ON SOME PREDICTION METHODS FOR CATEGORICAL DATA

Abstract

Good prediction methods are important in many fields where qualitative variables are involved. The criterion of a good prediction method, used in this paper, is the average mean squared error. This criterion is used to compare and derive prediction methods, when the variable of interest is binary. The methods considered here are based on the maximumlikelihood estimators of the expectation of the binary varible, for which we want to make a prediction. Derivations and simulations are made for the case where we have one qualitative background variable. It is for example demonstrated that, when the ordinary chi-squared test is used for choosing between two prediction methods, it should not be adopted on a conventional low level of significance (e.g. 5%).