中華心理學刊 民99，52卷，1期，95-112
Chinese Journal of Psychology 2010, Vol.52, No.1, 95-112

Yen-Ho Chen（Department of Psychology, National Taiwan University）；Yung-Fong Hsu（Department of Psychology, National Taiwan University）

Abstract

In psychophysical research of detection and discrimination, non-parametric adaptive methods, including the　fixed and non-fixed step-size methods, have been used extensively for threshold estimation through a combination of decreasing and increasing stimulus steps (see Leek, 2001; Treutwein, 1995, for reviews). In recent years, researchers have focused on the asymptotic and small-sample properties of some of the methods via computer simulations (e.g., Faes et al., 2007; Garcia-Perez, 1998, 2001). In this article we systematically investigate via simulations the large- and small-sample properties of some of the non-parametric adaptive methods in the yes-no detection task. The convergence for different starting values, relative step sizes, and response criteria is systematically investigated. The results show that in both the large- and small-sample conditions, the accelerated stochastic approximation (ASA) (Kesten, 1958) performs well with medium to large (initial) step size. A fixed-step-size method called the biased coin design (BCD) (Durham & Flournoy, 1993, 1995) with small step size is also recommended. Furthermore, our simulation results show that for small sample size, it is also feasible to apply ASA first, then followed by BCD in the sequence as the combined method.

Keywords: biased coin design, non-parametric adaptive methods, stochastic approximation, threshold estimation, transformed up-down