Inference for psychometric functions in the presence of nonstationary behavior

Measuring sensitivity is at the heart of psychophysics. Often, sensitivity is derived from estimates of the psychometric function. This function relates response probability to stimulus intensity. In estimating these response probabilities, most studies assume stationary observers: Responses are expected to be dependent only on the intensity of a presented stimulus and not on other factors such as stimulus sequence, duration of the experiment, or the responses on previous trials. Unfortunately, a number of factors such as learning, fatigue, or fluctuations in attention and motivation will typically result in violations of this assumption. The severity of these violations is yet unknown. We use Monte Carlo simulations to show that violations of these assumptions can result in underestimation of confidence intervals for parameters of the psychometric function. Even worse, collecting more trials does not eliminate this misestimation of confidence intervals. We present a simple adjustment of the confidence intervals that corrects for the underestimation almost independently of the number of trials and the particular type of violation.