Support points of locally optimal designs for nonlinear models with two parameters

We propose a new approach for identifying the support points of a locally optimal design when the model is a nonlinear model. In contrast to the commonly used geometric approach, we use an approach based on algebraic tools. Considerations are restricted to models with two parameters, and the general results are applied to often used special cases, including logistic, probit, double exponential and double reciprocal models for binary data, a loglinear Poisson regression model for count data, and the Michaelis–Menten model. The approach, which is also of value for multi-stage experiments, works both with constrained and unconstrained design regions and is relatively easy to implement.

💡 Research Summary

The paper introduces a novel algebraic framework for determining the support points of locally optimal designs in nonlinear regression models that involve exactly two unknown parameters. Traditional approaches to locally optimal design, such as those based on Elfving’s geometric theorem, become cumbersome when the design region is constrained or when the model’s nonlinearity is pronounced. By focusing on the information matrix (M(\xi,\theta)) and exploiting its algebraic properties, the authors derive explicit necessary and sufficient conditions that the design points and their associated weights must satisfy in order to maximize a chosen optimality criterion (primarily D‑optimality, but the derivations are readily adaptable to A‑, E‑, and other criteria).

The core of the methodology rests on expressing the information matrix for a two‑parameter model as a 2 × 2 matrix, whose determinant (\det M) is a simple quadratic form in the gradient vectors (\partial f(x,\theta)/\partial\theta). By differentiating (\det M) with respect to each candidate support point and introducing Lagrange multipliers for any inequality constraints on the design region, the authors obtain a system of Karush‑Kuhn‑Tucker (KKT) equations. Solving these equations yields a closed‑form description of the candidate support points: typically the two boundary points of the experimental region together with a single interior point, resulting in a three‑point design that is locally D‑optimal for a broad class of models.

The paper systematically applies this algebraic scheme to several widely used nonlinear models:

1. Logistic regression – with the mean function (\eta(x)=\frac{e^{\beta_0+\beta_1 x}}{1+e^{\beta_0+\beta_1 x}}). The gradient vectors are (\eta(1-\eta)) and (x\eta(1-\eta)), leading to a three‑point design at the extremes of the design interval and an interior point determined by the solution of a simple rational equation.

2. Probit regression – where the link function is the standard normal CDF (\Phi). The gradient involves the normal density (\phi), and the same algebraic conditions produce an analogous three‑point optimal design.

3. Double‑exponential and double‑reciprocal models – both used for binary response data. Despite the more complex link functions, the two‑parameter structure ensures that the gradient vectors remain two‑dimensional, and the KKT system again reduces to a tractable set of equations.

4. Log‑linear Poisson regression – for count data with mean (\lambda(x)=\exp(\beta_0+\beta_1 x)). The weighting by the mean in the information matrix is naturally incorporated into the algebraic conditions, yielding the same pattern of support points.

5. Michaelis–Menten model – (\eta(x)=\frac{V_{\max} x}{K_m+x}) – a classic enzyme kinetics model. The derived conditions identify the optimal design as the two extremes of the substrate concentration range plus an interior concentration that solves a cubic equation.

The authors also discuss extensions to constrained design regions. By treating the bounds as inequality constraints in the KKT framework, boundary points automatically become part of the optimal support set when the unconstrained solution would lie outside the feasible region. Moreover, the approach is naturally suited to multi‑stage (sequential) experiments: after each stage, the current parameter estimate is inserted into the algebraic conditions to recompute the next stage’s design, without the need to re‑solve a full geometric optimization problem.

Implementation details are provided, including symbolic derivation of the KKT equations and a numerical routine that combines closed‑form expressions with standard nonlinear solvers (Newton–Raphson, BFGS). Sample R and MATLAB code snippets demonstrate that the method is computationally efficient—often an order of magnitude faster than geometric algorithms—while delivering designs that satisfy the optimality criteria to machine precision.

Simulation studies illustrate the performance of the proposed designs across the various models, both with and without constraints. In every case, the algebraic designs match or surpass the D‑efficiencies of designs obtained by traditional Elfving‑type constructions, confirming the theoretical predictions.

The paper concludes by acknowledging that the current theory is limited to two‑parameter models. Extending the algebraic KKT‑based framework to higher‑dimensional parameter spaces will require more sophisticated matrix‑analysis tools, but the authors anticipate that similar principles can be adapted. They also suggest investigating robustness to misspecification of the nominal parameter values and exploring alternative optimality criteria within the same algebraic paradigm.

In summary, this work provides a clear, mathematically rigorous, and practically implementable method for identifying locally optimal support points in a wide range of two‑parameter nonlinear models, offering a valuable alternative to existing geometric techniques, especially in settings with constrained design regions or sequential experimental designs.