Empirical likelihood for linear models (Q1184201): Difference between revisions

If one wants to test the hypothesis that independent random vectors \(Z_ 1, \dots,Z_ n\) are identically distributed with expected value \(\mu\) while otherwise this common distribution remains unspecified, one can use what the author calls the ``empirical likelihood ratio statistic'' \[ R(\mu) = \max \left\{ \prod^ n_{i = 1} np_ i \mid p_ i \geq 0, \quad \sum^ n_{i = 1} p_ i = 1,\;\sum^ n_{i = 1} p_ iz_ i = \mu \right\}. \] As the theorem of the paper shows, the same statistic can be used for not necessarily identically distributed \(Z\)'s with common expectation under the hypothesis, and the limit distribution of \(-2 \log R (\mu)\) is \(\chi^ 2 (q)\) (where \(q\) is \(\dim \mu)\), under some mild assumption on, vaguely speaking, nondegeneracy of the covariance matrix of the \(Z\)'s and existence of \(E \| Z_ i \|^ 4\). Statistics like \(R(\mu)\) could be naturally considered under various forms of linear constraints in place of \(\sum p_ i Z_ i = \mu\). The paper specifically considers regression models like \(Y_ i = X_ i' \beta + \varepsilon_ i\) with mild assumptions on the distribution of the \(\varepsilon_ i\)'s. In this case the constraints take the form \[ \sum p_ i x_ i (y_ i - x_ i' \beta) = 0. \] The paper contains also a section on ANOVA, one called ``variance modelling'' and one with a numerical example.