Distinguishing Evidence Accumulation from Response Bias in Categorical Decision-Making

1
Distinguishing Evidence Accumulation from
Response Bias in Categorical Decision-Making
Vincent P. Ferrera1,2, Jack Grinband1,2, Quan
Xiao1,2, Joy Hirsch2, Roger Ratcliff3 (1) Mahoney
Center for Brain and Behavior Research, (2)
Center for Neurobiology and Behavior Columbia
University, (3) Department of Psychology, Ohio
State University
548.26
Task Performance Trials were sorted according to
stimulus speed and cue color. For each
condition, performance was quantified as the
percentage of trials on which the subject
categorized the stimulus as fast. Reaction
time was calculated as the interval between
motion stimulus onset and the subjects
response.
Discussion For the variable-criterion speed
categorization task, the model attributed the
subjects performance to variation in sensory
evidence rather than response bias. Drift rate
was found to vary almost linearly as a function
of the relative distance of the stimulus from the
category boundary. The same stimulus speed can
give rise to different drift rates depending on
the criterion. This indicates that the drift
rate is not a function of stimulus properties per
se, but reflects the relationship between the
stimulus and the category boundary. We have
previously found (1) that categorization tasks
such as that presented here are relatively
insensitive to the effects of stimulus
probability. Stimulus probability effects can be
quite large in other tasks and there is some
evidence that these effects are better modeled as
changes in response bias rather than changes in
drift rate (2). These results suggest that
perceptual categorization is based on a
representation that encodes the relationship of
the stimulus to the category boundary. They also
suggest that the process of categorization can be
separated from response selection at the
behavioral level. These studies provide a basis
for identifying neural activity related to
categorical decision-making using single cell
recordings or functional imaging.
Drift Rate Model parameters for drift rate and
response bias were estimated by fitting the
proportion of fast responses and reaction time
distributions for correct and error trials for
each condition (stimulus speed x cue). The
fitting process determines, for each conditions,
the starting point of the diffusion process
(bias) and drift rate. Variations in response
bias tend to produce large changes in the leading
edge of the reaction time distribution, whereas
variations in drift rate have a much smaller
effect on the leading edge.
Introduction Categorical decision-making is a
critical aspect of sensory-motor behavior. Yet
little is known about how sensory representations
are transformed into categorical or
decision-based representations. To investigate
this process, subjects performed a speed
categorization task with a randomly varying
category boundary separating slow and fast.
We used a diffusion model to fit each subjects
performance (accuracy and reaction times). The
model can help determine if subjects reacted to
the shifting category boundary by biasing their
responses, or, alternatively, by adjusting their
internal decision criterion.
Speed Categorization Task Human and monkey
subjects performed a task in which they judged
the speed of a moving random-dot pattern as fast
or slow relative to a variable criterion speed.
The criterion speed changed randomly from trial
to trial and was indicated by a cue presented at
the beginning of each trial. The stimulus
probabilities were adjusted so that for each
criterion the correct response was just as likely
to be fast as slow. Humans responded with a
button press, while monkeys used a touch panel.
Auditory feedback was provided at the end of each
trial to indicate correct or incorrect responses.

Figure 2. (Left) Human psychometric functions
averaged across observers (n8). (Middle) Human
reaction times (Right) Monkey psychometric
functions.
Diffusion Model Diffusion models treat a decision
process as the accumulation of evidence toward a
threshold. These models attribute differences in
performance across task conditions to changes in
either the rate of evidence accumulation (drift
rate) or a bias toward one or the other response.
Figure 4. Drift rates determined by fitting
diffusion model to data in Fig. 2.
Bias The bias in the starting point of the
diffusion process was estimated for each cue
(speed criterion). The bias was small and was in
the same direction for both criteria.

• References
• Grinband J, Hirsch J, Ferrera VP. A neural
  representation of categorization uncertainty in
  the human brain. Neuron. 2006 249(5)757-63.
• 2. Ratcliff R, Van Zandt T, McKoon G.
  Connectionist and diffusion models of reaction
  time. Psychol Rev. 1999 106(2)261-300.

Acknowledgements Supported by NIH-MH59244 (VPF),
R37-MH44640 (RR), T32-EY013933, T32-MH15174
Figure 5. Response biases determined by fitting
diffusion model. X-axis is bias for trials with
the slower criterion speed, y-axis is bias for
the faster criterion.
Figure 3. (Top) Accumulated evidence as a
function of time for single trial. (Bottom)
Diffusion model fits shape of psychometric
function and reaction time distribution.
Figure 1. (Top) Illustration of trial events
during task. (Bottom) The full range of stimulus
speeds was divided into slow and fast
categories by one of two criterion speeds.