Bootstrapping robust measures of correlation /
The purpose of the present investigation was to determine which of several bootstrap correction formulas yield more accurate confidence intervals when applied to robust measures of association. The recently developed percentage bend correlation, along with the Winsorized correlation, were selecte...
| Main Author: | |
|---|---|
| Format: | Thesis Book |
| Language: | English |
| Published: |
[Place of publication not identified] :
[publisher not identified] ;
2000.
|
| Subjects: | |
| Online Access: | http://proxy.library.tamu.edu/login?url=http://proquest.umi.com/pqdweb?did=728415871&sid=1&Fmt=2&clientId=2945&RQT=309&VName=PQD |
| Summary: | The purpose of the present investigation was to determine which of several bootstrap correction formulas yield more accurate confidence intervals when applied to robust measures of association. The recently developed percentage bend correlation, along with the Winsorized correlation, were selected for consideration. The ordinary Pearson correlation coefficient and its transformation served as points of comparison. In addition, the outlier deletion strategies employed in calculating the two robust measures were varied in an attempt to establish optimal settings for each index. In terms of bootstrapping, four of the more popular well-studied techniques were explored: the percentile bootstrap, the adjusted bootstrap, the bias-corrected (BC)accelerated (BC[a]) bootstrap, and the bias-corrected and bootstrap. Monte Carlo simulation experiment was conducted in which Type I error probability, bias, efficiency, and interval length were correlation and bootstrap methods. Results for three of the four criteria (efficiency was virtually a constant) confirmed the superior resiliency of the robust measures under distributional nonnormality, including mixed and contaminated distributional conditions. Type I error probabilities generally fell closer the nominal rate, bias was minimal, and the ''true'' (Monte Carlo simulated) interval lengths were more faithfully reproduced by the percentage bend and Winsorized correlations. The robust measures compared favorably when bivariate normal conditions were place as well. Neither robust correlation clearly outperformed the other, and results indicated a general similarity across criterion measures regardless of the calibration setting applied to each estimator. Additionally, the Fisherian transformation of the ordinary correlation coefficient did not appreciably improve either Type I error rate or bias. Unexpectedly, the four bootstrap techniques achieved roughly equivalent outcomes. The sample size correction (i.e., the adjusted bootstrap) may have actually inflated bias and Type error rate, while the more complex BC and BC[a] procedures failed to offer sizable improvements in interval accuracy and thus are probably not worth the elaborate calculations involved. |
|---|---|
| Item Description: | Vita. "Major Subject: Educational Psychology". |
| Physical Description: | xvi, 305 leaves : illustrations ; 28 cm. Issued also on microfiche from University Microfilm Inc. |
| Bibliography: | Includes bibliographical references (leaves 286-304). |