, residual differences are unchanged. For likelihood-based analyses, transformations of the response must be reflected in the likelihood and its density scale; changing the response units does not license changing the probability model.

Diagnostics, limitations, and common mistakes

Scaling is a computational choice, not evidence that a model is biologically adequate. Useful checks include confirming the transformed and back-transformed predictions numerically at several concentrations, checking units in every conversion, and confirming that reported parameters have been returned to the intended clinical units. If parameter values or predictions change materially when only units or equivalent coordinates change, investigate numerical convergence, implementation, and parameter constraints. This is a warning sign, not proof of a particular failure.

Numerical conditioning must also be distinguished from identifiability. If all observations lie in a concentration range where the curve is nearly linear, rescaling may make computations tidier but does not supply information about the plateau or EC50. Similarly, reparameterizing a strongly correlated pair of parameters may alter the geometry seen by an algorithm without resolving the underlying confounding. The limitations of the data remain.

Common mistake: treating scaled parameters as if they retained original units. A fitted \(c_{50}=2\) means \(EC_{50}=2C_s\), not necessarily 2 concentration units.
Common mistake: scaling concentrations but not \(EC_{50}\). The Emax fraction changes if only one is converted. Both must use the same concentration scale.
Common mistake: assuming better numerical ranges guarantee reliable estimates. Scaling can address avoidable numerical imbalance; it cannot correct sparse concentration coverage, weak identifiability, or model misspecification.
Common mistake: assuming an alternative coordinate changes the pharmacologic meaning. A reciprocal or log coordinate is a computational representation. Interpret and report the corresponding EC50 in the original units.

Key takeaway: Positive unit scales and consistent parameter transformations preserve every Emax prediction. Choose scales that are numerically convenient, verify equivalence by back-transformation, and interpret conclusions in the original concentration and response units.

Check Your Understanding

Answer each question, then compare with the solutions below.

If \(C_s=5\) mg/L and a fitted scaled \(c_{50}=1.4\), what is \(EC_{50}\) in mg/L?
Why must the same concentration scaling be applied to \(C\) and \(EC_{50}\)?
For \(c=2\) and \(c_{50}=0.5\), calculate the Emax fraction and the equivalent reciprocal-coordinate value using \(\kappa=1/c_{50}\).
Does centering and scaling response guarantee that EC50 and Emax are identifiable? Explain.
For an unweighted least-squares fit, what happens to the sum of squared residuals when all response residuals are divided by \(R_s\), and does the minimizing parameterization change if the transformation is applied consistently?

Solutions. (1) \(EC_{50}=C_s c_{50}=5(1.4)=7.0\) mg/L. (2) The fraction \(C/(EC_{50}+C)\) is dimensionless only when numerator and denominator use the same concentration units; scaling only one changes the curve. (3) The fraction is \(2/(0.5+2)=2/2.5=0.8\). Since \(\kappa=1/0.5=2\), the reciprocal form gives \(2(2)/(1+2(2))=4/5=0.8\). (4) No. Scaling changes numerical coordinates, not concentration coverage or the information in the observations; weak identifiability can remain. (5) The sum of squared residuals is multiplied by \(1/R_s^2\). With a fixed positive \(R_s\) and consistent transformation, its minimizer in prediction space is unchanged.