文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 自然科学 > 数学 > BayesSDT Software for Bayesian inference with signal detection theory
侵权投诉

BayesSDT Software for Bayesian inference with signal detection theory

BAYESIAN SDT 451

Table 1

Signal Detection Theory Terminology

Response“Yes”“No”

Signal TrialHitMiss

Noise TrialFalse alarm

Correct rejection

hit and false alarm rates and thus maps onto the counts in Table 1. In Figure 1, the hit rate, h, is shown as the propor-tion of the signal distribution above the criterion k. Simi-larly, the false alarm rate, f, is the proportion of the noise cation in psychology—for instance, as models of human cognition (e.g., Anderson, 1991; Griffiths & Tenenbaum, 2005; Kemp, Bernstein, & Tenenbaum, 2005; Sanborn, Griffiths, & Navarro, 2006; Tenenbaum & Griffiths, 2001) and as methods for relating psychological models to data (e.g., Kuss, Jäkel, & Wichmann, 2005; Lee, 2008; Lee & Webb, 2005; Navarro, Griffiths, Steyvers, & Lee, 2006; Pitt, Myung, & Zhang, 2002; Rouder, Lu, Speckman, Sun, & Jiang, 2005).

In the context of SDT, Bayesian advantages include avoiding the need for edge corrections that can be a prob-distribution above the criterion k. The usefulness of SDT is that, through this relationship, it is possible to take the sort of data in Table 1 and convert the counts of hits and false alarms into psychologically meaningful measures of discriminability and bias. Discriminability is a measure of how easily signal and noise trials can be distinguished. Bias is a measure of how the decision-making criterion being used relates to the balanced criterion.

A Graphical Model for Bayesian Inference

Bayesian inference. The Bayesian approach to statisti-cal inference (see, e.g., Gelman, Carlin, Stern, & Rubin, 2004; Jaynes, 2003; Lee & Wagenmakers, 2005) has a number of desirable properties in inferring the parameters of SDT models from data. These advantages stem from the basic Bayesian philosophy of always representing what is known and unknown about parameters of interest using probability distributions. The distributions provide com-plete representations of uncertainty and can be updated in a coherent and principled way using probability theory. Bayesian methods have recently found widespread appli-

x

BayesSDT Software for Bayesian inference with signal detection theory

y

tisneD y)

tiyl/ixb(agbol or P y

d /2

kd

Strength

Figure 1. Equal-variance Gaussian signal detection theory framework.

lem in standard analyses (see Snodgrass & Corwin, 1988),1 and posterior distributions are not constrained to take par-ticular parametric forms; thus, they better represent param-eter values—like very high hit or false alarm rates—that are near the boundary of the theoretically possible range. Bayesian methods are also automatically exact for any sam-ple size and are sensitive to how many data are available in making inferences. Additional potential advantages of Bayesian methods for SDT, which go beyond the function-ality of the software presented in the present article, include hierarchical extensions to model individual differences in participants and in items (e.g., Lee & Paradowski, 2007; Rouder & Lu, 2005) and the straightforward consideration of alternative unequal variance and other extended SDT models without requiring analytic results for inference.Graphical model. Graphical models provide a convenient formalism for expressing many Bayesian models (see Griffiths, Kemp, & Tenenbaum, in press; Jordan, 2004; Lee, 2008, for two psychological intro-ductions and one statistical introduction). In graphical modeling, a graph is created in which nodes represent parameters and data. The structure of the graph then in-dicates dependencies between the parameters and data, with children depending on their parents. The practical advantage of graphical models is that sophisticated and relatively general-purpose Markov chain Monte Carlo (MCMC) algorithms exist that can sample from the full joint posterior distribution of the parameters conditional on the observed data. Our software relies on WinBUGS (Spiegelhalter, Thomas, & Best, 2004), which uses a range of MCMC computational methods, including adap-tive rejection sampling, slice sampling, and Metropolis–Hastings (see, e.g., Chen, Shao, & Ibrahim, 2000; Gilks, Richardson, & Spiegelhalter, 1996; MacKay, 2003) to perform posterior sampling.

Figure 2 shows the graphical model we use for mak-ing inferences about the SDT parameters in Figure 1. The convention is used that observed variables (i.e., data) have shaded nodes, whereas unobserved variables (i.e., the pa-rameters to be inferred from data) are not shaded. In ad-dition, continuous variables have circular nodes, whereas discrete variables have square nodes. Finally, stochastic variables in our model are shown with single-bordered nodes, whereas deterministic variables (i.e., those that are defined simply as functions of other variables) are shown as double-bordered nodes.

The best way to understand the graphical model is to begin with the discriminability d and bias c parameters.

上一页第2页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email