Identification in the Linear Errors in Variables Model

From MaRDI portal





Consider the following multiple linear regression model with errors in variables: \(y_ j=\xi^ T\!_ j\beta +\epsilon_ j\), \(x_ j=\xi_ j+\nu_ j\), \(j=1,...,n\), where \(\xi_ j\), \(x_ j\), \(\nu_ j\), and \(\beta\) are k-vectors, \(y_ j\), \(\epsilon_ j\) are scalars. The \(\xi_ j\) are unobserved variables: instead the \(x_ j\) are observed. The measurement errors \(\nu_ j\) are unobservable as well and we assume \(\nu_ j\sim N(0,\Omega)\) for all j. The \(\epsilon_ j\) are assumed to follow a \(N(0,\sigma^ 2)\) distribution. The \(\nu_ j\) and \(\epsilon_ j\) are mutually independent of \(\xi_ j\). The \(\xi_ j\) are considered as random drawings from some, as yet unspecified, multivariate distribution with zero mean. (In the usual terminology this means that we deal with the structural version of the model.) It is fairly easy to show that if \(\xi_ j\) is drawn from a multivariate normal distribution the parameter vector \(\beta\) is not identified. For the case \(k=1\) \textit{O. Reiersøl} [ibid. 18, 375-389 (1950; Zbl 0040.225)] has shown that normality of \(\xi_ j\) is the only distributional assumption which spoils identification. Here we generalize his result to the case where k may be larger than one.




Cited in
(26)








This page was built for publication: Identification in the Linear Errors in Variables Model

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3330327)