The absence of underprediction does not imply the absence of measurement bias.
The absence of underprediction does not imply the absence of measurement bias.
复制标题
不存在低估并不意味着不存在测量偏差。
DOI：
10.1037/a0014992
复制
发表时间：
2009
期刊：
影响因子：
--
通讯作者：
Millsap,RogerE
中科院分区：
文献类型：
--
作者：
Wicherts,JelteM;Millsap,RogerE
Sackett、Borneman 和 Connelly（2008 年 5 月至 6 月）最近讨论了一些经常针对在选拔中使用认知测试而提出的批评。一种批评涉及针对社会特定群体的认知能力测试中的测量偏差问题。 Sackett 等人（2008）指出，“如果没有额外的信息，人们无法确定[测试分数]的平均差异是否反映了所测量的发展能力的真实差异或该能力测量中的偏差”（第222页）。他们对测量偏差的讨论似乎表明，通过研究跨组标准的差异预测，可以准确地检测测试中的测量偏差。在这篇评论中，我们认为这种说法是不正确的。事实上，十多年来人们都知道，差分回归测试通常不能诊断测量偏差（Millsap，1997，1998，2008）。差分预测意味着根据能力测试分数预测标准分数（例如平均绩点、工作表现评级）时各组之间存在差异。通过将标准分数与能力测试分数相关的回归线中的组差异，可以在回归上下文中揭示差异预测。当测试测量的构建相同但来自不同组的两个人在测试中获得相同分数的概率不同（即，他们具有不同的预期测试分数）时，就存在测量偏差。如果上述两个人在测试中获得任何分数的概率相同，则测试被认为没有测量偏差或测量不变性（Mellenbergh，1989）。 Sackett 等人（2008）同意测量不变性的定义，我们也是如此。通过采用项目响应理论或验证性因子分析 (CFA) 等测量模型，可以直接在项目、包裹或子测试级别研究测试中的测量不变性。在这些测量模型中，可以统计测试将潜在变量与测试分数相关联的参数组的相等性（Meredith，1993；Millsap 和 Everson，1993）。另一方面，跨组的相同测试标准回归的证明不足以建立测量不变性。例如，很容易证明，在测试和标准的公共因素模型下，即使测试标准的回归在各组之间是相同的，测量偏差也可以表现为测量截距的组差异（Millsap，2008）。
Sackett, Borneman, and Connelly (May–June 2008) recently discussed several criticisms that are often raised against the use of cognitive tests in selection. One criticism concerns the issue of measurement bias in cognitive ability tests with respect to specific groups in society. Sackett et al.(2008) stated that “absent additional information, one cannot determine whether mean differences [in test scores] reflect true differences in the developed ability being measured or bias in the measurement of that ability”(p. 222). Their discussion of measurement bias appears to suggest that measurement bias in tests can be accurately detected through the study of differential prediction of criteria across groups. In this comment, we argue that this assertion is incorrect. In fact, it has been known for more than a decade that tests of differential regression are not generally diagnostic of measurement bias (Millsap, 1997, 1998, 2008).Differential prediction implies differences across groups in the prediction of criterion scores (eg, grade point average, ratings of job performance) from ability test scores. Differential prediction can be revealed in the regression context by group differences in the regression lines relating the criterion scores to the ability test scores. Measurement bias exists when two individuals who are identical on the construct (s) measured by a test but who are from different groups have different probabilities of attaining the same score on the test (ie, they have different expected test scores). A test is considered free of measurement bias, or measurement invariant, if the two persons described above have the same probability of attaining any score on the test (Mellenbergh, 1989). Sackett et al.(2008) subscribe to this definition of measurement invariance, as do we. Measurement invariance in the test can be studied directly at the item, parcel, or subtest level by adopting measurement models such as those from item response theory or confirmatory factor analysis (CFA). Within these measurement models, the equality over groups of parameters that relate latent variables to test scores can be tested statistically (Meredith, 1993; Millsap & Everson, 1993). On the other hand, the demonstration of identical test-criterion regressions across groups is not sufficient to establish measurement invariance. For example, it is easily shown that under a common factor model for the test and criterion, measurement bias can be manifested in group differences in measurement intercepts even if the regression of the criterion on the test is identical across groups (Millsap, 2008).