Conditional expectations and residual analysis for the linear model
The author presents the review of some important results and ideas which have led to the current interest in Gibbs sampling. The linear model \(Y=X\beta+\varepsilon\), \(\varepsilon\sim MVN(0,V)\), \(Y=(Y_{1},\ldots,Y_{n})\), and notions of marginal and conditional residuals are considered. If \(\beta\) is estimated as \(\hat\beta=(X^{t}V^{-1}X)^{-1}X^{t}V^{-1}Y\), then \(\hat e=Y-\hat Y\) is a vector of marginal residuals, where \(\hat Y=x_{i}\hat \beta\). Writing \(\hat\beta_{(i)}\) as the vector of coefficients obtained from \((Y_{\setminus i},X_{(i)})\) (i.e. when the \(i\)th element is omitted from the estimating set) we may write \[ \tilde Y_{(i)}=x_{i}\hat\beta_{(i)}+(v_{i}^{t}V^{-1})_{(i)} (Y_{\setminus i}-X_{(i)}\hat\beta_{(i)}) \] and correspondingly define \(\tilde e_{(i)}=Y_{i}-\tilde Y_{(i)}\) as the \(i\)th leave-one-out conditional residual. If \(\tilde e=(\tilde e_{(1)},\ldots,\tilde e_{(n)})^{t}\) then we have the identity \(V^{-1}\hat e=D^{-1}\tilde e\), where \(D\) is diagonal with elements \(d_{i}={\text{Var}}(\tilde e_{(i)})\). This model is illustrated with reference to multivariate data and to a repeated measures analysis.
This page was built for publication: Conditional expectations and residual analysis for the linear model
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4231227)
