Signal detection

Background
Much of cognitive psychology involves gathering data from experimental participants.
Gathering good data is not always easy, especially when an experiment uses a variety of
people as participants. Researchers must carefully design an experiment to be certain that
participants are following the instructions and are motivated to try their best. Even despite
these efforts, experimental results can be contaminated by individual differences if the
researcher does not properly analyze the data.

For example, consider two participants in visual detection of a faint target. The researcher
wants to explore a property of the visual system, so he/she presents a visual stimulus and asks
the participants to report whether they saw the target. After 50 trials, Participant A reports
seeing the target 25 times and Participant B reports detection 17 times. Did Participant A do
better? Not necessarily, perhaps Participant A is simply more prone to report seeing the target
whereas Participant B is more conservative and requires more evidence before reporting a
target. That is, the two participants may have equivalent visual systems, but differences in
their criterion for reporting. Reports of simple detection do not allow the researcher to
compare participants' results.

A better experiment is a modification of the one above and has two kinds of trials, one with
the target present and one with the target absent. Again, the participants report whether they
saw the target. There are four statistics to be calculated from this experiment. (1) A hit is
when the participant correctly detects the target. (2) A miss is when the target was there but
the participant did not detect it. (3) A false alarm is when the participant reports seeing the
target when it was not actually there. (4) A correct rejection is when the participant correctly
reports that the target was not present.

Stimulus Present Stimulus Absent

(Signal + Noise) (Noise Only)
Respond
Hit False Alarm
Present
Respond
Miss Correct Rejection
Absent

Suppose that after 100 trials (50 for target present and 50 for target absent) the researcher
again finds that on the trials in which the target was in fact present, Participant A reports
seeing it 25 times and Participant B 17 times. Who is doing better? It depends on the
frequency of false alarms. If Participant A has 25 false alarms and Participant B has 5 false
alarms, then B is better than A at distinguishing the trials in which the target is present from
the trials in which the target is absent. That is, in this case, A tends to guess that the target is
there, but she/he is wrong (a false alarm) as often as she/eh is correct (a hit). B is more
selective about saying he/she detects the target, but rarely says the target is there when it is
not. Thus, B is doing better.

This type of analysis suggests that you need to consider two numbers, hits and false alarms,
to really be able to compare performance across participants. Fortunately, you can combine
the numbers in a careful way to produce a single number that gives an indication of the
sensitivity of the participant to the presence of the target. The calculation is structured so that,
with certain assumptions, it will not matter whether a participant takes a conservative or
liberal approach to claiming to detect the target. The most common measure of sensitivity is
called d' (d-prime), and a common measure of bias (whether the person took a conservative
or liberal approach) is called C.

A discussion of the algorithms for calculating sensitivity is beyond the scope of this
experiment (see Macmillan & Creelman, 1991, for further discussion).

This lab provides an experiment that measures sensitivity (d') and bias (C).

Instructions
If you have logged in, you'll see a black rectangle below. Make sure that you can see the full
area before you begin the lab.

A group of randomly placed dots (sort of like a "star field") will appear. The number of
random dots varies from trial to trial. Also, on some trials (target present) an additional set of
ten dots arranged in a straight line that slants downward from left to right is randomly placed
among the dot field. On the other trials (target absent), the line is not included. Your task is to
report whether the target — the set of ten dots arranged in a line — is present or absent.
There are 60 trials.

At the end of the experiment, you will be asked if you want to save your data to a set of
global data. After you answer the question, a new Web page window will appear that
includes a debriefing, your data, your group's data, and the global data.

Tablet Specific Details

If you are using a tablet, tap the Start Next Trial button to start a trial. If you think the target
is present, tap the Present button. If you think the target is absent, tap the Absent button.

Computer Specific Details

If you are using a computer, click the Start Next Trial button to start a trial. If you think the
target is present, click the Present button. If you think the target is absent, click the Absent
button.

Lab
What methods did we employ in this experiment?

One each trial, a bunch of randomly placed noise dots (sort of like a "star field") were shown.
The number of random dots varied from trial to trial. Also, on half of the trials (target
present), an additional set of ten dots – arranged in a straight line that slants downward from
left to right – was randomly placed among the dot field. On the other trials (target absent), the
line was not included. Your task was to report whether the target is present or absent.
The independent variables in this experiment were the number of noise dots in the star field
and the presence or absence of the target dots. The dependent variable was whether your
response was correct or incorrect.

What do we predict participants will do? Why?

The table below lists several detection statistics for different numbers of noise dots. Hits are
the proportion of trials in which the target was present and it was reported as present. False
alarms are the proportion of trials in which the target was absent, but it was reported as
present. Correct rejections are the proportion of trials in which the target was absent and it
was reported as absent. Misses are the proportion of trials in which the target was present, but
it was reported as absent.

The most important statistic is d'. It gives a measure of your ability to discriminate displays in
which the target was present from displays in which the target was absent. You should find
that d' decreases as the number of noise dots increases. The statistic C refers to a bias to say
the target was present or absent. A negative value of C indicates a liberal bias whereas a
positive value of C indicates a conservative bias.

How robust is this effect? Are there limits to this effect?

Detecting a signal among noise is common to many tasks in psychology. In addition to

demonstrating the effect of noise on your ability to detect the target, this lab shows how you
need to be careful about measuring human performance. The signal detection analysis
provides a powerful method of exploring the ability of people to distinguish a signal from
noise.

Student
Information

Data Summary for PSY4081-1653

2022-03-20 [Link] Australian Eastern Daylight Time

Date Completed
(Victoria)

Time Spent On Lab 17 minutes

Detection Measures
Number of d' C Hits False Correct Misses
Dots Alarms Rejections

144 6.626 -0.987 1.000 0.000 1.000 0.000

400 5.142 -1.729 1.000 0.200 0.800 0.000

1000 1.028 -0.767 0.900 0.600 0.400 0.100

Trial Number Trial Type Num. Dots Response Accuracy

1 Signal + noise 1000 Absent 0
2 Noise only 1000 Absent 1
3 Noise only 400 Absent 1
4 Signal + noise 144 Present 1
5 Signal + noise 1000 Present 1
6 Noise only 1000 Present 0
7 Noise only 144 Absent 1
8 Noise only 144 Absent 1
9 Signal + noise 400 Present 1
10 Signal + noise 400 Present 1
11 Noise only 400 Absent 1
12 Signal + noise 144 Present 1
13 Noise only 400 Absent 1
14 Noise only 1000 Absent 1
15 Noise only 144 Absent 1
16 Signal + noise 144 Present 1
17 Signal + noise 1000 Present 1
18 Noise only 1000 Absent 1
19 Noise only 400 Absent 1
20 Signal + noise 144 Present 1
21 Noise only 144 Absent 1
22 Signal + noise 144 Present 1
23 Noise only 1000 Present 0
24 Noise only 144 Absent 1
25 Signal + noise 144 Present 1
26 Signal + noise 400 Present 1
27 Signal + noise 400 Present 1
28 Signal + noise 1000 Present 1
29 Signal + noise 144 Present 1
30 Noise only 400 Absent 1
31 Noise only 1000 Present 0
32 Signal + noise 1000 Present 1
33 Signal + noise 1000 Present 1
34 Noise only 1000 Absent 1
35 Noise only 144 Absent 1
36 Noise only 400 Absent 1
37 Signal + noise 400 Present 1
38 Signal + noise 400 Present 1
39 Noise only 400 Present 0
40 Noise only 400 Absent 1
41 Signal + noise 400 Present 1
42 Signal + noise 1000 Present 1
43 Signal + noise 400 Present 1
44 Noise only 144 Absent 1
45 Signal + noise 1000 Present 1
46 Signal + noise 1000 Present 1
47 Noise only 400 Absent 1
48 Noise only 144 Absent 1
49 Noise only 1000 Present 0
50 Signal + noise 400 Present 1
51 Signal + noise 144 Present 1
52 Signal + noise 1000 Present 1
53 Noise only 400 Present 0
54 Noise only 144 Absent 1
 55 Noise only 144 Absent 1
56 Noise only 1000 Present 0
57 Noise only 1000 Present 0
58 Signal + noise 144 Present 1
59 Signal + noise 144 Present 1
60 Signal + noise 400 Present 1

Group data

Data for group PSY4081

Data as of 2022-03-20 [Link] Australian Eastern Daylight Time (Victoria)

The table below lists several detection statistics for different numbers of noise dots. The most
important statistic is d'. It gives a measure of your ability to discriminate displays in which
the target was present from displays in which the target was absent. You should find that d'
decreases as the number of noise dots increases. It also reports C, a measure of bias (negative
= liberal; positive = conservative).

Means based on data from 1269 participants.

Number of d' C Hits False Correct Misses

Dots alarms rejections

144 dots 6.103 -1.136 0.996 0.050 0.950 0.004

400 dots 4.960 -1.052 0.972 0.127 0.873 0.028

1000 dots 1.164 0.216 0.620 0.265 0.735 0.380

Standard deviations based on data from 1269 participants.

Number of d' C Hits False Correct Misses

Dots alarms rejections

144 dots 0.884 0.400 0.023 0.087 0.087 0.023

400 dots 1.582 0.745 0.053 0.141 0.141 0.053

1000 dots 0.980 0.495 0.167 0.193 0.193 0.167

Global Data

Global Data

Data as of 2022-03-20 [Link] Australian Eastern Daylight Time (Victoria)

The table below lists several detection statistics for different numbers of noise dots. The most
important statistic is d'. It gives a measure of your ability to discriminate displays in which
the target was present from displays in which the target was absent. You should find that d'
decreases as the number of noise dots increases. It also reports C, a measure of bias (negative
= liberal; positive = conservative).

Means based on data from 46785 participants.

Number of d' C Hits False Correct Misses

Dots alarms rejections

144 dots 5.811 -1.147 0.990 0.074 0.926 0.010

400 dots 4.543 -1.057 0.960 0.167 0.833 0.040

1000 dots 0.927 0.095 0.614 0.325 0.675 0.386

Standard deviations based on data from 46785 participants.

Number of d' C Hits False Correct Misses

Dots alarms rejections

144 dots 1.188 0.513 0.045 0.114 0.114 0.045

400 dots 1.732 0.789 0.071 0.160 0.160 0.071

1000 dots 0.938 0.589 0.180 0.208 0.208 0.180