We collected samples at two different sites: Site 1 (collected at the Icahn School of Medicine at Mount Sinai) and Site 2 (collected at the University of Michigan). The sample consisted of 43 OCD patients (18 from Site 1 and 25 from Site 2) and 34 HCs (16 from Site 1 and 18 from Site 2). Fourteen OCD patients were unmedicated (uOCD, 5 from Site 1 and 9 from Site 2) and 29 were taking psychoactive medication [medicated OCD (mOCD), 13 from Site 1 and 16 from Site 2] including serotonin-reuptake inhibitors (SRIs) (n = 27), lisdexamfetamine (n = 1), and clomipramine (n = 1). The three final groups (uOCD, mOCD, and HC) were matched for age, education, and sex (see Table 1).

Participants were assessed for Axis I disorders using structured clinical interviews [for Site 1, using the Mini-International Neuropsychiatric Interview (30); for Site 2, using the Structured Clinical Interview for DSM diagnoses (31)]. HCs were excluded for lifetime diagnosis of Axis I disorder. All OCD patients met DSM-IV criteria for current OCD, excluding primary hoarding subtypes. Patients were excluded for lifetime presence of psychosis, bipolar disorder (by structured clinical interview), and major developmental or neurological disorder (by self-report). Twenty-four patients carried OCD as their only major Axis I diagnosis at the time of assessment (see Table 1 for comorbidities). Symptom severity in OCD was assessed using the Yale-Brown Obsessive–Compulsive Scale (Y-BOCS) (32). All participants gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the institutional review boards at both the Icahn School of Medicine at Mount Sinai and University of Michigan Medical School sites.

The task was based on that used by Tom et al. (25) and contained 256 trials across 3 runs. On each trial, participants were presented with a gamble showing a 50/50 chance of gaining one amount of money or losing another amount (Figure 1). Possible gains and losses ranged from $5 to $20 in $1 increments, and all possible combinations were presented across the experiment. On each trial, participants decided to accept or reject playing the gamble using button press responses. For half of participants, a decision to accept was made by the left button press (and decision to reject by right button press), whereas for the other half, a decision to accept was made by the right button press (and reject by left button press). Both gain and loss values switched between the left and right sides of the gamble equally throughout the experiment. Outcomes were not presented on each trial; instead, participants were told that one accepted gamble would be randomly selected and played (with a coin toss) at the end of the experiment. Subjects received a real monetary bonus based on participation at the end of the task.

Although the fundamental approach across study sites was the same, there were minor differences in task implementation. First, Site 1 presented the gamble for 3,000 ms, whereas Site 2 presented the gamble for 2,500 ms. Second, Site 1 participants made a binary choice of either accepting or rejecting the gamble, whereas Site 2 participants used four buttons to further qualify their choices based on confidence [accept strongly, accept weakly, reject strongly, and reject weakly; see Ref. (25)]. As in prior studies using this four-response approach (11, 22, 25, 26, 29), this was done to encourage participants to make more deliberate decisions, but data were recoded into a binary accept–reject variable (irrespective of strongly or weakly) for all analyses. Third, participants performed the task in a behavioral testing room for Site 1 and in an fMRI scanner for Site 2 (neuroimaging data to be reported separately). Due to these differences, study site was specified as an additional factor for all analyses.

Age and education were compared in two 2 (Site: 1/2) × 3 (Group: uOCD/mOCD/HC) ANOVAs. Sex was compared between groups and study sites using Pearson chi-square. Y-BOCS scores were compared between OCD patient groups with a 2 (Site: 1/2) × 2 (Group: uOCD/mOCD) ANOVA.

The effects of group and study site on percentage of reject choices were analyzed using a 2 (Site: 1/2) × 3 (Group: uOCD/mOCD/HC) ANOVA, with follow-up t-tests further interrogating the direction of significant effects. Note that because participants had to choose between rejecting or accepting each gamble, these choices were mutually determined, and there was no need to run an additional analysis of percentage of accept choices (which would provide redundant information). An analysis of decision times [reaction times (RTs)] to make choices was conducted with a 2 (Choice type: accept/reject) × 2 (Site) × 3 (Group) mixed model ANOVA.

To investigate loss aversion, for each participant, a logistic regression analysis (with no intercept) was conducted using gain values (+5 through +20) and loss values (−5 through −20) as predictors of the (log)likelihood of accepting vs. rejecting a gamble. From these subject-level regressions, beta weights were derived reflecting the relationship between increasing losses and the increasing likelihood of rejecting the gamble (β_loss) and the relationship between increasing gains and increasing likelihood of accepting the gamble (β_gain). Greater beta values for the loss compared to gain predictor are indicative of loss aversion. The loss aversion coefficient λ was computed as the ratio of β_loss to β_gain, with λ > 1 reflecting greater sensitivity to increasing losses (i.e., greater likelihood to reject) compared to sensitivity to increasing gains (i.e., greater likelihood to accept) (loss aversion), λ < 1 reflecting greater sensitivity to increasing gains than losses, and λ = 1 reflecting equivalent sensitivity to increasing losses and gains. This approach is consistent with that used in the literature (11, 22, 25, 26, 29) and makes the simplifying assumption of equal decision weights of 0.5 for gains and losses and linearity of the value function for the small values used in the study. At the group level, loss aversion coefficients were compared to the reference value of 1.0 (no loss aversion) using one-sample t-tests and were compared between the groups using a 2 (Site: 1/2) × 3 (Group: uOCD/mOCD/HC) ANOVA, with follow-up t-tests interrogating the direction of significant effects. Given that loss aversion coefficients are frequently positively skewed, we also repeated the above analysis of loss aversion using log(λ) values.

In this report, we do not consider other aspects of prospect theory such as probability weighting and editing. Although the design did not include gain-only or loss-only gambles and thus could not assess risk aversion separately from loss aversion in one model, we did conduct follow-up analyses to disentangle these two effects. For all gambles where gain = |loss| (16 trials, i.e., 50% chance of gaining 5 and 50% chance of losing 5), a risk aversion-only model has gain^α = |loss|^α (where α is the curvature of the value function) and that equality holds for all α values. For example, a risk aversion-only model would predict that a gamble with a 50% chance of gaining 5 and a 50% chance of losing 5 should be as equally acceptable as a gamble with a 50% chance of gaining 20 and a 50% chance of losing 20, so accept and reject rates should not change as gain/|loss| values increase across gambles where gain = |loss|. By contrast, the loss aversion-only model predicts increased rejection of gambles as gain/|loss| values increase even when gain = |loss|. For example, a loss aversion-only model would predict that a gamble with a 50% chance of gaining 20 and a 50% chance of losing 20 will be less acceptable—rejected more often—than a gamble with a 50% chance of gaining 5 and a 50% chance of losing 5. To test whether a risk aversion-only or loss aversion-only model best fits our data, we ran follow-up logistic regressions using gain/|loss| values (5 through 20) as predictors of the likelihood of rejecting (vs. accepting) a gamble for only those gambles where gain = |loss|. For this analysis, a positive parameter estimate would reflect an increasing likelihood of rejection (decreasing likelihood of acceptance) as gain/|loss| values increase, which, as described above, can be explained by a loss aversion model but not risk aversion model. By contrast, a parameter estimate at or near zero would indicate no relationship between gain/|loss| values and likelihood of rejection and would fit with a risk aversion model but not a loss aversion model. Subject- and group-level data were included in a generalized linear mixed model regression analysis that compared parameter estimates between the groups using the HC group as the reference.

Demographic and clinical information are presented in Table 1. There were no significant differences in age between the groups (F_2,76 = 0.1, p = 0.89) or study sites (F_1,76 = 2.9, p = 0.10) and no interaction between factors (F_2,76 = 1.2, p = 0.32). There were also no differences between the groups in years of education (F_2,76 = 0.2, p = 0.84) and no interaction between factors (F_2,76 = 1.5, p = 0.24), although participants in Site 1 had completed more years of education (16.7) than those in Site 2 (15.3) (F_1,76 = 6.3, p = 0.01). The proportion of women to men was not significantly different between groups (χ² = 3.6, p = 0.17) or study sites (χ² = 2.1, p = 0.15).

There was a main effect of group on Y-BOCS scores: unmedicated OCD patients had significantly greater symptom severity (Y-BOCS score = 24.0) compared to medicated OCD patients (19.7) [F_1,42 = 7.4, p = 0.01, Cohen’s d (effect size) = 0.79]. There was no effect of study site (F_1,42 = 0.6, p = 0.46) or interaction between group and site (F_1,42 = 1.7, p = 0.21) on Y-BOCS scores.

Overall, participants accepted significantly fewer gambles (39.2%) than they rejected (60.8%). In the analysis of percentage of reject choices, there was a main effect of group (F_2,71 = 4.6, p = 0.01), with uOCD patients rejecting significantly more gambles (accepting significantly fewer gambles) (percent reject for uOCD: 69.7%, SEM = 3.9) than mOCD patients (57.5%, SEM = 2.2, t₄₁ = 2.9, p = 0.005, d = 0.92) and HCs (56.8%, SEM = 2.2, t₄₆ = 3.1, p = 0.004, d = 0.94) (Figure 2). There was a main effect of study site (F_1,71 = 4.6, p = 0.04, d = 0.56), with participants rejecting more gambles overall in Site 2 (62.6%, SEM = 2.3) than Site 1 (55.3%, SEM = 1.8). The interaction between group and site was not significant (F_2,71 = 0.8, p = 0.45), indicating that uOCD patients’ propensity to reject more gambles (accept fewer gambles) than mOCD and HC was not different across sites.

For the analysis of decision times (RT to make choices) (Figure 3), there was an interaction between choice type (accept/reject) and group (F_2,71 = 9.0, p < 0.001). Unmedicated OCD patients were significantly faster to reject (1,220 ms, SEM = 42.9) than accept (1,366 ms, SEM = 76.2) (t₁₃ = 2.7, p = 0.02, d = 0.72), whereas HCs were significantly faster to accept (1,228 ms, SEM = 45.3) than reject (1,279 ms, SEM = 44.5) (t₃₃ = 2.6, p = 0.02, d = 0.44) and mOCD patients showed no differences in RT between choices (1,316 ms, SEM = 50.7 and 1,320 ms, SEM = 47.2 for accept and reject choices, respectively, t₂₈ = 0.2, p = 0.86). There was also a main effect of study site (F_1,71 = 31.9, p < 0.001, d = 1.46): participants were faster overall to make choices in Site 1 (1120 ms, SEM = 38.8) than Site 2 (1416 ms, SEM = 26.6). This main effect of site was qualified by an interaction with choice type (F_1,71 = 7.1, p = 0.01), indicating that the differences in RT between accept and reject choices was not the same for both sites. There was no three-way interaction between choice type, group, and site (F_2,71 = 0.1, p = 0.94).

All three groups had loss aversion coefficients (λ) that were significantly greater than 1.0 (i.e., indicative of significant loss aversion) (uOCD: λ = 1.57, SEM = 0.16, t₁₃ = 3.7, p = 0.002, d = 1.0; mOCD: λ = 1.16, SEM = 0.05, t₂₈ = 3.4, p = 0.002, d = 0.62; HC: λ = 1.13, SEM = 0.05, t₃₃ = 2.5, p = 0.02, d = 0.43). In the ANOVA comparing study site and group, there was a significant main effect of group (F_2,71 = 6.4, p = 0.003, Figure 4): uOCD patients exhibited greater λ values than mOCD patients (t₄₁ = 2.5, p = 0.02, d = 0.92) and HC (t₄₆ = 3.5, p = 0.001, d = 0.96), who were not different from each other (t₆₁ = 0.5, p = 0.65). Thus, not only were uOCD patients more likely to reject (less likely to accept) a gamble overall (Figure 2), the difference in slopes of the lines predicting rejection from loss value and acceptance from gain value was larger in the uOCD group compared to the other two groups. Note that the finding of increased λ is not equivalent to the finding of increased percentage of reject choices, because the λ analysis characterizes choices based on the change in loss/gain value, whereas the analysis of percentage reject choices does not take values into account.

Greater λ values in uOCD appears to be driven by both a steeper increase in the likelihood to reject with increasing losses (Figure 5, left) as well as a flatter increase in the likelihood to accept with increasing gains (Figure 5, right). Note that although the λ coefficient is a relative measure between the loss and gain slopes, the pattern shown in Figure 5 need not necessarily be the case and that a greater λ could be hypothetically driven by a flatter slope for gains and no change in slope for losses, or a steeper slope for losses and no change in slope for gains.

There was also a main effect of study site on λ (F_1,71 = 6.7, p = 0.01, d = 0.59), with participants in Site 2 showing greater λ values overall (λ = 1.32, SEM = 0.44) than those in Site 1 (λ = 1.10, SEM = 0.25) (Figure 4). The two-way interaction between site and group was not significant (F_2,71 = 0.99, p = 0.38), indicating that λ was greater for uOCD patients at both sites. The consistency of the λ findings across sites was confirmed by examining the main effect of group (uOCD/mOCD/HC) on λ separately for Site 1 and Site 2. The significant effect of group on λ was replicated across sites (Site 1: F_2,31 = 3.3, p = 0.05; Site 2: F_2,40 = 5.3, p = 0.009). Analysis of log(λ) values revealed the same effects, with a main effect of group (F_2,71 = 5.6, p = 0.005) and study site (F_1,71 = 5.4, p = 0.02), but no interaction between factors.

In our follow-up regression analysis, we examined the relationship between the likelihood of rejecting a gamble and the gamble’s gain/|loss| value for the subset of gambles where gain = |loss|, which allowed us to examine whether the effects of value on choice behavior were more consistent with a risk aversion-only model (which would predict no effect of value on behavior [parameter estimate not statistically different from zero] in this subset of gambles) or loss aversion-only model (which would predict a positive effect of value on behavior [parameter estimate statistically greater than zero] in this subject of gambles). There were significantly different parameter estimates for uOCD compared to HC (the reference) but not mOCD compared to HC. Overall, the parameter estimate for HC was not significantly different from zero (β = 0.03, p = 0.27), indicating that the likelihood of rejecting a gamble was not related to gain/|loss| values among this subset of gambles in the HC group. Medicated OCD patients were not significantly different from HC (β = 0.05, p = 0.27). By contrast, uOCD showed parameter estimates that were significantly greater than HC (β = 0.14, p = 0.02), indicating that as the gain/|loss| value of the gamble increased, uOCD were significantly more likely to reject the gamble. Overall, these data are consistent with our findings from the full sample of gambles showing an increased λ in uOCD, and importantly indicate that the λ effects cannot be explained by group differences in risk aversion only.

Even though there was not a significant difference in the distribution of men vs. women between the three groups, women did make up a larger proportion of the uOCD group than the other groups. Given prior research suggesting that women are more loss and risk averse than men (21, 33, 34), we investigated whether sex could affect our results. We performed an exploratory analysis comparing men and women on the loss aversion λ within each of the groups using independent samples t-tests. There were no statistically significant differences between men and women in λ, although, numerically, in all three groups, women actually showed lower λ values than men (uOCD: men: λ = 1.9, SEM = 0.2, women: λ = 1.5, SEM = 0.1; mOCD: men: λ = 1.2, SEM = 0.9, women: λ = 1.1, SEM = 0.9; HC: men: λ = 1.2, SEM = 1.0, women: λ = 1.1, SEM = 0.8). When collapsing across mOCD and HC participants (where the numbers of males and females were relatively equal), an independent samples t-test examining sex differences in λ revealed a trend effect (t₆₁ = 1.96, p = 0.06) whereby females (n = 34, λ = 1.1, SEM = 0.04) showed lower λ values than males (n = 29, λ = 1.2, SEM = 0.1). Critically, the group difference in λ between uOCD, mOCD, and HC remained significant in an ANCOVA specifying sex as a covariate (F_2,73 = 10.1, p < 0.001, covariate-adjusted means, HC = 1.1, SEM = 0.1; mOCD = 1.2, SEM = 0.1; uOCD = 1.6, SEM = 0.1). The effects of group (F_2,70 = 7.6, p = 0.001) and site (F_1,70 = 5.1, p = 0.03) also remained significant in an ANCOVA using group and study site as factors and sex as a covariate. These data indicate that the greater loss aversion among uOCD was not likely due to the greater proportion of women comprising that group.

Given that uOCD had higher symptom severity scores than mOCD patients, we conducted a post hoc comparison of λ just between uOCD and mOCD, controlling for Y-BOCS scores (and Site). Results indicated that uOCD still showed significantly higher λ values than mOCD even after statistically controlling for Y-BOCS scores (F_1,38 = 5.86, p = 0.02).

Measures of trait anxiety were collected at both sites using the Spielberger State-Trait Anxiety Inventory (STAI) (35). Three participants did not complete the scale. Of the remaining participants (n = 74), there was no correlation between self-reported trait anxiety and λ (r = 0.11, p = 0.35). In ANCOVAs controlling for trait anxiety, group differences in λ remained highly significant in models that included (F_2,57 = 7.5, p = 0.001) and omitted (F_2,70 = 9.0, p < 0.001) study site as a factor. Unfortunately, we did not obtain any other symptom measures that were the same across the two sites. Both sites did acquire measures of depression severity, although Site 1 used the Beck Depression Inventory (BDI) (36) and Site 2 used the Hamilton Depression Rating Scale (HDRS) (37). In an exploratory analysis examining the correlation between λ and depression severity within each site separately, results indicated that neither the BDI score at Site 1 (r = −0.15, p = 0.40) nor the HDRS score at Site 2 (r = 0.12, p = 0.46) was correlated with λ. These data suggest that the results are not likely to be driven by differences in general anxiety or depression severity.

Within each group, Pearson’s correlations examined the relationships between decision times to accept and reject gambles separately with the percentage of gambles that were accepted as well as λ. Due to the small sample size of the uOCD group, analyses were repeated using non-parametric Spearman’s rank correlations. For mOCD patients and HCs, there were no significant relationships using either correlational analysis approach between decision times and percent accept choices or λ. Unmedicated OCD patients showed a negative correlation between the percentage of gambles that were accepted and RT when accepting gambles (Pearson’s r = −0.85, p < 0.001; Spearman rho = −0.85, p < 0.001), indicating that patients who accepted fewer gambles (rejected more gambles) had slower decision times when they did accept them. Similarly, there was a positive correlation between loss aversion λ and RT when accepting gambles (Pearson’s r = 0.78, p = 0.001, Spearman rho = 0.84, p < 0.001). These correlations remained significant when controlling for study site (partial correlation r = −0.86, p < 0.001 and r = 0.76, p = 0.003, respectively). RT when rejecting gambles was not correlated with percentage of gambles that were accepted or λ in uOCD. There were no correlations between the above behavioral variables and Y-BOCS scores within either the uOCD or mOCD groups separately or in the full patient sample.

Dosage information was available for 25 of the 27 mOCD patients who were taking SRIs. We computed fluoxetine equivalence dosages and correlated these values with behavioral and clinical variables. SRI dosage was negatively correlated with Y-BOCS (Pearson’s r = −0.46, p = 0.02, Spearman’s rho = −0.40, p = 0.05), indicating that higher dosage was associated with lower symptom severity as might be expected. Dosage was not correlated with any other behavioral variable.