Lattice conditional independence models for contingency tables with non-monotone missing data patterns (Q1297583)

For monotone missing data in contingency tables with multinomial distribution the likelihood function can be factored and so explicit MLEs can be obtained. For non-monotone missing data such factorization is generally impossible and iterative procedures are required. However, they may not converge to a unique solution, the result may depend on the initial values, or the rate of convergence may be extremely slow. The authors show that, under a minimal set of lattice conditional independence (LCI) restrictions, the likelihood function can be factorized similarly as in the monotone case and explicit MLEs of the parameters can be obtained. The LCI assumptions can be tested by an explicit likelihood ratio test which is also derived in this paper.