The Impact of Conditional Scores on the Performance of DETECT.
Saved in:
| Title: | The Impact of Conditional Scores on the Performance of DETECT. |
|---|---|
| Language: | English |
| Authors: | Zhang, Yanwei Oliver, Yu, Feng, Nandakumar, Ratna |
| Peer Reviewed: | N |
| Page Count: | 52 |
| Publication Date: | 2003 |
| Document Type: | Reports - Research Speeches/Meeting Papers |
| Descriptors: | Estimation (Mathematics), Nonparametric Statistics, Scores, Simulation |
| Abstract: | DETECT is a nonparametric, conditional covariance-based procedure to identify dimensional structure and the degree of multidimensionality of test data. The ability composite or conditional score used to estimate conditional covariance plays a significant role in the performance of DETECT. The number correct score of all items in the test (T) and the number correct score of remaining items (S), other than the two items in consideration, are two natural candidates for computing conditional covariances. However, these conditional scores produce biased estimates in finite samples. Some type of correction is required in computing the estimates of conditional covariances. This study investigated the effect of centering and/or averaging T and S as bias correction methods. This process resulted in six different estimates of conditional covariances for use in the DETECT procedure, and 72 types of test data were simulated to vary in sample size, test length, degree of multidimensionality, and distribution of items into clusters. The impact of the six estimates on the performance of DETECT were studied for three aspects: Dmax value, r ratio, and the percentage of items correctly classified into clusters. The results show that the centered conditional score S performed best. The next best index was the average of T and S with centering, followed by the average of T and S without centering. (Contains 8 tables, 11 figures, and 11 references.) (Author/SLD) |
| Entry Date: | 2004 |
| Accession Number: | ED478170 |
| Database: | ERIC |
| Abstract: | DETECT is a nonparametric, conditional covariance-based procedure to identify dimensional structure and the degree of multidimensionality of test data. The ability composite or conditional score used to estimate conditional covariance plays a significant role in the performance of DETECT. The number correct score of all items in the test (T) and the number correct score of remaining items (S), other than the two items in consideration, are two natural candidates for computing conditional covariances. However, these conditional scores produce biased estimates in finite samples. Some type of correction is required in computing the estimates of conditional covariances. This study investigated the effect of centering and/or averaging T and S as bias correction methods. This process resulted in six different estimates of conditional covariances for use in the DETECT procedure, and 72 types of test data were simulated to vary in sample size, test length, degree of multidimensionality, and distribution of items into clusters. The impact of the six estimates on the performance of DETECT were studied for three aspects: Dmax value, r ratio, and the percentage of items correctly classified into clusters. The results show that the centered conditional score S performed best. The next best index was the average of T and S with centering, followed by the average of T and S without centering. (Contains 8 tables, 11 figures, and 11 references.) (Author/SLD) |
|---|
