Validating (easy) measures to combine speed and accuracy

实验心理学研究的主要部分产生了两个性能指标：按下两个按钮之一的速度（平均响应时间，平均RT）和准确性（正确百分比，PC）（两个选择强迫选择任务，2AFC）。并不总是清楚的是，这两个依赖性措施中哪一个最相关，因此应该进入统计分析，以及将相应的另一个措施视为偶然是否安全。有问题的是，平均RT和PC之间存在着强烈的、非任意的关系，因为速度的增加通常伴随着准确性的相应降低。综合措施的目的是控制这种速度-精度权衡（SAT），以反映“真实”的性能。过去已经提出了几种将两个性能方面联合收割机结合起来的措施，并在研究中大量使用。本项目旨在测试其有效性和范围。这将通过大规模的模拟使用已建立的计算决策模型和目的收集的经验数据与性能和SAT水平的实验操作来完成。其目的是确定（并进一步发展）这些措施，有效地控制SAT和最好地反映整体性能的2AFC任务，从而为未来的实验心理学研究提供指导方面的依赖措施的选择。一个潜在的结果是，没有一个检查措施证明适合于这一目的。WP 1将专注于对漂移扩散模型模拟数据的参与者内配对比较（因为我们已经发表的研究涵盖了参与者间设计）。WP 2将使用各种模型和模型版本，这些模型和模型版本实现了各种建议产生SAT的机制，以检查我们的结论是否适用于其他数据生成算法。WP 3将开发实验方法，以经验方式诱导期望的SAT水平，并产生大规模的基准数据集。WP 4将仔细研究早期WP中出现的困难，并试图解决这些问题。如果适用，WP 4还将进一步开发任何有前途的综合绩效指标。

The major share of experimental psychological studies produces two performance measures: the speed (mean response times, mean RTs) and accuracy (percentage correct, PC) of pressing one of two buttons (two-alternative forced choice tasks, 2AFC). It is not always clear, which of the two dependent measures is most relevant and should therefore be entered into statistical analyses and whether it is safe to treat the respective other measure as incidental. Problematically, there exists a strong, non-arbitrary relationship between mean RTs and PCs in that increasing speed usually comes with a corresponding decrease in accuracy. Combined measures aim to control for this speed-accuracy tradeoff (SAT), in order to reflect “true” performance. Several measures that combine both performance aspects have been suggested in the past and are heavily used in research. The present project aims to test their validity and scope. This will be done via large-scale simulations using established computational decision models and purpose-collected empirical data with experimental manipulations of performance and SAT levels. The aim is to identify (and further develop) those measures that effectively control for SATs and best reflect overall performance on 2AFC tasks and thereby to provide guidance for future experimental psychological studies with regard to the choice of dependent measures. One potential outcome is that none of the examined measures proofs suitable for this purpose. WP 1 will focus on pairwise within-participants comparisons on data simulated with the drift-diffusion model (because between-participants designs were covered by our already published study). WP 2 will use various models and model versions that implement various mechanisms that have been suggested to produce SATs to examine whether our conclusions generalize to other data-producing algorithms. WP 3 will develop experimental approaches to induced desired levels of SATs empirically and produce a large-scale benchmark data set. WP 4 will take a closer look on difficulties that have emerged in earlier WPs and try to resolve these. If applicable, WP 4 will also further develop any promising combined performance measures.