東華大學圖書館 |

Language: English

Help

回圖書館首頁

手機版館藏查詢

Back

Switch To: Labeled | MARC Mode | ISBD

Discriminating between measures of d...

Karafa, Matthew Thomas.

Linked to FindBook

Google Book

Amazon

博客來

Discriminating between measures of discrimination: A comparison of ROC area to alternatives.

Record Type:	Electronic resources : Monograph/item
Title/Author:	Discriminating between measures of discrimination: A comparison of ROC area to alternatives./
Author:	Karafa, Matthew Thomas.
Description:	107 p.
Notes:	Source: Dissertation Abstracts International, Volume: 64-07, Section: B, page: 3032.
Contained By:	Dissertation Abstracts International64-07B.
Subject:	Biology, Biostatistics. -
Online resource:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=3100011

Discriminating between measures of discrimination: A comparison of ROC area to alternatives.
Karafa, Matthew Thomas.

Discriminating between measures of discrimination: A comparison of ROC area to alternatives. - 107 p.

Source: Dissertation Abstracts International, Volume: 64-07, Section: B, page: 3032.

Thesis (Ph.D.)--Case Western Reserve University (Health Sciences), 2003.

Discrimination and calibration are fundamental to prediction accuracy, with medical studies often needing to discriminate events from non-events. This study examines the relationship between calibration and discrimination using three existing measures of discrimination ability. We simulated datasets with 50 events plus 50 non-events, with probabilistic judgments assigned. Prediction distributions among events and non-events were gradually changed from 0.5 for both groups to 1.0 for events and 0.0 for nonevents. Then, we generated 1000 bootstrap samples at each separation level and calculated several discrimination and accuracy measures, using the 2.5th, 50 th (median), and 97.5th percentiles of these bootstraps to describe their sampling distributions. Minimum detectable discrimination was defined as the smallest separation in mean predictions with a lower bound that excluded 0.0 (nil) and maximum detectable discrimination as the largest mean prediction difference in with upper bound excluding 1.0 (perfect). As the mean predictions for events groups separate (discrimination improves), calibration initially worsens but then improves once the event group predictions are non-overlapping. As mean predictions diverge, Normalized Discrimination Index (NDI) exhibits an S-shaped curve, Somer's D displays a logarithmic relationship, limiting to 1.0, and Slope Index (SI) exhibits a linear increase. Observed maximum detectable discrimination was a function of prediction variance for Somer's D and NDI. The unadjusted SI exhibits the expected linear increase but does not account for the prediction variance. With increasing prediction separation, (1) Variance adjusted SI exhibits similar maxima to NDI and Somer's D, (2) Misclassification adjusted SI has a slightly non-linear relationship, without evidence of relationship to prediction variance, and (3) Noise adjusted SI has a linear relationship and exhibits a moderate penalty for variance increases. However, alone none of these measures are sufficient to describe prediction scheme discrimination ability. The three main measures should be examined when evaluating prediction models and diagnostic test efficacy.Subjects--Topical Terms:

1018416
Biology, Biostatistics.

Discriminating between measures of discrimination: A comparison of ROC area to alternatives.
LDR:03085nmm 2200265 4500 001 1860509
005 20041028080312.5
008 130614s2003 eng d
035 $a (UnM)AAI3100011
035 $a AAI3100011
040 $a UnM $c UnM
100 1 $a Karafa, Matthew Thomas. $3 1948141
245 1 0 $a Discriminating between measures of discrimination: A comparison of ROC area to alternatives.
300 $a 107 p.
500 $a Source: Dissertation Abstracts International, Volume: 64-07, Section: B, page: 3032.
500 $a Adviser: Neal V. Dawson.
502 $a Thesis (Ph.D.)--Case Western Reserve University (Health Sciences), 2003.
520 $a Discrimination and calibration are fundamental to prediction accuracy, with medical studies often needing to discriminate events from non-events. This study examines the relationship between calibration and discrimination using three existing measures of discrimination ability. We simulated datasets with 50 events plus 50 non-events, with probabilistic judgments assigned. Prediction distributions among events and non-events were gradually changed from 0.5 for both groups to 1.0 for events and 0.0 for nonevents. Then, we generated 1000 bootstrap samples at each separation level and calculated several discrimination and accuracy measures, using the 2.5th, 50 th (median), and 97.5th percentiles of these bootstraps to describe their sampling distributions. Minimum detectable discrimination was defined as the smallest separation in mean predictions with a lower bound that excluded 0.0 (nil) and maximum detectable discrimination as the largest mean prediction difference in with upper bound excluding 1.0 (perfect). As the mean predictions for events groups separate (discrimination improves), calibration initially worsens but then improves once the event group predictions are non-overlapping. As mean predictions diverge, Normalized Discrimination Index (NDI) exhibits an S-shaped curve, Somer's D displays a logarithmic relationship, limiting to 1.0, and Slope Index (SI) exhibits a linear increase. Observed maximum detectable discrimination was a function of prediction variance for Somer's D and NDI. The unadjusted SI exhibits the expected linear increase but does not account for the prediction variance. With increasing prediction separation, (1) Variance adjusted SI exhibits similar maxima to NDI and Somer's D, (2) Misclassification adjusted SI has a slightly non-linear relationship, without evidence of relationship to prediction variance, and (3) Noise adjusted SI has a linear relationship and exhibits a moderate penalty for variance increases. However, alone none of these measures are sufficient to describe prediction scheme discrimination ability. The three main measures should be examined when evaluating prediction models and diagnostic test efficacy.
590 $a School code: 0499.
650 4 $a Biology, Biostatistics. $3 1018416
650 4 $a Statistics. $3 517247
690 $a 0308
690 $a 0463
710 2 0 $a Case Western Reserve University (Health Sciences). $3 1250782
773 0 $t Dissertation Abstracts International $g 64-07B.
790 1 0 $a Dawson, Neal V., $e advisor
790 $a 0499
791 $a Ph.D.
792 $a 2003
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=3100011