Predicting Free Choice with fMRI: An attempt
Masterthsese van
Jimmy J. IJsbrandij, 10196595
Universiteit van Amsterdam
Onder begeleiding van dr. Yaïr Pinto
Abstract
Current study will focus on the age old question whether free will exists and how conscious we are of the choices we make. There are indications that our choices are made, up to several seconds before we become aware of them. The decision is thus made subconsciously after which it reaches
consciousness. Here, we provide a deeper insight into the mechanics behind those subconscious processes and how they are influenced by an emotional or motivational questioning. Functional magnetic resonance imaging (fMRI) has been carried out, but unfortunately didn’t result in any usable functional data. Instead, we based our study on behavioral data and found that 8 out of 19 participants answered in a (statistically significant) predictable manner, regardless of the type of question. This may be because of a one-sidedly skewed input, however. Presumably, there are also more complex interactions between the different kinds of tasks, but the current study is not
sufficient to handle these questions. We will end with some suggestions for improvement so that we are able to approach these kinds of questions in the future.
Introduction
Is there such a thing as free will? With every choice we make, it feels as if we made it consciously and that we therefore have utter control over our behavior. But is this actually true? Perhaps, man has no influence over its actions at all and we are mere spectators of this movie we call life, where every occurrence is predetermined and we become conscious of those as they pass by. Possibly our choices are fixed and by becoming aware of them, it feels as if we made the decisions consciously. So the most important question is what precedes what, either the decision making process or the conscious intention. It is possible to measure the brain processes that code for the decision process and compare those to the moment a person reaches conscious awareness of the outcome. This has been done by Libet et al. in 1982 and they showed that using EEG, task related brain activity can be measured that precedes a spontaneous or a planned action. They showed participants a clock face with a single moving hand, which participants could stop at any given moment by pressing a button. After pressing, participants reported the moment they first felt the urge to press (W) by moving the hand to the position it had at that time. The EEG showed an increasing negative charge that
preceded W which was called the readiness potential (RP) and is believed to code for the brain's readiness to take action. Two or three variants have been distinguished that are each elicited by a different kind of task. RP type I occurred when a participant carried out a planned action, when they aimed for a preset time on the clock, while type II occurred when participants performed a task in which they were required to act spontaneously, either as a response to an unexpected stimulus or out of free will. Type I is characterized by a slower build-up and can be measured during the 800 ms. before W, followed by the action 200 ms. later. Type II (and possibly III) has a steeper slope, and precedes W by 500 ms., while the time between W and the action remained the same. Activity was most prominent in central areas of the scalp, particularly vertex Cz and central Cc (C3 or C4),
contralateral to the hand used by the participant. These areas are located above the preSMA and primary motor cortex, which suggests that the RP component may be a preparation for the motor output and thus only partly describes the process of generating a conscious action. In research conducted by Soon et al. (2008) fMRI was used to trace the origin of the decision further back in time by looking at frontal regions known to be involved in planning and decision making, including the medial and lateral frontopolar cortex. This allowed for an earlier prediction of the timing of the decision. Further, participants were given multiple options to choose from in so that the option itself could be differentiated. They found that in these frontal areas as well as in the precuneus/posterior cingulate cortex, patterns of activity could be differentiated between both options, up to 10 seconds before the decision was made. That means that as early as 10 seconds before people become aware of their decision, the decision has been made. In a subsequent study by Soon et al. (2013), the choice participants made was extended to a more abstract kind of task, in order to know whether a non-motoric decision could be predicted. Here, participants could choose to either add or subtract two numbers and carry out their decision. Again, the precuneus/posterior cingulate and medial
frontopolar cortex were engaged in the process and signified the outcome approximately 4 seconds before the action took place, though the choice itself could not be accurately predicted.
In these studies, all choice options were meaningless. Pressing either left or right, or either adding or subtracting two numbers in the studies by Soon et al. had no personal value to the participant and the outcome has no consequences at all. However, the expected outcome to a problem is of course crucial to the decision, and learning what brain regions participate in this kind of decision making may be very useful. Therefore, our study will focus on the influence of an emotional or a motivational
component on the decision making process, by proposing questions with emotional or motivational aspects and have relevant outcomes to the participant. Participants will be provided moral dilemmas where they choose between two ways to resolve a conflict, which will elicit an emotional response. To create a motivational incentive, participants will play a game where a possible monetary reward depends on the choice that has been made. Both types of questions will be compared to Libet's clock task, where participants were able to respond freely. Whereas in the original experiment participants only had to respond at any moment they desired, here they will also be able to act out the decision with either their left or right hand, to create an arbitrary choice with no emotional or motivational aspect. The Libet task will thus be used to correct for brain activity related to making meaningless decisions and the subsequent motor output. Further, in Libet trials participants are able to respond spontaneously, while in other tasks people need to accumulate information over time after which they are able to respond at the prompt. To minimize this difference, participants will be able to respond as soon as they made the decision, even before the prompt. This provides an indication of when people become aware of the decision.
Brain activity will be determined with fMRI to find a BOLD variant of the readiness potential along with other processes that occur during simple choices and have a better understanding of their spatial origins. Those will be compared to activity during more complex decision tasks with either a motivational or emotional incentive.
Hypotheses
For all tasks, activity in the preSMA and motor cortex is to be expected, similar to RP. Additional activity in the frontal regions, including frontopolar cortex, dorsolateral prefrontal cortex or orbitofrontal region, in the motivational task and dilemmas. Activity in these frontal regions should precede RP, since this reflects the activity related to the motor output. The actual choice is expected to be made in frontal regions and activity here may be of similar shape to RP, slowly increasing towards a threshold after which either the motor output sequence is started (RP) or people become aware of the decision right away. This will tell us whether the occurrence of RP is necessary to achieve a conscious state or if it can be achieved by local activity elsewhere, like the aforementioned frontal regions. The former situation implies a strong connection between activity in motor output regions and consciousness and if the latter case is true, it leaves some questions about the function of RP, because a conscious state has already been achieved at that time, if RP indeed does occur after frontal activation.
Activity related to different choice options will be compared within tasks to pinpoint the exact moment the decision was made subconsciously, by determining at what moment both patterns of activity significantly differ. This moment will be related to the moment people become aware of the decision, shortly before making the movement. According to Libet's research, this final delay
between reaching awareness and initiating a move equals about 0.2 seconds (and will henceforth be called reaction time). Earlier research has revealed a substantial delay between the subconscious decision and the moment of reaching awareness which is also dependent on the type of task, ranging from 10 seconds in a simple motor task (Soon et al., 2008) to 4 seconds in task involving a more abstract decision (Soon et al., 2013). We do not yet understand the origin of this difference between tasks, but we will try to replicate the results by offering both a motor decision and a cognitive or emotional decision in the same study. However, our research is constrained by the fact that a lot of
information needs to be conveyed to the participants in order for them to make a decision, which cannot be presented all at once. Instead, information will be spread out over multiple slides by means of texts and images. This will make it harder to specify what piece of information the decision is based upon, though analysis of behavioral data may provide some insights. Timing of mental processes and the difference between the time of the decision and reaching awareness should be largely identical among participants. Also, participants are told to respond as soon as they made their choice to minimize differences in reaction time.
The emotional or motivational content of the tasks may not only determine the answer in the current trial, but also influence the decision in subsequent trials of any kind. Therefore, we will look at trial-to-trial effects on choice and the accompanying brain activity that may persist. The inter-trial-interval would normally mitigate this influence and activity would return to baseline. However, one or more emotionally unsettling trials may understandably elicit a certain mood in participants that makes the participant act as if they are primed to answer in a predetermined way. fMRI may elucidate this persisting brain activity, as a deviation from the actual baseline, and function as an insight into the neural correlates of the elicited mood, though it may not directly predict the outcome of the decision by itself, because personal factors are of great importance. For instance, considering emotionally burdening of the dilemmas, people are expected to find the answer which conveys the least amount of suffering. Therefore, they could pick the option with the better trade-off, thus saving as many people as possible, or they choose not to take action and dissociate from the problem. In the motivational task, people will be given a choice in a social game that leads to either a small or a large amount of money. The outcome of one game (one trial) may have consequences in social behavior and the decision one makes in subsequent trials, due to emotions elicited by co-players, and may also transfer to dilemma trials, where participants decide over the fate of other people. The fact of winning or losing a game could also have a role in deciding the outcome to a dilemma, possibly through loss aversion. Again, this is difficult to predict because of the subjective view on relative "loss" in dilemmas.
Methods Participants
Participants were acquaintances of the researchers and were mostly students, from varying
disciplines, and none were excluded. Twenty people took part in this study, of which 14 were female and 6 male. Ages ranged from 20 to 29 years old. The dominant hand of each participant is unknown. Every participant took part in a practice session to get used to the different tasks and to ask
questions. Stimuli
The tasks were programmed in MatLab (multiple versions) and tested for comprehensiveness on several people. All trials from different tasks were offered in random order and were grouped into either 5 or 9 blocks (see Data Collection) of about 8 minutes each to form a session.
Libet trials
The task that was originally carried out by Libet et al. was replicated here. Participants were shown a clock face numbered 1-12, with 5 minute marks. A single clock hand rotates clockwise, starting from a random position and completes the circle in 2.56 seconds. The participant had to wait until the hand made one full lap and then press a button at any moment they desired, remembering the position of the hand at the moment they felt the urge to press. They were urged to press the button as spontaneously as possible and they were able to respond with either their left or right hand. After pressing, the hand will continue to move for about two seconds to mask the exact moment the participant pressed. Next, the participant was shown a new clock face with a hand that they could move, and were told to move the hand exactly to where it was at the moment they felt the urge to press the button.
Split-or-Steal trials
The split-or-steal (SoS) task was derived from the television show Golden Balls, a British game show that was aired between 2007 and 2009 by ITV. In the final round of this show, two contestants play the prisoner's dilemma game (one iteration) for the accumulated prize money. In the prisoner's dilemma, both players have the option to cooperate (split) or defect (steal). If both players
cooperate, the money will be split equally among them. When one of the players cooperates and the other defects, the one who defects will receive all the money. And when both players choose to steal, neither of them will have the money. Before they make the choice, the players are given the opportunity to talk the other contestant to come up with a strategy.
We took still frames of the contestants of the show's episodes and noted everything that was said as well as the decision of both players, though some quotes and decisions were altered to equalize the ratio between cooperating and defecting players. Later, quotes were adjusted again to fit in the context of the trust game. This is again a two player game where both are able to make money, though the roles are somewhat different. Our participant will be the first player and starts the round with 10 points that he or she can share with the second player. They can do so be giving either 2 or 6 points. The second player will receive this amount tripled (8 or 18) and either cooperates, and gives back either half of his money, or defects by giving only one sixth. The participant may choose once
each trial to give either a small amount or a large amount, based on the picture of the contestant and the adjusted quotes, mostly advocating their own trustworthiness. By giving a small amount, the risk is significantly smaller, but so is the potential profit. The opposite goes for giving a large amount. After the decision, the trial ends and a slide will be shown indicating the points won in this trial and the new total amount of points. Participants were told that these points would be converted into a real monetary reward. This prospect of a potential reward would motivate people to carefully consider each case to maximize their profit, in contrast to Libet tasks, where the expected value of both options is equal and its required spontaneous motor action takes no consideration of options at all.
Dilemma trials
In these trials, participants were given a moral dilemma and they were prompted to either take the action described or refuse to do so by pressing the corresponding left or right button. Most dilemmas were adapted from Lotto et al. (2013), few others were from Koenigs et al. (2007) or were made up by us. Dilemmas were (re)stated so that they were covered by 5 slides, containing 10 to 16 words each. Slides were presented for 0.3 seconds per word on the slide, with a minimum of 4 seconds. On the sixth slide, the question was briefly reiterated followed by the word "Choose", to have the participant make a decision. Earlier responses were allowed because the time of the response will be used to estimate the time the decision was made. Though, responses made at the 4th slide or earlier were discarded, since participants may have missed too much vital information to base their decision on. All dilemmas were based on the trolley problem and required participants to choose between taking a proposed action or refusing to do so. Questions vary in context and mostly involve sacrificing or killing one person or a small number of people, for a certain trade-off of saving more people. These dilemmas are subdivided by self-involvement, stating whether the participant's own life is at risk or not, and by the type of action that is required. That is, in some dilemmas participants will be given the choice to kill a person as a means to save another (instrumental) and in other cases the death of one person is the byproduct of saving another person (incidental). For examples of these kinds of questions, see table 1, adapted from Lotto et al. The last category of dilemmas includes fillers, that involve moral problems other than people dying, with topics such as committing fraud, poisoning a person non-lethally, or buying a blu-ray player.
Data Collection
The fMRI sessions were carried out at the Spinoza Center's 3T scanner near the AMC in Amsterdam. Sessions initially lasted 100 minutes and participants fulfilled 2 sessions, but the long sessions were too tiring for participant. In order to prevent diminished scan quality, scanning time was
redistributed into 3 sessions of 70 minutes each. Every session was composed of a number of blocks, lasting 8 minutes each, in which multiple trials of all three conditions (tasks) were presented in mixed order. Long sessions consisted of 9 blocks, while the shorter sessions consisted of 5 blocks. Because of an error in saving the output of Libet trials, 5 people were asked to return and partake in another 90 minute session of Libet trials only.
Table 1: Examples of the 5 types of dilemmas. In self-involvement dilemmas, your own life is at risk, in contrast to other-involvement dilemmas. Further, another person's death can either be of instrumental value to saving other people, or it may be an incidental byproduct. In filler dilemmas there are no lethal casualties.
Dilemma Scenario Resolution
Instrumental
Self-involvement
You are the fourth in a team of five mountaineers involved in a climb. The head of the team has just secured himself to the rock face when the second in the team starts to slide, pulling you, and the others, with him. You all fall for tens of meters and stop suspended above a crevasse. Your weight is too much and the rope is not going to resist for long.
To lighten the load, you cut the rope which links you to the last climber. You know that he will fall into the crevasse and die, but you and the other two climbers will survive.
Instrumental Other-involvement
You are carrying out research into the behavior of lions in the Savannah of Central Africa. From an observation tower, you can see four people hide behind a bush. They are being followed by a lion and are trying to get away from it. The lion has seen them and is heading for the bush. Another person has been able to climb the observation tower.
You push off the person who has climbed the tower so that the lion is drawn towards him. You know that this person will be mauled to death, but the other four will have time to escape.
Incidental
Self-involvement
You are in the head office of your bank together with four other people. Suddenly, the director calls you because he has discovered a bomb in an office on the ground floor. He knows you are a bomb disposal expert and asks you to defuse it. You realize immediately that there is not enough time to evacuate the people in the bank before the bomb explodes.
You throw the bomb into the basement where there is the security vault. You know that the explosion will kill the security guard in the vault, but you and the other four people will be saved.
Incidental Other-involvement
You are a building worker who is maneuvering a crane on a building site. You have just started your day on the site, when you realize that the cable of the crane is about to break. Attached to the cable is an enormous steel beam which is directly above a crew of six who are working on the outside of a building in construction.
You move the arm of the crane a short distance to another area of the site. You know that there is a worker there who will be crushed by the steel beam and will die, but the other six workers will be unhurt. Filler Because of the economic crisis of the last year,
the company you work for has closed and you have lost your job. Recently you have been looking for a new job, but without success. You realize that you need some experience in
computer technology and are convinced that you will be employed much more easily if this
experience is on your Curriculum Vitae.
You insert false information in your CV about your ability in information technology. You know that in this way you will be considered above the other more qualified candidates and get the job.
All measurements were made in multi-echo format, which is widely used to improve the signal to noise ratio. However, the analysis package that was used here is unable to handle multi-echo data so all data was converted to single-echo format. Pre-processing was carried out using Steven Scholte's fMRI pre-processing pipeline (University of Amsterdam). This encompasses some crucial steps to prepare the data before analysis can take place. First, all data points of a single functional scan, corresponding to one session, are combined into a single file and aligned to a structural T1 scan to create a continuous activity pattern. This is repeated for every session of every participant. Next, a brain extraction tool detects different kinds of tissues on the T1 and removes everything except for grey and white matter and trims the functional data accordingly. Following, a motion correction detects head motion and reduces the resulting drifts in functional activity. Because the fMRI scanner measures activity sequentially, one voxel at a time, a slice time correction is needed to compensate for the latency between measurement of different spatial locations in the brain. The time required to scan a single brain once is called the TR (repetition time) and equals 2.375 seconds for this scanner. To distinguish between given inputs and underlying functional brain activity, a multi-voxel pattern analysis (MVPA) is carried out in Matlab. By training the machine on a particular set of data it is able to abstract existing patterns, which are tested on new data. Brain activity leading up to a conscious thought or a motor action will be analyzed by training the machine to distinguish between the two options in every kind of task (condition), which will each have its own pattern of activity. This is repeated 8 times per trial, for 5 timeslots before the button press and 3 timeslots after the button press, each with a length of 1 TR, resulting in a scope of 11.875 seconds of activity before the button press and 7.125 seconds after the button press. For every run of the MVPA, 90% of the available data is used for training and 10% is used for testing. In the testing phase, a certain pattern is offered to the MVPA and the pattern is classified into one of two models and after many repeats, a prediction accuracy is calculated. To cross validate this number, the process is repeated tenfold, creating models based on a different 90% each time. The process is carried out for each of the 8 TR's and all trials, blocks and sessions are combined accordingly, subdivided per condition and participant, which will result in a prediction accuracy for each TR, for all 3 conditions, for all participants.
Behavioral data
For Libet trials, a lateralization index will be calculated to identify a potential disparity between left and right presses. In split-or-steal trials it will be determined whether people are able to choose the option with the better outcome, i.e. choosing to give a large amount when the opponent is going to split or choosing a small amount when the opponent is going to steal. Further, the effect of winning or losing in one trial onto subsequent SoS trials will be explored. In Dilemma trials, responses to the different kinds of categories will be contrasted and a correlation between the average response and the trade-off will be calculated. Finally, the influence of one dilemma trial to the next will be
determined. A support vector machine (SVM) will be used to detect patterns of responses in
successive trials and thus non-random answering, by regarding 2 trials (input) and predicting the next (output). The SVM is another machine learning technique that will be trained on random input-output pairs and will generate a prediction accuracy. This process is repeated 1000 times, supplying new input-output pairs every time, to create a probability distribution of possible prediction accuracies under the condition of random pairings. Then the SVM will be trained on the actual pairs and again a prediction accuracy will be calculated, as a percentage of the total amount of predictions. This score will be compared to the probability distribution and considered significant when it's higher
than 95% of all prediction accuracy scores, meaning that the chance of an equally high score or higher score is ≤5% in random input-output pairings.
To approximate the possibility of priming between SoS trials and dilemmas, answers given in one SoS trial or dilemma, will be correlated to the next answer to a trial of the other condition. The possible interaction between the outcome of an SoS trial to the choice in subsequent SoS trials or dilemmas will also be considered. Libet trials are excluded here since those are not expected to have any influence. Finally, for each participant we will calculate a correlation between the percentage of questions that has been answered confirmatory (thus carrying out the suggested action) to the percentage of SoS trials in which the larger amount of money was donated.
Results Libet
The behavioral data showed a great discrepancy between left and right button presses in Libet trials. However, to be able to use MVPA this ratio should be equal. Otherwise, the prediction will be biased towards the more strongly represented option and either left or right button presses will be favored above the other. It is therefore impossible to create models corresponding to left or right presses. Further, due to an irresolvable error in data acquisition, no function fMRI data can be used at all. The exact cause of the problem is not yet known, but the error is expected to have occurred during multi-echo to single-multi-echo conversion. With fMRI data missing, it is not possible to replicate RP and its relation to W, though W can be determined in relation to the button press. On average, W occurs 959.4 ms (s.d. 604.7 ms.) before the button press.
Split-or-steal
Participants answered correctly on split-or-steal trials with an accuracy of 54.7%, and scored significantly above chance (t(55) = 2.69, p < 0.01). An answered is considered correct when either both players cooperate or both players defect, yielding the optimal outcome for the participant. The higher amount was offered in 49% of all trials. Unfortunately, the influence of subsequent SoS trials could not be established. This is mainly due to the design of this project, in which trials of different conditions are offered in random order to the participants. It is therefore impossible to determine which of the previous trials contributed to the decision in any given trial and the strength of each of those effects.
Moral Dilemmas
Dilemmas were overall answered confirmatory at a rate of 40.5%. This percentage was also
calculated for each dilemma separately and compared with the results of Lotto et al. using a χ² test. The results between both studies differ significantly ( = 88,83, p < 0,01). A one-way ANOVA was used to determine there's no significant effect between the category of dilemma (instrumental self-involvement, instrumental self-involvement, incidental self-self-involvement, incidental
other-involvement en filler) and the chance of a confirmatory answer (F(4,70) = 1.05), p > 0.05). To establish the effect of trade-off in the percentage of answering "yes" on a dilemma, dilemmas were grouped according to the ratio of people saved relative to people sacrificed with fillers defined as having a trade-off of zero. All dilemmas with equal ratios (1 saved for 1 sacrificed) up to a ratio of 6 to 1 had its own category and the 8 dilemmas with a higher saved-to-sacrificed ratio had its own category (figure 1). A student's t-test showed the absence of a significant correlation between trade-off and chance of a confirmatory answer (r = 0.19, t = 0.48, = 2.45, p > 0.05).
0.00 10.00 20.00 30.00 40.00 50.00 60.00 70.00 80.00 90.00 100.00 0 1 2 3 4 5 6 7 A ff ir m ativ e r e sp o n se % Trade-off
Figure 1: No significant correlation between trade-off (mean per category ± SD) and the chance of a confirmatory answer when asked to perform an action.
Combined conditions
An SVM was used to determine whether responses can be predicted across an entire session, regardless of the condition. In 8 out of 19 participants the prediction accuracy falls within the upper 5% of the area of the probability distribution and thus significantly differs from the distribution's mean. This is depicted in figure 2.
Figure 2: Probability distributions of prediction accuracies per participant and the actual prediction accuracy (red vertical lines, SD in pink). Red text indicates significant results (p < 0.5).
Prediction accuracies were plotted against a lateralization index, defined as the difference between left and right presses, divided by the total number of responses (see figure 3) and correlate
significantly (Pearson's r = 0.943, p < 0.001), indicating that predictability of responses is due to a unilateral input. However, Soon et al. preselected participants based on the tendency to alternate responses by excluding participants with a lateralization index above 0.3. In this study, 12 out of 19 participants would be excluded, removing all significantly predictable response patterns and eliminating the correlation between the lateralization index and prediction accuracy.
Figure 3: The lateralization index of given answers across all sessions per subject and related SVM prediction accuracy. Participants with red markings indicate significantly predictable responses in the SVM. Red vertical bar marks the cut-off point in lateralization index used by Soon et al. (2008).
Unfortunately it is impossible to test for carryover effects between SoS trials and dilemmas, such as the effect of losing a large amount of money in SoS trial on the decision in subsequent dilemmas. This is due to the design of this study in which trials of different conditions are presented in random order and it is therefore not possible to tell which trial or trials contributed to the current decision. No significant correlation has been found within subjects between the chance of trusting a larger amount of money and the chance of answering "yes" in dilemmas ((r = -0.15, t = -0.67, = 2.09, p > 0.05). Also see figure 4 below.
0 10 20 30 40 50 60 70 0 20 40 60 80 100 % A ff ir m ativ e % Split
Figure 4: No significant correlation between the percentage of SoS trials in which the large amount was trusted and the percentage of dilemma trials that were answered affirmatively.
Discussion Libet
Based on debriefing reports, the provided tasks proved to be confusing to subjects despite the obligatory practice session. The Libet task in particular was believed to be the most difficult since it required participants to gain insight into their own motives, by remembering the moment they became aware of wanting to perform a certain action and remembering the position of the clock hand at that time. It is the introspection that makes it most difficult for participants since that is something people don't usually do. Further, there was a lack of clarity in how to react in trials of this condition. Participants were told to use "either hand" when responding and so some participants had very high lateralization indexes, resulting in useless data to train the pattern classifier on because it will always predict the most salient response and overlook potential other, more complex patterns. Therefore it should have been stressed that both types of inputs, both left and right, were needed to base our model on, but that the order should not be planned.
Libet's W could only be determined in relation to the moment of the button press and this turned out to be considerably longer than anything found in previous research. Here, W occurred 959.4 ms. (s.d. 604.7 ms.) before the action, compared to about 200 ms. found by Libet et al. (several
publications). The only difference between the Libet study and this study could be of importance. Here, it has been decided to make the clock hand continue moving for another 2 seconds after the button has been pressed. This would make it more difficult for people to see the position of the clock hand at the moment they pressed the button because knowing this would likely influence the
participant's indication of W. Quite possibly, people compensate for not knowing when they pressed, only knowing that the urge to move occurs before moving, placing it further back in time than they would, had the clock hand stopped immediately. This has some implications for the integrity of W, a measure that is already controversial for its reliance on introspection, because it may be estimated based on the position of the clock hand's final position or the perceived moment of the button press. Split-or-Steal
The purpose of split-or-steal trials is to differentiate between functional brain activity of choices with possible rewards versus arbitrary choices to identify regions that integrate the motivational
component in decision making. Unfortunately, with fMRI data missing, we were not able to locate this functional difference. Behavioral data showed that participants were able to guess the
opponents' action with accuracy above chance level. They donated the high amount of money to a trustworthy opponent or a low amount to a non-trustworthy opponent 54.7% of the time, despite the fact that some of the opponents' choices have been altered to equalize the ratio of cooperating and defecting opponents. One explanation is that some trials are easier than others, accompanying a higher percentage of correct guesses than other trials. Or perhaps participants tend to base their judgment upon the previous trials, by choosing to give a larger amount to the next opponent after playing against a multitude of defecting opponents, and not acting out of emotional incentive by "playing safe" after losing some trials or risking more money after some successful trials. Again, these kinds of effects cannot be established with certainty since there are too many trials that could potentially influence the current trial. On the other hand, some participants complained that it was unclear how they performed in SoS trials, because the slide went by too fast. The slide, containing information about the acquired number of points in this round and the new total number of points,
did not explicitly mention the opponent's action (cooperating or defecting). Instead, the number of added points was shown from which participants were able to deduce the opponent's action, though the rates at which points are given based on both players' actions were not properly explained to participants. In combination with the relatively short exposure of the slide (about 2 seconds), it’s unlikely that subjects were able to know whether the opponent split or stole and that any emotional response was elicited due to winning or losing, resulting in the absence of a carryover effect between trials due to emotional decision making.
A final word on split-or-steal trials concerns the variable that is actually being measured. The main question has been in what way the decision making process in SoS trials differs from that of Libet trials. In SoS trials, a choice is offered between two options with outcomes of different values, which are unknown to the subjects, while in Libet trials, the expected outcome is the same for both
options. The difference between both conditions and associated brain activity would therefore be a process of estimating the expected value of each option and comparing both options within an SoS trial would expose this difference in expected value between those options. However, it remains unclear what choices are based on, which can be very different for each individual. Participants could, for example, make a decision based on results of previous trials, the appearance of the opponent, an estimation of trustworthiness, their own current mood and perhaps many other factors. It may therefore be useful to assess the participant's motivations by asking them in between trials or by supplying a questionnaire after fMRI sessions, possibly with an additional personality test. Dilemmas
When comparing different categories, personal involvement and the difference between
instrumental or incidental death seem to be irrelevant in making a decision. What does stand out is the degree to what people answer filler dilemmas in accordance to one another. The category of filler dilemmas has questions with the highest and lowest percentages of people agreeing to perform an action. It is also the one category where people refuse to perform the action unanimously,
yielding a score of 0%. Again, the mean percentage did not differ from other categories and the standard deviation was very similar to that of other categories. Discussion on this topic remains speculation, though it could signify an existing underlying mechanism. Filler dilemmas are by necessity more mundane than others, and therefore people are much more familiar to these situations. Firstly, we don’t get to decide about life or death on a daily basis and secondly there is a much more clear consensus about the socially desirable, morally justifiable and legally permitted action in given filler situations. It is therefore plausible that people generally act according to these rules. It also implies that choices in the other kinds of dilemmas are harder to predict since people are unfamiliar to these situations and on a side note, it may not reflect what people would actually do in that situation. In absence of clear rules, people look for other facts to base their decision on such as information about the context provided in the question, their intuition, and probably previous trials but there won't be a systematic approach and people may ignore facts such as personal involvement and trade-off. So there will be more disagreement between subjects and the answers will be more difficult to predict at population level, but patterns of answering on these kinds of dilemmas may provide a personal profile of the variables a decision is based on. These personal profiles could indicate if some participants did acknowledge the trade-off of a dilemma or whether the untrained participant had too little time to extract this information from the slides. Even though the slides were tested for readability, perhaps participants should be given more time to think.
Multiple conditions
In 8 out 19 tested participants, a pattern classifier was able to accurately predict the next answer based on the previous two given answers, regardless of the type of trial. Predictability is significantly correlated to the lateralization index, which in turn is most likely due to biased responses in Libet trials, because responses in the other two conditions were more equally distributed (49/51 in SoS trials and 40.5/59.5 in dilemmas). A pattern classifier should indicate whether a predictable response pattern can be confirmed between trials within the same condition, which would in fact be a
carryover effect within conditions, arguably with a subconscious cause.
Other carryover effects between SoS trials and dilemmas could not be established here because of a lack of data yet an abundance of variables. Any given answer or outcome of a trial could potentially influence any other trial in this mixed conditions design. To answer these kinds of questions some adjustments need to be made in the way trials are offered to subjects. The number of trials of a single session could be reduced and minimize the cumulative effect of many previous trials onto the next. Any combination of SoS trials and dilemmas should be provided in pairs, separated from other sets by means of a distractor task to help subjects focus on the current pair of tasks.
Conclusion
Despite lacking all functional data, some questions could be answered by using behavioral data. It may however still be useful to replicate this study with the adjustments that were mentioned above. Said research could contribute to a better understanding of the link between the type of choice and the localization of the related brain activity, allowing for earlier predictions and a better scope of the subconscious processes that evolve into a conscious thought. The next question will be what these subconscious processes encompass and why it takes a considerable time before the outcome reaches consciousness, as was stated by Soon et al.
Even though it remains unclear how subconscious activity becomes a conscious thought, we do know that a conscious choice is always preceded by subconscious evaluations. This doesn’t necessarily mean there is no such thing as free will. Back in 1982, Libet et al. hinted that decisions are made consciously by demonstrating that a decision could be vetoed after it reached conscious awareness and shows that actions are not entirely subconsciously executed and there is room for self reflection before the operation occurs.
References
Bechara, A., Damasio, H., & Damasio A.R. (2000). Emotion, Decision Making and the Orbitofrontal Cortex. Cerebral Cortex, 10, 295-307.
Greene, J.D., Sommerville, R.B., Nystom, L.E., Darley, J.M., & Cohen, J.D. (2001). An fMRI Investigation of Emotional Engagement in Moral Judgment. Science, 293, 2105-2108. Koechlin, E., & Hyafil, A. (2007). Anterior Prefrontal Function and the Limits of Human
Decision-Making. Science, 318, 594-598.
Libet, B. (1985). Unconscious Cerebral Initiative and the Role of Conscious Will in Voluntary Action. The Behavioural and Brain Sciences, 8:4, 529-566.
Libet, B., Gleason, C.A., Wright, E.W., & Pearl, D.K. (1983). Time of Conscious Intention to Act in Relation to Onset of Cerebral Activity (Readiness Potential). Brain, 106, 623-642. Lotto, L., Manfrinati, A., & Sarlo, M. (2013). A New Set of Moral Dilemmas: Norms for Moral
Acceptability, Decision Times, and Emotional Salience. Journal of Behavioural Decision Making, 27, 57-65.
Pereira, F., Mitchell, T., & Botvinick, M. (2008). Machine Learning Classifiers and fMRI: A Tutorial Overview. NeuroImage, 45, S199-S209.
Soon, C.S., Brass, M., Heinze, H.J., & Haynes, J.D. (2008). Unconscious Determinants of Free Decision in the Human Brain. Nature Neuroscience, 11, 543-545.
Soon, C.S., He, A.H., Bode, S., & Haynes, J.D. (2013). Predicting Free Choices for Abstract Intentions. PNAS, 110, 6217-6222.
