Seventeen participants underwent two fMRI scanning sessions, and performed two experimental tasks, called CONTINGENT and RANDOM (see Supplementary Material for Instructions to the participants). Though the same participants took part in both tasks, the instructions emphasized the different source of the wealth shocks. The time interval between the tasks varied from 1.5 h to 15 days. The order of CONTINGENT and RANDOM for this study was not counterbalanced, with CONTINGENT preceding RANDOM. Two participants' data was removed due to inconsistent behavioral responses and scanner related issues. The remaining 15 participants' data was analyzed. The participants were five male and ten female students at the University of Tübingen, aged 18–30 (average age = 26.6 years). They were recruited for the experiment with a paper advertisement. All participants were right-handed, with no present or previous history of psychiatric/neurological disorders (Wittchen et al., 1997), no medical condition affecting the central nervous system, and no substance dependence or abuse. All participants had scores greater than 28/30 points on the mini-mental state examination (Folstein et al., 1975). Written informed consent was obtained from each participant, and approval for the research was obtained from the ethics committee of the Faculty of Medicine, University of Tübingen, Germany. Participants were informed that a minimum of 5 Euros and a maximum of 60 Euros per task would be paid for their participation in both experimental tasks.

In each of the experimental tasks, CONTINGENT and RANDOM, participants' decisions (to accept or reject a gamble) and their decision times (time from onset of stimulus till the decision was made) for each trial were recorded. Based on a participant's 256 decisions, we computed a behavioral loss aversion coefficient (LAC) for each participant. These behavioral parameters were analyzed using a repeated measures ANOVA.

A standard EPI sequence 3T TIM Trio whole-body scanner (Siemens, Erlangen, Germany), with the following parameters was used: TR = 1,500 ms, TE = 30 ms, number slices = 16. A T1-weighted structural scan was collected from each participant using the following pulse sequence parameters: MPRAGE, matrix size = 256 × 256, 160 partitions, 1 mm³ isotropic voxels, TR = 2,300 ms, TE = 3.93 ms, TI = 1100 ms, α = 8°. The above protocols were repeated in the four experimental runs, which constituted each of the two tasks. The functional images were analyzed using General Linear Model of SPM8.

While supine in the fMRI scanner, each participant performed the two experimental tasks (CONTINGENT and RANDOM) in two separate experimental sessions. The CONTINGENT task was conducted first and then followed by the RANDOM task. Each task, CONTINGENT and RANDOM, consisted of 4 runs. Each run consisted of 64 trials of 5 s each, presented in immediate succession. The 5 s duration of each trial consisted of a 4-s decision phase followed by a 1- s fixation cross (inter-trial gap).

Participants were given 20 Euro as initial endowment at the beginning of each experimental task. In each trial, a 50/50 gamble was presented, representing a 50% chance of a monetary loss and a 50% chance of a gain. For example, one trial consisted of a 50% chance of losing 19 Euro and a 50% chance of gaining 38 Euro, relative to the initial endowment of Euro 20 (Euro 1 equaled approximately US$1.20 at the time the experiment was conducted). The payoffs in the gambles were chosen to include every element in an 8 × 8 lossxgain matrix, as in Tom et al. (2007) and Pammi et al. (2015). Loss values ranged from 5 to 20, with a step size of 2, and gain values ranged from 10 to 40, with a step size of 4 (values are in terms of Euros).

Participants were told that each trial was independent of the other trials. Their task was to accept or reject the gamble presented in each trial. Accepting a gamble would yield participants either the loss or gain amounts specified in that trial. In the example trial shown in Figure 1, the loss and gain values are −19 and +38, respectively, from the initial endowment. Rejecting the gamble would result in the initial endowment of Euro 20. Participants were required to make their choice within 3 s after the arrow turned to pink. A fixation-cross against the black screen acted as an inter-trial gap. Participants were instructed to start thinking about their choice to accept or reject from the onset of the stimulus, but could confirm their choice only after the arrow turned to pink. If they took longer than 3 s to make their decision, the stimulus presentation program proceeded to the next trial after recording that there was no decision.

In the CONTINGENT task, the initial endowment was Euro 20 for the first run and the wealth available for subsequent runs was obtained from the outcome of previous runs. In this task, participants were instructed that at the end of each run, six trials would be randomly chosen to count toward their payoff. They would receive the average payoff of these six trials. Therefore, participants were aware that the average payoff at the end of each run could yield a better or worse payoff than their wealth at the beginning of the run. They were instructed that at the end of each run, verbal feedback about the success or failure of the gamble would be provided, and that the amount remaining after the resolution of the gambles would be carried over to subsequent runs.

We determined the amounts for all participants in the same manner, and those amounts were added or subtracted from participants' cash balance at the end of each run. Participants in the CONTINGENT task started with Euro 20 as an initial endowment. At the end of the first run they were made to lose Euro 10, leaving them with Euro 10 for the second run. At the end of second run they were made to win 20 Euros, leaving them with Euro 30 for the third run. From the third run to fourth run, they were made to retain the same amount, Euro 30.

In the RANDOM task, participants were instructed that they would perform trials in each run with different initial balances of wealth at the beginning of each run and that these wealth levels would be fixed by the experimenter. The starting values for each run were 20, 10, 30, and 30 Euros, representing therefore a decrease, increase, and maintenance of wealth from one run to the next. Importantly, although the computation of gains and losses was performed at the end of each run, participants were informed that no feedback would be provided about the success or failure of their gambles after each run, and rather that feedback would be provided only at the end of the fourth run.

To make both tasks comparable (Figure 1), we manipulated the changes in the endowment value for both the tasks (CONTINGENT and RANDOM) in such a way that the same values (i.e., 20, 10, 30, 30 Euros) for each run were used. Participants were also informed that a minimum of 5 Euros and a maximum of 60 Euros per task would be paid for their participation in the experiment. However, for ethical reasons, at the end of the experiment, and irrespective of their performance, the maximum amount of money was paid to each participant. The experimental design software package, COGENT 2000 (developed by the Functional Imaging Laboratory, University College London), was used for the experimental design and data acquisition.

Behavioral loss aversion coefficients were computed by conducting ordinary least squares regressions on each individual decision. A participant's decision to accept or reject a gamble, and the magnitudes of the gain and the loss were the dependent variables. Dividing the beta coefficient on losses by that on the gains yielded a loss aversion coefficient (LAC). The LAC was estimated for each run and each individual separately. To compute the coefficients, we ran OLS regressions with the following specification:

where Y_it = 1 if subject i chooses the risky option in trial t, Loss_t is the possible loss if the gamble is chosen at time t, and Gain_t is the possible gain if the gamble is chosen at time t. The regression is estimated separately for each participant i. The ratio of the coefficients “−β₁/β₂” is taken as the Loss Aversion Coefficient (LAC) for the individual. We also used the pooled data for all subjects to estimate an aggregate LAC. This single model yielded β₁ and β₂ values and based on which the loss aversion coefficients was computed for each run. The estimated loss aversion coefficients for a representative individual were 2.06, 2.54, 2.55, and 2.75 respectively, in the four runs of CONTINGENT. They were 2.58, 2.63, 2.63, and 2.76 in the four runs in RANDOM.

The loss x gain matrix of 8 × 8 (as in Tom et al., 2007) was reduced to 4 × 4 for the behavioral analysis. The decision times (DT) for each run were also recorded. The analyses of LAC and DT aimed at quantifying differences, between the CONTINGENT and RANDOM tasks, in the behavioral response to the wealth changes during each task. To obtain wealth-change-related effects, we computed differences in these measures between consecutive runs. The Run2-1, Run3-2, Run4-3 differences for both behavioral parameters, namely LAC and DT, were computed as Run2 values subtracted from Run1, Run3 values subtracted from Run2, and Run4 values subtracted from Run3, respectively. Factorial models were constructed as 2 factors (task: CONTINGENT and RANDOM) x 3 factors (consecutive run differences: Run2-1, Run3-2, Run4-3) for both LAC and DT.

Functional images were analyzed using SPM8 (Wellcome Department of Imaging Neuroscience, University College London) utilizing standard first and second level models. The functional imaging data were pre-processed using standard procedures such as realignment and unwarping, spatial normalization, and smoothing (with an 8 mm Gaussian kernel). In the first level analysis, the decision phase was modeled using decision times measured from the stimulus onset to the confirmation of response. Thus, modeling the decision phase with decision times allowed us to use for naturally occurring jitter. Separate first level models were constructed for CONTINGENT and RANDOM for every participant. Because CONTINGENT and RANDOM were performed in different scanning sessions, we modeled them separately. We then constructed a Random effects second level model pooling the data from both tasks. Each first-level design matrix was constructed using functional images from all four runs for decision phase duration. In both the brain activation and the behavioral data analysis the onset and offset of each trial is an event regressor, and the gain and loss magnitudes of the lottery available in that trial is a parametric modulator. The dependent variable, the choice to accept or reject, was obtained from the reduced payoff matrix of size 4 × 4 (Tom et al., 2007). The parametric modulation procedure allowed us to dissociate the brain activations related to potential loss and gain magnitudes. The model also included motion parameters as regressors for each run.

Separate models were constructed for each of the two tasks. This allowed us to identify the brain regions whose activations correlated with neural loss aversion in each task separately. This also permitted us to identify activation patterns that correlate with loss aversion, independently of the origin of the wealth shock. Following Dreher (2007) and Tom et al. (2007), we obtained brain regions that responded parametrically with losses and gains for each task. A two (task: CONTINGENT and RANDOM) × four (Neural LACs: Run1, Run2, Run3 and Run4) factorial design at the second level (Glascher and Gitelman, 2008) was constructed by pooling neural loss aversion related contrast images from each first level model. This factorial model resulted in brain activations pertaining to loss aversion systems for the two tasks and modulations across the four runs. In order to find a common or shared network of brain regions subserving both tasks and both positive and negative wealth shocks, we constructed a contrast by equally weighting all eight conditions in the factorial model. In order to obtain brain activations due to a task (CONTINGENT vs. RANDOM), we constructed differential statistical contrasts between tasks for each of the four runs.

Neural loss aversion (Dreher, 2007; Tom et al., 2007) was computed in the first-level models for the brain regions responding with decreasing activity in the loss modulator and increased activity in the gain modulator. The neural loss aversion was calculated for each of the two tasks separately. The neural loss aversion at every voxel was obtained by subtracting the parameter estimate (slope) of the gain response from that of the loss response (negative slope). The mean beta values corresponding to the neural loss aversion as reported in Figures 3, 4 and Table 1 were extracted over a spherical volume of 8 mm ROI from the 2 × 4 factorial model and these values for used for the computation of correlations and quantification of each of the tasks.

In addition, we also calculated correlations between the neural-loss-aversion-related brain activations (ROIs reported in Table 1) and the behavioral loss aversion coefficients (LAC) for each run and condition separately, with each individual as the unit of observation. In addition, correlations were performed for the areas observed common in both the tasks (Figure 3). The mean beta values of the General Linear Model (GLM) were extracted over a spherical volume of 8 mm from the 2 × 4 factorial model and these values were used for the computation of correlations and quantification of each of the tasks. This results are reported in Figure 5.

The behavioral loss aversion coefficients (LAC) and decision times (DT) were analyzed using a 2 factor (Experimental task: CONTINGENT and RANDOM) × 3 factor (Main effect of run: Run 2-1, Run 3-2, Run 4-3) repeated measures ANOVA.

The second level 2 × 4 factorial design yielded common and differential neural systems underlying the two tasks, as well as the loss, gain or no-change conditions present in the four runs. The common activations for neural loss aversion were found in the right amygdala, bilateral striatum, right ventral striatum and dorsal anterior cingulate cortex (Figure 3). This common network can be interpreted as describing the neural substrate underlying loss aversion. In line with the behavioral results, differential activations between the CONTINGENT and RANDOM treatments were found only for Run2 (Table 1 with threshold T > 4.22). The brain activations in the ventral prefrontal cortex/rostral anterior cingulate cortex (BA 47/32), right primary somatosensory cortex (S1)/postcentral (BA 3) and right superior occipital cortex (BA 19/18) are shown in Figure 4 and Figure S1. These brain activations also survived the small volume correction and hence corrected for multiple comparisons. The brain activations and the activity profiles across runs and tasks reveal the engagement of these areas in run 2 under CONTINGENT, after the negative wealth shock perceived to be a consequence of the subject's own decisions.

We also performed a brain-behavior correlational analysis (BBC) between the ROIs, consisting of the loss aversion related brain areas reported in Figures 3, 4 (right amygdala, right striatum, bilateral striatum, right ventral striatum and dorsal anterior cingulate cortex, ventral prefrontal cortex/rostral anterior cingulate cortex, right primary somatosensory cortex/postcentral and right superior occipital cortex), and the behavioral loss aversion coefficients (LAC) for the pooled data for the four runs. The BBCs revealed significant correlations only in the ventral prefrontal cortex/rostral anterior cingulate cortex (peak activations reported in the Figures 3, 4 and Table 1). The BBCs (Figure 5) are significant in CONTINGENT (R = 0.254, p = 0.049) though not in RANDOM (R = 0.047, p = 0.717). This result reveals the role of the ventral prefrontal cortex/rostral anterior cingulate cortex (BA 47/32), specific to the CONTINGENT task.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.