Please enable JavaScript.
Coggle requires JavaScript to display documents.
SIGNAL DETECTION THEORY - Coggle Diagram
SIGNAL DETECTION THEORY
Introduction
Noises make perception imperfect
Internal/External noises
Problem in real situations
SDT quantify sensitivity e bias in perception
SDT basic concepts
Sensitivity
Capacity to discriminate signal from noise
How improve?
Training
Increasing external signal
Indicate with d
Detectability
Discriminability
Distinguishability
Distance signal/noise
d = 0
Internal Perceptual signal
Two states, four possibile outcomes
Hit: Yes, Signal
Miss: No, Signal
False Allarm: Yes, Noise
Correct Rejection: No, Noise
Expressed in proportion (rates) or percentage
Sensitivity increase hit and correct rejection
Criterion and response bias
Liberal criterion
Lower criterion (accepts weak signals)
Biased towards yes
Increasing hit, reducing correct rejections
Conservative criterion
Higher criterion (accepts strong signals
Biased towards no
Increasing correct rejections, reducing hits
Criterion (c) or bias (β)
Pay-off matrix
Losses (costs)
More perceived than actual
Gains (benefits)
Individual/Society interests
Likelihood of occurence
Trasforms pay-off values in expected values
An ideal observer considers these values
Quantifying sensitivity and bias
1. Noise distribution
Standard Normal distribution
Mean = 0
Standard Deviation (z-score) = distance from mean
2. The noise + signal distribution
Adding noises we create another graphic
This graphic overlap the other graphic
Creation of four areas under the curve
Creation of uncertainty
Increasing strenght of the signal, reducing overlap
Increasing distance signal-noise = increasing sensitivity
Establishing a criterion
Weaker than criterion value
No
Stronger than criterion value
Yes
3. Computing d'
d' = z of p(hits) - z of p(false allarms)
Necessity of using Z tables
4. Computing criterion, c or β
c (criterion)
c = - (z[p(hits)] + z[p(false alarms)]) / 2
c = 0 neutral
c > 0 conservative
rightwards shift
c < 0 liberal
leftwards shift
More simple to calculate
β (bias)
Based on the height of the curve
β = f[z|p(hits)] / f[z|p(false alarms)]
β = 0 neutral
β > 0 conservative
height N+S > height S
β < 0 liberal
height N+S < height S
Comparable with optimal β
5. The ROC curve
Nearest to 0;0 conservative
Nearest to 1;1 liberal
More distance from diagonal, more sensitivity