This paper discusses the development of a PCA-like method being able to capture the structure of incomplete multivariate data without any statistical assumption such as a multivariate normal distribution or a random missing process. This method, purely descriptive, is derived from a lower rank approximation of a data matrix with missing values. Parameters are estimated by the Newton-Raphson method in order to minimize the least squares criterion with respect to observed values. Two examples of educational measurement are added to demonstrate practial use of the method.
ASJC Scopus subject areas