Cognitive Psychology 95 (2017) 116

Contents lists available at ScienceDirect

Cognitive Psychology
journal homepage: www.elsevier.com/locate/cogpsych

Breaking the rules in perceptual information integration

Maxim A. Bushmakin a,b,, Ami Eidels c, Andrew Heathcote c,d
a
Department of Psychological and Brain Sciences, Indiana University, USA
b
Volen National Center for Complex Systems, Brandeis University, USA
c
School of Psychology, The University of Newcastle, Australia
d
School of Medicine, The University of Tasmania, Australia

a r t i c l e i n f o a b s t r a c t

Article history: We develop a broad theoretical framework for modelling difficult perceptual information
Accepted 25 March 2017 integration tasks under different decision rules. The framework allows us to compare coac-
tive architectures, which combine information before it enters the decision process, with
parallel architectures, where logical rules combine independent decisions made about each
perceptual source. For both architectures we test the novel hypothesis that participants
break the decision rules on some trials, making a response based on only one stimulus even
though task instructions require them to consider both. Our models take account of not
only the decisions made but also the distribution of the time that it takes to make them,
providing an account of speed-accuracy tradeoffs and response biases occurring when
one response is required more often than another. We also test a second novel hypothesis,
that the nature of the decision rule changes the evidence on which choices are based. We
apply the models to data from a perceptual integration task with near threshold stimuli
under two different decision rules. The coactive architecture was clearly rejected in favor
of logical-rules. The logical-rule models were shown to provide an accurate account of
all aspects of the data, but only when they allow for response bias and the possibility for
subjects to break those rules. We discuss how our framework can be applied more broadly,
and its relationship to Townsend and Nozawas (1995) Systems-Factorial Technology.
2017 Elsevier Inc. All rights reserved.

1. Introduction

Human choices often depend on combining noisy signals from multiple sources. When approaching an intersection on a
dark and rainy night, for example, a driver must determine whether traffic lights are red or green and whether pedestrians
are crossing. If the light is red, or if pedestrians are crossing, braking is required, following an OR decision rule for the
presence of either one signal or the other. Once stopped, and before continuing a trip, the driver must confirm that the light
is green and that no pedestrians are in their path, following an AND decision rule requiring both one signal and the other.
The OR rule allows processing to be terminated after only one signal is detected, whereas the AND rule requires processing
both signals, leading Townsend (1974) to describe them as stopping rules requiring, respectively, first-terminating and
exhaustive processing of stimuli. Stopping rules have been studied in many areas of human cognition, from categorization
(Fific, Little, & Nosofsky, 2010), to consumer choices (Fific & Buckmann, 2013), and memory- and visual-search tasks
(Fific, Townsend, & Eidels, 2008; Sternberg, 1966).

Corresponding author at: Volen National Center for Complex Systems, Brandeis University, USA
E-mail address: [email protected] (M.A. Bushmakin).

http://dx.doi.org/10.1016/j.cogpsych.2017.03.001
0010-0285/ 2017 Elsevier Inc. All rights reserved.
2 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116

Early investigations of stopping rules manipulated the number of items in either memory or a visual display and focused
on the slope of the response time (RT) function as the number of items increases (Sternberg, 1966, 1969; see Algom, Eidels,
Hawkins, Jefferson, and Townsend (2015), for a review). However, later work by Townsend and colleagues (e.g., Townsend &
Ashby, 1983; Townsend & Colonius, 1997) showed that this approach was flawed, because different rules could predict the
same pattern of slopes. Subsequently, Townsend and Nozawa (1995) showed that AND and OR stopping rules do make
unique predictions in a design often called the double-factorial paradigm that manipulates the relative salience of
two or more signals.
It has often been found using the double-factorial paradigm that human observers do appear to apply the stopping rule
appropriate to the task at hand (e.g., Eidels, Townsend, & Algom, 2010b; Fific, Nosofsky, & Townsend, 2008; Fific et al., 2008;
Fific et al., 2010; Little, Nosofsky, & Denton, 2011; Moneer, Wang, & Little, 2016). However, double-factorial designs are not
always easy to implement, and Townsend and Nozawas (1995) analysis does not take account of the effects of mixtures of
stopping rules (i.e., applying different rules on different trials). Cousineau and Shiffrin (2004) found evidence for such incon-
sistent rule application in a difficult visual-search task. On a proportion of trials it appeared that only one of the two items in
a display was processed fully, with participants guessing about the other item.
The use of an inappropriate stopping rule can have detrimental effects on behavioural performance. For example, exhaus-
tively processing under an OR requirement requires more effort and slows response with no benefit to accuracy. However,
inappropriate stopping may also have benefits. For example, failing to process exhaustively under an AND requirement, and
either ignoring the remaining items, or making guesses about them as in Cousineau and Shiffrins (2004) study, makes
responding faster and easier, although it will also cause errors. Depending on the value a participant places on speed over
accuracyperhaps as a function of trying to complete an experiment with minimal effort, or in order to have time to attempt
more decisions in a fixed time periodrule breaking may be an attractive and even optimal strategy (Bogacz, Brown,
Moehlis, Holmes, & Cohen, 2006).
In the present paper we focus on a perceptual information integration task. Such tasks are of interest because humans are
often in a position where they need to integrate information within a sensory modality (Bushmakin & James, 2014; Eidels,
Donkin, Brown, & Heathcote, 2010a) as well as across modalities (Alais & Burr, 2004; McGurk & MacDonald, 1976; Stevenson
et al., 2014a). Moreover, the irregularities in perceptual integration have been linked to certain disorders like autism and
schizophrenia (Stevenson et al., 2014b; Williams, Light, Braff, & Ramachandran, 2010).
Despite both empirical evidence for rule breaking in other tasks, and in at least some cases there being a rational basis for
participants to purse a rule-breaking strategy, previous analyses of information integration have not, to our knowledge,
considered whether decision makers might sometimes not abide by the rules that experimenters use to score the accuracy
of their performance. The present paper develops a framework that takes into account the possibility of rule breaking for
tasks requiring both AND and OR rules, and the possibility that a participant can use a mixture of rule-following and
rule-breaking strategies. We apply models derived from this framework to the performance of each participant separately
to account for the possibility that there will be individual differences in the factors that cause rule breaking, such as the value
placed on effort or the speed vs. accuracy.
The nature of the two stopping rules we investigate makes it important to take account of tradeoffs between the speed
and accuracy of different responses. These OR and AND designs have built-in biases towards one response or another.
Consider the driver OR example presented above; attempting to drive through the intersection, the driver needs to stop if
she detects a pedestrian approaching, a red light, or both. In contrast there is only one case in which she can go: if there
is a green light and no pedestrians. These contingencies are known to create biases in responding (e.g., Mordkoff &
Yantis, 1991). All other things being equal, it is likely that a bias towards responding YES will develop under OR instruc-
tions and a bias to respond NO under AND instructions, because a YES response is required more often in the former case
and a NO response more often in the latter case. Attempts to remove those biases by unbalanced presentation of stimuli (e.g.,
more no-target trials in AND task or more double-target trials in an OR task) create other contingencies and hence other
potential biases. In light of this, our experiments used balanced stimulus presentation, and we take account of the potential
biases in our models.
In summary, two questions remain unanswered: (1) how well do people abide by AND and OR decision rules, and the
related questions of how they manage the associated trade-off between speed and accuracy; and (2) how they select an
appropriate level of response bias under each rule. To address these questions, we also investigated two other fundamental
questions about the effect of decision rules on the inputs to, and architecture of, the decision process.
First, do response-rule instructions affect only the decision process itself, or do they affect the inputs to the decision pro-
cess, that is, the evidence on which decisions are based? In particular, is the evidence accumulation rate for each signal the
same across decision rules? This question is novel, perhaps because most past analyses have not tried to simultaneously
account for responding under two different decision rules by the same participants, as we do here. In contrast, the second
question has been the subject of intense scrutiny: is evidence combined before it enters a single decision process (sometimes
called a coactive architecture; e.g., Little, Nosofsky, Donkin, & Denton, 2013; Miller, 1978, 1982), or are separate decisions
made about each signal and later combined by logical rules (e.g., Eidels et al., 2010a).
To address all of these questions, we develop a unified theoretical framework that expands on Brown and Heathcotes
(2008) Linear Ballistic Accumulator (LBA) model of choice RT and on its extension to a logical-rule model of the OR task
developed by Eidels, Donkin et al. (2010) and Eidels, Townsend et al. (2010). We generalized Eidels et al.s model of the
OR task to the AND task, and, for both AND and OR models, we also allowed for the possibility that participants sometimes
 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116 3

broke the rules and processed only one signal. We also extended Brown and Heathcotes standard LBA to provide coactive
AND-task and OR-task models. The coactive models assume that evidence from a number of sources (signals) is combined
before entering a single LBA decision process with the possibility of rule breaking. We compared these models to test the
effects of response-instructions on architecture (coactive vs. parallel) and rule breaking, and used parameter estimates asso-
ciated with the models to test for their effects on inputs and response bias. Together the set of models provides a powerful
and comprehensive framework for answering questions about noisy perceptual information integration.
We apply our framework to data reported by Eidels, Townsend, Hughes, and Perry (2015), where the same participants
made difficult AND and OR decisions about the presence or absence of one or two near-threshold dots of light, one positioned
above and one below fixation. Eidels et al. (2015) examined this data with an accuracy-based measure, the no-response
probability contrast (NRPC; Mulligan & Shaw, 1980). They also analyzed RT data for correct responses collected from the
same participants in another experiment with supra-threshold stimuli using Townsend and Nozawas (1995) nonparametric
Systems Factorial Technology (SFT). After describing our parametric model-based analyses and applying it to Eidels et al.s
(2015) experiment with near-threshold stimuli, we discuss the relationship of our results to their NRPC and SFT based
results.

2. A Modelling framework for perceptual integration

Our implementation of coactive models uses a standard two-accumulator LBA (see Fig. 1), where each accumulator cor-
responds to a potential response (i.e., YES or NO) and pools inputs from each position (i.e., potential dot locations: above and
below fixation). On each trial accumulators begin with evidence totals drawn independently from a uniform distribution
with range (0, A) and a rate of evidence accumulation that is drawn independently from a Gaussian distribution with mean
v and standard deviation sv. Evidence accrues linearly, and the response selected corresponds to the accumulator that first
reaches its threshold, b. Differences in thresholds between accumulators mediate response bias. RT is the sum of the
response-selection time and the time for perceptual encoding of the stimulus and response production, ter (ter = te + tr,
Fig. 1) which is assumed to be the same for both accumulators and greater than zero.
All modelsboth coactive and logical rulewere subject to some common parameter constraints. They assumed ter was
the same across stimulus conditions defined by the number of targets. The starting point distribution parameter A was
allowed to change between conditions yielding two parameters in AAND and AOR. Threshold (b) was re-parameterized in
terms of the distance from the top of the start-point distribution to the threshold, B = bA, so that the restriction B > 0
enforced the condition that evidence never starts above the threshold (Brown & Heathcote, 2008). We also allowed B to vary
between accumulator (i.e., Y = YES or N = NO) and condition to account for the possibility of bias, yielding four parameters:
BY,AND,, BN,AND, BY,OR, and BN,OR. Finally, we fixed sv = 1 for the accumulator that mismatched the stimulus (i.e., the YES accu-
mulator if no target was present and the NO accumulator if a target was present) in order to make the model identifiable (see
Donkin, Brown, & Heathcote, 2009). The sv parameter for the other accumulator was estimated from the data, consistent
with previous LBA applications (e.g., Heathcote & Love, 2012; Rae, Heathcote, Donkin, Averell, & Brown, 2014).
The stimulus encoded for each position provides evidence for the presence and absence of a target. We examined both
coactive and logical rule models in which rates could differ for upper and lower positions; however, according to our
model-selection methods, these models did not provide sufficient improvement in fit over models that assumed same rates
to justify the associated doubling in the number of estimated rate parameters.1 Hence we assume in all of the results reported
here the same rates for upper and lower positions.
Rates can differ depending on the response accumulator and whether a target stimulus is actually absent (A) or present (P) at
a given position. Hence, there are four types of rates estimated: two when a target is present, for the YES accumulator (vY,P) and
for the NO accumulator (vN,P), and two when a target is absent, again one for the YES accumulator (vY,A) and one for the NO accu-
mulator (vN,A). Hence, vY,P and vN,A are rates corresponding to the correct response, whereas vY,A and vN,P indicate rates for the
accumulators corresponding to an incorrect response. When a target is present, the ordering vY,P > vN,P results in above chance
responding, as the YES accumulator tends to reach its threshold more quickly than the NO accumulator. Similarly, when the
target is absent, vN,A > vY,A implies above chance responding. These inequalities would be expected to hold under a wide range
of reasonable assumptions. This would be the case, for example, if the YES accumulator rate is an increasing function of how far
luminance of the stimulus (l) is above a threshold (v), rate = g(l
v), and NO accumulator rate is an increasing function of how
far luminance is below that threshold, rate = h(v
l), where g() and h() are monotonic increasing functions.

2.1. Coactive models

For the coactive models, pooling of inputs from each position was achieved by simply adding rates associated with upper
and lower target positions. Assuming inputs of equal magnitude for each position, the inputs to YES and NO accumulators

1
The coactive model with separate rates for top and bottom stimuli, had AIC =
1614 and BIC =
324 for the joint parameterization (16 parameters) and
AIC =
8087 and BIC =
6324 for the separate parameterization (26 parameters). Logical rule models with separate rates for top and bottom stimuli, had
AIC =
9528 and BIC =
8441 for the joint parameterization (16 parameter) and AIC =
10,014 and BIC =
8252 for the separate parameterization
(26 parameters). These models were also worse according to the BIC and AIC methods of model selection relative to the rule breaking models with the same
accumulation rate for both locations.
4 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116

Fig. 1. The standard LBA model, with separate accumulators for target (Y = YES) and non-target (N = NO) responses.

will be 2vY,P and 2vN,P respectively when both targets are present. When only one target is present, regardless of where it is
positioned, the inputs to YES and NO accumulators are, respectively, vY,P + vY,A and vN,P + vN,A. When both targets are absent,
the inputs to YES and NO accumulators are 2vY,A and 2vN,A, respectively. In all cases the standard deviation of the sum is 2sv2.
For example, suppose vY,P = vN,A = 1 and vY,A = vN,P = 0 (where vY,P > vN,P and vN,A > vY,A, as required for above chance accu-
racy). When there are two targets, the input to the YES accumulator is larger than the input to the NO accumulator (2 vs. 0),
whereas when there are no targets present the opposite is true (0 vs. 2). When only one target is present both accumulators
have the same input, but an appropriate level of response bias can still enable accurate responding for AND and OR decision
rules, because these rules are linearly separable. For an AND decision rule, for example, the required response is NO for a
single target, so the response threshold must be lower for the NO accumulator. For an OR decision rule, in contrast, the
threshold must be lower for the YES accumulator.
All models (both coactive models and the logical rule models detailed below) were fit to each participants data
separately by maximizing the sum of the logarithm of the likelihoods (L) for each observed response-RT pair. Given param-
eters h = (b, A, v, sv), and t = RT
ter, the density function f(t|h) gives the instantaneous probability that an accumulator
reaches its threshold at time t, and the survivor function S(t|h) gives the probability that an accumulator has not reached
threshold at time t. Denoting the coactive model by the subscript CO, we have:
LYESjCO hYES ; hNO jt f YES tjhYES  SNO tjhNO 1

LNOjCO hYES ; hNO jt f NO tjhNO  SYES tjhYES 2

In words, these equations say that the likelihood (which is just the probability density function with the conditioning of t
and h reversed) of a given response at time t equals the density (f) of the corresponding accumulator reaching its threshold at
time t times the probability (given by the survivor function, S) that the other accumulator has not reached its threshold. Fit-
ting a separate model to either the AND or OR condition required estimation of nine parameters: four vs two Bs (for YES and
NO accumulators to allow for response bias) and one A, sv, and ter.
We fit two types of coactive model. The first assumed only the A parameters and the two B parameters (one each for YES
and NO accumulators) differed between the AND and OR instruction conditions (i.e., a joint model, with 12 parameters:
2  A, 4  B, 4  v, ter, and sv). This joint coactive model instantiates the assumption that decision-rule instructions only
change the decision process, not the inputs to the decision process. We also fit a more flexible model in which all parameters
could differ between AND and OR conditions (i.e., two separate fits, with 18 parameters in total), instantiating the assump-
tion that the instructions could also change the nature of the evidence on which decisions are based.
 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116 5

2.2. Logical-rule models

The logical rule models have two pairs of accumulators, one pair taking input from the lower position (LT, lower target)
and one from the upper (UT, upper target). Unlike the coactive models, evidence from the two positions is not combined
prior to the decision. Rather, a separate decision is made for each position concerning the presence or absence of a target
at that location, and the outcome of each sub-decision is then combined logically. For both AND and OR response rules,
the likelihood of a YES or NO response at time t can be worked out using multiplication (to get the probability that one event
and another occurs) and addition (to get the probability that one event or another occurs). For a YES response in the AND
case, both NO accumulators must be below threshold (i.e., they are both survivors), which occurs with probability
SNO,LT(t)  SNO,UT(t). There are two possible cases for the YES accumulators (i.e., one case or another). For the first case, the
lower YES accumulator has previously reached threshold, with probability FYES,LT(t) = 1
SYES,LT(t), and the upper YES
accumulator reaches threshold at time t, with instantaneous probability fYES,UT(t). The second case swaps the roles of upper
and lower YES accumulators. The products of these terms are then summed as shown in the following equation, where for
brevity t is omitted:
LYESjAND SNO;LT  SNO;UT  f YES;LT  F YES;UT f YES;LT  F YES;UT 3

For a NO response both YES accumulators cannot be above threshold, which is most easily calculated by subtracting the
probability that both are above threshold (i.e., FYES,LT x FYES,UT) from one. Again there are two cases, where the lower NO accu-
mulator finishes before the upper NO accumulator, or vice versa. Hence the likelihood of a NO response is:
LNOjAND 1
F YES;LT  F YES;UT  f NO;LT  SNO;UT f NO;UT  SNO;LT 4
The OR rule, described in detail in Eidels, Donkin et al. (2010) and Eidels, Townsend et al. (2010), yields equations of a very
similar form, but swapping the structure of YES and NO equations and the roles of YES and NO accumulators within each:
LYESjOR 1
F NO;LT  F NO;UT  f YES;LT  SYES;UT f YES;UT  SYES;LT 5

LNOjOR SYES;LT  SYES;UT  f NO;LT  F NO;UT f NO;UT  F NO;LT 6

In an analogous manner to the coactive models (and with the same number of parameters), we fit the logical rule models
either separately to each AND and OR conditions, or jointly, assuming the same values for input-related parameters across
the two conditions (i.e., the joint parameterization assuming no effect of decision rule on inputs).

2.3. Rule-breaking models

We extended both coactive and logical-rule models by assuming that participants only follow the rules on some trials,
which occurs with probability p. When they break the rules, with probability (1
p), they either process only the lower posi-
tion, with probability q, or the upper position, with probability (1
q). In the separate fits, this adds two parameters to each
of the AND and OR conditions. In the joint fits, we allowed separate values of p for each decision-rule condition, but a com-
mon value of q.2
The likelihoods of YES and NO responses in the rule-breaking (RB) case are:
LYESjRB q  f YESjLT  SNOjLT 1
q  f YESjUT  SNOjUT

LNOjRB q  f NOjLT  SYESjLT 1
q  f NOjUT  SYESjUT

For the rule-breaking coactive (RB-CO) model:

LYESjRB
CO p  LYES 1
p  LYESjRB 7

LNOjRB
CO p  LNO 1
p  LNOjRB 8

As in the rule-following model, AND and OR have identical likelihood equations and are differentiated by parameter
estimates.
For the rule-breaking logical-rule models:
LYESjRB
RULE;AND p  LYESjAND 1
p  LYESjRB 9

LNOjRB
RULE;AND p  LNOjAND 1
p  LNOjRB 10

LYESjRB
RULE;OR p  LYESjOR 1
p  LYESjRB 11

2
Allowing separate values of q for the two rule conditions produced unstable estimates due to multiple local minima and was rejected by model selection.
Additionally, all of the subjects demonstrated same preference for one location over the other, so for simplicity we report only results where a common value of
q was assumed.
6 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116

LNOjRB
RULE;OR p  LNOjOR 1
p  LNOjRB 12

Each rule-breaking model requires estimation of two more parameters than the corresponding rule-following model.
In the following analysis of Eidels et al.s (2015) data, we report the results for eight types of models: a factorial combi-
nation of coactive vs. logical-rule, rule-following vs. rule-breaking, and separate vs. joint parameters. Fits were obtained by
maximizing the likelihoods in (1)-(12).

3. Example application

Eidels et al.s (2015) experiments are similar to those used by Townsend and Nozawa (1995) to develop SFT, with the dots
appearing 1 above or below fixation 0.5 s after the offset of a fixation point. SFT provides a non-parametric method of iden-
tifying not only stopping rules, but also cognitive architecture (e.g., serial vs. parallel processing) and the capacity constraints
(e.g., limited vs. unlimited capacity).
For their first study, Eidels et al. (2015) focused on an SFT analysis in a paradigm using both dimmer or brighter super-
threshold dots. SFTs mean- and survivor-interaction contrast analyses, which are diagnostic of architecture and stopping
rule, clearly rejected serial processing and favoured a parallel-processing architecture in both conditions, with each follow-
ing the appropriate rule, first-terminating for OR instructions and exhaustive for AND (see Houpt and Townsend (2010), for a
converging analysis).
For our purposes here, Eidels et al.s (2015) first study introduces a challenge, since the use of super-threshold stimuli
caused accuracy to be near ceiling. This makes it difficult to disentangle differences between YES and NO responses caused
by bias as opposed to effects caused by a difference in the quality of evidence for each alternative. Response bias seems likely
as all four stimulus conditionsno target (NT), lower target only (LT), upper target only (UT), and a double target (DT)were
equi-probable. Hence, a NO response was required on 75% of trials under AND instructions, making it likely that some par-
ticipants developed a bias to respond NO, whereas a YES response was required on 75% of trials under OR instructions, mak-
ing a YES bias likely.
Fortunately, when accuracy is lower, bias can be identified because it has a negatively correlated effect on speed and
accuracy (i.e., it causes a speed-accuracy trade off), whereas changes in evidence quality have a positively correlated effect.
For the same reason, it is difficult to fit evidence-accumulation models such as the LBA to high accuracy data, because such
models estimate separate parameters related to bias (i.e., the threshold amount of evidence required to make each response)
and evidence (i.e., the rate at which evidence enters accumulators associated with each choice). Given these considerations,
the analysis reported here used data from Eidels et al.s (2015) second study, which re-tested the same participants as the
first, but calibrated dot luminance for each participant to a near-threshold level in order to achieve 85% - 95% accuracy.

4. Results

4.1. Model selection

We measured model misfit using deviance (D), which is minus two times the likelihood that was maximized to estimate
the best-fitting parameters. Deviance differences between nested models (e.g., joint vs. separate parameterizations; rule-
following vs. rule-breaking) have a v2(df) distribution, where df (degrees of freedom) equals the difference in the number
of parameters estimated for each model. Model selection was accomplished using the Akaike information criterion (AIC)
P
and the Bayesian information criterion (BIC) aggregated over the i = 1 . . . S participants (S = 9): AIC Si1 Di 2S  p,
PS PS
and BIC i1 Di S  p  ln i1 ni , where ni is the number of data points for the ith participant and p is the number
of model parameters. The term after the plus sign in each case is a penalty for model complexity as measured by the number
of model parameters (i.e., p parameters for each of the S subjects). The complexity penalty means that AIC and BIC can prefer
worse fitting (i.e., higher deviance) models with fewer parameters to better fitting models with more parameters if the extra
parameters do not improve fit sufficiently. The penalty is larger for BIC than AIC (as n is large), and so BIC prefers simpler
models.
Table 1 shows a very clear rejection of the coactive models, which always fit much worse than the corresponding (i.e.,
same number of parameters) logical-rule models. Indeed, the advantage is so great that even the simplest logical-rule model
fits better than the most flexible coactive model. Further, none of the three measures in Table 1 supported a coactive model
in general and for any individual participant (supplementary materials Tables 1S and 2S, including JS and LB, for whom Eidels
et al. (2015) found evidence for a co-active architecture). Hence, we focus only on the logical-rule models in further analyses.
AIC selected the most complex model; the model with both a separate parameterisation and rule-breaking, but BIC
selected a simpler model with the joint parameterization. However, when the joint parameterization was enforced, the
decrease in fit was significant for both the rule-following models, v2(54) = 605, p < 0.001, and the rule-breaking models,
v2(63) = 404, p < 0.001. The same was true when rule-following is enforced: for the joint parameterization, v2(27) = 749,
p < 0.001, and for the separate parameterization, v2(36) = 549, p < 0.001.
Fig. 2 compares the fit of the four logical-rule models to accuracy data, and reveals the source of the misfit for rule-
following models: an inability to explain lower accuracy in the single target conditions in the OR task, particularly when
 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116 7

Table 1
Model selection results. Number of parameters (p) is specified per participant. The best AIC and BIC values are in bold font.

Model class Rules breaking Parameterization Parameters per participant (p) Summed deviance AIC BIC
Logical rule No Joint 12
9289
9073
8257
Separate 18
9892
9564
8342
Yes Joint 15
10,042
9769
8751
Separate 22
10,441
10,040
8547
Coactive No Joint 12 1015 1181 1997
Separate 18
7769
7608
6386
Yes Joint 15
6124
6208
5189
Separate 22
8521
8158
6665

Rulefollowing joint Rulefollowing seperate

100

100
O A A
A
O
O A A
A O
A
A O
A
A O A O
O O
O O
O
O
% Correct

% Correct
O O
75

75
A A
A
A
50

50

DT T1 T2 NT DT T1 T2 NT

Rulebreaking joint Rulebreaking seperate

100

100

O
O A
A
A
A
O
O A A
A
O A O
O
O
A
A A
A
O O
O O
% Correct

% Correct

O O
75

75

O O

A
A A
A
50

50

DT T1 T2 NT DT T1 T2 NT

Fig. 2. Logical rule-following and rule-breaking model fits to accuracy data from the four within-subject conditions: NT = non-target, LT = lower target, one
dot below, UT = upper target, one dot above, DT = double target, both dots. Larger O and A symbols indicate, respectively, average observed accuracy in
the OR and AND tasks. Solid lines (with smaller O symbols) represent predictions for the OR condition and dotted lines (with smaller A symbols)
predictions for the AND condition.

the target was in the lower location (LT); and to a lesser degree in both decision-rule conditions when the target was in the
upper location (UT). The misfit to the AND data is more subtle, being mainly in the UT condition. Fig. 3 shows fits of the joint
and separate parameterizations of rule-breaking models to RT distributions, represented by three percentiles indicating the
fastest (10th percentile), middle (50th, i.e., median), and slowest (90th) responses. Both joint and separate parameterizations
of rule-breaking models provide a good account of the RT data, with the exception of some over-prediction of slow responses
in the single target UT condition. The fit of the joint parameterization model is only a little worse, indicating that using the
same rate parameters across AND and OR instructions mainly affects the account of accuracy.
Overall these results support logical rule-breaking models. In order to see if there was any inconsistency in the degree of
rule breaking between instruction conditions, we examined the evidence for rule breaking in separate fits to each instruction
condition for this model. We found that for OR instructions, rule breaking was favoured over the rule following by both AIC
(
6730 vs.
6226) and BIC (
5982 vs.
5612). However, under AND instructions, both AIC (
3314 vs.
3342) and BIC
8 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116

(
2565 vs.
2729) supported rule following. To follow up on this resultand to examine individual differenceswe calcu-
lated separately for each condition the posterior probability, based on both AIC and BIC (Wagenmakers & Farrell, 2004), for
the rule-following and rule-breaking models. As shown in Table 2, the general picture favoured rule breaking in the OR con-
dition and rule-following in the AND condition, but there were also some marked individual differences.
In particular, in the OR condition both AIC- and BIC-based results strongly favoured rule-breaking for all participants
except WY, who was classified as clearly rule following by both BIC and AIC. For the AND condition, in contrast, the two mea-
sures consistently indicated strong evidence for rule breaking in only participant WY. For six others, the results consistently
supported rule following. For the remaining two participants, AIC preferred the more complex rule-breaking model, whereas
BIC preferred the simpler rule-following model. However, even for the participant classified as rule breaking in the AND con-
dition, rule breaking was estimated as occurring on only a small number (2%) of trials. Estimates of the probability of rule
breaking were much higher for all participants in the OR condition except WY, as shown in Table 3. Table 3 also shows that
where rule breaking occurred it was almost always because participants processed only the top position, with the only
exception being participant AW, who processed only the bottom position on about one quarter of the 14% of trials where
they broke the rules.
Overall, our results suggest that the simpler separate parameterization of the rule-following model provides a good
description of all participants in the AND condition, whereas the more complex separate parameterization of the rule-
breaking model is required for the OR condition, although the level of rule breaking was quite minimal for one participant.
Plots of these models (i.e., rule breaking for OR, rule following for AND, with separately estimated drift rates for each con-
dition) to individual participant accuracy (Fig. 4) and RT distribution (Fig. 5) data confirmed they provide a good fit, with the
exception of participant LB in the AND condition. However, the systematic misfit for LB was hardly improved even with the
greater flexibility afforded by rule breaking; thus for this participant and condition, the misfit indicates that none of the mod-
els considered here was able to provide an entirely satisfactory account.

4.2. Parameter estimates

4.2.1. Evidence accumulation rates

We next performed an analysis of the parameter estimates for the models shown in Figs. 4 and 5. Given the rejection by
model selection of the joint-parameterization model (which assumes that AND and OR instructions do not influence the
inputs to evidence accumulation), we first examined rate estimates. Supporting the model-selection results, there was a sig-
nificant main effect of instruction, F(1, 8) = 13.5, MSE = 1.77, p = 0.006, and an interaction between the instruction and
match (i.e., matching vs. mismatching accumulator, where the matching accumulator is the target accumulator for a target
stimulus and a non-target accumulator for the a non-target stimulus) factors, F(1, 8) = 25.3, MSE = 0.4, p = 0.001.
In order to understand these results, it is useful to focus on the difference between rates for the two levels of the match
factor (i.e., the rate for the matching accumulator minus the rate for the mismatching accumulator). This difference indexes
the quality of the information extracted from the stimulus, with larger values associated with more accurate responding.
The two-way interaction indicated quality was greater on average in the OR condition (a difference of 3.5 on average) than
the AND condition (a difference of 2). This advantage for the OR condition was very consistent, being present for every
participant.3
The three-way interaction between the match, instruction and target (presence vs. absence) factors was also significant, F
(1, 8) = 5.9, MSE = 1.7, p = 0.04. To explore this interaction we conducted two separate two-way ANOVAs on rates for each
instruction condition. In the OR condition a significant interaction between the match and target factors, F(1, 8) = 6.4,
MSE = 0.36, p = 0.03, marked higher quality information on the target absent condition (4.01) than the target present
(2.98) condition. This difference in quality was very consistent at the individual level, with a substantial advantage for all
but two participants, RM and AW, who had close to equal quality. In the AND condition, in contrast, the difference for target
absent (1.96) and target present (2.02) were very similar, and the corresponding interaction did not approach significance, F
(1, 8) = 0.03, MSE = 0.24, p = 0.85. However, there were still considerable individual differences in the quality of information
between the two conditions, ranging from participant MB, whose target-absent difference was 1.7 larger than their target
present difference, to participant RM, with the opposite order held by 1.5.

4.2.2. Thresholds
Estimates of the LBA threshold parameter (B) indicated that participants were less cautious (i.e., required less evidence
before making a response) in the OR condition (B = 0.62) than the AND condition (B = 0.87), F(1, 8) = 5.7, MSE = 0.1,
p = 0.04. This difference was quite consistent, being displayed by all but participant RS. There was also a strong interaction,
F(1, 8) = 18.1, MSE = 0.02, p = 0.003, because under AND instructions the threshold was lower for the NO accumulator (0.74)
than the YES accumulator (1.0), whereas the reverse was true under OR instructions (0.68 vs. 0.56 respectively). This pattern
of results of response bias is exactly what would be expected if participants (quite reasonably) set a lower threshold for the
more common response in each condition (i.e., 75% NO in the AND condition and 75% YES in the OR condition). It was very

3
Due to dropouts, session counterbalancing was not achieved, with AW, LB and WY doing the AND session first, whereas the remaining six participants did
the OR session first. The OR advantage was on average 2.3 for the former group and 1.1 for the latter, so it is possible that at least some part of the OR advantage
was due to, for example, a learning effect.
 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116 9

1.0
Rulefollowing joint Rulefollowing joint

1.0
O
A A O
A
A A
A A
0.8

0.8
A O O
O
RT (s)

RT (s)
A
A O O O
O
0.6

0.6
O
A
A A
A A
A
A
A
O
O O O
O
O
O O
A
A A
A A O
0.4

0.4
A
O
O O O
O

DT LT UT NT DT LT UT NT

Rulebreaking joint Rulebreaking joint

1.0

1.0
O
A
A
O
A
A A
A A
0.8

0.8
A O
RT (s)

RT (s)
O O
A O
A O O O
0.6

0.6
A
A A O
A A
A
A
A O
O O O
O
O
O O
A
A A
A A O
0.4

0.4
A
O
O O O
O

DT LT UT NT DT LT UT NT

Rulefollowing separate Rulefollowing separate

1.0

A 1.0 O
A O
A
A A
A A
0.8

0.8

A
O
O
RT (s)

RT (s)

O
A
O O O
O
0.6

0.6

O
A
A A
A A
A
A A
O
O O O
O
O
O O
A
A A
A A O
0.4

0.4

A
O
O O O
O

DT LT UT NT DT LT UT NT

Rulebreaking separate Rulebreaking separate

1.0

1.0

A
O
O
A
A A
A A
0.8

0.8

A
O
RT (s)

RT (s)

O O
A
A O O O
O
0.6

0.6

A
A A
A A
O
A
A A
O
O O O
O
O
O O
A
A A
A A O
0.4

0.4

A O
O
O O O

DT LT UT NT DT LT UT NT

Fig. 3. Logical rule-following and rule-breaking separate and joint parameterization model fits to response RT distribution data (lower line = 10th
percentile, middle line = 50th percentile, upper line = 90th percentile) from the four within-subject conditions: NT = non-target, LT = lower target, one dot
below, UT = upper target, one dot above, DT = double target, both dots. Larger O and A symbols indicate, respectively, average observed RT percentiles in
the OR and AND tasks. Solid lines with smaller O symbols represent predictions for the OR condition and solid lines with smaller A symbols represent
the predictions for the AND condition.

Table 2
Posterior model probabilities across the set of eight models based on BIC for each participant for the separate parameterization rule breaking and rule following
model. Probabilities are rounded to two decimal places.

Models BJ RS JS MB RM LB JG WY AW
AIC AND Rule following 1 0.9 0.88 1 0.13 0 .16 0 0.99
Rule breaking 0.01 0.1 .12 0 0.87 1 0.84 1 0.01
OR Rule following 0 0 0 0 0 0 0 0.98 0
Rule breaking 1 1 1 1 1 1 1 0.02 1
BIC AND Rule following 1 1 1 1 0.94 0.01 0.95 0.02 1
Rule breaking 0 0 0 0 0.06 0.99 0.05 0.98 0
OR Rule following 0 0 0 0 0 0 0 1 0.05
Rule breaking 1 1 1 1 1 1 1 0 0.95
10 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116

Table 3
Parameter estimates for each participant of the probability of rule following (p) and the probability of processing the lower position given rule breaking (q) for
the OR condition.

BJ RS JS MB RM LB JG WY AW
p 0.79 0.68 0.88 0.53 0.54 0.74 0.61 1 0.84
q 0 0.04 0 0.01 0 0.07 0.05 0.02 0.25

BJ MB JG
100

100

100
O
O A
A O
A
O A
A
A A
A
O O
O A
A A
A A
A
A O
O A
A A O O
O O
O O
O O
O
O
O
O
% Correct

% Correct

% Correct
O
A
75

75

75
O
O
A A
A A
50

50

50
DT LT UT NT DT LT UT NT DT LT UT NT

RS RM WY
100

100

100
O
O A
A
O A
A O A A
A
O O O A
A O O
A A
A O O A
O
O A O
A
A
% Correct

% Correct

% Correct
A A O A
A O O
75

75

75
A O
O

A
O
O
O
O
50

50

A
A 50

DT LT UT NT DT LT UT NT DT LT UT NT

JS LB AW
100

100

100

O
O A
A A
A O
O A
A A
A
O O
O A
A
O
A
O A
O
A A O
O O
A O
O O
O
O
O A A
A
O
O O
% Correct

% Correct

% Correct

A
A
75

75

75

O
O

A A
50

50

50

A
A

DT LT UT NT DT LT UT NT DT LT UT NT

Fig. 4. Rule-breaking separate-parameterization model fits to accuracy data from the four within-subject conditions: NT = non-target, LT = lower target, one
dot below, UT = upper target, one dot above, DT = double target, both dots. Larger O and A symbols indicate, respectively, observed accuracy in the OR
and AND tasks. Solid lines (with smaller O symbols) represent predictions of the rule-breaking model for the OR condition and dotted lines (with smaller
A symbols) predictions of the rule-following model for the AND condition.

consistent at the individual level, being displayed by every participant except LB, who had the opposite pattern of bias in the
AND condition, the same condition where the model displayed systematic misfit.
The two remaining parameters, ter (non-decision time) and A (start-point variability), did not differ significantly between
instruction conditions: F(1, 8) = 0.3, MSE = 0.008, p = 0.61, and F(1, 8) = 3.8, MSE = 0.38, p = 0.09, respectively.

5. Discussion

In this paper we have developed a parametric framework for understanding the integration of information from noisy
perceptual signals. The general framework uses mixtures of independent racing evidence-accumulation processes. In the
detailed implementation of the framework we developed here, we assumed that these processes followed Brown and
Heathcotes (2008) LBA model. For the data set we examined in detail here (Eidels et al., 2015), these assumptions
 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116 11

BJ MB JG

1.6

1.6

1.6
1.2

1.2

1.2
RT (s)

RT (s)

RT (s)
O
0.8

0.8

0.8
O O
O O
O O O
O O
O O
O O
O O
O O
O O
O O O
O O
O O
O O
O O
O O
O O
O
O O
O
O
O O O O O O O O O O
O O
O
0.4

0.4

0.4
O O O O O O O
O
O
O O
O O
O O
O O O
O

DT LT UT NT DT LT UT NT DT LT UT NT

RS RM WY
1.6

1.6

1.6
O
1.2

1.2

1.2
RT (s)

RT (s)

RT (s)
O O
O
O O
O
0.8

0.8

0.8
O
O O
O
O
O O
O O
O O
O O
O O
O
O
O O
O O
O O
O
O
O O
O O
O O
O O
O O
O
O O
O
O
O O
O O
O O
O
O O O O O O
0.4

0.4

0.4
O O O
O O
O O O O
O O O

DT LT UT NT DT LT UT NT DT LT UT NT

JS LB AW
1.6

1.6

1.6
1.2

1.2

1.2
O
RT (s)

RT (s)

RT (s)
O

O O
O
0.8

0.8

0.8
O
O O
O O
O O O
O
O O
O O O O
O O
O O O
O
O
O O
O O
O O
O O
O O
O O
O O
O O
O
O O
O O
O
O O
O
O O O
0.4

0.4

0.4
O
O O
O O
O O
O
O O
O O
O O
O O O

DT LT UT NT DT LT UT NT DT LT UT NT

BJ MB JG
1.6

1.6

1.6
1.2

1.2

1.2
RT (s)

RT (s)

RT (s)

A
A A
0.8

0.8

0.8

A A A A
A A
A A A A A
A A
A
A A
A A
A A
A A
A
A A A
A
A
A A
A A
A A A A
A A
A A
A A A
A A A
A
A A A A
0.4

0.4

0.4

A A
A A
A A
A A
A A
A A
A A
A A A
A

DT LT UT NT DT LT UT NT DT LT UT NT

RS RM WY
1.6

1.6

1.6

A
A
A A A
1.2

1.2

1.2

A
RT (s)

RT (s)

RT (s)

A
A A A A A A
A A A A
A A A
0.8

0.8

0.8

A A A
A A
A A
A
A A A
A A A
A
A
A
A A
A A
A A
A A
A A
A A
A A
A A
A A
A A
A A
A
A A A A
0.4

0.4

0.4

A A A A A
A A
A

DT LT UT NT DT LT UT NT DT LT UT NT

JS LB AW
1.6

1.6

1.6

A
1.2

1.2

1.2

A A A
RT (s)

RT (s)

RT (s)

A
A A
A
A
A
A
0.8

0.8

0.8

A
A
A
A A
A
A
A A
A A
A
A A
A A
A A
A A
A A A A A A
A
A
A A
A A
A A
A A
A A
A
A
A
A
A
A
A
A A
A
A
A A A A A
0.4

0.4

0.4

A A A
A A A A

DT LT UT NT DT LT UT NT DT LT UT NT

Fig. 5. Rule-breaking separate-parameterization model fits to response RT distribution data (lower line = 10th percentile, middle line = 50th percentile,
upper line = 90th percentile) from the four within-subject conditions: NT = non-target, LT = lower target, one dot below, UT = upper target, one dot above,
DT = double target, both dots. Larger O and A symbols indicate, respectively, observed RT percentiles in the OR and AND tasks. Solid lines (with smaller
O symbols) represent predictions of the rule-breaking model for the OR condition and dotted lines (with smaller A symbols) predictions of the rule-
following model for the AND condition.

proved sufficient to provide an accurate and detailed description of all aspects of each participants performance, includ-
ing the frequency of choices and the distribution of response time (RT), with only one exception, which we discuss
further below.
12 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116

For Eidels et al.s (2015) data, our analysis provided clear answers to four questions concerning the integration of
information from pairs of near-threshold stimuli with response rules of the OR (i.e., respond YES if one signal or another
was present) or AND (i.e., respond YES if one signal and another is present) type. The first question concerns whether
information from each stimulus is combined before it enters a single decision processa coactive architectureor whether
separate decisions are made about each signal, with these decisions then combined by logical rules. Our model fits decisively
rejected the coactive architecture in favor of logical rules. These findings largely agree with Eidels et al.s analyses based on a
contrast between the probability of responding NO in different conditions, which also rejected coactive processing for all
participants in the OR condition and all but two in the AND condition.
The second question is, how well do people abide by AND and OR decision rules? Past attempts to model information
integration have assumed participants are compliant, but we found otherwise. Under OR instructions, for about one quarter
of trials on average, participants only processed one stimulus. However, there was substantial individual variation, with rule
breaking occurring on anything from 4% to 45% of trials for different participants. For the same participants under AND
instructions most were best described as never breaking the rules, with at best two of the nine participants failing to process
both stimuli on 3% or less of trials. Thus, although there were individual differences, there was also quite a systematic effect
of the type of response rule on the occurrence of rule breaking. We discuss the implications of these novel findings below.
Our third question was whether response rules affect only the decision process itself, or whether they also affect the
inputs to the decision process, that is, the evidence on which decisions are based. Our results supported the latter answer,
with two clear differences related to the quality of the information extracted from the stimuli. First, quality was higher in the
OR condition than the AND condition. Second, in the OR condition, information quality for target-absent displays was sys-
tematically higher than for target-present displays, whereas in the AND condition there was no systematic difference. The
advantage for the models that allowed response rules to affect rates was very consistent, occurring for all participants in
all conditions.
The findings with respect to our third question are related to our fourth questionhow participants managed the tradeoff
between the speed and accuracy of respondingas we found that decision-processes thresholds also differed systematically
between AND and OR conditions. In particular, most participants set a lower threshold in the AND than OR condition. This
difference largely counteracted the better quality of information in the OR condition, so overall accuracy in the OR condition
(91%) was only marginally greater than in the AND condition (88%). We also found that most participants selected an appro-
priate level of response bias that favoured the more commonly correct response (i.e., YES in the OR condition, and NO in the
AND condition). We now discuss the implications of all of our findings for understanding information integration more gen-
erally, both with noisy near-threshold stimuli and with super-threshold stimuli.

5.1. Understanding information integration

The question of whether information integration involves coactivation versus an independent parallel architectureas is
assumed by our logical-rule modelshas traditionally been addressed by non-parametric analyses focused on RT in para-
digms where accuracy is near ceiling. Millers (1978, 1982) seminal studies found violations of the RT-based race model
inequality, consistent with coactive processing of visual and auditory signals, but later studies have failed to support this
conclusion (Alais & Burr, 2004; Wuerger, Hofbauer, & Meyer, 2003). Mordkoff and Danek (2011) suggested that coactivation
occurs only when redundant targets (i.e., two features with an OR response rule) are part of the same object. Eidels et al.
(2015) hypothesized that processing architecture might also depend on the difficulty of the required discriminations. In par-
ticular, they suggested that, even though the signals in their second experiment were spatially separated, a coactive archi-
tecture might be favoured because pooling near-threshold stimuli might improve discrimination. This possibility could have
important implications in real-world information integration tasks where, as suggested by the example with which we
began the paper, integration of noisy perceptual information can be required.
Unfortunately, the study of near-threshold stimuli is complicated by the fact that differences between conditions in accu-
racy as well as RT often occur. Accuracy differences can potentially confound tests based solely on RT, such as the non-
parametric race model inequality and related non-parametric techniques using the full distribution of RT, as developed in
Townsend and Nozawas (1995) Systems Factorial Technology (SFT). In order to address data where accuracy is well below
ceiling, Eidels et al. (2015) extended seminal work by Mulligan and Shaw (1980), and developed a test of coactivation based
purely on the probability of responding NO in different conditions. In close agreement with our findings, this testwhich is
also parametric as it assumes choices arise from a latent Gaussian distributionrejected coactivation for all but two partic-
ipants in the AND condition of their second experiment.
Taken together, these results indicate that near-threshold stimuli do not necessarily, or even not very often, cause coactive
processing. Future work could examine whether this finding generalizes to other stimuli and paradigms, following up, for
example, on evidence for coactivation with integral but not separable dimensions (Fific et al., 2008; Little et al., 2013) and
on Mordkoff and Daneks (2011) suggestions that coactivation occurs when redundant targets (i.e., two features with an OR
response rule) are part of the same object. For such purposes, the parametric framework we developed here has a number of
advantages. Like Eidels et al.s (2015) test of coactivation, our framework requires strong distributional assumptions (Jones
& Dzhafarov, 2014). However, in contrast to their test, these assumptions make testable predictions that can be evaluated
by the goodness-of-fit of the models (Heathcote, Brown, & Wagenmakers, 2014). Our framework also has the advantage that
it addresses all aspects of the data, including the probability of both NO and YES responses as well as the full distribution of
 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116 13

RT (see Townsend and Altieri (2012) for a similarly comprehensive but non-parametric approach). A further advantage is that it
is able to address the possibility that, at least on some occasions, participants break the rules of the information integration task.
Why might Eidels et al.s (2015) participants have broken the rules and processed only one position on some trials? An
enabling condition for this behavior is that this can be done while maintaining above chance accuracy, which is the case for
both the AND and OR tasks. If participants processing only one position responded YES based on detecting a single target and
NO otherwise, they will be 75% accurate in both the OR task (making an error only when the other position contained a tar-
get, which occurs on 25% of trials) and the AND task (making an error only when the other position contains a non-target,
which again occurs on 25% of trials). Thus, the penalty for rule breaking is not particularly large; at the average rate observed
in Eidels et al.s OR condition, it adds only around 6% extra errors. Given that processing only one stimulus might speed at
least a subset of responses (e.g., a NO response in the OR condition and a YES response in the AND condition, as both other-
wise require classification of stimuli at both locations), this type of rule breaking provides a mechanism for achieving a
speed-accuracy tradeoff.
One puzzle raised by our results is why rule breaking rarely occurred in the AND condition. Perhaps this was because
participants construed the task as being about detecting the presence of dots, although logically it is equally about detecting
their absence. Participants may have perceived the AND task as requiring the performance of two such target detections. This
possibility suggests that rule breaking may be less prevalent when non-targets are defined by the presence rather than
absence of a particular feature. Future research might examine this possibility using a variety of stimuli.
Given the possibility that rule breaking can mediate a speed-accuracy tradeoff, future research might also look at the
effect of instructions to respond rapidly. Such instructions are likely to also cause changes in thresholds for decisions about
individual features, and even potentially in the rate of evidence accumulation (e.g., Rae et al., 2014). The analysis framework
developed here, which can accommodate such changes, will be appropriate for future analyses of such paradigms.
The occurrence of rule breaking, at least in the OR task, raises a potential concern that it has confounded earlier analyses
of perceptual integration. This may not be a serious concern if rule breaking is associated with only near-threshold stimuli,
since most prior research has used super-threshold stimuli. Super-threshold stimuli are associated with near-ceiling accu-
racy, which is not consistent with the errors caused by rule breaking. However, the calculations reported abovenamely that
only around 6% errors are added by rule breaking on 25% of trialsindicate that fairly high accuracy does not exclude rule
breaking, at least on a sizeable minority of trials.
A potential experimental solution to the problem of rule breaking is to use an exclusive or (XOR) response rule. The XOR
task requires a YES response when one or the other target is present, and a NO response when both targets are present or
when neither target is present. Processing only one feature under these instructions leads to chance performance. Heathcote
et al. (2015) also suggested an XOR response rule could be useful in removing response bias, which they also found in AND
and OR tasks. As discussed previously, when the four conditions occur with equal frequency under AND and OR rules, NO and
YES responses, respectively, are correct on 75% of trials, which induces response bias. One solution could have been to induce
unequal condition frequencies so that correct YES and NO responses are equally likely. However, tampering with the relative
frequencies of trial types may induce sequential dependencies that can cause other problems (i.e., anticipation of upcoming
trials). Therefore, Heathcote et al. proposed using an XOR rule, which naturally has 50% correct YES and NO responses under
equal trial-types frequencies.
The XOR task can also be useful to the non-parametric SFT, since the latter makes use of measures composed of responses
that occur with different frequencies that can be distorted by response bias. Townsend and Eidels (2011) suggested as a rem-
edy composing SFT contrasts from a mixture of different AND- and OR-rule responses that are correct with equal frequencies
in their respective tasks (e.g., NO in AND and YES in OR). Unfortunately, this strategy may be confounded by other factors,
such as any a priori tendency to prefer one response, or overall differences in speed between the sessions in which AND and
OR tasks were performed. The XOR rule does not have these potential problems, and can achieve the goal of equating correct
response frequency within a single task.
Perhaps our most surprising finding relates to the strong and systematic differences in the inputs to the decision process
(i.e., the rates of evidence accumulation) induced by differences in response rules. Non-systematic differences between par-
ticipants might be explained by fluctuations in the average level of attention occurring because the AND and OR tasks were
performed in different sessions. However, this cannot explain why the quality of information about the stimuli was system-
atically higher in the OR session than in the AND session, and also why the quality of information about non-targets was
systematically higher than the information about targets in the OR session, whereas it did not differ systematically in the
AND session. To understand these findings it is important to note that, although the stimuli themselveswhich did not differ
between sessionsare one major determinant of evidence accumulation rates, rates are also influenced by factors related to
attention (e.g., working memory capacity, Schmiedek, Oberauer, Wilhelm, S, & Wittmann, 2007) and response instructions
(e.g., speeded responding instructions can reduce quality; Rae et al., 2014; Starns, Ratcliff, & McKoon, 2012). Our results indi-
cate that response-rule instructions can also affect the rate of evidence accumulation. Thus, these findings are at least not
inconsistent with previous applications of evidence accumulation models, but further work is clearly required to understand
why they occur.
Although the application of the framework we developed was largely successful, two limitations are of note. First, we
assumed unlimited processing capacity, in the sense that we assumed the same rate of evidence accumulation for each posi-
tion regardless of whether the other position contained a target or not (see Eidels, Donkin et al. (2010) and Eidels, Townsend
et al. (2010), for a demonstration of the relationship between capacity and evidence accumulation rates). SFT provides a
14 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116

non-parametric RT-based method of testing this assumption, which was applied to Eidels et al. (2015) data from their super-
threshold experiment. For the OR task they found all participants had more capacity than if the two signals shared a fixed
pool, although it was a little less than unlimited. In the AND condition, in contrast, there were strong individual differences:
four participants had severely limited capacity (i.e., less than fixed), one was intermediate, and four had super-capacity (i.e.,
more resources for two than one target), at least for slower responses. Although these findings might cast some doubt on our
unlimited-capacity assumption, the fact that we mostly obtained good fits provides some reassurance that it is reasonable
for Eidels et al.s near-threshold experiment.
A second limitation relates to the misfit observed for participant LB in the AND condition. This misfit underlines the point
that evidence accumulation models such as the LBA, are constrained to only accommodate a limited range of patterns of
behavior (Heathcote et al., 2014), and that the same is true of our general framework. Although this is a good characteristic
of our approach, because it allows critical evaluation via goodness-of-fit, it also raises the question of how the observed mis-
fit might be explained. There are many possibilities: data may have not conformed to the unlimited capacity assumption,
participants may have broken the response rules in a different way to that which we assumed (e.g., applied an OR rule in
the AND condition), their data may have been contaminated by mind wandering (e.g., Mittner et al., 2014) or the assump-
tions made by the LBA may be inappropriate.
Fortunately, if misfit were to occur on a sufficiently wide-spread basis in future data, the general framework which we
have proposed can be elaborated to accommodate such possibilities, including incorporation of alternatives to the LBA
(e.g., Heathcote & Love, 2012; Leite & Ratcliff, 2010; Logan, Van Zandt, Verbruggen, & Wagenmakers, 2014; Terry et al.,
2015; Van Zandt, Colonius, & Proctor, 2000). Similarly, extensions of the framework to the integration of information from
more than two sources, and to perform hierarchical Bayesian rather than maximum-likelihood estimation (e.g., where data
per participant are limited), are relatively straightforward. In order to facilitate applications of our framework for under-
standing information integration we provide in Supplementary Materials the code required to obtain likelihoods in the
open-source R language (R Core Team, 2015), along with examples of how to perform maximum-likelihood fitting to the data
we examined in the present paper.

6. Concluding remarks

We developed a comprehensive parametric framework for identifying important processing attributes such as architec-
ture (in particular, parallel-independent vs. coactive) and stopping rule (exhaustive vs. minimum time). The approach com-
plements existing non-parametric tools that focused exclusively on either RT (Millers, 1982; Townsend & Nozawa, 1995,
race model inequality) or accuracy (the no response probability contract, Eidels et al., 2015; Mulligan & Shaw, 1980), but
which do not combine the two dependent measures (see also Townsend & Altieri, 2012, and Townsend, Houpt, & Silbert,
2012, for other approaches that combine these measures with respect to a single aspect of processing). Our approach uses
an evidence accumulation model, LBA (Brown & Heathcote, 2008), as a basic building block, and offers a principled way for
combining a number of these building blocks to construct a system capable of complex decisions. In this paper we used com-
binations that form coactive and independent logical-rule process models, using either OR or AND rules, but future studies
could form almost an endless number of combinations to suit many tasks and conditions.
The approach we presented offers a number of advantages over alternatives: (i) simultaneous fitting of OR and AND deci-
sions (given the right empirical data), (ii) consideration of speed-accuracy trade-off and, related, (iii) decomposition of com-
plex decisions to latent variables, such as evidence accumulation rate, bias, and non-decision processes. Finally, (iv) it
allowed a rule-breaking parameter, to indicate whether participants were conforming to the tasks demands. The results
with respect to the latter were quite marked. Participants in the data we analysed consistently ignored the rule dictated by
the OR task. Eidels et al.s (2015) dot-detection task requires divided attention to two spatial locations, each of which could
contain a target. Yet, with some frequency, participants processed only one location and not the other, despite the fact such
behavior leads to reduced accuracy.
The latter result has profound theoretical implications for perceptual judgments and models of complex decision-making.
The prevalent view in simple decision-making is that errors are intrinsic to the decision process. In evidence accumulation
models, such as the LBA or Ratcliffs (1978) diffusion model, errors occur when a noisy accumulation process reaches thresh-
old associated with an incorrect response. The exact characteristics of these errors, and how models could produce errors
that are either slower than correct responses on some instances or faster on other instances had been a matter of close scru-
tiny (e.g., Ratcliff & Rouder, 1998). Our modelling results reveal an interesting phenomenon, another source of errors driven
by participants not fully conforming to the task instructions. Such failures in processing have been noted in other complex
decision tasks (e.g., Failure to Engage in task switching: de Jong, 2000; Poboka, Karayanidis, & Heathcote, 2014). The
broader implication is, therefore, that such failures should be more widely considered in extending models of simple deci-
sion making to more complicated choices.

Acknowledgments

This research was funded in part by the NSF Brain-Body-Environment Systems IGERT grant to Randy Beer and in part by
the Faculty Research Support Program, administered through the Office of the Vice President of Research at Indiana
 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116 15

University, Bloomington, and an Australian Research Council (ARC) Professorial Fellowship for Heathcote and ARC Discovery
Project grants for Eidels and Heathcote.

Appendix A. Supplementary material

Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.cog-
psych.2017.03.001.

References

Alais, D., & Burr, D. (2004). No direction-specific bimodal facilitation for audiovisual motion detection. Cognitive Brain Research, 19(2), 185194.
Algom, D., Eidels, A., Hawkins, R. X., Jefferson, B., & Townsend, J. T. (2015). Features of response times: Identification of cognitive mechanisms through
mathematical modeling. The Oxford Handbook of Computational and Mathematical Psychology, 6398.
Bogacz, R., Brown, E., Moehlis, J., Holmes, P., & Cohen, J. D. (2006). The physics of optimal decision making: A formal analysis of models of performance in
two-alternative forced-choice tasks. Psychological Review, 113, 700765.
Brown, S. D., & Heathcote, A. (2008). The simplest complete model of choice response time: Linear ballistic accumulation. Cognitive Psychology, 57, 153178.
Bushmakin, M. A., & James, T. W. (2014). The influence of feature conjunction on object inversion and conversion effects. Perception, 43(1), 3142.
Cousineau, D., & Shiffrin, R. M. (2004). Termination of a visual search with large display size effects. Spatial Vision, 17(4), 327352.
de Jong, R. (2000). An intention-activation account of residual switch costs. In S. Monsell & J. Driver (Eds.), Control of cognitive processes: Attention and
performance XVIII (pp. 357376). Cambridge, MA: MIT Press.
Donkin, C., Brown, S. D., & Heathcote, A. (2009). The overconstraint of response time models: Rethinking the scaling problem. Psychonomic Bulletin & Review,
16(6), 11291135.
Eidels, A., Donkin, C., Brown, S. D., & Heathcote, A. (2010a). Converging measures of workload capacity. Psychonomic Bulletin & Review, 17, 763771.
Eidels, A., Townsend, J. T., & Algom, D. (2010b). Comparing perception of Stroop stimuli in focused versus divided attention paradigms: Evidence for
dramatic processing differences. Cognition, 114(2), 129150.
Eidels, A., Townsend, J. T., Hughes, H. C., & Perry, L. A. (2015). Evaluating perceptual integration: Uniting response-time-and accuracy-based methodologies.
Attention, Perception, & Psychophysics, 77, 659680.
Fific, M., & Buckmann, M. (2013). Stopping Rule Selection (SRS) theory applied to deferred decision making. In M. Knauff, M. Pauen, N. Sebanz, & I.
Wachsmuth (Eds.), Cooperative minds: Social interaction and group dynamics. Proceedings of the 35th annual conference of the cognitive science society
(pp. 22732278). Austin, TX: Cognitive Science Society.
Fific, M., Little, D. R., & Nosofsky, R. M. (2010). Logical-rule models of classification response times: A synthesis of mental-architecture, random-walk, and
decision-bound approaches. Psychological Review, 117(2), 309348.
Fific, M., Nosofsky, R. M., & Townsend, J. T. (2008). Information-processing architectures in multidimensional classification: A validation test of the systems
factorial technology. Journal of Experimental Psychology: Human Perception and Performance, 34(2), 356375.
Fific, M., Townsend, J. T., & Eidels, A. (2008). Studying visual search using systems factorial methodology with targetDistractor similarity as the factor.
Perception & Psychophysics, 70, 583603.
Heathcote, A., Brown, S. D., & Wagenmakers, E.-J. (2014). The falsifiability of actual decision-making models. Psychological Review, 121, 676678.
Heathcote, A., Coleman, J. R., Eidels, A., Watson, J. M., Houpt, J., & Strayer, D. L. (2015). Working memorys workload capacity. Memory & Cognition, 43(7),
973989.
Heathcote, A., & Love, J. (2012). Linear deterministic accumulator models of simple choice. Frontiers in Psychology, 3.
Houpt, J. W., & Townsend, J. T. (2010). The statistical properties of the Survivor Interaction Contrast. Journal of Mathematical Psychology, 54(5), 446453.
Jones, M., & Dzhafarov, E. N. (2014). Unfalsifiability and mutual translatability of major modeling schemes for choice reaction time. Psychological Review, 121
(1), 132.
Leite, F. P., & Ratcliff, R. (2010). Modeling reaction time and accuracy of multiple-alternative decisions. Attention, Perception, & Psychophysics, 72(1), 246273.
Little, D. R., Nosofsky, R. M., & Denton, S. E. (2011). Response-time tests of logical-rule models of categorization. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 37(1), 1.
Little, D. R., Nosofsky, R. M., Donkin, C., & Denton, S. E. (2013). Logical rules and the classification of integral-dimension stimuli. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 39(3), 801820.
Logan, G. D., Van Zandt, T., Verbruggen, F., & Wagenmakers, E.-J. (2014). On the ability to inhibit thought and action: General and special theories of an act of
control. Psychological Review, 121(1), 6695.
McGurk, H., & MacDonald, J. (1976). Hearing lips and seeing voices. Nature, 264(5588), 746748.
Miller, J. (1978). Multidimensional samedifferent judgments: Evidence against independent comparisons of dimensions. Journal of Experimental
Psychology: Human Perception and Performance, 4, 411422.
Miller, J. (1982). Divided attention: Evidence for coactivation with redundant signals. Cognitive Psychology, 14, 247279.
Mittner, M., Boekel, W., Tucker, A., Turner, b., Heathcote, A., & Forstmann, B. U. (2014). When the brain takes a break: A model-based analysis of mind
wandering. Journal of Neuroscience, 34(49), 1628616295.
Moneer, S., Wang, T., & Little, D. R. (2016). The processing architectures of whole-object features: A logical-rules approach. Journal of Experimental
Psychology: Human Perception and Performance, 42, 14431465.
Mordkoff, J. T., & Danek, R. H. (2011). Dividing attention between color and shape revisited: Redundant targets coactivate only when parts of the same
perceptual object. Attention, Perception, & Psychophysics, 73, 103112.
Mordkoff, J. T., & Yantis, S. (1991). An interactive race model of divided attention. Journal of Experimental Psychology: Human Perception and Performance, 17
(2), 520.
Mulligan, R. M., & Shaw, M. L. (1980). Multimodal signal detection: Independent decisions vs. integration. Perception & Psychophysics, 28, 471478.
Poboka, D., Karayanidis, F., & Heathcote, A. (2014). Extending the failure-to-engage theory of task switch costs. Cognitive Psychology, 72, 108141.
R Core Team (2015). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. <http://www.R-project.
org/>.
Rae, B., Heathcote, A., Donkin, C., Averell, L., & Brown, S. (2014). The hare and the tortoise: Emphasizing speed can change the evidence used to make
decisions. Journal of Experimental Psychology: Learning, Memory, and Cognition, 40(5), 12261243.
Ratcliff, R. (1978). A theory of memory retrieval. Psychological Review, 85(2), 59108.
Ratcliff, R., & Rouder, J. N. (1998). Modeling response times for two-choice decisions. Psychological Science, 9, 347356.
Schmiedek, F., Oberauer, K., Wilhelm, O., S, H.-M., & Wittmann, W. W. (2007). Individual differences in components of reaction time distributions and
their relations to working memory and intelligence. Journal of Experimental Psychology: General, 136(3), 414429.
Starns, J. J., Ratcliff, R., & McKoon, G. (2012). Evaluating the unequal- variability and dual-process explanations of zROC slopes with response time data and
the diffusion model. Cognitive Psychology, 64, 134.
Sternberg, S. (1966). High-speed scanning in human memory. Science, 153, 652654.
Sternberg, S. (1969). The discovery of processing stages: Extensions of Donders method. Acta Psychologica, 30, 276315.
16 M.A. Bushmakin et al. / Cognitive Psychology 95 (2017) 116

Stevenson, R. A., Ghose, D., Fister, J. K., Sarko, D. K., Altieri, N. A., Nidiffer, A. R., ... Wallace, M. T. (2014b). Identifying and quantifying multisensory
integration: A tutorial review. Brain Topography, 27(6), 707730.
Stevenson, R. A., Siemann, J. K., Woynaroski, T. G., Schneider, B. C., Eberly, H. E., Camarata, S. M., & Wallace, M. T. (2014a). Evidence for diminished
multisensory integration in autism spectrum disorders. Journal of Autism and Developmental Disorders, 44(12), 31613167.
Terry, A., Marley, A. A. J., Barnwal, A., Wagenmakers, E.-J., Heathcote, A., & Brown, S. D. (2015). Generalising the drift rate distribution for linear ballistic
accumulators. Journal of Mathematical Psychology, 68, 4958.
Townsend, J. T., & Altieri, N. (2012). An accuracyresponse time capacity assessment function that measures performance against standard parallel
predictions. Psychological Review, 119, 500516.
Townsend, J. T., & Ashby, F. G. (1983). Stochastic modeling of elementary psychological processes. Cambridge University Press.
Townsend, J. T., & Colonius, H. (1997). Parallel processing response times and experimental determination of the stopping rule. Journal of Mathematical
Psychology, 41(4), 392397.
Townsend, J. T., & Eidels, A. (2011). Workload capacity spaces: A unified methodology for response time measures of efficiency as workload is varied.
Psychonomic Bulletin & Review, 18(4), 659681.
Townsend, J. T., Houpt, J. W., & Silbert, N. H. (2012). General recognition theory extended to include response times: Predictions for a class of parallel
systems. Journal of Mathematical Psychology, 56, 476494.
Townsend, J. T. (1974). Issues and models concerning the processing of a finite number of inputs. In B. H. Kantowitz (Ed.), Human information processing:
Tutorials in performance and cognition. Potomac, Md: Erlbaum.
Townsend, J. T., & Nozawa, G. (1995). Spatio-temporal properties of elementary perception: An investigation of parallel, serial, and coactive theories. Journal
of Mathematical Psychology, 39(4), 321359.
Van Zandt, T., Colonius, H., & Proctor, R. W. (2000). A comparison of two response time models applied to perceptual matching. Psychonomic Bulletin &
Review, 7(2), 208256.
Wagenmakers, E. J., & Farrell, S. (2004). AIC model selection using Akaike weights. Psychonomic Bulletin & Review, 11(1), 192196.
Williams, L. E., Light, G. A., Braff, D. L., & Ramachandran, V. S. (2010). Reduced multisensory integration in patients with schizophrenia on a target detection
task. Neuropsychologia, 48(10), 31283136.
Wuerger, S. M., Hofbauer, M., & Meyer, G. F. (2003). The integration of auditory and visual motion signals at threshold. Perception & Psychophysics, 65(8),
11881196.