Universidad Autónoma de Madrid, Madrid, Spain.
Behav Res Methods. 2020 Feb;52(1):1-22. doi: 10.3758/s13428-018-1181-x.
Sometimes the reports of primary studies that are potentially analyzable within the signal detection theory framework do not report sample statistics for its main indexes, especially the sample variance of d'. We describe a procedure for estimating the variance of d' from other sample statistics (specifically, the mean and variance of the observed rates of hit and false alarm). The procedure acknowledges that individuals can be heterogeneous in their sensitivity and/or decision criteria, and it does not adopt unjustifiable or needlessly complex assumptions. In two simulation studies reported here, we show that the procedure produces certain biases, but, when used in meta-analysis, it produces very reasonable results. Specifically, the weighted estimate of the mean sensitivity is very accurate, and the coverage of the confidence interval is very close to the nominal confidence level. We applied the procedure to 20 experimental groups or conditions from seven articles (employing recognition memory or attention tasks) that reported statistics for both the hit and false alarm rates, as well as for d'. In most of these studies the assumption of homogeneity was untenable. The variances estimated by our method, based on the hit and false alarm rates, approximate reasonably to the variances in d' reported in those articles. The method is useful for estimating unreported variances of d', so that the associated studies can be retained for meta-analyses.
有时，潜在可在信号检测理论框架内进行分析的主要研究报告并未报告其主要指标（尤其是 d'的样本方差）的样本统计信息。我们描述了一种从其他样本统计信息（特别是观察到的击中率和虚报率的均值和方差）估计 d'方差的程序。该程序承认个体在敏感性和/或决策标准方面可能存在异质性，并且不采用不合理或不必要的复杂假设。在本文报道的两项模拟研究中，我们表明该程序会产生一定的偏差，但在元分析中使用时，它会产生非常合理的结果。具体来说，平均敏感性的加权估计非常准确，置信区间的覆盖范围非常接近名义置信水平。我们将该程序应用于七篇文章中的 20 个实验组或条件（涉及识别记忆或注意任务），这些文章报告了击中率和虚报率以及 d'的统计信息。在这些研究中，大多数假设都是不可接受的。我们基于击中率和虚报率估计的方差与这些文章中报告的 d'的方差相当接近。该方法可用于估计未报告的 d'方差，以便可以保留相关研究进行元分析。
