A maximum-likelihood method for estimating thresholds in a yes-no task.

Abstract

A maximum-likelihood procedure for estimating threshold values in a yes-no task is presented. In computer simulations of this procedure, it is demonstrated that the variability of the threshold estimates is little affected by the density of the hypotheses tested for a fixed range, or by serious misestimates of the slope of the psychometric functions. The threshold value is also largely independent of the starting value of the signal. The standard deviation of the threshold estimates appears to decrease with the square root of the number of trials, with a 2- to 3-dB standard deviation possible if only 12 trials are used in the threshold estimates. Data are presented using human listeners tested on 5 days. Two threshold estimates, based on 12 trials, were made at each of the six audiometric frequencies on each day. The mean data appear sensible, and the standard deviation of the measured thresholds is about 3 dB. Using this procedure, it takes less than 3 min to measure the audiogram for a single ear.