最穩定的後設認知監控力指標之探討：指標決定監控力是否具有一致性和一般性

The Best Measure of The Accuracy of Metacognitive Monitoring: Measures Determining the Consistency and Generality of Metacognition

本研究以實徵的方式比較六個後設認知監控力指標的穩定性，包括機率分數平均值（PS）、準度指標（CI）、偏誤指標（Bias）、區辨力指標（DI）、調整／常態化區辨力指標（ANDI）以及Goodman-Kruskal r相關係數（G指標），共採用三個穩定性的檢核準則，其為不隨題目難易度變動的穩定性、單一領域內的穩定性以及跨領域的穩定性。本研究以59位大學生為受試者，受試者重複參與三個分屬不同領域的實驗（文字再認、人臉再認、一般性常識測驗），各實驗皆採用受試者間的單因子設計（題目難易度：簡單／適中／困難），且皆採用信心判斷準確性研究範式來測量受試者的監控力準確性。本研究的三個實驗結果顯示：ANDI和G是不隨題目難易度變動的穩定監控力指標，PS（上橫線－）、Bias、CI和DI可以捕捉到單一領域內監控力的穩定性，PS（上橫線－）、Bias、CI和DI可顯示出跨領域的監控力穩定性。整合三個準則的結果，筆者所檢視的六個監控力指標中，沒有一個指標是全然穩定的監控力指標，因此發展一個新的且具穩定性的監控力指標有其必要性。本研究結果支持監控力應是一個具有一致性的一般性能力，先前有關後設認知監控力究竟是一個領域獨特的能力或是跨領域能力的爭議至少部分是源於測量指標的不同。

The Goodman-Kruskal gamma correlation coefficient (denoted as G) was evaluated by Nelson (1984) as the best measure for assessing the accuracy of metacognitive monitoring. Consequently, it was widely adopted by many researchers in the field of metacognition. Recently, some researchers found that the value of the G measure could not accurately reflect the individual's metacogntion (Schwartz & Metcalfe, 1994), it varied with item difficulty (Weaver & Bryant, 1995), and it was unstable within a single domain (Thompson & Mason, 1996). Meanwhile, some scholars used the mean probability score, bias, the calibration index, the discrimination index, and the adjusted normalized discrimination index (denoted separately as PS-, Bias, CI, DI, and ANDI) to evaluate a subject's accuracy of metacognitive monitoring, (e.g., Koriat & Goldsmith, 1996; Maki, 1998; Schraw, Dunkle, Bendixen, & Roedel, 1995). With these vaiours measures, one enigmatic issue concerns which measure best reflects the accuracy of metacognitive monitoring. In constructing a test, one needs to establish the validity and the reliability of that test. Same conceptions apply to the measures of metacognitive monitoring. From the viewpoint of definitions and mathematical formulas, it is apparent that each of thest existing measures possesses construct validity. Nevertheless, few studies examined thoroughly the reliability of these meausures. The present study, thus, empirically compared the stability of these six measures in terms of three criteria: The stability across item difficulty, the stability within a single domain, and the stability across domains. Three experiments, each with a single factor design (item difficulty: easy/medium/difficult), were conducted to assess the stability of these six measures. Fifty-nine college students repeatedly participated in three experiments. Although these experiments belonged to different domains (the word recognition test, the face recognition test, and the general knowledge test), they all adopted the confidence-judgment accuracy paradigm to measure the subject's accuracy of metacognitive monitoring. The experiments of word recognition and face recognition were conducted on IBM-compatible PCs. Each of these two experiments began by asking participants to memorize a set of items, then followed by a two-alternative recognition test. As to the general knowledge experiment, it was a one-stage recognition test. For each recognition item, regardless of the experiment, the subject had to choose the correct answer from the two alternatives, then gave a confidence rating (in the range of 50% to 100%) for the chosen answer to be regarded as correct. For each experiment, the values of PS-, Bias, CI, DI, ANDI, and G were computed. A Kruskal-Wallis test was then conducted to examine the effect of item difficulty on each of those measures. The Spearman correlation of each index was computed from split-halves of each test to evaluate the stability of each index in a single domain. Spearman correlations were also computed among experiments to reflect stability across domains. Results from the experiments showed that the values of the ANDI and G did not change with item difficulty. Those of PS-, Bias, CI, and DI were stable within a domain. The values of PS-, Bias, CI, and DI showed stability across domains. In conclusion, none of the examined measures was an entirely stable measure for the accuracy of metacognitive monitoring. Consequently, it is necessary to develop a new stable measure to assess the accuracy of metacognitive monitoring. In addition, results from various measures (PS-, Bias, CI, and DI) indicated stability over time and among tasks, implying the existence of a general and consistent metacognitive ability. The present study, thus, suggests that previous controversy about the nature of the metacognitive ability is partly due to different measures.