Mutual interference in working memory updating: A hierarchical Bayesian model

We propose a hierarchical Bayesian model for working memory updating. This model accounts for both the accuracy of the responses and the reaction times (RT) in the memory updating paradigm, which is a commonly used paradigm to measure working memory capacity. We adapt a mutual interference model from Oberauer and Kliegl (2006) to explain responses. Oberauer and Kliegl (2006) used a Boltzmann equation framework based on the activation levels of items stored in working memory to quantify the probability of correct response at the final recall step after memory updating. We expand the original framework with a Markov chain structure, so that the model accounts for the probabilities of all possible responses, correct or incorrect, at both the intermediate steps during memory updating and the final recall step after memory updating. We use a Wald diffusion process to characterize RT, where the drift rate parameters are associated with the activation levels of items in working memory. This model allows us to investigate the mechanisms underlying choices and RTs in the memory updating paradigm under a joint theoretical framework. A simulation study shows the effectiveness of this model, and posterior predictive distributions and out-of-sample validations show that this model gives a good account of empirical working memory updating findings. We apply the model to two published data sets. The first data set, from Oberauer and Kliegl (2001), examined age differences in working memory. Results from our model reveal an increased level of mutual interference, less use of memory trace information, and potentially less pre-activation of memorized items in older adults compared to younger adults. The second data set, from De Simoni and von Bastian (2018), investigated transfer effects of working memory training. Results from our model reveal a potential transfer effect in the speed of information accumulation, where training in one working memory task may improve the information processing speed in another.