Abstract
Despite disadvantages such as inefficient use of trials, the method of constant stimuli remains popular in psychophysical measurement. It is simpler to implement, less taxing on observers, and more immune to lapse errors than adaptive techniques such as staircases or QUEST. The challenge then is to find an optimal set of stimulus levels for each subject in order to estimate a full psychometric function (PF). Efficiency is important when limited trials are available due to time constraints or subject fatigue. We evaluated a hybrid procedure where the testing levels are specified in terms of the mean and spread of a parameterized PF. With only the minimum and maximum stimulus levels detailed beforehand, the PF parameters are estimated online during the testing procedure and guide stimulus placement. Unlike other adaptive methods that eventually repeat testing near the subject's threshold, stimuli are chosen randomly from the specified points on the PF, mitigating expectation bias and subject fatigue. First, using Monte Carlo simulations, we explored optimal stimulus placement for known PFs. We validated previous literature showing that parameters can be estimated equally well with 2-5 carefully placed stimulus locations, with some variation based on the total number of trials. Approximately 100 trials were required to ensure that 95% of the simulations were within 5% of the veridical parameter values. Surprisingly, when the PF was initially unknown and progressively estimated during a simulated run, 100-150 trials were sufficient to reliably converge to the veridical parameter values. Experiments with human observers, including some with significant variability, confirmed that parameter estimation was less accurate when trials remained near subject thresholds, versus those that spread more broadly across the PF. In summary, this method provides a compromise between the flexibility and optimality of adaptive procedures and the simplicity and robustness of the method of constant stimuli.
Meeting abstract presented at VSS 2014