452 LEE
k 1
2d` cB d`c
d` Gaussian(0, 1/2)c Gaussian(0, 2)
h &¤¥1³¦2d` c´
µf &¤¥¦ 12d` c³´
µ
H Binomial(h, S)F Binomial(f, N)
Figure 2. Graphical model for inferring discriminability (d ), criterion (k), two measures of bias (c and ;), and hit (h) and false alarm (f) rates from H observed hits out of S signal trials and F observed false alarms out of N noise trials.
In order to relate these parameters to the observed data, d and c can be reparameterized (under the model given by SDT) into hit and false alarm rates h and f. This means that there is a deterministic relationship between h and f on the one hand, and between d and c on the other, given by
h & 12d` c
(1)
f & 2d` c
(2)
where (·) is the standard Gaussian cumulative density function.
The number of signal trials, S, and the number of noise trials, N, are observed data, as are the hit count H and the false alarm count F. The hit and false alarm counts follow a binomial distribution depending on both the hit and false alarm rates, and the number of signal and noise trials, so that H Binomial(S, h) (3)and
F Binomial(N, f). (4)
In this way, SDT relates discriminability d and bias c pa-rameters to the observed data so that posterior inferences
can be drawn. From the posterior distributions of d and c, the posterior distributions for the criterion k and alter-native bias measure ; can be calculated according to the standard results
k d`2
c (5) ; cd . (6)

The final component of the Bayesian graphical model involves prior assumptions. We choose priors on discrim-inability and bias that correspond to the assumption of uniform priors for the hit and false alarm rates, consistent with standard practice. It can be proven2 that the required distributions are d Gaussian(0,2)
(7)
and
c Gaussian(0,1 2 ). (8)
WinBUGS implementation. The graphical model
shown in Figure 2 is straightforward to implement in WinBUGS, using the script in Table 2. The required in-puts are the hit, false alarm, miss, and correct-rejection counts, and the outputs are samples from the posterior distribution of the d , h, f, k, c, and ; parameters of the SDT model.3
The BayesSDT Software Package
The BayesSDT software package consists of a num-ber of MATLAB scripts and functions, as well as Win-BUGS scripts. It can be downloaded from doc.guandang.net .uci.edu/~mdlee/. BayesSDT requires that MATLAB and WinBUGS 1.4.2 or later be installed. WinBUGS is available from doc.guandang.net. Bayes-SDT also requires the MatBugs MATLAB function avail-able at doc.guandang.net matbugs.html, which allows MATLAB and WinBUGS to pass information to each other.
The BayesSDT software provides two MATLAB in-terfaces with the WinBUGS implementation of the SDT graphical model. The easiest to use is a graphical user in-terface (GUI), but a MATLAB function call is also pro-vided to allow more than four data sets to be analyzed simultaneously.
Graphical user interface. The BayesSDT GUI is started by running the BayesSDT_GUI MATLAB script. A demonstration of the interface is displayed in Figure 3, showing the entry of three illustrative data sets for anal-ysis. The first is the “many” data set, involving 70 hit counts and 50 false alarm counts from 100 signal and 100 noise trials. The second is the “few” data set, involving 7 hit counts and 5 false alarm counts from 10 signal and 10 noise trials. The idea is that the many data set has the same rate of hits and false alarms as does the few data set, to illustrate how Bayesian analysis represents the decrease in uncertainty with additional data. The third, “perfect” data set involves 10 hits and 0 false alarms from 10 signal and 10 noise trials. The purpose of this data set is to illus-trate how Bayesian analysis deals with edge effects.
As Figure 3 shows, the GUI allows a label to be entered for each data set. The color, width, and style properties of the lines used to draw the posterior distribution for each data set can also be specified using standard MATLAB options. The help plot command in MATLAB pro-vides a full list. Checkboxes allow the selection of the SDT parameters to be analyzed. Finally, the number of poste-rior samples to be generated—as well as the number of bins used to display the final posterior densities—can be
Bayesian Inference for Econometric Models using Empirical Likelihood Functions: Extended Abstract Hyungsik Roger Moon and Frank Schorfheide Introduction Estimators ba...
Bayesian analysis of aut... 暂无评价 12页 免费 Bayesian Detection of In....Semiparametric bayesian inference for time series with mixed spectra A Bayesian...
免费 BayesSDT Software for Ba... 暂无评价 7页 免费喜欢此文档的还喜欢 Asymptotically...Higher Accuracy for Bayesian and Frequentist Inference Large Sample The...
Bayesian Inference for Spiking Neuron Models with a Sparsity Prior Generalized linear models are the most commonly used tools to describe the stimulus selectiv...
BayesGCM Software for Ba... 暂无评价 10页 免费 Bayesian Inference for C....Despite their commonality and well-developed underlying theory, there appear ...
1 Likelihood ratios and Bayesian inference for Poisson...[0,T ] and ρ denotes the “signal to noise...the MongeKantorovitch measure transportation theory. ...
Simultaneous PIXE and RBS data analysis using Bayesian Inference with the DataFurnace...backscattering, elastic recoil detection, and non-resonant nuclear reaction...
Bayesian Methods for Multimedia Signal Processing [...Moreover, software tools for automated code ...for Bayesian inference, namely Markov Chain Monte ...
Theory Change and Bayesian Statistical Inference_专业资料。This paper addresses...BayesSDT Software for ... 暂无评价 7页 免费 A Challenge for Statis... ...
STATISTICAL INFERENCE,... 暂无评价 44页 免费 BayesGCM Software for ... 暂无评价 10页 免费 Approximate Bayesian I... 27页 免费 Bayesian Inference wit.....
热门文档