Some General Quantitative Considerations for the Statistical Analysis of Isoperformance Curves
Abstract
In this paper we present a recommended quantitative approach for analyzing the concept of isoperformance. The ideas outlined here rely upon the Bayesian version of model evaluation. We define models as hypotheses about the probabilities of subjects being categorized by a combination of predictor variables and criterion variables. From this foundation, a computational formula is derived whose value can be compared to ax2 distribution. For example, we are often interested in calculating the probability of a subject failing during some phase of flight training given that we have information on certain predictor variables. We would like to ascertain whether the extra information contained in such predictor variables is useful. If it is useful, then it enables us to predict the probability of failure for any given student. This ability to predict a change in the probability of failure, either in the upwards or downwards direction, is very helpful to managers and decision makers in the training community. In addition, these techniques can help answer the question of whether a candidate for flight training can "trade-off" a high score on one predictor variable for a low score on a different predictor variable. In particular, we would like to investigate the possibility of trading off different classes of predictor variable. In particular, we would like to investigate the possibility of trading off different classes of predictor variables, say cognition information processing variables and personality variables, and still achieve the same level of performance. The maximum entropy principle is used as a systematic disciplined approach to find parsimonious models.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 1999
- Accession Number
- ADA531669
Entities
People
- D. J. Blower
Organizations
- Naval Aerospace Medical Research Laboratory