Likelihood Imputation

From MaRDI portal





Denote by \(\{f(x,\theta),\;\theta\in\Theta\}\) some parametric model of \(X\) with parameter \(\theta\), both may be vector-valued. This model is said to have a latent structure if the complete data \(x\) can only be observed through a statistic \(y=Y(x)\), where \(x\) is said to be latent. Given the data \(y\) under the latent structure model, an imputation is any latent \(x^*\) such that \(y=Y(x^*)\). The author introduces the residual likelihood to describe the information that is in \(x^*\) but outside of \(y\). In section 2 he defines the \(\delta\)-value as a measure of information contained in the residual likelihood and the likelihood imputation which minimizes this \(\delta\)-value. Consistency properties of an imputation procedure are discussed as well as algorithm implementation. Section 3 presents two examples with data from a panel survey and linear regression which illustrate various aspects of likelihood imputation in comparison with standard conditional-mean single imputation. Section 4 initiates some applications to sensitivity analysis, especially important in the case of incomplete data.











This page was built for publication: Likelihood Imputation

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4231313)