Signal detection theory Flashcards
Psychophysical experiments were done bc people were interested in determining absolute tresholds of perception. Which types of experiments where there? Give examples in an audio experiment.
- method of limits: beeps up/down from start point until pp responds to hear them.
- method of constant stimuli: beeps are of random frequency
- method of adjustment: pp can adjust a continuous signal themself until they hear it
- adaptive testing: like 1, but then after an approximation is made, only sweeps around the treshold are done to save time.
What are issues with psychophysical threshold experiments? When the experiments were adjusted to incorporate the critique, what astonishing result was found that could not be explained by classical psychophysical theory and thus lead to a paradigm shift?
- in all trials stimulus present, pp could just be saying yes without hearing anything and you’d never know . -> needs trials with stimulus not present
- when they did such ‘signal detection experiments’, it was found sometimes pp reported hearing a signal even though none was there. (false alarm)
- also pp could be hearing something but not resond bc theyre not sure -> misses
What conclusions did the alternative signal detection theory have about the nature of tresholds, after this paradigm shift? What is the main idea of SDT?
- absolute tresholds dont really exist
- detecting a signal depends on other things than just sensitivity, like people’s response criterion (liberal vs conservative strategies), signal intensity, amount of noise, etc.
In which situations can SDT be applied as a model?
- binary decision making
- about the true state of the world
- where there is a certain degree of uncertainty
How can you use SDT to investigate the true sensitivity (“threshold” / perceptual process) of a participant (or a machine’s ‘statistical prediction rule’)? Which graphical presentation is used for it? What measures do you need at LEAST?
- ROC (Receiver Operating Characteristic) curve for pp is constructed by manipulating the payoffs/costs associated with each trial type (hit, miss, FA, CR). this will give different points (criterions) upon the curve
- x axis is P(false alarm) and y axis is P (hit), you need at least these, the other 2 are complementary.
- amount of ‘curve’, d’ , is a measure of sensitivity
How can you use SDT to investigate the decisional process (criterion/response bias) of a particpant? Which graphical presentation is used for it?
- two Normal curves, noise and signal + noise
- x-axis Intensity of perceptual effect (evidence)
- y axis probability
- Xc (criterion) value depending on liberal (more to left)/neutral/conservative, separating yes from no responses (right is yes), giving P of hit/P of FA ratio.
- C is measured relative to neutral point, in standard units Z from this point (pos when right, = conservative)
What is the discriminability index d’? How is it calculated? How to influence it? What is so great about it?
d’ = separation (m2 - m1) / spread ( standard deviation)
- separation is increased with signal strength and decreased if the amount of noise is high.
- spread
- can change depending on: stim I./salience, giving memory aids of what signal looks like/training
- is independent from a persons criterion
What is B (beta)?
- similar to C (criterion/bias).
- is equal to (affects) = Ratio of P ( of this X if signal) to P (of this X if noise). These Ps are the value on the Y axis in the curves.
- is also EQUAL TANGENT SLOPE ROC
- is 1 at neutral (intersection), >1 if conservative (to right)
- is an implicit rule set by observer, which can be calculated from data
How to construct optimal B:
- in case of no differential payoffs to the 4 outcomes AND 50/50 signal probability?
- in case of different signal probabilities?
- in case of differential payoffs?
- B = 1 (minimizes errors)
- B = P (N) / P(s) (likelihood of a noise vs a signal trial)
- B = profit maximizing: ( P(N) / P (S) ) * ( v(CR) + c(FA) ) / ( v(HI) + c(MI) )
When situations change, B adjustment by humans is sluggish. Why?
People cannot judge probability well (but are better in tracking payoffs)
Which assumptions about the N and S + N curves are made?
- Internal noise Normal Gaussian distr
- Same noise distr. for N and for N + S trials
- noise variance is independent from the (strength of the) signal (i.e., fixed).
- S + N dont interact, but instead adds up linearly (shifts rightward) from the N curve
- decision is based on 2 consecutive processing steps (perception + decision)
What are 2 ways in which measures of sensitivity can be made when only few data points/limited time are available?
- 1 data point on ROC: calculate area p(a) = (p(h) + (1 - p(fa))) / 2 (triangle)
- use confidence ratings (sure signal, uncertain, sure no signal) and calculate two betas, one liberal beta by which uncertain trials count towards a signal trial and a conservative beta where uncertain trials count towards noise trial.
What are 3 applications of SDT?
- comparing different decision making strategies (e.g. based on CT scan only, or based on a clinical interview) on sensitivity (which is better?)
- determine optimal decision criterion based on probabilities of hits/misses etc. for the context
- researching whether interventions or conditions have different sensitivity or lead to a shift criterion
What does the AREA under the curve represent when it is:
- left of C, of the N + S curve or N curve res.p
- right of C, of the N + S curve or N curve resp.
S+N: left of C = p(miss), right of C = p(hit)
N: left of C = p(correct rejection), right of C = p(false alarm)
Riecke study: early studies found certain parameters influence whether or not the continuity illusion is heard. What effect do soft vs loud noises have?
soft noises -> high SNR -> continuity illusion not heard
loud noises -> low SNR -> continuity illusion heard
