Delta-Method Uncertainty for Target-Effect Doses

Why Uncertainty in a Target Dose Needs Its Own Calculation

A dose chosen to attain a specified effect is not usually a directly estimated model parameter. It is a function of several estimated parameters: the baseline effect, the dose-response magnitude, and the dose scale, for example. Consequently, uncertainty in those parameters must be propagated through the equation that defines the target dose. The delta method provides a local, first-order approximation to that uncertainty.

The key feature is the dose-response gradient at the target. If the response changes sharply with dose, a modest error in predicted effect tends to imply a modest error in the dose needed to attain the target. If the response is nearly flat, even a small effect error can translate into a large dose error. The calculation therefore depends not only on parameter standard errors, but also on the local slope of the fitted dose-response curve.

Definition: Let \(\theta\) be a vector of dose-response parameters, let \(m(d,\theta)\) be the model-predicted effect at dose \(d\), and let \(T\) be a fixed target effect. A target dose \(d^*(\theta)\) is a solution to \(m(d^*,\theta)=T\). The delta-method variance approximates the sampling variance of the estimated target dose \(\hat d^*=d^*(\hat\theta)\) by propagating the estimated parameter covariance through the gradient of \(d^*(\theta)\).

Deriving the Target-Dose Gradient

Suppose the parameter estimator \(\hat\theta\) is approximately multivariate normal in repeated samples, with mean near the true parameter \(\theta\) and covariance matrix \(V\). Assume that the target equation has a unique interior solution near the true values, that \(m\) is differentiable in dose and parameters there, and that the dose-response slope is nonzero. The slope condition is essential: if the curve is locally flat, a small change in effect may require a very large change in dose.

Write the target equation as \(F(d,\theta)=m(d,\theta)-T=0\). Differentiating this identity with respect to the parameter vector gives the dose gradient. Here, \(\nabla_\theta m\) is the column vector of partial derivatives of the model-predicted effect with respect to the parameters, holding dose fixed, and \(\partial m/\partial d\) is the derivative with respect to dose.

\frac{\partial m}{\partial d}\nabla_\theta d^*+\nabla_\theta m=0 \qquad\Longrightarrow\qquad \nabla_\theta d^*=-\frac{\nabla_\theta m}{\partial m/\partial d}

All derivatives in this expression are evaluated at \(d=d^*\) and the parameter values of interest. The minus sign is part of the result: a parameter change that raises the predicted effect at a fixed dose will generally reduce the dose required to reach the target when the dose-response curve is increasing.

A first-order Taylor expansion of \(d^*(\hat\theta)\) around \(\theta\) is \(\hat d^*\approx d^*+(\nabla_\theta d^*)^\mathsf{T}(\hat\theta-\theta)\). Taking the variance of this linear approximation yields the standard delta-method result:

\widehat{\operatorname{Var}}(\hat d^*)\approx (\nabla_\theta d^*)^\mathsf{T}\widehat V(\nabla_\theta d^*) \qquad\text{and}\qquad SE(\hat d^*)=\sqrt{\widehat{\operatorname{Var}}(\hat d^*)}

The covariance matrix \(\widehat V\) must use the same parameter order and parameterization as the gradient. Off-diagonal covariances matter: in general, the variance is not obtained by adding only the parameter-specific variance contributions. A Wald-style interval can then be formed as \(\hat d^*\pm z_{1-\alpha/2}SE(\hat d^*)\), when a normal approximation for the estimated dose is adequate. Its units are dose units.

Worked Emax Calculation

For the increasing Emax model, let \(E_0\) be baseline effect, \(E_{\max}\) the maximum increase above baseline, and \(ED_{50}\) the dose producing half of that maximum increase. Assume these parameters are positive where appropriate and the target lies strictly between baseline and the plateau. Solving the target equation gives an explicit target dose and makes its gradient transparent.

m(d,\theta)=E_0+\frac{E_{\max}d}{ED_{50}+d}, \qquad d^*=ED_{50}\frac{T-E_0}{E_0+E_{\max}-T}

For parameter order \(\theta=(E_0,E_{\max},ED_{50})^\mathsf{T}\), set \(A=T-E_0\) and \(B=E_0+E_{\max}-T\). Differentiating the solved dose gives:

\nabla_\theta d^*= \begin{pmatrix} -\dfrac{ED_{50}E_{\max}}{B^2}\\[4pt] -\dfrac{ED_{50}A}{B^2}\\[4pt] \dfrac{A}{B} \end{pmatrix}

Worked Example: Delta-Method Uncertainty for an Emax Target Dose

Hypothetical model. Suppose effect is measured in arbitrary response units and dose in mg. Use \(E_0=10\), \(E_{\max}=100\), \(ED_{50}=20\) mg, and a fixed target \(T=60\). Assume the estimated parameter covariance is diagonal, with standard errors 3 response units for \(E_0\), 10 response units for \(E_{\max}\), and 4 mg for \(ED_{50}\). The diagonal assumption is made for this calculation; it is not a general property of Emax estimates.

First check that the target is attainable under the fitted curve: \(10<60<10+100=110\). The target dose is \(d^*=20(60-10)/(10+100-60)=20\) mg. Thus the target equation has a positive finite solution.

At these estimates, \(A=50\) response units and \(B=50\) response units. The gradient components are \(-20(100)/50^2=-0.800\) mg per response unit for \(E_0\), \(-20(50)/50^2=-0.400\) mg per response unit for \(E_{\max}\), and \(50/50=1.00\) mg per mg for \(ED_{50}\). With a diagonal covariance matrix, the approximate variance is:

\widehat{\operatorname{Var}}(\hat d^*)= (-0.800)^2(3^2)+(-0.400)^2(10^2)+(1.00)^2(4^2) =5.76+16.00+16.00=37.76\ \mathrm{mg}^2

Therefore \(SE(\hat d^*)=\sqrt{37.76}=6.145\) mg. A 95% Wald-style interval is \(20\pm1.96(6.145)\), or approximately \([7.96,\ 32.0]\) mg. This interval summarizes approximate sampling uncertainty in the estimated dose that solves the model's target equation. It is not an interval for individual patients' responses or a prediction interval for their outcomes.

Examples of the Gradient and Its Limits

Worked Example: A Linear Dose-Response Model

Hypothetical model. Suppose the fitted mean response is \(m(d,\theta)=\beta_0+\beta_1d\), with dose in mg and response in response units. Let \(\hat\beta_0=5\), \(\hat\beta_1=2\) response units per mg, and target \(T=25\) response units. Assume the parameter estimates have standard errors 2 response units and 0.20 response units per mg, respectively, and covariance zero.

The target dose is \(d^*=(T-\beta_0)/\beta_1=(25-5)/2=10\) mg. Differentiating that expression gives gradient components \(-1/\beta_1=-0.500\) mg per response unit and \(-(T-\beta_0)/\beta_1^2=-5.00\) mg per (response unit per mg). The delta-method variance is \((-0.500)^2(2^2)+(-5.00)^2(0.20^2)=1.00+1.00=2.00\) mg², so the standard error is 1.414 mg. A 95% Wald interval is \(10\pm1.96(1.414)=[7.23,12.8]\) mg. The result is conditional on the linear model being suitable in this dose range and the stated covariance being valid.

Worked Example: A Target Near the Emax Plateau

Hypothetical calculation. Retain the Emax estimates and parameter standard errors in the preceding Emax example, but set the target to \(T=108\) response units. This target is still below the fitted plateau of 110, so it has a finite positive solution. However, \(A=98\) and \(B=2\), giving \(d^*=20(98)/2=980\) mg.

The gradient is \(-20(100)/2^2=-500\) mg per response unit for \(E_0\), \(-20(98)/2^2=-490\) mg per response unit for \(E_{\max}\), and \(98/2=49\) mg per mg for \(ED_{50}\). Under the same diagonal covariance assumption, the approximate variance is \((-500)^2(3^2)+(-490)^2(10^2)+49^2(4^2)=26{,}298{,}416\) mg², giving an approximate standard error of 5,128 mg. That standard error is far larger than the estimated dose itself. The calculation signals extreme local sensitivity; a symmetric normal interval would be a poor summary of a positive dose and could extend below zero. It does not make the target dose estimate useful merely because a formula produced a number.

Assumptions, Diagnostics, and Interpretation

The approximation is most defensible when the parameter estimator has a reliable covariance estimate, the model is differentiable near the fitted values, and the target dose is a unique interior root with a clearly nonzero slope. In likelihood-based fitting, the covariance estimate is often obtained from an information matrix or another justified estimator; its adequacy depends on the data and model, not on the delta method itself. The target \(T\) is treated as fixed here. If the target is estimated or uncertain, its uncertainty and covariance with the model parameters require additional terms.

Before reporting a result, check the root and the calculation:

Substitute \(\hat d^*\) into the fitted response equation and verify that it returns the target, within numerical tolerance.
Check that the target is attainable and that the solution lies in the intended dose range. For the increasing Emax model, this requires \(E_0<T<E_0+E_{\max}\).
Evaluate the local dose-response slope. A small \(|\partial m/\partial d|\) inflates the dose gradient and can make the approximation unstable.
Use the full estimated covariance matrix when parameters are correlated, and confirm that its ordering and units match the gradient.
Compare the standard error with the estimated dose and inspect whether a proposed normal interval is compatible with the dose domain and the model's root conditions.

The delta-method interval is a confidence interval for an estimated target dose, not a prediction interval for a future patient's response. Confidence intervals versus prediction intervals for dose decisions address different sources of uncertainty. The delta method is attractive because it is fast, explicit, and interpretable: its gradient shows which parameter directions most affect the dose. Its limitation is the local linear approximation. Near a plateau, a boundary, or a region where the target root appears or disappears under small parameter changes, the distribution of the dose estimate may be strongly skewed or may not be well represented by a single local normal approximation. In such settings, a Wald interval should not be treated as automatically reliable; likelihood-based or resampling approaches are alternatives to consider, with their own assumptions and diagnostics.

Common Pitfall

Common mistake: “The Emax model has a finite target dose, so the delta-method standard error must be a reasonable measure of its uncertainty.” A finite root only establishes that the fitted curve crosses the target. It does not establish that the crossing is stable. Near the plateau, the dose-response slope can be very small, making \(\nabla_\theta d^*\) enormous. Check the gradient, standard error relative to the dose, and root conditions; if these indicate strong nonlinearity, qualify or avoid a symmetric Wald interval.

Check Your Understanding

Work through each question, then compare your reasoning with the solutions below.

For a target equation \(m(d,\theta)=T\), write the general gradient of the target dose in terms of the dose-response slope and parameter gradient.
In the linear model \(m(d)=\beta_0+\beta_1d\), let \(\beta_0=4\), \(\beta_1=3\), and \(T=19\). Find the target dose.
For the Emax model with \(E_0=8\), \(E_{\max}=72\), \(ED_{50}=15\) mg, and \(T=44\), find the target dose. Is the root positive and finite?
Explain why an Emax target close to \(E_0+E_{\max}\) can have a very large dose standard error.
What does a delta-method confidence interval for a target dose describe, and what does it not describe?

Solutions. (1) Implicit differentiation gives \(\nabla_\theta d^*=-(\nabla_\theta m)/(\partial m/\partial d)\), evaluated at the target root. (2) \(d^*=(19-4)/3=5\) mg. (3) The target is between baseline 8 and plateau \(8+72=80\), and \(d^*=15(44-8)/(8+72-44)=15(36)/36=15\) mg, a positive finite root. (4) The fitted curve is shallow near its plateau, so \(|\partial m/\partial d|\) is small; dividing by that slope makes the dose gradient large and inflates \((\nabla_\theta d^*)^\mathsf{T}\widehat V(\nabla_\theta d^*)\). (5) It approximates sampling uncertainty in the estimated dose solving the specified model target equation. It does not describe future individual response variability or predict the response of a patient given a dose.