Lattice-ordered conditional independence models for missing data (Q1186630): Difference between revisions

Let \(x_ 1,\dots,x_ n\) be i.i.d. \(N_ p(\mu,\Sigma)\), \(\mu\) and (positive definite) \(\Sigma\) unknown, and let some of the components of one or more of the \(x_ j\) be missing. When the missing data pattern is monotone (= nested), an approach consists in reducing the problem to the analysis of several normal linear regression models by stepwise conditioning. For example, if \(p=3\) and observations are available only on the pairs of variables 1,2 and 1,3, the joint likelihood function \(f\) may be factored as \(f=f(1)f(2\mid 1)f(3\mid 1)\), which is equivalent to the conditional independence of 2 and 3 given \(1(2\amalg 3\mid 1)\). The authors state: ``When the missing data pattern is non-monotone, however, such analysis is impossible. It is shown that every missing data pattern naturally determines a set of lattice-ordered conditional independence restrictions which, when imposed upon the unknown covariance matrix \(\Sigma\), yields a factorization of the joint likelihood function as a product of (conditional) likelihood functions of normal linear regression models just as in the monotone case. From this factorization the maximum likelihood estimators of \(\mu\) and \(\Sigma\) (under the conditional independence restrictions) can be explicitly derived.