Estimating from Misclassified Data

The problem examined is that of estimating category probabilities using sample items which may have been misclassified. The items being classified into categories are taken as subjects being interviewed, and the subjects may be hostile. A “hostile” subject is one whose response to a question posed in an interview does not correspond to the true or factual situation.The procedure recommended for problems where hostility is inhereuL requires that an assessment be made, for each subject, of the probability that that subject is hostile. These probabilities are then combined with the actual subject responses and the anticipated lying biases to yield maximum likelihood estimators of the category probabilities as solutions of a concave programming problem. Thus, an effort is made to provide better estimators in situations where the data is unreliable, but external information can be brought to bear. A Bayesian interpretation of the misclassification probabilities is presented, and numerical examples are given to illustrate application of the procedure and the effect of ignoring the misclassified items problem.