334
PE. Nachtigall et al.
report "signal present" only when the sensation is particularly intense, and
others may be biased to report its presence at lower magnitudes. This difference could imply that the sensitivities of the two individuals are different, but changing the size or amount of reward given for signal-present
responses versus those for signal-absent responses can dramatically change
the individual from extreme conservatism to extreme liberalism (e.g.,
Schusterman et al. 1975; Schusterman 1980). Similarly, changing the proportion of trials containing a signal for which the signal-present response
is correct and rewarded can produce analogous changes in the likelihood
that the subject will make a signal-present response. Increasing the proportion of signal-present trials increases the likelihood that the subject will
respond "signal present," even when it is not actually present. Traditional
threshold theory is inconsistent with this finding because it predicts that
signals below the threshold are too weak to result in a sensation and so it
is difficult to see how responses to them could depend on changes in factors
separate from sensation.
Signal detection theory provides a means to measure the effects of such
biasing factors as signal probability and payoff. It requires the investigator
to keep track of both the correct responses when a signal is actually present
(hits) and the incorrect responses when the signal is absent (false alarms).
More liberal subjects will not only generate more hits, they will also generate more false alarms. Less liberal subjects will generate fewer hits and
fewer false alarms. Hence, a given level of sensitivity can generate either
more or fewer correct positive responses. Signal detection theory represents
a perceiver by two parameters, one for the individual's sensitivity, that is,
its ability to detect the signal, and another for bias, that is, the perceiver's
tendency to say signal-present or signal-absent.
There are several versions of signal detection analyses that differ in their
assumptions about the underlying distribution of sensory effects. The simplest version to explain assumes that the signal-present and signal-absent
trials each result in a perceived sensation that is normally distributed about
a mean corresponding to one of the two trial types. On signal-absent trials
(i.e., on trials on which the signal has actually not been presented) the sensation is a sample from the so-called noise-only distribution. On signalpresent trials (i.e., on trials on which the signal has actually been presented)
the sensation is a sample from the so-called signal-plus-noise distribution.
The signal-plus-noise distribution is assumed to be identical to the noiseonly distribution except for its greater mean. The separation between the
two distributions depends on the strength of the signal. By definition, the
detector knows only the sensation level resulting from this sample, not
the identity of the distribution that produced it. Because these two distributions overlap, there is some uncertainty about the specific distribution
from which the sample came. If the two distributions overlap completely,
no systematic discrimination is possible. The further they are separated, the
more accurate can be the discrimination between signal-present and signal-
PE. Nachtigall et al.
report "signal present" only when the sensation is particularly intense, and
others may be biased to report its presence at lower magnitudes. This difference could imply that the sensitivities of the two individuals are different, but changing the size or amount of reward given for signal-present
responses versus those for signal-absent responses can dramatically change
the individual from extreme conservatism to extreme liberalism (e.g.,
Schusterman et al. 1975; Schusterman 1980). Similarly, changing the proportion of trials containing a signal for which the signal-present response
is correct and rewarded can produce analogous changes in the likelihood
that the subject will make a signal-present response. Increasing the proportion of signal-present trials increases the likelihood that the subject will
respond "signal present," even when it is not actually present. Traditional
threshold theory is inconsistent with this finding because it predicts that
signals below the threshold are too weak to result in a sensation and so it
is difficult to see how responses to them could depend on changes in factors
separate from sensation.
Signal detection theory provides a means to measure the effects of such
biasing factors as signal probability and payoff. It requires the investigator
to keep track of both the correct responses when a signal is actually present
(hits) and the incorrect responses when the signal is absent (false alarms).
More liberal subjects will not only generate more hits, they will also generate more false alarms. Less liberal subjects will generate fewer hits and
fewer false alarms. Hence, a given level of sensitivity can generate either
more or fewer correct positive responses. Signal detection theory represents
a perceiver by two parameters, one for the individual's sensitivity, that is,
its ability to detect the signal, and another for bias, that is, the perceiver's
tendency to say signal-present or signal-absent.
There are several versions of signal detection analyses that differ in their
assumptions about the underlying distribution of sensory effects. The simplest version to explain assumes that the signal-present and signal-absent
trials each result in a perceived sensation that is normally distributed about
a mean corresponding to one of the two trial types. On signal-absent trials
(i.e., on trials on which the signal has actually not been presented) the sensation is a sample from the so-called noise-only distribution. On signalpresent trials (i.e., on trials on which the signal has actually been presented)
the sensation is a sample from the so-called signal-plus-noise distribution.
The signal-plus-noise distribution is assumed to be identical to the noiseonly distribution except for its greater mean. The separation between the
two distributions depends on the strength of the signal. By definition, the
detector knows only the sensation level resulting from this sample, not
the identity of the distribution that produced it. Because these two distributions overlap, there is some uncertainty about the specific distribution
from which the sample came. If the two distributions overlap completely,
no systematic discrimination is possible. The further they are separated, the
more accurate can be the discrimination between signal-present and signal-