Reduced loss aversion in pathological gambling and alcohol dependence is associated with differential alterations in amygdala and prefrontal functioning

Diagnostic criteria for pathological gambling and alcohol dependence (AD) include repeated addictive behavior despite severe negative consequences. However, the concept of loss aversion (LA) as a facet of value-based decision making has not yet been used to directly compare these disorders. We hypothesized reduced LA in pathological gamblers (PG) and AD patients, correlation of LA with disorder severity, and reduced loss-related modulation of brain activity. 19 PG subjects, 15 AD patients and 17 healthy controls (HC) engaged in a LA task in a functional magnetic resonance imaging setting. Imaging analyses focused on neural gain and loss sensitivity in the meso-cortico-limbic network of the brain. Both PG and AD subjects showed reduced LA. AD subjects showed altered loss-related modulation of activity in lateral prefrontal regions. PG subjects showed indication of altered amygdala-prefrontal functional connectivity. Although we observed reduced LA in both a behavioral addiction and a substance-related disorder our neural findings might challenge the notion of complete neuro-behavioral congruence of substance-use disorders and behavioral addictions.


Administered questionnaires
PG subjects were diagnosed using the German short questionnaire for gambling behavior questionnaire (Kurzfragebogen Spielsucht, KFG) (cutoff ≥ 16) (Petry and Baulig, 1996), internal consistency, i.e. Cronbach's Alpha = 0.79, retest reliability 2 weeks = 0.80 (Petry, 1996). According to the KFG 4 subjects displayed mild, 14 subjects medium and 1 subject severe PG. Otherwise, any known history of a neurological disorder or a current psychological disorder (except substance abuse and tobacco dependence) as assessed by the screening of the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I) (First et al, 2002) lead to exclusion from the study. For matching purposes subjects completed the Wechsler Intelligence Test for Adults (WAIS) matrices test (Wechsler, 1997) and they were asked to indicate age, smoking status, amount of personal debt and monthly personal income. Furthermore, they were asked to indicate their level of education and handedness (Oldfield, 1971). For further characterization of the three groups subjects also completed Beck's Depression Inventory (BDI-II) (Beck et al, 1996) and the Barratt Impulsiveness Scale Version 10 (BIS-10) (Patton et al, 1995).

Reaction times
Reaction times were submitted into a linear mixed effects model with random effects (Bates et al, 2015b), where centralized gain, centralized loss, absolute Euclidean distance were fixed effects and also allowed to vary randomly per subject, using the lmer function in lme4 in R. In a second model fixed effects of gain and loss were modulated by group. Both models were compared using the anova function in R performing a Chi-Square-Difference test. Post-hoc ttests were performed using Satterthwate's approximations implemented in lmerTest (Kuznetsova et al, 2016). 5
To perform model comparison we estimated each model using the glmer function in lme4 (Bates et al, 2015b) in each group separately or with all groups together using group as a between subject fixed effect, respectively. From the glmer models we could simply note down the Aikaike Information Criterion (AIC) values and computed mean AIC values, so that all reported AIC values are always "mean AIC per subject" values. Only this way AIC values can be compared between groups, because the groups have different sizes (Table S1).

6
The la model had the lowest mean group AIC value (i.e. best model), also reflected in the likelihood ratio tests comparing all models to lae (Table S1). We thus chose for the analyses in the main text the la model. We used mixed effects modeling because it yields more robust single-subject parameter estimates and also mixed effects modeling is designed to estimate group fixed effects (Bates et al, 2015a).
We computed ′ per model and correlated them. The lambdas of la correlated well with 's of all other considered models ( Table S2). We also computed mean values per group and model and performed group comparisons. For this we extracted the fixed effects and random effects and added them and computed one values per subject and model. Note that this is a different but very fast method to estimate the fixed effect of loss aversion ( ) and get standard errors of the parameters. This method was only used here, not in our main analysis, where we bootstrapped parametrically the p-values for group comparisons of fixed effects of . Note that all models yielded the same expected group differences (Table S3).   (Andersson et al, 2001). The T1 image was co-registered to the unwarped mean GE-EPI image using affine spatial transformation. The T1 image was then segmented into tissue classes and transformed into the Montreal Neurological Institute-standard space (MNI). This process yielded linear and non-linear parameters for the transformation between individual and standard space, which were applied to all unwarped EPI images. EPI images were resampled to a voxel size of 3.5mm x 3.5mm x 3.5mm. Finally, these images were spatially smoothed with an isotropic Gaussian 9 kernel (full-width-at-half maximum 8mm). Additionally, we used the VBM8 toolbox (Kurth et al, 2010) to segment T1 images into tissue classes. Gray matter tissue probability maps (TPMs) were than warped into standard space, spatially smoothed and down sampled to a voxel size of 3.5mm x 3.5mm x 3.5mm to match the resolution of our functional images. These gray matter TPMs then represent local gray matter volume or local gray matter density (GMD) (Good et al, 2002), irrespective of overall brain size, and lend themselves for BPM analysis.

The fMRI single-subject model
Additionally, the head motion parameters obtained during motion correction were entered into the model to account for signal fluctuations caused by the interaction of movement and susceptibility (Morgan, Dawant, Li, & Pickens, 2007). After high pass filtering (cut off frequency = 1/128 Hz) and the elimination of high frequency noise by autoregressive (AR (1)) modeling, the General Linear Model (GLM) was fit to the preprocessed EPIs using a restricted maximum likelihood algorithm. Only gray matter voxels according to the SPM12 gray matter template (p > 0.2) were considered.

12
For VTA we used a probabilistic ROI of the midbrain (Murty et al, 2014). These authors constructed a midbrain mask based on hand-drawn VTA-substantia-nigra-midbrain masks of 50 healthy subjects. For DRN we used an 8mm radius sphere around the MNI coordinate [-2, -32, -16] (Pedroni et al, 2011). Note that both areas are quite large with respect to the actual size of the mentioned nuclei to account for inter-individual differences. These masks for VTA and DRN were chosen because these areas are not part of the SPM12 atlas, nor the AAL atlas. For DLPFC we used the WFU pick atlas to select Brodman areas (BA) 8,9,10 and 46 (dilated in 2D, i.e. in-plane, by 1 voxel) (Collins, 2001;Draganski et al, 2008;Maldjian et al, 2003) within the middle frontal gyrus according to the AAL atlas (Tzourio-Mazoyer et al, 2002). For VLPFC we used BA 44, 45, 47 (dilated in 2D, i.e. in-plane, by 1 voxel) within the inferior frontal gyrus (Badre and Wagner, 2007;Danker et al, 2008;Gold et al, 2006) (Figure S1). The complete NOI can be found as .nii file in the Supplementary Online Material.

The rBPM analysis
Robust biological parametric mapping (rBPM in toolbox BPMe) was used running on SPM5 (Casanova et al, 2007;Yang et al, 2011) and results were evaluated in SPM8. Note that BPMe is only available for SPM5 but results may be evaluated in SPM8 but not in SPM12.

Functional connectivity
We fit new single subject models. Specifically, for every seed region we expanded the standard single subject model by interaction terms multiplying the time series of the respective seed region and each parametric modulator (McLaren et al, 2012). All the other terms in the single subject model, including motion parameters as covariates of no interest, stayed the same. We then submitted the contrast images pertaining to the interaction terms for gain and loss to second-level T-tests comparing PG and AD to HC, respectively. In the ANDREA model (Fig. S2), when LA exists, the amygdala sends a salience signal to OFC which is stronger for losses than for gains. This enhances the represented loss value over the represented gain value in OFC. Lack of LA may thus emerge from a more efficacious transmission of the amygdala salience signal for gains. We thus expected a functional connectivity which grows more strongly for increasing gains in both PG and AD subjects compared to HC subjects.
According to (Basten et al, 2010) (Fig. S2), the VMPFC is said to be a comparator region integrating cost signals from amygdala and gain signals from VS. We hence computed a gPPI analysis on single subject level with amygdala as seed region and used the VMPFC ROI for small volume correction and expected HC to show stronger functional connectivity from amygdala to VMPFC with respect to growing losses than both PG and AD subjects.
Found group differences in functional connectivity were checked for stability against adjusting for age using ancova analysis in SPM. Only results are reported which survived adjustment for age.
16 Figure S2: Network models for gPPI analyses. The ANDREA model (left, adapted from (Litt et al, 2008)) and the model by Basten et al. (right, adapted from (Basten et al, 2010). The network models were used as hypotheses generators regarding differences in functional connectivity between PG, AD and HC subjects. The arrows mean functional connections. Next to the arrows it is stated whether the connection processes gain or loss signals.

Reaction times
Inclusion of group into the behavioral model was significant, Δdf = 6, p(ΔChi 2 ) = 0.023, ΔAIC = 2. The HC group showed a mean reaction time (rt) of 1.27s, the AD group of 1.54s and the PG group of 1.39s. HC's rt was shorter than that of AD subjects (HC < AD, p = 0.030). AD patients showed a stronger increase in rt with growing losses than HC subjects (β = 0.019, p = 0.019), also PG subjects showed this (β = 0.018, p = 0.018). With increasing gains, PG subjects showed a stronger decrease in rt compared to HC (β = -0.011, p = 0.033) ( Figure S3). Adjusting for age by allowing age to impact the fixed intercept and the rt within each group, yielded the same results, except the overall mean difference in rt of HC vs. AD and HC vs. PG was rendered insignificant. Figure S3: Reaction times. Depicted are mean reaction times (time until decision is made) per group and condition in seconds with bootstrapped 95% confidence intervals. MeanRT is the mean reaction time at presentation of mean gain and mean loss in the proposed gamble. Bygain shows how this meanRT changes when gain increases by 5 euros. Byloss shows how reaction time changes when losses increase by 5 euros. Note that PG and AD subjects change their reaction times as a function of gain and loss but not HC subjects.

Debt
We have checked the relationship of debt (yes/no) (28 yes, 19 no, 4 NA) and loss aversion. The median LA for no debt was 1.64 and for debt 0.97. This difference was significant (Kruskall-Wallis test, p = 0.02). We fit our original model (group explaining behavioral gain and loss sensitivity) and the alternative model (debt (yes/no) explaining behavioral gain and loss sensitivity), while excluding in both cases the 4 subjects which did not provide information on their debt. Model comparison showed that the group model was still slightly better than the debt model: Δdf = 4, Chi 2 = 11.4, p = 0.022, ΔAIC = 3.5 (AIC of group model better than that of debt model). We could not usefully correlate the amount of debt with behavioral LA because we had too many missings (15 NA) in the variable "amount of debt". This is because 15 subjects declined to answer this question. 20 Figure S4: Functional connectivity group differences. A,B: PG subjects show a stronger functional connectivity from Amygdala to posterior OFC with regards to growing gains. It seems they transmit the amygdala signal with respect to gains more and more efficaciously to OFC, when gains increase, while HC subjects do so less or even do the reverse. C: With growing losses HC subjects show stronger connectivity increase from left amygdala to VMPFC than PG subjects. It seems they transmit the amygdala signal with respect to losses more and more efficaciously to OFC, when losses increase, while HC subjects do so less. D: The same is true for the functional connectivity from left posterior OFC to DRN.