An Empirical Test of the Weighted Effect Approach to Generalized Prediction Using Recursive Neural Nets

All Contents

The requirement of a strict and fixed distinction between dependent variables and independent variables, together with the presence of missing data, typically imposes considerable problems for most standard statistical prediction procedures. This paper describes a solution of these problems through the "weighted effect" approach in which recursive neural nets are used to learn how to compensate for any main and interaction effects attributable to missing data through the use of an "effect set" in addition to the data of actual cases. Extensive simulations of the approach based on an existing psychological data base showed high predictive validity, and a graceful degradation in performance with an increase in the number of unknown predictor variables. Moreover, the method proved amenable to the use of two-parameter logistic curves to arrive at a three way "low," "high," and "undecided" decision scheme with a-priori known error rates.