Analyzing cellwise weighted data

Abstract: Often the rows (cases, objects) of a dataset have weights. For instance, the weight of a case may reflect the number of times it has been observed, or its reliability. For analyzing such data many rowwise weighted techniques are available, the most well known being the weighted average. But there are also situations where the individual cells (entries) of the data matrix have weights assigned to them. The purpose of this note is to provide an approach to analyze such data. We define a cellwise weighted likelihood function, that corresponds to a transformation of the dataset which we call unpacking. Using this weighted likelihood one can carry out multivariate statistical methods such as maximum likelihood estimation and likelihood ratio tests. We pay particular attention to the estimation of covariance matrices, because these are the building blocks of much of multivariate statistics. An R implementation of the cellwise maximum likelihood estimator is provided, which employs a version of the EM algorithm. Also a faster approximate method is proposed, which is asymptotically equivalent to it.