Principal Component Analysis of Categorical Data, with Applications to Roll-Call Analysis
So far, many ad hoc techniques have been proposed to compute maxium likelihood estimates for various specific models. Some work well, some don't. Our purpose in this presentation is to present a general approach based on quadratic majorization. This class of algorithms has the desirable property that it computes maximum likelihood estimates by solving a sequence of least squares problems, which are generally much simpler. It also produces an algorithm which is globally convergent.