Laboratory studies using signal detection paradigms produce bias-free estimates of listener sensitivity. Forced-choice procedures, trial-by-trial feedback, equal a priori probability of signal occurrence, and a bal- anced pay-off matrix assess sensitivity with response bias forced to zero. That is well-suited to the evaluation of computational models; however, real-world listening tasks don’t fit traditional paradigms. Real-world listening situations do not lend themselves to trial-by-trial feedback. Davis (2015) investigated the effects of incomplete feedback on response bias for a simple tone-in-noise experiment. Feedback ranged from no feedback to eight conditions with feedback for some Signal – Response combinations, to full feedback. Davis provided a descriptive analysis for individual subjects. Liu (2020) used Bayesian modeling to account for participant variability in Davis’s results. Liu reported that complete feedback drives the response criteria toward the optimum, and incomplete feedback conditions result in various degrees of deviation from the optimal criterion. Plotting Liu’s results as a Z-score ROC highlights the importance of providing feedback for Hits. Feedback for Signal presentation trials, for YES response trials, or for trials that result in Correct Responses led to minimal bias in the Davis study. These results have implications for the design of laboratory simulations of realistic listening tasks.

This content is only available via PDF.