Article Critique
bretmorse27
Article1.pdf
How Does the Emotional Experience Evolve? Feeling Generation as Evidence Accumulation
Ella Givon, Ayelet Itzhak-Raz, Anat Karmon-Presser, Gal Danieli, and Nachshon Meiran Ben-Gurion University of the Negev
How do people answer the question “How do you feel?” In the present work, participants were given 2 tasks in each trial. They first indicated whether a picture made them feel pleasant (or was supposed to be felt as pleasant, in another group), and then made gender decisions regarding faces. Evidence accumulation modeling showed that (a) reporting genuine feeling is qualitatively different from reporting the supposed feeling; (b) reporting one’s feeling is remarkably similar to gender decisions; and (c) evidence regarding negative feelings accumulates more quickly than in positive feelings. These results support the assumption that when asked, participants report genuine as opposed to supposed feelings and strengthen the analogy between feeling reports and perceptual decisions.
Keywords: emotional feelings, evidence-accumulation modeling, valence
You drive to work, expecting a very important meeting you planned months ahead, and suddenly realizing that the cars ahead of you are slowing down into what initially seems like a heavy traffic jam that would certainly make you miss the meeting. Your heart pounds heavily and you start thinking that it would be a disaster if this meeting would be missed. Your partner calls and asks if you feel good and you immediately reply “not at all.”
In the current work, we aimed to understand how this report of feeling emerges. In other words, we asked how people become aware (conscious) of their emotional experience. We did so by extending previous work from our lab (Karmon-Presser, Sheppes, & Meiran, 2018). In that work, the authors continued theoretical suggestions (James, 1884; Sokolov & Boucsein, 2000; Thagard & Aubie, 2008) which draw an analogy between feeling generation and perceptual decision. Karmon-Presser, Sheppes, and Meiran (2018) extended these suggestions by testing whether a well- established model of perceptual decisions, signal-detection theory (SDT; Macmillan & Creelman, 2004) is an appropriate model for emotional-feeling reports. According to this theorizing, evidence regarding emotion is noisy and (conscious-reportable) feelings emerge once the evidence is sufficiently strong and crosses a decision boundary. However, Karmon-Presser et al. (2018) used SDT, and this model does not take into account the time domain,
namely the time taken to complete the perceptual decision and the rate by which emotional evidence accumulates. We reasoned that if feeling reports truly resemble perceptual decisions, this analogy between emotional feeling and perception should also be reflected in the temporal domain.
Evidence accumulation models (EAMs; Ratcliff & Smith, 2004; Teodorescu & Usher, 2013) fill this gap by explaining how per- ceptual decisions unfold in time. The core assumption of EAMs is that the representation of stimuli is inherently noisy, such that in order to make a decision, one must repeatedly sample evidence regarding the stimulus. A perceptual decision is made once suffi- cient evidence has accumulated. EAMs explain both accuracy and the precise shape of the reaction time (RT) distributions for correct and incorrect answers and as such are considerably more con- strained than SDT. In the following sections, we review literature related to the main areas which are dealt with in this work includ- ing the emotional subjective experience and EAMs.
Feeling: The Conscious Emotional Subjective Experience
Following past research (e.g., Scherer, 1987), some emotion researchers define emotion by their components. Among the com- ponents of emotions are autonomic changes, facial and bodily expressions, thought and action tendencies, and cognitive apprais- als. A highly important component of emotion is the associated conscious experience—the feeling. Here we specifically focus on the awareness of the emotion, for example, being aware of the fact that I feel bad, as reflected in feeling reports.
One cannot overestimate the importance of conscious emotional feelings: Feelings indicate personal relevance (Ekman & David- son, 1994), guide decision making (Clore, Gasper, & Garvin, 2001), and help in understanding the experience of others (Lambie & Marcel, 2002). Additionally, feeling is strongly related to emo- tion regulation, considering that one’s ability to choose optimally which emotion regulation strategy to use is highly depended on one’s ability to recognize her own emotion (Barrett, Gross, Chris-
This article was published Online First March 7, 2019. Ella Givon, Ayelet Itzhak-Raz, Anat Karmon-Presser, Gal Danieli, and
Nachshon Meiran, Department of Psychology and Zlotowski Center for Neuroscience, Ben-Gurion University of the Negev.
This research was supported by a research grant from the Israel Science Foundation (381/15). We thank Assaf Kron and Oxana Itekes for providing the instructions used in Itkes et al. (2017), and Maya Weismann for her help in data collection.
Correspondence concerning this article should be addressed to Ella Givon, Department of Psychology and Zlotowski Center for Neurosci- ence, Ben-Gurion University of the Negev, Beer-Sheva, Israel. E-mail: [email protected]
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
Emotion © 2019 American Psychological Association 2020, Vol. 20, No. 2, 271–285 1528-3542/20/$12.00 http://dx.doi.org/10.1037/emo0000537
271
tensen, & Benvenuto, 2001). The ability to feel varies among people. Evidence for individual differences in the ability to feel can be found in research concerning feeling-related constructs such as emotional awareness (Lane & Schwartz, 1987), emotion differentiation (Barrett et al., 2001), and affective clarity (Lisch- etzke, Cuccodoro, Gauger, Todeschini, & Eid, 2005).
In the current work, we follow the dimensional approach and focus on the valence dimension (e.g., Barrett, 2006; Kron, Gold- stein, Lee, Gardhouse, & Anderson, 2013). Nevertheless, we ac- knowledge the fact that this approach has been criticized (e.g., Zachar & Ellis, 2012), and thus adopt it only as a starting point. Additionally, while using established norms of valence from the Nencki Affective Picture System (NAPS) database (Marchewka, Żurawski, Jednorǒg, & Grabowska, 2014), we compared negative and positive valence as being part of a bipolar scale (Feldman Barrett & Russell, 1998). While our norms are driven from a bipolar scale, our procedure involves a yes/no report and, as such, does not require a full commitment to the bipolar approach. Nev- ertheless, we acknowledge the fact that the choice to use a bipolar scale is also controversial (e.g., Cacioppo & Berntson, 1994). Lastly, since the construct of the emotional experience is charac- terized by heated debates such as the aforementioned ones, we decided to use what seems to be the simplest components of feeling (i.e., valence).
Evidence Accumulation Models (EAMs)
EAMs, whose roots are already half a century old, have become popular in recent years (Bogacz, Brown, Moehlis, Holmes, & Cohen, 2006; Bogacz, Usher, Zhang, & McClelland, 2007; Brown & Heathcote, 2008; Ratcliff & Rouder, 1998; Ratcliff & Smith, 2004; Stone, 1960; Usher & McClelland, 2001; van Ravenzwaaij, Mulder, Tuerlinckx, & Wagenmakers, 2012). These models, as- suming that the evidence for decisions is noisy, can thus be viewed as an extension of SDT. However, in contrast to SDT, EAMs take into account the time domain. EAMs do so by using accuracy and RT data to uncover psychological processes that underlie decision making. All EAMs share a fundamental principle of sequential sampling of evidence regarding choice alternatives. This sampling resolves the noise-related uncertainty through averaging external and internal evidence over time, until the amount of evidence favoring a given choice reaches a threshold. This noise reduction and reliability increase is analogous to how reliability of a psy- chological test score increases when the number of test items increases. In this regard, pieces of evidence can be viewed as test items. EAMs may be viewed as a robot that has controllers (parameters) that specify its behavior. This robot generates re- sponses and RTs for those responses. When “fitting” an EAM, we essentially “play with” the controllers until its responses and RTs closely resembles the data of responses and RTs from actual participants. When resemblance is sufficiently high, we conclude that actual participants probably process the task in the same way that the robot does. This is sometimes referred as “reverse engi- neering.”
Among the various EAMs, we chose to use the linear ballistic accumulator model (LBA; Brown & Heathcote, 2008) in the current study (the reasons for choosing this specific model are discussed in the following section). As we view it (see Sternberg & Backus, 2015), all EAMs, including the LBA, treat RT as
combined of three processing stages: an early processing stage in which low-level features are extracted, a decision stage, and a motor preparation stage in which the decision leads to the pro- gramming of the motor response (see Figure 1a). The core of the model is the elaborated description of the decision stage (see further elaboration below). To get its gist, we will use a sink-filling contest metaphor1 (see Figure 1b) according to which there are two sinks, each being filled with water at a constant rate by a faucet. The decision is based on a race between the sinks in which the sink to become full first is winning. This race begins when there is already water in the sinks.
The generic LBA has five parameters, which are illustrated in Figure 1c in relation to our experimental paradigm. In our para- digm, participants were requested to make a binary choice regard- ing what they feel.2 Accordingly, the model describes the decision by assigning one evidence accumulator (a “sink with its faucet”) for each response (“pleasant” or “unpleasant”). These accumula- tors gather evidence continuously (“filled with water”) until one of them reaches a threshold (which is analogous to the size of the sink) and takes control over the response. The decision stage is described mainly by two key parameters: the first is the drift-rate (v), the mean rate of evidence accumulation (power of the water flow), related to signal-to-noise ratio. Drift-rate represents the quality of emotional information derived from the stimulus. LBA assumes that per trial, the evidence accumulation rate is constant. Accordingly, drift-rate is defined as the slope of evidence as a function of time (see arrows in Figure 1c), with a steeper slope indicating a higher accumulation rate. The quality of the informa- tion derived from the stimulus depends jointly on the objective properties of the stimulus (e.g., its strength) and on the efficacy of the perceptual system (Ratcliff & Smith, 2004). The second pa- rameter is the response threshold (b), which is the total amount of evidence that is required to make a decision (the dashed line in Figure 1c, and the size of the sinks in Figure 1b). The total amount of evidence required for a decision is mainly controlled by the participant and reflects speed–accuracy trade-off. Specifically, a high threshold (large sink) reflects demand for a large amount of evidence and it accounts for a cautious response style. Concomi- tantly, a low threshold reflects demand for a small amount of evidence and it accounts for a hasty response style. Evidence accumulation does not necessarily begin from zero, because accu- mulators might have some initial evidence before accumulation starts (the amount of water in the sink before the race begins), reflecting a bias in favor of corresponding choice. This is repre- sented in Figure 1c by the fact that the arrows (i.e., accumulators) do not start from zero. The amount of initial evidence is repre- sented by the starting-point (SP) parameter, which is the highest possible amount of initial evidence present in the accumulator before evidence begins to accumulate. In other words, the amount of initial evidence in the accumulators is between zero and SP, and these values are assumed to come from a rectangular distribution (see the gray rectangle in Figure 1c). Two additional parameters are SV, the standard deviation of the normal distribution from which the drift-rate in a given trial is drawn, and the (constant) nondecision time, which refers to the time taken for other pro-
1 We thank Ami Eidels for suggesting this metaphor. 2 LBA can deal with n-choices and is not limited to binary choices.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
272 GIVON, ITZHAK-RAZ, KARMON-PRESSER, DANIELI, AND MEIRAN
cesses that are not a part of the decision itself (i.e., low-level feature extraction and motor preparation, see Figure 1a). Finally, the overt response (i.e., the chosen key press) is determined by the accumulator to first reach the threshold, and the RT is the decision- time, (Threshold – [SP/2]) / Drift_Rate, plus the nondecision time.
Why We Chose the LBA?
One classification of EAMs is into two groups: competitive versus independent (e.g., Teodorescu & Usher, 2013). In compet-
itive models, the response threshold is based on a comparison between accumulators (e.g., max-minus next; Krajbich & Rangel, 2011) or on a shared accumulator for all responses as in the very popular drift diffusion model (DDM; Ratcliff & McKoon, 2008). In both, the max-minus next and the drift diffusion model, evi- dence that strengthens the probability for Response A necessarily weakens the probability for Response B. Independent (race) mod- els assign a separate evidence accumulator for each response alternative and these accumulators run independently toward a
Figure 1. (a) Processing stages in the linear ballistic accumulator (LBA), with the first and last stages represented by the nondecision time parameter and all three add up to reaction time (RT). (b) A sink-filling contest metaphor describing the decision stage. Each sink is filled with water at a different rate, and the contest begins when sinks already have water in them. The first sink to become full is the winning sink. (c) A more formal description of the decision stage (X-axis represents time and Y-axis represents evidence; i.e., the current amount of water in the sink). Panel c represents an exemplar trial with two accumulators (sinks), one for each of “pleasant” and “unpleasant” response. In a given trial, the initial level of evidence in each accumulator (the initial amount of water in the sink) is randomly drawn from a rectangular distribution ranging between zero and starting-point (SP). Evidence accumulator rate (the slope of the arrows; i.e., the power of the water flow) is randomly drawn in each trial from a normal distribution with � � drift-rate and � � SV. The first arrow to cross threshold (the dashed line, representing the total amount of evidence required to make a decision; i.e., the size of the sink) takes control over the response. Panel c is an adaptation of Figure 3 “The simplest complete model of choice response time: Linear ballistic accumulation,” by S. D. Brown and A. Heathcote, 2008, Cognitive Psychology, 57, pp. 153–178. Copyright 2007 by Elsevier Inc. Adapted with permission. See the online article for the color version of this figure.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
273HOW DOES THE EMOTIONAL EXPERIENCE EVOLVE?
common threshold and not interacting with one another. There is a growing number of studies using the (independent) LBA model. Donkin, Brown, Heathcote, and Wagenmakers (2011) found that although the drift diffusion model, a competitive model, and the LBA, an independent model, rely on different frameworks, they lead to very similar conclusions. Because the LBA is a simpler (theoretically and practically) and because of the sufficient simi- larity between the models, we chose to use it in the current study.
LBA and Feeling Reports
In relation to feeling reports, drift-rate can be viewed as repre- senting (a) the emotional intensity of the stimulus; (b) the respon- siveness of the emotional systems to that stimulus (e.g., to what degree the heart-rate or cognitive evaluations change as a result of exposure to the stimulus); and (c) participants’ ability to recognize these internal changes. Emotional evidence comprises detection of autonomic reactivity, posture, facial expression, cognitive evalua- tions, and action tendencies (see Figure 2a). There are trial-to-trial changes in drift-rate, resulting in variability which is described by the SV parameter. SV is related to stimulus variance (i.e., not all the pleasant stimuli are equally pleasant), and to variability in responsiveness (i.e., the same stimulus might be experienced as more or less pleasant at different occasions), among other factors. The emotional evidence adds to preexisting evidence in the accu- mulators (related to the SP parameter) and it accumulates contin- uously until its amount exceeds the value of the threshold param- eter. Regarding our paradigm, emotional stimuli are complex pictures, requiring considerable low-level (nonemotional) percep- tual processing. In addition, making a key-press response requires motor programing. The duration of these two processes sums up to nondecision time.
LBA as all EAMs is fitted to the proportion of errors and to the RT distributions for correct responses and errors. In this regard, it is important to specify how we define errors in an emotional task. The usual applications of EAMs are in tasks in which there are objectively defined stimulus attributes, and accordingly an objec- tive correct answer. We instead used the published norms (NAPS; Marchewka et al., 2014) as a substitute of the objectively correct answer. Our rationale was based on theories that view the emo-
tional reaction as being related to prediction (Bach & Dayan, 2017; Barrett, 2017). To illustrate this issue, consider the discrepancy between two types of hotel ratings: the objective rating (known as the stars system) and the normative subjective rating often used in online booking websites. From our impression, many people are more likely to trust average ratings than they trust the objective stars-grade of the hotel. We suggest that the average rating of others’ experience may be perceived as more informative and a better predictor of their future satisfaction from the hotel. This observation is somewhat related to the “wisdom of the crowd” phenomenon (Galton, 1907), suggesting that the average estimate taking from many independent observers is remarkably precise. Nonetheless, we take a cautious approach in our usage of norms by creating a rough division of the stimuli into two disparate groups (negative and positive). While doing so, accuracy is not measured literally as rating the exact value of the norm, but rather a basic classification of the stimuli into one out of two well-separated valence-groups. We define “feeling errors” as responses that are not in line with the normative valence ratings (i.e., normatively negative stimuli judged as pleasant and normatively positive stim- uli judged as unpleasant).
A reasonable objection would be that feelings cannot be mea- sured objectively or intersubjectively, given their deeply subjective nature. We have two (admittedly imperfect) replies. The first, and the crucial one, concerns with the way we think of many types of feelings, as being related to the interpersonal. This certainly holds true for anger, shame, and so forth, but is also true for seemingly intrapersonal feelings such as disgust, because the outer expression of disgust can alarm against the threat of pathogens (Schaller & Park, 2011). Appreciating the reciprocity between feelings and interpersonal relationship (e.g., Lambie & Marcel, 2002; Lazarus, 2006), relational meanings determine which emotions are experi- enced and displayed in social interactions. If the significance of emotions is mainly expressed through the relational domain, then one being accordant to based-on-others emotional norms would result in better functioning, well-being, and interpersonal relation- ship. We thus suggest that “feeling errors” would be described as follows: It is not that one does not know what s/he feels, it is that what s/he feels is intersubjectively wrong. In fact, the approach
Figure 2. A schematic demonstration of feeling generation as a process of emotional evidence accumulation. The left panel (a) is based on Figure 1 (a, b) from Karmon-Presser et al. (2018) and represents the different types of emotional evidence that can arise in response to an external stimulus. The right panel (b) represents the accumulation of emotional evidence until (when hitting the decision threshold), a conscious emotional experi- ence (“feeling”) comes to be.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
274 GIVON, ITZHAK-RAZ, KARMON-PRESSER, DANIELI, AND MEIRAN
that feelings may be wrong is now more widely accepted (e.g., Mayer, Salovey, Caruso, & Sitarenios, 2003; Oishi, Diener, & Lucas, 2007; Rottenberg & Gross, 2003). This approach is also reflected in how psychopathology is defined. For example, a person suffering from claustrophobia truly experiences intense fear in closed spaces, but we would regard this reaction as dispropor- tionate. Our second response is related to statistical issues as being reflected in our paradigm. A person might experience what may be seen as a peculiar emotion such as liking venomous snakes. How- ever, if this peculiarity is unsystematic, and is not seen in every domain, the statistics associated with using a large enough stimu- lus set would cancel it out. It is only when the peculiarity is pervasive that it would reflect in our estimates (and we suggest, rightfully so).
The Current Study
The scenario described in the beginning of our article includes various components of the emotional state, such as autonomic changes (heartbeats) and cognitive appraisals (it would be a disas- ter if this meeting would be missed). In this work, we focus on the conscious emotional experience (“I do not feel good at all”). We are even more specific than that, because the conscious experience may be evaluations of the world (e.g., “The world is such a terrible place;” Lambie & Marcel, 2002) and we suggest a process model which describes only the summary evaluation of one’s emotional state reflected for example, in answering the question “How do you feel?” We rely on Oxford’s dictionary definition of “process” as a series of actions or steps taken in order to achieve a particular end. Accordingly, we propose that the generation of a feeling is a three-step process beginning with early perceptual processes, con- tinuing with evidence accumulation and ending with motor prep- aration (see Figure 1 and also e.g., Sternberg & Backus, 2015). The LBA model which we adopted provides a detailed description of Step 2—that is, evidence accumulation. In the present work, we follow Satpute, Shu, Weber, Roy, and Ochsner (2013), who ac- knowledged the importance of studying the mechanisms underly- ing reporting emotional states (“I feel good!”). However, while Satpute et al. (2013) examined the underlying functional neuro- anatomy, the aim of the current study was to describe the process itself. Lastly, we study emotion reports as a proxy for studying feelings because, as pointed out by Block (2005), reports about one’s conscious experience are commonly considered as the “gold standard” of consciousness, the aspect that can be empirically studied. We elaborate on this issue in the Discussion section.
Recent research recognized the similarity between emotion and cognition, being nonseparable mental processes (e.g., Duncan & Barrett, 2007). Likewise, we propose to address feeling generation as being a close analogue of perceptual decision-making and analyze it accordingly. In detail, in order for a person to be aware of a certain feeling, s/he is required to “look inward” and collect information such as bodily sensations, cognitions, action drives, and bodily expressions. Using EAM terms, one is accumulating samples of “in-self information,” until reaching sufficient emo- tional evidence and making a decision—which feeling one has (see Figure 2 for a graphic illustration of the proposed process).
Previous research has already applied EAM in studying clinical phenomena and emotional-related constructs (e.g., Pe, Vandeker- ckhove, & Kuppens, 2013; White, Ratcliff, Vasey, & McKoon,
2009; White, Ratcliff, Vasey, & McKoon, 2010; White, Skokin, Carlos, & Weaver, 2016). Perhaps the most closely related work is by White, Skokin, Carlos, and Weaver (2016; see also White, Liebman, & Stone, 2018). In that work, participants were asked to classify words as threatening (made them feel worried) or not, and RT and accuracy data were modeled using both SDT and drift diffusion model. However, it is difficult to tell whether participants report their genuine feelings as opposed to telling what they believe to be the valence of the stimulus. A recent article by Itkes, Kimchi, Haj-Ali, Shapiro, and Kron (2017) highlights this point. In that work, the authors empirically dissociated between affective and semantic valence. Affective valence concerns report of one’s own feeling while semantic valence concerns report about the value of the stimulus. Additionally, while White et al. (2016) assessed model fit, it is unclear if EAMs (drift diffusion model being one example) are as suitable for emotional reports as they are for perceptual decisions.
In the present work, we tried to address these two questions. First, following Itkes et al. (2017), we compared between reports concerning one’s own emotion and semantically based reports regarding the normative emotional response. A significant differ- ence between these two report types would support the assumption that, when asked, participants report their genuine own feelings. Second, we validated the use of EAM for emotional tasks through a comparison with a benchmark perceptual decision task. This comparison allows us to conclude that the model fit of an emo- tional experience task is remarkably similar to that in the bench- mark task, thus supporting the analogy between own feeling re- ports and perceptual decisions. Additionally, we provide an initial validation of the model itself by testing for selective influence on the drift-rate parameter (e.g., Rae, Heathcote, Donkin, Averell, & Brown, 2014). Last, we used the LBA to compare the dynamics of negative and positive feelings.
Experiment 1
The design of the present experiment was informed by two pilot experiments. To deal with the question regarding whether partic- ipants report genuine versus expected feelings, we randomly as- signed participants to one of two groups, differing solely in the instructions regarding how to respond to the emotional stimuli. While participants in the self-focused group were instructed to response only according to what they feel (“Does this photo makes you feel pleasant?”), participants in the stimulus-focused group were instructed to decide what they think is the appropriate feeling to feel in response to the stimulus, or how most of the people would respond to it (“Is this photo supposed to create a pleasant feeling?”). Meaningful differences between the groups would pro- vide support for the claim that when participants report own feelings (the self-focused group) they employ different processing than when they report expected feelings. Such evidence supports the hypothesis that they report own feelings, although it does not prove it.
To deal with our second question whether EAM is a suitable model for feeling generation, we created a nonemotional two- choice RT task which is widely modeled with EAM and used it as a benchmark. The nonemotional task, the faces task, was chosen so that it would yield similar RTs and error rates as the emotional task. We then asked whether the two tasks can be described by
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
275HOW DOES THE EMOTIONAL EXPERIENCE EVOLVE?
similar EAM parameter values? Similar parameters for the two tasks would support our analogy between feeling generation and perceptual decision making.
Our experimental approach was to let participants decide if a stimulus generates a pleasant feeling. We masked the goal of the experiment to encourage participants in the self-focused condition to report their true feelings and discourage them reporting what the stimulus should make one feel. Accordingly, the goal of the experiment was described as one focusing on the influence of emotions on the faces task, which was regarded as the main task. Specifically, in each trial, we presented an emotional stimulus before a neutral stimulus (from the faces task) that requires a “classical” perceptual decision.
Method
Participants. Fifty-Nine Israeli ungraduated students of Ben- Gurion University (38 females, mean age: 23.35) participated in the experiment for course credit. The research was approved by the department’s ethics committee. All participants declared being native speakers of Hebrew and having normal or corrected-to- normal vision. We assured that participants understood the instruc- tions properly by asking each participant during debriefing to explain in his or her words what he/r were asked to do. Further- more, all participants reported believing that the faces task was the main task, providing evidence that the masking of the goal of the experiment worked as expected.
Emotional task. Participants viewed 200 emotion-eliciting photos with established norms (NAPS; Marchewka et al., 2014). We ensured that there would be equal representation of various content-categories (animals, objects, faces, people, and landscapes). Unbeknown to the participants, the photos were taken from two different categories according to their valence (slightly negative, va- lence range: 3.5– 4.5, Mean(negative) � 4.01, SD(negative) � .301; positive, valence range: 6.5–7.5, Mean(positive) � 6.94, SD(positive) � .273). Under each photo appeared a yes/no question regarding the pleasantness feeling (i.e., whether the stimulus elicits a pleasant feel- ing?). Accuracy and RTs were collected, and errors were defined as responding “yes” to a normatively negative photo or responding “no” to a normatively positive photo. RT was defined as the time from picture presentation until the key press.
Faces task. Participants viewed 200 pictures of male and female faces, evenly divided between genders. The face database was provided by the Max Planck Institute for Biological Cyber- netics in Tuebingen, Germany (Blanz & Vetter, 1999; Troje & Bülthoff, 1996). Participants were required to decide whether the face is of a female or a male. Accuracy and RTs were recorded, and errors were defined as responding “male” to a picture from the female category and vice versa. RT was defined as the time from picture presentation until the key press. Formerly research has stated the application of EAM to various simple two-choice RT tasks, were one is asked to discriminate perceptual stimuli and name to which response category they belong (Brown & Heath- cote, 2005; Ratcliff & McKoon, 2008; Usher & McClelland, 2001).
Procedure. The experiment was programmed in OpenSesame (Mathôt, Schreij, & Theeuwes, 2012), was run on computers with 17” screens, and lasted approximately 30 min. Participants were randomly assigned to one of two conditions: self-focused or
stimulus-focused. We created a modified version of the instruc- tions that were previously used to dissociate affective and semantic valence (Itkes et al., 2017). The experiment was identical for both groups except for the instructions and the question that appeared below the emotional photos (“Does this photo make you feel pleasant?” for the self-focused condition; “Does this photo sup- posed to create a pleasant feeling?;” for the stimulus-focused condition). Participants responded by pressing the “l”/“k” keys (“no”/“yes”) in the emotional task, and the “s”/“d” keys (“male”/ “female”). Which pair of keys went with each task was counter- balanced across participants.
The following description is relevant for all participants, across groups. The experimenter explained the instructions to the partic- ipant and stayed in the room during a short training phase (five trials) to ensure the participant understands the tasks, and then left the room. Participants performed five equivalent blocks of trials. Blocks began with an instruction reminder screen. Each block consisted of 40 trials. A trial consisted of two sequences: a se- quence for the emotional task consisting of a fixation screen lasting 995 ms (wide or long empty frame, congruent to the shape and size of the emotional photo), followed by an emotional photo with the pleasant feeling question at the bottom of the screen. The display remained visible until a response was made. Then, there was a similar sequence, now with a face stimulus (see Figure 3). After finishing the experiment, participants were shortly debriefed and released.
Independent variables. We created an additional variable for the emotional task, which we named “intensity.” To this end, we divided the emotional stimuli into two categories (across positive and negative): extreme-valence (7�valence �7.5 or 3.5�va- lence �4) and medial-valence stimuli (4�valence �4.5 or 6.5�valence �7). The extreme-valence stimuli are presumably more clearly positive or negative, as their distance from the neutral valence is larger. To conclude, our experimental design involved four independent variables: group (self-focused or stimulus- focused), task (emotional or faces), valence (positive or negative), and intensity (extreme-valence or medial-valence). All variables except for the group variable were manipulated within-subjects.
A bird’s eye view of the analytic approach. Our major goal in this work was to test the appropriateness of LBA as a model of feeling reports. To this end, we adopted the usual modeling stan- dards according to which models need to (a) show good fit to data; (b) have better fit relative to alternative models; and (c) show “selective influence” (Rae et al., 2014)—that is, their parameters should change predictably with manipulations. All these three steps require model fitting.
Figure 4 presents a schematic description of the idea behind model fitting. Essentially, participants produce N responses, each being associated with accuracy and RT. These data are represented as two RT distributions, one for correct responses (blue solid line) and one for errors (red solid line). Each run of the LBA machinery also produces a response with RT and accuracy. Specifically, in each such run, one accumulator wins. The identity of the winning accumulator determines response accuracy and the time taken for it to win the race determines RT. N runs of this machinery thus generate simulated data that can also be described as two RT distributions (in dashed blue/red lines). The proportion of correct trials as well as the shape of the RT distributions change with changes in the LBA parameters. During the course of model
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
276 GIVON, ITZHAK-RAZ, KARMON-PRESSER, DANIELI, AND MEIRAN
fitting, the LBA parameters are changed until the simulated data (expected) become as similar as possible to those of the actual (observed) data. When a sufficiently high fit is achieved, one can make the inference that humans possibly behave like the LBA does.
A bird’s eye view of the current study. Our independent variables allowed us to answer our research questions (see more below). For each independent variable, in Stage I, we created a set of models, such that the comparison between the respective model fits allowed us to tell which LBA parameter was associated with the given independent variable. Specifically, we designed a model in which, for example, there was a different threshold level for each level of the independent variable. This model was then compared with the null model, assuming no differences in any parameters between levels of the independent variable. A signifi- cantly better fit of the threshold model relative to the null model supports the hypothesis that the threshold is associated with the independent variable. This procedure was repeated for all the LBA parameters. In Stage II, we used the LBA parameters showing significant relations to the independent variable (as determined in Stage I). Here we designed models in which the levels of the independent variable differed in more than one LBA parameter. The best of all the models from Stages I–II (in terms of its fit) was selected as the one best describing the associated independent variable. This entire process was re- peated for all independent variables. Tables 1–5 describe the models set for each independent variable.
Results
Analysis. Before modeling, we omitted trials with RTs shorter than 200 ms, as well as trials with RTs higher than 2.5 SD above the participant’s mean RT. The LBA model was fitted by using rtdists package (Singmann et al., 2017) from R software (R Core Team, 2013). Comparison of means employed Bayes factors (BF) which were calculated using the freely available JASP software (The JASP Team, 2015) with the default priors.
Because we had several questions, we chose to present our results divided into three sections: Section 1 addresses the question of reporting genuine self-emotion; Section 2 addresses the ques- tion of the suitability of EAM for modeling feeling generation; Section 3 presents fresh findings that were made possible due to the use of EAM as a model of feeling generation.
Section 1: Is it feeling? Before examining the models, we looked at the accuracy rates. It interested us because it is reason- able to assume that participants in the stimulus-focused group, who were instructed to respond according to what they believe reflects the common response, would respond more normatively. The results indicated decisive support for this predicted trend with accuracy rates being 0.829 and 0.898 in the self-focused and stimulus-focused conditions, respectively (BF10 � 1.555e � 23).
We reasoned that meaningful differences between self-focused and stimulus-focused would be in line with our assumption that participants in the self-focused group report their genuine feelings. We adopted the approach of model comparison to test our hypoth-
Figure 3. An illustration of a single trial. In the second slide, the text says: “Does this photo make you feel pleasant?” From “Face recognition under varying poses: The role of texture and shape,” by N. F. Troje and H. H. Bülthoff, 1996, Vision Research, 36, pp. 1761–1771. Copyright 1996 by the Max-Planck Institute for Biological Cybernetics in Tuebingen, Germany. Adapted with permission. See the online article for the color version of this figure.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
277HOW DOES THE EMOTIONAL EXPERIENCE EVOLVE?
esis regarding differences between groups. For the preliminary analysis, we ran a set of models (using the data of the emotional task), each allowing a distinct parameter to vary between groups. The null model (Model 1), is a model were all parameters were determined to be equal in the two groups. Five alternative models were compared with the null model, each assuming that all the parameters are equal in the two groups except for one parameter
which was allowed to differ between groups. These models in- cluded SP between groups model (Model 2), threshold between groups model (Model 3), nondecision time between groups model (Model 4), drift-rate between groups model (Model 5), and SV between groups model (Model 6). For each model, we estimated the BIC (Bayesian Information Criterion; Schwarz, 1978). The BIC is an acceptable statistic for model comparison because it considers the number of free parameters (by penalizing for extra parameters). BIC was calculated using the following formula:
Figure 4. A schematic illustration of model fitting. On the left side, participants execute N trials of the binary (pleasant/unpleasant) choice task—thus produce the observed data that are summarized as reaction time (RT) distributions (in solid lines). On the right side, N simulation runs of the linear ballistic accumulator (LBA) machinery generate equivalent (expected) data (summarized by RT distributions presented in dashed lines). By changing the LBA parameters, the expected proportion of errors and shape of RT distributions become similar to those of the observed data. Model fit is the degree of similarity between the observed data and the expected data. See the online article for the color version of this figure.
Table 1 BIC Values for All Models Comparing Groups (Stimulus-Focused vs. Self-Focused)a
Model Parameter Number of free
parameters BIC
1 Null model 5 36,907.497 2 SP 6 36,793.023 3 Threshold 6 36,835.508 4 Nondecision time 6 36,884.901 5 Drift-rate 6 36,915.645 6 SVb 6 36,738.724 7 SP and SV 7 36,748.046 8 SP, SV, and threshold 8 36,757.369 9 SP, SV, threshold, and
nondecision time 9 36,766.692
a SP-difference in starting point; SV-difference in the standard deviation of the drift-rate. b Model 6 (allowing SV to vary between groups; see bold) was the chosen model, the one to yield the best BIC. BIC � Bayesian Information Criterion.
Table 2 BIC Values for All Models Comparing Tasks (Emotional vs. Faces)a
Model Parameter Number of free
parameters BIC
1 Null 5 29,315.847 2 SPb 6 26,749.737 3 Threshold 6 27,441.506 4 Nondecision time 6 28,402.339 5 Drift-rate 6 28,718.660 6 SV 6 28,578.223 7 All parameters 10 26,756.394 8 SP and threshold 7 26,753.386
a SP-difference in starting point; SV-difference in the standard deviation of the Drift-Rate. b Model 2 (allowing only SP to vary between tasks; see bold) was the chosen model, the one to yield the best fit. BIC � Bayesian Information Criterion.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
278 GIVON, ITZHAK-RAZ, KARMON-PRESSER, DANIELI, AND MEIRAN
BIC � �2logLikelihood � log(N) � P
where N is the number of total observations and P is the number of free parameters per model. BIC is a comparative value, and models with lower BIC are considered superior. We were especially in- terested in �BIC because it is directly translated into BF, when BF � e 0.5��BIC (Neath & Cavanaugh, 2012). In other words, the �BIC values were used for statistical inference regarding differ- ences in parameter values between conditions. We computed the �BIC for each comparison of an alternative model to the null model. We considered a �BIC score between 6 to 10 to represent strong evidence (BF � 20 –150), and a �BIC score higher than 10 to represent very strong evidence for substantial difference be- tween models (BF �150; Raftery, 1995). Four out of five alter- native models produced better fit values (lower BICs) compared with the null model, which indicates that each parameter (except for the drift-rate parameter) can explain a meaningful difference between groups. Next, we constructed a set of models, each responding to a different combination of parameters (based on the preliminary analysis): SP and SV between groups model (Model 7); SP, SV, and threshold between groups model (Model 8); and SP, SV, threshold, and nondecision time between groups model (Model 9). Nevertheless, Model 6 generated the best (lowest) BIC
value (BIC � 36,738.724; see Table 1 for fits values of all models regarding the group variable and Table 6 for parameter values of the chosen model). To conclude, the groups differ in the SV parameter. Specifically, the drift-rate of the self-focused group was drawn from a distribution with a larger variance as compared with the stimulus-focused group (.509 and .387, respectively; BF10 � 4.452e � 36, see Figure 6a).3 Given the lack of significant differ- ences in mean drift-rate, this finding implies that in the self- focused group, some stimuli were associated with exceptionally slow drift-rates (i.e., difficulty in telling what the feeling is) while others were associated with exceptionally quick drift-rates (i.e., easily identified feeling). We remind readers that there was only a rather slight task difference between the groups. Finding a group difference in both correct proportion and the process (difference in the SV parameter), despite of that fact, implies that reporting one’s own feeling is a different task than reporting the “objective” valence of a photo. This implication accords with our assumption that participants in the self-focused group reported their genuine feeling. At minimum, we can state that participants in the self- focused group did not report the “objective” valence of the photo. Further implications regarding the differences between groups are reviewed in the Discussion section.
Given that the following questions involve own emotion, all the subsequent analyses were conducted only on the data from the self-focused group.
Section 2: Is EAM suitable for feeling generation? This section tackles the main issue of this article, which is the impli- cation of evidence accumulation models on the process of feeling. We chose to deal with this question using two approaches: The first was to compare the emotional task with a well-established
3 We also estimated the model separately for each participant and ran a usual t-test on the SV parameter. The result of this t-test was broadly consistent with that from the model comparison, despite being based on much less stable estimates. Specifically, when all the parameters were estimated (resulting in 59 [participants] 5 [LBA parameters] � 295 estimated-parameters), the SV comparison was indecisive, possibly be- cause of consequent parameter estimation instability. However, when we tried to stabilize the estimates by fixing SP at the value found in the model comparison part (resulting in 59 4 � 1 (SP) � 237 estimated parame- ters), we found a significant null hypothesis testing result, t(57) � 2.122, p � .038, an indecisive two-sided Bayesian t-test, BF10 � 1.685, as well as a significant one-sided effect, BF10 � 3.275.
Table 3 BIC Values for All Models Comparing Intensities (Medial- Valence vs. Extreme-Valence)a
Model Parameter Number of free
parameters BIC
1 Null 5 17,962.242 2 SP 6 17,959.065 3 Threshold 6 17,969.686 4 Nondecision time 6 17,967.567 5 Drift-rate 6 17,900.129 6 SV 6 17,919.941 7 SV and drift-rateb 7 17,889.884
a SP-difference in starting point; SV-difference in the standard deviation of the drift-rate. b Model 7 (allowing SV and drift-rate to vary between intensities; see bold) was the best model for describing the data. BIC � Bayesian Information Criterion.
Table 4 BIC Values for All Models Comparing Valences (Negative vs. Positive) in Experiment 1a
Model Parameter Number of free
parameters BIC
1 Null 5 17,962.242 2 SP 6 17,949.437 3 Threshold 6 17,970.709 4 Nondecision time 6 17,966.570 5 Drift-rate 6 1,7812.438 6 SV 6 17,913.396 7 SV, SP, and drift-rate 8 17,819.354 8 SV and drift-rateb 7 17,811.001
a SP-difference in starting point; SV-difference in the standard deviation of the drift-rate. b Model 8 (allowing SV and drift-rate to vary between valences; see bold) was the best model for describing the data, although it is not significantly different from Model 5. BIC � Bayesian Information Criterion.
Table 5 BIC Values for All Models Comparing Valences (Negative vs. Positive) in Experiment 2a
Model Parameter Number of free parameters BIC
1 Null 5 6,259.267 2 SP 6 6,230.279 3 Threshold 6 6,263.589 4 Nondecision time 6 6,264.354 5 Drift-rateb 6 6,115.133 6 SV 6 6,255.128 7 SP and drift-rate 7 6,122.945
a SP-difference in starting point; SV-difference in the standard deviation of the drift-rate. b Model 5 (allowing only drift-rate to vary between va- lences; see bold) was the best model for describing the data. BIC � Bayesian Information Criterion.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
279HOW DOES THE EMOTIONAL EXPERIENCE EVOLVE?
nonemotional benchmark task. The second was based on the method of selective influence. Details concerning both approaches are presented subsequently.
For creating a validated reference-point, we adopted a nonemo- tional, two-choice RT task of the sort for which EAM has been shown to be applicable. In this task, participants were requested to discriminate between male and female faces as quickly and accu- rately as possible. We compared data from the emotional task to data from the faces task, only for self-focused participants. We reasoned that potential similarity between both tasks in terms of the model and its parameters would strengthen our view of asso- ciating EAM with feeling generation.
To address this question, we created a set of models, each allowing a different parameter to vary between tasks. As in the analyses in Section 1, the first model was the null model, where all parameters were set to be equal across tasks, and it served as the comparison model for all the alternative models: SP between tasks model (Model 2), threshold between tasks model (Model 3), non- decision time between tasks model (Model 4), drift-rate between tasks model (Model 5), and SV between tasks model (Model 6). Again, each model generated a BIC value and we calculated �BIC for each comparison of an alternative model to the null model (see Table 2 for model comparison and Table 6 for parameter values of the chosen model). All five alternative models produced lower BIC values compared with the null model. Next, we constructed a model combining all significant parameters, all parameters be- tween tasks model (Model 7), and compared its BIC with the BICs of all other models. Additionally, we created a model which combines the two most influential parameters: SP and threshold between tasks model (Model 8). In line with our hypothesis, neither Model 7 nor Model 8 produced the best fit, rather it was Model 2 that allowed SP to differ between tasks. This result indicates that the tasks differ significantly in the SP parameter. Specifically, the emotional task had a larger SP as compared with the faces task (1.551 and .378, respectively, BF10�1.014e � 304, see Figure 6b).4 Given that SP specifies both the range (0-SP) and the mean value (SP/2), this difference between tasks has two implications. The first implication is that the accumulators asso- ciated with each decision tended to have a more variable level of initial evidence in the emotional task. The second (related) impli- cation is that there was more initial evidence on average in the emotional task. Both of these could have resulted from stronger and more durable carryover effects between trials in the emotional task, that is, the emotion experienced in the previous trial contin- ued to exert influence in the following trial. Another possibility is that when the task is felt as relatively more subjective (as presum-
ably happens in the emotional task), the role of priors (here, the initial level of evidence, i.e., SP) becomes relatively more prom- inent.
We remind the readers that the faces task and the emotional task involved completely different stimuli, responses, and instructions. Finding that these tasks produce similar results and model fits and differ only in one parameter (SP), provides some support for our analogy between perceptual decision-making and feeling genera- tion, including the applicability of EAM.
In addition, we wished to assess the fit of the models graphically and therefore further demonstrate the similarity between the emo- tional task and the benchmark task. Figure 5 shows a version of quantile probability plot (see Brown & Heathcote, 2008), in which RTs and cumulative probabilities are presented for the .1, .3, .5 (median), .7, and .9 quantiles. Error rates are represented through the relative heights of the correct and incorrect responses. This is done by using defective cumulative distributions in which each correct quan- tile is multiplied by the correct-responses rate while each incorrect quantile is multiplied by the error-rate. We plotted predictions and observed data separately for our three conditions: emotional task of the stimulus-focused group, emotional task of the self-focused group, and faces task of the self-focused group. Here, the horizontal discrep- ancy denotes the lack of fit in RT, while the vertical discrepancy denotes lack of fit with respect to error rate. As evident in the figure, the fit for error rates was excellent and that for RT was generally good except for the slowest RT quantiles and more so for errors. This is a quite typical picture (e.g., Brown & Heathcote, 2008). Importantly, the fit does not appear visually to be better for faces than for the emotional task.
An essential aspect of establishing our model entails creating predictions regarding the influence of manipulations on model parameters, so called “selective influence” (Rae et al., 2014). In the current study, the intensity manipulation (creating extreme vs. medial valence stimuli) was predicted to selectively affect the drift-rate parameter. This prediction is based on the nature of drift-rate as reflecting the quality of the evidence which the stim- ulus provides. As before, we used model comparison while creat- ing different models regarding the intensity variable. The first model was the null model (where all parameters are set to be equal across intensities), which repeatedly served as the comparable
4 As before, we ran t-tests on the individually-assessed parameters. These tests of comparing the SP parameter between tasks were significant, including in null hypothesis testing, t(28) � 10.28, p � .001, and in the Bayesian t-test, BF10 � 3.660e � 8, two-sided.
Table 6 Parameter Values for All Chosen Modelsa
Model SP Threshold Nondecision
time Drift-rate SV
SV between groups 2 2.441 .231 .98 .387 (stimulus) .509 (self) SP between tasks 1.551 (emotional) .378 (faces) 2.045 .189 .897 .457 SV and drift-rate between intensities 1.571 2.082 .196 .876 (medial) .98 (extreme) .495 (medial) .436 (extreme) SV and drift-rate between valences 1.439 2.021 .142 .98 (negative) .814 (positive) .432 (negative) .474 (positive) Drift-rate between valences
Experiment 2 .842 1.302 .05 .674 (negative) .594 (positive) .128
a SP-starting point; SV-standard deviation of the drift-rate.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
280 GIVON, ITZHAK-RAZ, KARMON-PRESSER, DANIELI, AND MEIRAN
model for all the alternative models: SP between intensities model (Model 2), threshold between intensities model (Model 3), nonde- cision time between intensities model (Model 4), drift-rate be- tween intensities model (Model 5), and SV between intensities model (Model 6). Model 5 is in the focus of our interest since we predicted it will produce the best fit. As shown in Table 3, only two models were significantly better than the null model: Model 5 (drift-rate) and Model 6 (SV). It is important to note that all parameters that are unrelated to rate of evidence accumulation where not influenced by the variance in intensity. Next, we ran Model 7, in which SV and drift-rate were allowed to differ be- tween intensities. Model 7 was the best model, and the one to produce the lowest BIC (see Table 3 for model comparison and Table 6 for parameter values of the chosen model). This result validates the model especially considering both drift-rate and SV are related to the rate of evidence accumulation. We should em- phasize that the difference between drift-rates was in the expected direction, such that extreme-valence stimuli obtained higher drift- rate, compared with medial-valence stimuli (.98 and .876, respec- tively, BF10 � 5.156e � 15, see Figure 6c).
5 The fact that intensity manipulation did not have an influence on other parameters, mainly the threshold parameter, served as a sanity check for the model, because we obtained the predicted results.
Section 3: What additional information can EAM provide regarding feeling generation? Following the establishment of EAM as a viable model of feeling generation, we will now present some new findings that EAM made possible. For this purpose, we examined the valence variable, distinguishing negative from pos- itive feelings. As before, we used the method of model compari- son, and the results of the models can be seen in Table 4 (see also Table 6 for parameter values of the chosen model). When com-
paring the generation of negative and positive feelings, the most interesting finding is the difference in the rate of evidence accu- mulation. In accordance with the extant literature stating that “bad is stronger than good” (Baumeister, Bratslavsky, Finkenauer, & Vohs, 2001), we received higher drift-rate for negative feelings, compared with positive feelings (.98 and .814, respectively, BF10 � 6.943e � 32, see Figure 6d).
6 In other words, evidence regarding negative feelings accumulates more quickly than that of positive feelings. Further considerations regarding differences be- tween negative and positive feelings are discussed below.
Experiment 2
According to the results from Experiment 1, the rate of evidence accumulation is higher for negative feelings as compared with posi- tive feelings. Following this finding, we wanted to make sure that our results reflect genuine difference in the processes of feeling genera- tion, and it is not due to a possible motivation to quickly remove negative stimuli. Such motivation could have contributed to the re- sults because the stimuli disappeared from the screen upon the key press. Thus, quicker processing of negatively valenced stimuli could
5 As before, we ran t-tests on the individually-assessed parameters. These tests of comparing the drift-rate parameter between intensities were significant, including in null hypothesis testing, t(28) � 5.538, p � .001, and in the Bayesian t-test, BF10 � 3197, two-sided.
6 As before, we ran t-tests on the individually-assessed parameters. These tests of comparing the drift-rate parameter between valences were significant, including in null hypothesis testing, t(28) � 2.439, p � .021, and in the Bayesian t-test, BF10 � 2.42, two-sided. For one-sided Bayesian t-test, we received BF10 � 4.775.
Figure 5. Graphical assessment of models fit. linear ballistic accumulator (LBA) predictions (circles) and data values (triangles) are presented for both correct (upper lines) and incorrect (lower lines) responses, for .1, .3, .5, .7, and .9 quantiles.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
281HOW DOES THE EMOTIONAL EXPERIENCE EVOLVE?
have helped shortening the exposure to these stimuli.7 To deal with this alternative explanation, we designed Experiment 2. This experi- ment was strictly identical to the self-focused group of Experiment 1, except for the duration of the emotional stimuli. While in Experiment 1 emotional stimuli remained until response was made, in Experiment 2 emotional stimuli remained for a constant time, regardless of par- ticipant’s responses.
Method
Participants. Twenty participants, similar in attributes to those in Experiment 1 (17 females, mean age: 23.3) took part in the experiment for course credit. Four participants were excluded from the analysis for technical reasons (i.e., those four subjects obtained high rate of responses that were executed after the disappearance of the stimulus, and the computer program failed recording this type of response).
Procedure. The procedure was identical to the procedure of the self-focused group in Experiment 1, except for the duration of the emotional stimuli, which was set to 2,500 ms.
Results
Experiment 2 involved three independent within-subject variables: task (emotional or faces), valence (positive or negative), and intensity (extreme-valence or medial-valence). For this preliminary analysis, we omitted trials with RTs shorter than 200 ms or longer than 2,500 ms (reflecting the fact that slower responses would have been exe- cuted after the disappearance of the stimuli).
We were mostly interested in the valence variable. Specifically, we wanted to examine whether the difference between positive and negative valence in terms of the drift-rate parameter remains in the same direction as in Experiment 1. Nevertheless, we wanted to first assure that we can replicate our findings regarding the com-
parison between the faces and the emotional tasks. We, indeed, replicated the difference between the two tasks. Specifically, we ran the same set of models as described in Experiment 1 and the best model was the model that allowed SP and Threshold to differ between the tasks. It had the lowest BIC value of 10,147.42 as compared with BIC value of 10,156.13 in the next-best model. In the chosen model, the difference in SP parameter was in the same direction as in Experiment 1, presenting larger SP for the emo- tional task as compared with the faces task (.773 and .377, respec- tively, BF10�1.014e � 304, see Figure 6e).
8
To answer the main question of Experiment 2, we again used a model comparison approach while creating different models re- garding the valence variable (see Table 5 for model comparison and Table 6 for parameter values of the chosen model). Not only that drift-rate was the parameter to differentiate between valences, it was in the same direction as in Experiment 1, presenting higher drift-rate for negative feelings, compared with positive feelings (.674 and .594, respectively, BF10 � 1.987e � 31, see Figure 6f).
9
As noted by one reviewer, the fact that the stimuli were not removed from the screen after the key press, does not rule out the possibility that participants stop attending to the stimuli after making a response. We can thus cautiously conclude that the advantage for negative feelings in terms of the drift-rate parameter is not (only) due to a motivation to quickly remove negative
7 We thank Yoav Kessler for pointing out this possibility. 8 As before, we ran t-tests on the individually-assessed parameters.
These tests of comparing the SP parameter between tasks were significant, including in null hypothesis testing, t(14) � 4.576, p � .001, and in the Bayesian t-test, BF10 � 79.88, two-sided.
9 As before, we ran t-tests on the individually-assessed parameters. These tests of comparing the drift-rate parameter between valences were significant, including in null hypothesis testing, t(15) � 3.002, p � .009, and in the Bayesian t-test, BF10 � 6.005, two-sided.
Figure 6. Differences in parameters between conditions. (a) Difference in the SV parameter between groups, Experiment 1. (b) Difference in the starting-point (SP) parameter between tasks, Experiment 1. (c) Difference in the drift-rate parameter between intensities, Experiment 1. (d) Difference in the drift-rate parameter between valences, Experiment 1. (e) Difference in the SP parameter between tasks, Experiment 2. (f) Difference in the drift-rate parameter between valences, Experiment 2.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
282 GIVON, ITZHAK-RAZ, KARMON-PRESSER, DANIELI, AND MEIRAN
stimuli from the screen. Moreover, the replicability across exper- iments contributes to the trustworthiness of the findings.
Discussion
The goal of the present study was to answer the question how the report of feeling emerges? We decomposed the processes that are involved in the emergence of feeling using EAM, which is typically used to describe perceptual choices. We asked partici- pants to make a binary choice of how they feel (pleasant or unpleasant) in response to emotional stimuli before performing a choice RT task (deciding whether a face is of a male or a female). Among various EAMs, we chose the LBA model (Brown & Heathcote, 2008).
The first question we asked is whether participants report their own genuine feelings as opposed to expected feelings. The differ- ence in accuracy rates between groups (self-focused vs. stimulus- focused) was in the expected direction, with less normative ratings when reporting own feeling as compared with what is believed to be the normative response. This finding was accompanied by evidence for a process difference (in the SV parameter) between the groups, which further supports our assumption that these are two qualitatively different tasks. The second question regarded the suitability of EAM for feeling generation. Here, we validated the model in two ways. The first, and the crucial way, relates to the similarities between the emotional task and the benchmark (faces) perceptual-decision task. The second way was through selective influence, in which as we expected, intensity manipulation exclu- sively affected the drift-rate parameter. In doing so, we followed Heathcote, Brown, and Wagenmakers’s (2015) guidelines for proper modeling practices. Lastly, difference in the drift-rate pa- rameter between positive and negative feelings suggests that evi- dence regarding negative feelings is accumulated relatively more efficiently. All these analyses lead us to conclude that feeling reports are a very close analogue of perceptual decisions. Unlike Karmon-Presser et al. (2018), who based their conclusions on response identity, the present investigations take into account time, and even the exact shape of the RT distributions and thus provide a much stronger support for the analogy between feeling reports and perceptual decisions.
Our results have several implications regarding previous litera- ture. Itkes et al. (2017) from whom we adopted the self versus stimulus focused manipulation, found that self-reported feelings attenuated with repeated exposure, while stimulus-focused reports did not. Our results further support the dissociation between these two conditions, in providing evidence for greater normativity and smaller SV in the stimulus-focused condition. Mathematically, the larger SV in the self-focused condition indicates that there are trials with exceptionally slow evidence accumulation rate and other trials with exceptionally quick evidence accumulation rate. Psychologically, this observation could be explained post hoc as follows.10 Stimuli that are relatively normatively homogenous, may be more variable subjectively due to, for example, personal meaning and past experience. For example, most people would regard stimuli related to war as negatively valenced. However, for someone who has actually experienced war, these stimuli would be experienced as especially intensively negative. A related nonmu- tually exclusive account is that in the stimulus-focused group, responses were based only on cognitive (“objective”) evaluations,
such as “this snake can kill.” However, in the self-focused group, responses possibly required additional type of evidence, such as the experience of the bodily states (e.g., pounding heart). Some- times, this additional evidence precedes the cognitive evaluation, and in these cases, the evidence accumulation rate would be exceptionally quick. At other times, the additional evidence lags, resulting in an especially slow evidence accumulation rate. These speculations suggest several implications. One is that reporting own feelings is less stable and more variable than a mere “cold” cognitive evaluation. The other implication, as suggested by others (Blouw, Solodkin, Thagard, & Eliasmith, 2016; James, 1884; Pollatos, Gramann, & Schandry, 2007), is that the awareness of bodily states may be critical in the experience of emotion.
Finding that negative feelings are processed more efficiently than positive feelings is aligned with the extant literature suggest- ing that “bad is stronger than good” (Baumeister et al., 2001). According to this literature, people tend to perceive, process, and remember negative events and stimuli better than positive ones. Our result goes further as it explains the underlying mechanism of this tendency. It indicates that the quality of emotional information driven from negative stimuli is higher (less noisy, perhaps) than from positive stimuli. It is important to note that in our task, negative stimuli were less deviant from neutral as compared with positive stimuli. Thus, the advantage of negative stimuli may actually be larger than estimated here.
Interestingly, White, Liebman, and Stone (2018), who used the drift diffusion model, found a bias in favor of positive judgments. The closest relative of the bias in LBA is SP, because SP/2 is the mean level of evidence present in the accumulator prior to stimulus exposure. Our results also indicate that SP was significantly higher for positive pictures (Model 2, Table 4). However, this model was outperformed by another model in this set of comparisons.
The present work suggests several future directions for research. One is to manipulate the rate of evidence accumulation via the hypothesized sources of emotional evidence. These include, for example, manipulations of appraisals, facial expression, and so forth. Our theory makes a clear prediction in this regard that the drift-rate would be selectively influenced by such manipulations. Our model(s) can also be applied to the understanding of emotion regulation, relevant individual differences and psychopathology.
Limitations
Obviously, the conclusions drawn here are limited by the ex- perimental design, stimulus-set, and type of stimuli (pictures) that we used. The applicability of the conclusions to other settings should thus be examined in future studies. More broadly, when inquiring the conscious emotional experience via self-reports, we inevitably narrowed the exploration to reportable feelings. We rely on Shadlen and Kiani’s (2011) definition of consciousness as a decision to engage. These authors further establish this definition by showing that the neural mechanisms that underlie simple deci- sions share features with the neural mechanisms that characterize consciousness. Specifically, while consciousness includes differ- ent states and functions, acting with awareness of a purpose involves a decision to engage in a potential reportable narrative. We thus adopt the above description and elaborate it through
10 Part of this account is based on a suggestion made by one reviewer.
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
283HOW DOES THE EMOTIONAL EXPERIENCE EVOLVE?
defining feeling as a decision to engage in a reportable emotional narrative.
When dealing with reported emotional feelings, two possible errors can occur: not reporting a feeling you own, or reporting a feeling you do not own. Regarding the first type of error, we used stimuli that are likely to minimize it because they did not invoke socially unacceptable emotional reactions, such as in pornographic pictures or pictures that arouse morbid curiosity. Regarding the second type of error, we used a cover story and instructions that have probably minimized it.
Conclusions
The current study accords with other research, which adopted the analogy between emotional feelings and perceptual decisions. This line of research is promising and suggests novel insights. Our main contribution to this growing literature is in supplying evi- dence for the trustworthiness of reports as reflecting genuine feelings and in providing some validation to the analogy between emotional feelings and perceptual decision making.
References
Bach, D. R., & Dayan, P. (2017). Algorithms for survival: A comparative perspective on emotions. Nature Reviews Neuroscience, 18, 311–319. http://dx.doi.org/10.1038/nrn.2017.35
Barrett, L. F. (2006). Are emotions natural kinds? Perspectives on Psy- chological Science, 1, 28 –58. http://dx.doi.org/10.1111/j.1745-6916 .2006.00003.x
Barrett, L. F. (2017). The theory of constructed emotion: An active infer- ence account of interoception and categorization. Social Cognitive and Affective Neuroscience, 12, 1–23.
Barrett, L. F., Gross, J., Christensen, T. C., & Benvenuto, M. (2001). Knowing what you’re feeling and knowing what to do about it: Mapping the relation between emotion differentiation and emotion regulation. Cognition and Emotion, 15, 713–724. http://dx.doi.org/10.1080/ 02699930143000239
Baumeister, R. F., Bratslavsky, E., Finkenauer, C., & Vohs, K. D. (2001). Bad is stronger than good. Review of General Psychology, 5, 323–370. http://dx.doi.org/10.1037/1089-2680.5.4.323
Blanz, V., & Vetter, T. (1999). A morphable model for the synthesis of 3D faces. Proceedings of the 26th Annual Conference on Computer Graph- ics and Interactive Techniques (pp. 187–194). Austin, TX: ACM Press/ Addison-Wesley Publishing Co.
Block, N. (2005). Two neural correlates of consciousness. Trends in Cognitive Sciences, 9, 46 –52. http://dx.doi.org/10.1016/j.tics.2004.12 .006
Blouw, P., Solodkin, E., Thagard, P., & Eliasmith, C. (2016). Concepts as semantic pointers: A framework and computational model. Cognitive Science, 40, 1128 –1162. http://dx.doi.org/10.1111/cogs.12265
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 Re- view, 113, 700 –765. http://dx.doi.org/10.1037/0033-295X.113.4.700
Bogacz, R., Usher, M., Zhang, J., & McClelland, J. L. (2007). Extending a biologically inspired model of choice: Multi-alternatives, nonlinearity and value-based multidimensional choice. Philosophical Transactions of the Royal Society of London Series B, Biological Sciences, 362, 1655– 1670. http://dx.doi.org/10.1098/rstb.2007.2059
Brown, S., & Heathcote, A. (2005). A ballistic model of choice response time. Psychological Review, 112, 117–128. http://dx.doi.org/10.1037/ 0033-295x.112.1.117
Brown, S. D., & Heathcote, A. (2008). The simplest complete model of choice response time: Linear ballistic accumulation. Cognitive Psychol- ogy, 57, 153–178. http://dx.doi.org/10.1016/j.cogpsych.2007.12.002
Cacioppo, J. T., & Berntson, G. G. (1994). Relationship between attitudes and evaluative space: A critical review, with emphasis on the separa- bility of positive and negative substrates. Psychological Bulletin, 115, 401– 423. http://dx.doi.org/10.1037/0033-2909.115.3.401
Clore, G. L., Gasper, K., & Garvin, E. (2001). Affect as information. In J. P. Forgas (Ed.), Handbook of affect and social cognition (pp. 121– 144). Mahwah, NJ: Lawrence Erlbaum.
Donkin, C., Brown, S., Heathcote, A., & Wagenmakers, E.-J. (2011). Diffusion versus linear ballistic accumulation: Different models but the same conclusions about psychological processes? Psychonomic Bulletin & Review, 18, 61– 69. http://dx.doi.org/10.3758/s13423-010-0022-4
Duncan, S., & Barrett, L. F. (2007). Affect is a form of cognition: A neurobiological analysis. Cognition and Emotion, 21, 1184 –1211. http:// dx.doi.org/10.1080/02699930701437931
Ekman, P., & Davidson, R. J. (1994). The nature of emotion: Fundamental questions. New York, NY: Oxford University Press. Retrieved from http://PsycNET.apa.org/record/1995-97541-000
Feldman Barrett, L., & Russell, J. A. (1998). Independence and bipolarity in the structure of current affect. Journal of Personality and Social Psychology, 74, 967–984. http://dx.doi.org/10.1037/0022-3514.74.4.967
Galton, F. (1907). Vox Populi (The wisdom of crowds). Nature, 75, 450 – 451. http://dx.doi.org/10.1038/075450a0
Heathcote, A., Brown, S. D., & Wagenmakers, E.-J. (2015). An introduc- tion to good practices in cognitive modeling. In B. U. Forstmann & E.-J. Wagenmakers (Eds.), An introduction to model-based cognitive neuro- science (pp. 25– 48). New York, NY: Springer. http://dx.doi.org/10 .1007/978-1-4939-2236-9_2
Itkes, O., Kimchi, R., Haj-Ali, H., Shapiro, A., & Kron, A. (2017). Dissociating affective and semantic valence. Journal of Experimental Psychology: General, 146, 924 –942. http://dx.doi.org/10.1037/xge 0000291
James, W. (1884). What is an emotion? Mind, 9, 188 –205. http://dx.doi .org/10.1093/mind/os-IX.34.188
Karmon-Presser, A., Sheppes, G., & Meiran, N. (2018). How does it “feel”? A signal detection approach to feeling generation. Emotion, 18, 94 –115. http://dx.doi.org/10.1037/emo0000298
Krajbich, I., & Rangel, A. (2011). Multialternative drift-diffusion model predicts the relationship between visual fixations and choice in value- based decisions. Proceedings of the National Academy of Sciences of the United States of America, 108, 13852–13857. http://dx.doi.org/10.1073/ pnas.1101328108
Kron, A., Goldstein, A., Lee, D. H-J., Gardhouse, K., & Anderson, A. K. (2013). How are you feeling? Revisiting the quantification of emotional qualia. Psychological Science, 24, 1503–1511. http://dx.doi.org/10 .1177/0956797613475456
Lambie, J. A., & Marcel, A. J. (2002). Consciousness and the varieties of emotion experience: A theoretical framework. Psychological Review, 109, 219 –259. http://dx.doi.org/10.1037/0033-295X.109.2.219
Lane, R. D., & Schwartz, G. E. (1987). Levels of emotional awareness: A cognitive-developmental theory and its application to psychopathology. The American Journal of Psychiatry, 144, 133–143.
Lazarus, R. S. (2006). Emotions and interpersonal relationships: Toward a person-centered conceptualization of emotions and coping. Journal of Personality, 74, 9 – 46. http://dx.doi.org/10.1111/j.1467-6494.2005 .00368.x
Lischetzke, T., Cuccodoro, G., Gauger, A., Todeschini, L., & Eid, M. (2005). Measuring affective clarity indirectly: Individual differences in response latencies of state. Emotion, 5, 431– 445. http://dx.doi.org/10 .1037/1528-3542.5.4.431
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
284 GIVON, ITZHAK-RAZ, KARMON-PRESSER, DANIELI, AND MEIRAN
Macmillan, N. A., & Creelman, C. D. (2004). Detection theory: A user’s guide. London, UK: Psychology Press. http://dx.doi.org/10.4324/ 9781410611147
Marchewka, A., Żurawski, Ł., Jednorǒg, K., & Grabowska, A. (2014). The Nencki Affective Picture System (NAPS): Introduction to a novel, standardized, wide-range, high-quality, realistic picture database. Behav- ior Research Methods, 46, 596 – 610. http://dx.doi.org/10.3758/s13428- 013-0379-1
Mathôt, S., Schreij, D., & Theeuwes, J. (2012). OpenSesame: An open- source, graphical experiment builder for the social sciences. Behavior Research Methods, 44, 314 –324. http://dx.doi.org/10.3758/s13428-011- 0168-7
Mayer, J. D., Salovey, P., Caruso, D. R., & Sitarenios, G. (2003). Mea- suring emotional intelligence with the MSCEIT V2.0. Emotion, 3, 97– 105. http://dx.doi.org/10.1037/1528-3542.3.1.97
Neath, A. A., & Cavanaugh, J. E. (2012). The Bayesian information criterion: Background, derivation, and applications. Wiley Interdisciplin- ary Reviews: Computational Statistics, 4, 199 –203. http://dx.doi.org/10 .1002/wics.199
Oishi, S., Diener, E., & Lucas, R. E. (2007). The optimum level of well-being: can people be too happy? Perspectives on Psychological Science, 2, 346 –360.
Pe, M. L., Vandekerckhove, J., & Kuppens, P. (2013). A diffusion model account of the relationship between the emotional flanker task and rumination and depression. Emotion, 13, 739 –747. http://dx.doi.org/10 .1037/a0031628
Pollatos, O., Gramann, K., & Schandry, R. (2007). Neural systems con- necting interoceptive awareness and feelings. Human Brain Mapping, 28, 9 –18. http://dx.doi.org/10.1002/hbm.20258
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, 1226 –1243. http://dx.doi.org/10.1037/a00 36801
Raftery, A. E. (1995). Bayesian model selection in social research. Soci- ological Methodology, 25, 111–163. http://dx.doi.org/10.2307/271063
Ratcliff, R., & McKoon, G. (2008). The diffusion decision model: Theory and data for two-choice decision tasks. Neural Computation, 20, 873– 922. http://dx.doi.org/10.1162/neco.2008.12-06-420
Ratcliff, R., & Rouder, J. N. (1998). Modeling response times for two- choice decisions. Psychological Science, 9, 347–356. http://dx.doi.org/ 10.1111/1467-9280.00067
Ratcliff, R., & Smith, P. L. (2004). A comparison of sequential sampling models for two-choice reaction time. Psychological Review, 111, 333– 367. http://dx.doi.org/10.1037/0033-295X.111.2.333
R Core Team. (2013). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Retrieved from http://citeseerx.ist.psu.edu/viewdoc/download?doi�10 .1.1.470.5851&rep�rep1&type�pdf
Rottenberg, J., & Gross, J. J. (2003). When emotion goes wrong: Realizing the promise of affective science. Clinical Psychology: Science and Practice, 10, 227–232. http://dx.doi.org/10.1093/clipsy.bpg012
Satpute, A. B., Shu, J., Weber, J., Roy, M., & Ochsner, K. N. (2013). The functional neural architecture of self-reports of affective experience. Biological Psychiatry, 73, 631– 638. http://dx.doi.org/10.1016/j .biopsych.2012.10.001
Schaller, M., & Park, J. H. (2011). The Behavioral immune system (and why it matters). Current Directions in Psychological Science, 20, 99 – 103. http://dx.doi.org/10.1177/0963721411402596
Scherer, K. R. (1987). Toward a dynamic theory of emotion: The compo- nent process model of affective states. Geneva Studies in Emotion, 1, 1–96.
Schwarz, G. (1978). Estimating the dimension of a model. Annals of Statistics, 6, 461– 464. http://dx.doi.org/10.1214/aos/1176344136
Shadlen, M. N., & Kiani, R. (2011). Consciousness as a decision to engage. In S. Dehaene & Y. Christen (Eds.), Characterizing consciousness: From cognition to the clinic? (pp. 27– 46). Berlin, Heidelberg: Springer. http://dx.doi.org/10.1007/978-3-642-18015-6_2
Singmann, H., Brown, S., Gretton, M., Heathcote, A., Voss, A., Voss, J., & Terry, A. (2017). Package “rtdists”. Retrieved from http://healthstat .snu.ac.kr/CRAN/web/packages/rtdists/rtdists.pdf
Sokolov, E. N., & Boucsein, W. (2000). A psychophysiological model of emotion space. Integrative Physiological and Behavioral Science, 35, 81–119. http://dx.doi.org/10.1007/BF02688770
Sternberg, S., & Backus, B. T. (2015). Sequential processes and the shapes of reaction time distributions. Psychological Review, 122, 830 – 837. http://dx.doi.org/10.1037/a0039658
Stone, M. (1960). Models for choice-reaction time. Psychometrika, 25, 251–260. http://dx.doi.org/10.1007/BF02289729
Teodorescu, A. R., & Usher, M. (2013). Disentangling decision models: From independence to competition. Psychological Review, 120, 1–38. http://dx.doi.org/10.1037/a0030776
Thagard, P., & Aubie, B. (2008). Emotional consciousness: A neural model of how cognitive appraisal and somatic perception interact to produce qualitative experience. Consciousness and Cognition, 17, 811– 834. http://dx.doi.org/10.1016/j.concog.2007.05.014
The JASP Team. (2015). Software to sharpen your stats. APS Observer. Retrieved from https://www.psychologicalscience.org/observer/bayes- or-bust-with-new-softwares
Troje, N. F., & Bülthoff, H. H. (1996). Face recognition under varying poses: The role of texture and shape. Vision Research, 36, 1761–1771. http://dx.doi.org/10.1016/0042-6989(95)00230-8
Usher, M., & McClelland, J. L. (2001). The time course of perceptual choice: The leaky, competing accumulator model. Psychological Re- view, 108, 550 –592. http://dx.doi.org/10.1037/0033-295X.108.3.550
van Ravenzwaaij, D., Mulder, M. J., Tuerlinckx, F., & Wagenmakers, E.-J. (2012). Do the dynamics of prior information depend on task context? An analysis of optimal performance and an empirical test. Frontiers in Psychology, 3, 132.
White, C. N., Liebman, E., & Stone, P. (2018). Decision mechanisms underlying mood-congruent emotional classification. Cognition and Emotion, 32, 249 –258. http://dx.doi.org/10.1080/02699931.2017 .1296820
White, C., Ratcliff, R., Vasey, M., & McKoon, G. (2009). Dysphoria and memory for emotional material: A diffusion-model analysis. Cognition and Emotion, 23, 181–205. http://dx.doi.org/10.1080/02699930 801976770
White, C. N., Ratcliff, R., Vasey, M. W., & McKoon, G. (2010). Using diffusion models to understand clinical disorders. Journal of Mathemat- ical Psychology, 54, 39 –52. http://dx.doi.org/10.1016/j.jmp.2010.01.004
White, C. N., Skokin, K., Carlos, B., & Weaver, A. (2016). Using decision models to decompose anxiety-related bias in threat classification. Emo- tion, 16, 196 –207. http://dx.doi.org/10.1037/emo0000109
Zachar, P., & Ellis, R. D. (2012). Categorical versus dimensional models of affect: A seminar on the theories of Panksepp and Russell. Amster- dam, the Netherlands: John Benjamins Publishing. http://dx.doi.org/10 .1075/ceb.7
Received May 14, 2018 Revision received September 14, 2018
Accepted September 14, 2018 �
T hi
s do
cu m
en t
is co
py ri
gh te
d by
th e
A m
er ic
an P
sy ch
ol og
ic al
A ss
oc ia
ti on
or on
e of
it s
al li
ed pu
bl is
he rs
. T
hi s
ar ti
cl e
is in
te nd
ed so
le ly
fo r
th e
pe rs
on al
us e
of th
e in
di vi
du al
us er
an d
is no
t to
be di
ss em
in at
ed br
oa dl
y.
285HOW DOES THE EMOTIONAL EXPERIENCE EVOLVE?
- How Does the Emotional Experience Evolve? Feeling Generation as Evidence Accumulation
- Feeling: The Conscious Emotional Subjective Experience
- Evidence Accumulation Models (EAMs)
- Why We Chose the LBA?
- LBA and Feeling Reports
- The Current Study
- Experiment 1
- Method
- Participants
- Emotional task
- Faces task
- Procedure
- Independent variables
- A bird’s eye view of the analytic approach
- A bird’s eye view of the current study
- Results
- Analysis
- Section 1: Is it feeling?
- Section 2: Is EAM suitable for feeling generation?
- Section 3: What additional information can EAM provide regarding feeling generation?
- Experiment 2
- Method
- Participants
- Procedure
- Results
- Discussion
- Limitations
- Conclusions
- References
