Keywords

adaptive learning model, asymmetric error, choice-based samples, error costs

Abstract

This paper introduces an adaptive-learning model, EGB2, which optimizes over a parameter space to fit data to a family of models based on maximum-likelihood criteria. We also show how EGB2 can be modified to handle asymmetric costs of Type I and Type II errors, thereby minimizing misclassification costs. It has been shown that standard methods of computing maximum-likelihood estimators of qualitative-response models are generally inconsistent when applied to sample data with different proportions than found in the universe from which the sample is drawn. We investigate how a choice estimator, based on weighting each observation's contribution to the log-likelihood function, can contribute to estimator consi-~ - - -,, ~nd how this feature can be implemented in EGB2.