Abstract
A large body of work has used the change detection task to measure and understand the nature of visual working memory capacity. Across two experiments, we examine whether the nature of the latent memory signals used to perform change detection tasks are continuous vs. all-or-none, and consider the implications for proper measurement of performance in change detection. We compare a theory that views these signals as continuous in strength—signal detection theory—with a threshold-based theory (underlying "K" values) that views working memory as being capable of holding a fixed number of items that are either remembered or completely forgotten. In Experiment 1, we measure confidence and find evidence that visual working memory is continuous in strength, with strong support for equal variance signal detection models. In our preregistered and high powered Experiment 2, we tested a critical implication of this result without relying on model comparison or confidence reports. Participants with a conservative response bias were encouraged to be more neutral in their responses (i.e., respond “same” closer to 50%) and compared to a standard change detection task we found strong evidence that performance improved (by roughly 30%) when measured by K; despite no change in the underlying memory signals (Bayes factor of 24 to 1). By contrast, evidence favored the idea that d’ is fixed across these same instructional changes, demonstrating that it correctly separates response criterion from memory performance. Overall, our data suggests working memories are best thought of as continuous in strength and best analyzed in terms of signal detection measures. This calls into question a large body of work using threshold models, like K, to analyze change detection data since we show that this metric confounds response criteria with memory performance in standard change detection tasks.