Conditional non-independence of data in a practical bayesian decision task

Abstract The repetitive use of Bayes' theorem to aggregate data requires that the data be conditionally independent. The robustness of Bayesian analysis to violations of this assumption was tested by comparing Bayesian with other techniques in a practical prediction task. In classifying 861 MMPI profiles as either neurotic or psychotic, two variants of Bayesian analysis techniques, both of which ignored sizable conditional dependence in the data, had cross-validated hit rates as high as comparable analyses which correctly handled the interrelationships among the MMPI scales.