FORM OF EMPIRICAL ROCS IN DISCRIMINATION AND DIAGNOSTIC TASKS - IMPLICATIONS FOR THEORY AND MEASUREMENT OF PERFORMANCE
FORM OF EMPIRICAL ROCS IN DISCRIMINATION AND DIAGNOSTIC TASKS - IMPLICATIONS FOR THEORY AND MEASUREMENT OF PERFORMANCE
复制标题
DOI：
10.1037/0033-2909.99.2.181
复制
发表时间：
1986-03-01
影响因子：
22.4
通讯作者：
SWETS, JA
中科院分区：
文献类型：
--
作者：
SWETS, JA
实验心理学中的辨别任务和几个实际领域中的诊断任务都提供了一个经验相对操作特征(ROCs)的样本。在二次正态图上，这些说明性的ROC被一条变化斜率的直线很好地拟合。这一结果对绩效模型和指标具有根本性的影响。经验ROC的形式与信号检测理论提供的一种变准则模型是一致的，与该模型的其他版本以及主要阈值模型都是不一致的。这种形式支持使用从检测理论得出的某些区分精度指数，并证实了该遗产的其他一些指标的潜在不可靠性，更重要的是，可以证明可以暗示一个或另一个阈值模型的几个共同使用的精度指数的不可靠性。精确度的首选检测理论指标是适应经验ROC斜率变化的指标。这些指数实际上独立于可以单独编制索引的决策标准。恰当地测量准确度和决策标准使人能够在心理学研究中描绘出这些因变量中的哪些受到哪些自变量的影响，并评估诊断系统的有效性和准确性。
A sample of empirical relative operating characteristics (ROCs) is presented, drawn both from discrimination tasks in experimental psychology and from diagnostic tasks in several practical fields. These illustrative ROCs are seen to be fitted well by a straight line, of varying slope, on a binormal graph. This result has fundamental implications for models and indices of performance. The form of empirical ROCs is consistent with one version of the variable-criterion model provided by signal detection theory, and is inconsistent both with other versions of that model and with the main threshold models. That form supports the use of certain indices of discrimination accuracy derived from detection theory, and substantiates the potential unreliability of others of that heritage and, more importantly, the unreliability of the several accuracy indices in common use that can be shown to imply one or another threshold model. The preferred detection-theory indices of accuracy are ones that accommodate the varying slope of empirical ROCs. These indices are effectively independent of the decision criterion, which can be indexed separately. Measuring both accuracy and the decision criterion appropriately enables one to delineate which of these dependent variables is affected by which independent variables in psychological studies, and to assess the efficacy as well as the accuracy of diagnostic systems.