STATISTICAL TESTS FOR DETECTION MODELS.
Abstract
Methodology is given for statistical testing of detection models. The tests are intended for use in evaluating probabilistic models using data from operational exercises and are specifically designed to accommodate various features which arise in such data due to the fact that they are, in general, not controlled experiments. The principal new results in the report are the development of two statistical tests which are specifically designed for use when the exercises used in the tests do not constitute well defined experiments in the sense that the reconstruction of the runs subsequent to the detection event (if it occurs) is affected by the occurrence of the detection. An adaptation of the Bernoulli trials test is given as a test of the accuracy of a model in regard to the expected number of detections. The adaptation consists of revising the classical Bernoulli test statistic to apply to runs which are altered by the occurrence of detection ('truncated' runs). As a test of the ability of a model to predict the distribution of detections, generalizations of the classical Kolmogorov-Smirnov (K-S) test are developed, including a generalization which allows for truncated runs. Sample size requirements for the new tests are modest but they require machine computation. Computer programs are included for their implementation. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 19, 1970
- Accession Number
- AD0702497
Entities
People
- David C. Bossard