Optimal designs for model averaging in non-nested models (Q2051020): Difference between revisions

Consider an input-output system \(y=y(x)\) modeled by means of different regression functions \(y=\eta_s(x,\vartheta_s)\), \(s=1,\dots,r\), where \(\vartheta_s\) are unknown parameters to be estimated. Moreover, for each of the \(k\) input vectors \(x_i\), \(i=1,\dots,k\), there are \(n_i\) observations \(y_{ij}\), \(j=1,\dots,n_i\), of the output \(y_i=y(x_i)\), received with additive observational errors \(\epsilon_{ij}\). The errors are realizations of zero-mean, stochastic independent normal distributed random variables, having an unknown variance \(\sigma^2\). For each of the \(r\) regression models, using the observations \(y_{ij}\), the joint parameter \(\theta_s=(\vartheta_s,\sigma^2)\) is determined by maximum likelihood estimation. Each estimator \(\hat{\theta}_s\) of \(\theta_s\) yields then an estimate \(\hat{\mu}_s\) of a certain target parameter \(\mu_s=\mu(\theta_s)\). Given nonnegative weights \(w_s\), \(s=1,\dots,r\), \(\sum_{s=1}^r w_s =1\), of the model functions \(\eta_s\), the model average estimate \(\hat{\mu}\) of \(\mu\) is then defined by the weighted average of the estimates \(\hat{\mu}_s\). Asymptotic properties of the model average estimation procedure are considered. Moreover, optimal designs for model averaging are treated, and a simulation study is given.