Design of experiments with unknown parameters in variance

Let the observed \(k\times 1\) vector \(y\) have a normal distribution and \(E[y |x]=\eta(x,\theta)\), Var\([y |x]=S(x,\theta)\), where \(\eta(x,\theta)=(\eta_1(x,\theta),\ldots,\eta_{k}(x,\theta))\), \(S(x,\theta)\) is a \(k\times k\) matrix, \(x\) are independent variables (predictors) and \(\theta\in \Omega\subset R^{m}\) are unknown parameters. The authors propose the combined iteratively reweighted least-squares estimator \(\hat\theta_{N}=\lim_{t\to\infty}\theta_{t}\), where \(\theta_{t}=\arg\min_{\theta\in\Omega}v_{N}^2(\theta,\theta_{t-1})\), \[ v_{N}^2(\theta,\theta')=\sum_{i=1}^{N}[y_{i}-\eta(x_{i},\theta)]^{T} S^{-1}(x_{i},\theta')[y_{i}-\eta(x_{i},\theta)]+ \] \[ +2^{-1}\sum_{i=1}^{N}tr[\{[y_{i}-\eta(x_{i},\theta')] [y_{i}-\eta(x_{i},\theta')]^{T} -\Delta\eta(x_{i};\theta,\theta') -S(x_{i},\theta)\}S^{-1}(x_{i},\theta')]^2, \] \[ \text{with}\quad\Delta \eta(x_{i};\theta,\theta')= [\eta(x_{i},\theta)-\eta(x_{i},\theta')][\eta(x_{i},\theta) -\eta(x_{i},\theta')]^{T}. \] The asymptotic equivalence of the combined iterated estimator and the maximum likelihood estimator is proved. The additive structure of the corresponding information matrices allows the authors to apply convex design theory which leads to optimal design algorithms. As an illustration of the proposed approach the following examples are considered: multivariate linear regression with unknown but constant covariance matrices, optimal designs with cost constraints, and random coefficients models.

Recommendations

Cites work

Cited in

(16)

This page was built for publication: Design of experiments with unknown parameters in variance

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4797846)