Genetic associations between alcohol phenotypes and life satisfaction: a genomic structural equation modelling approach

Alcohol use (i.e., quantity, frequency) and alcohol use disorder (AUD) are common, associated with adverse outcomes, and genetically-influenced. Genome-wide association studies (GWAS) identified genetic loci associated with both. AUD is positively genetically associated with psychopathology, while alcohol use (e.g., drinks per week) is negatively associated or NS related to psychopathology. We wanted to test if these genetic associations extended to life satisfaction, as there is an interest in understanding the associations between psychopathology-related traits and constructs that are not just the absence of psychopathology, but positive outcomes (e.g., well-being variables). Thus, we used Genomic Structural Equation Modeling (gSEM) to analyze summary-level genomic data (i.e., effects of genetic variants on constructs of interest) from large-scale GWAS of European ancestry individuals. Results suggest that the best-fitting model is a Bifactor Model, in which unique alcohol use, unique AUD, and common alcohol factors are extracted. The genetic correlation (rg) between life satisfaction-AUD specific factor was near zero, the rg with the alcohol use specific factor was positive and significant, and the rg with the common alcohol factor was negative and significant. Findings indicate that life satisfaction shares genetic etiology with typical alcohol use and life dissatisfaction shares genetic etiology with heavy alcohol use.

Alcohol-related cohorts and phenotypes.AUD-related phenotypes.AUD case/control status came from the Million Veteran Program (MVP) dataset and was defined according to ICD-9 or ICD-10 codes for dependence or abuse diagnoses from Veteran's Affairs electronic health records (EHR).Participants with at least one inpatient and/or two outpatient alcohol-related ICD-9/10 codes (from 2000 to 2018) were considered to be AUD cases 9 .AUD case/control status was available for 267,391 participants in MVP (N = 55,584 cases, N = 218,807 controls).Alcohol dependence case/control data came from a PGC-SUD meta-analysis 14 , which included over 20 datasets.Cases were defined to be meeting criteria for a DSM-IV (and DSM-III-R for one study) diagnosis of alcohol dependence and all controls were alcohol exposed (N = 46,568; N = 11,569 cases, N = 34,999 controls).
Alcohol use-related phenotypes.Drinks per week (DPW), defined as the average number of drinks a participant reported drinking each week, was examined in a combined approach with GSCAN consortium and UK Biobank (UKB) 17 (N = 941,280); see Table 1.In studies that reported binned response ranges (e.g., 1-4 drinks), the midpoint of the range was utilized 17 .The AUDIT 29 was available in multiple forms and studies.First, the AUDIT total score (AUDIT-T), was available in the 23andMe dataset 30 for 20,328 participants.Second, data from the AUDIT-C subscale, which consists of three items measuring past-year typical quantity and frequency of drinking and frequency of heavy/binge drinking 31 , were available in two datasets: EHR data from the annual AUDIT-C assessment in MVP collected on individuals between 2007-2017 N = 206,254; Ref. 9 and as part of the full 10-item AUDIT in the UK Biobank N = 121,604; Ref. 21 .Third, the AUDIT-Problems (P) subscale, consisting of 7 items that focus on the problematic consequences of drinking, was used from the UK Biobank (N = 121,604).Finally, in MVP, a quantitative measure of maximum habitual alcohol use in a typical month MaxAlc; Ref. 32 was used to reflect typical/habitual maximum use (N = 126,936).

Life satisfaction phenotypes.
Summary statistics for the life satisfaction items were taken from the Social Science Genetic Association Consortium (SSGAC; https:// www.thess gac.org/).Four items captured the extent to which individuals felt satisfied with their family, friends, work, and finances (e.g., "In general, how satisfied are you with the work that you do?") with answer choices being 1-6 and higher scores indicating more satisfaction in each domain.These satisfaction items have been shown to be a good proxy for subjective well-being 33 .The summary statistics coming from genome-wide association studies for each of these four domains were used as indicators of the life satisfaction factor in the current analyses.Within the SSGAC website, satisfaction with family (n = 168,313), friends (n = 168,001), work (115,038), and finances (n = 169,051) were taken from UK Biobank 34 .
Genotyping, quality control, and imputation.Summary statistics used in these analyses have undergone quality control pipelines applied by the specific consortia (e.g., PGC quality control pipeline including filtering to remove SNPs with imputation information value < 0.90 and minor allele frequency/MAF < 0; 01; Sullivan, 2010).The analytic pipeline for these analyses incorporates additional filtering keeping approximately 1,200,000 SNPs for each phenotype with the exception of the MVP analyses which keep approximately 625,000; Genomic structural equation modeling.We conducted analyses using the GenomicSEM package in R (version 0.0.3;https:// github.com/ Genom icSEM/ Genom icSEM/ wiki).GenomicSEM uses a two-stage SEM approach (Grotzinger et al., 2019).In the first stage, the covariance matrix and sampling covariance matrix are estimated for each dataset (see Supplementary Table 1).In the second stage, a SEM is specified and parameters are estimated by minimizing the discrepancy between the model-implied genetic covariance matrix and the empirical covariance matrix.The fit of the model can then be evaluated using standard metrics, including the standardized root mean square residual (SRMR), model chi square, Akaike Information Criterion (AIC), and the Comparative Fit Index (CFI) 35,36 .Precomputed linkage disequilibrium (LD) scores were obtained from the 1000 Genomes Project, specifically the Europeans subsample (https:// data.broad insti tute.org/ alkes group/ LDSCO RE/ eur_w_ ld_ chr.tar.bz2).For case/control samples, liability scale estimates assumed a population prevalence of 15.9% for alcohol dependence and AUD 14 .

Data analytic plan.
Factor analyses: overview of different factor models to be tested.Some of the text from this section is overlapping with our prior published gSEM manuscript 26 .The first thing we did was we asked whether a common factor model in which all alcohol-related and life satisfaction indicators loading on the same factor (Model A) would fit the data well.If Model A fits best, it would mean that these items are part of one underlying latent factor (versus more than one).Next, we tested whether either a two-factor model with all alcohol items loading on one factor and life satisfaction loading on a second factor (Model B) or a correlated three-factor model allowing for separate life satisfaction, alcohol use, and AUD-related factors (Model C) would provide better fit.If these models fit best it would mean that (Model B) the genetic influences on all alcohol indicators and life satisfaction are distinct, or that (Model C) the genetic influences on the alcohol use indicators are distinct from the AUD indicators and the life satisfaction indicators.We then asked whether a more complicated, Bifactor model that allows for factors common and specific to life satisfaction and alcohol phenotypes (Model D) would fit the data well, and whether the inter-factor correlations between life satisfaction and alcohol use and life satisfaction and AUD would differ from one another when estimating a factor common to all alcohol phenotypes.If this Model D fits best, it suggests that there are genetic influences that are common to all alcohol items, but that there are also importance genetic influences that are specific to alcohol use, and AUD, and also that the genetic influences on life satisfaction are distinct from the alcohol items.In Bifactor models, all items are allowed to load on one common factor and on their specific group factors.The group factors are allowed to correlate with one another, but their correlations with the general factor are usually set to zero.Typically, within Bifactor models, each item loads on the common factor and one specific factor.However, here the 23andMe AUDIT-T item was initially allowed to load on both alcohol use and AUD, as it is composed of items related to both use (AUDIT-C) and problems (AUDIT-P).We hypothesized that, within this framework, the genetic correlation between positive items and AUD would be significant and negative, the correlation between life satisfaction and alcohol use would be significant and positive, and the genetic correlation between life satisfaction and the common alcohol factor would be non-significant.
Factor analyses: determining number of factors.To determine which model best fit the data, we examined the substantive interpretability of each model and its loadings, including the genetic associations between the life satisfaction factor and factors common and specific to alcohol use and AUD.We also examined goodness-of-fit indices with the standard cut-offs for good fit, including a CFI: ≥ 0.9 and SRMR ≤ 0.08 and lower AIC values suggesting better fit and parsimony 37,38 .We used the zero-order genetic correlations between life satisfaction and alcohol phenotypes generated from this same author group 39 to inform which alcohol items would load onto the alcohol use-related factor (i.e., drinks per week, AUDIT-C, and AUDIT-T) or the AUD-related factor (i.e., AUDIT-T, Max Alc, AUDIT-P, AUD, and alcohol dependence).

Results
Zero-order genetic correlations among life satisfaction items.In examining the genetic correlations among the four life satisfaction items, all were significantly correlated with one another (p < 0.001).The associations were all moderate to large (absolute value r g : 0.34-0.85).Thus, we proceeded with including all four items in a one factor model.This one factor model showed great fit to the data (χ: 54.78, AIC = 70.78,CFI = 0.97, SRMR = 0.08).All standardized loadings of items on this factor were stronger than + /− 0.4 (absolute value range: 0.50-0.88).
Estimating initial models: a change to included items.Although the plan had been to include all items described above, upon viewing the loadings of the three factor model, the AUDIT-T from 23andMe had a near zero loading (−0.01,NS) on the AUD factor.Since the AUDIT-T is comprised of items capturing alcohol use (AUDIT-C) and alcohol problems (AUDIT-P), we opted to omit this item from the model entirely, rather than allowing it to load at a near zero level on one of those two factors and potentially water down the factor it loads on.Thus, the steps that are described in the next section pertain to the models we are calling "final", in which all items described earlier-with the exception of AUDIT-T from 23andMe-were included.2) (Upon viewing the loadings of the three factor model, although all other loadings were stronger than + /− 0.4 (and p < .001),AUDIT-T from 23andMe had a near zero loading (-0.01,NS) on the AUD factor.Since the AUDIT-T is comprised of items capturing alcohol use and problems, we opted to omit this item from the model entirely.Thus, the steps that are described in this section pertain to the models we are calling "final", in which all items described earlier-with the exception of AUDIT-T from 23andMe-were included).The loadings indicated that this factor was driven by the alcoholrelated factors, while the loadings of the life satisfaction items were small and/or not statistically different from zero (Supplemental Table 2).
Two factor model/model B. The two-factor model (Model B) fit somewhat better, but not well (χ 2 = 882.33,AIC = 928.33,CFI = 0.90, SRMR = 0.14; Table 2).For Model B, all the alcohol and life satisfaction loadings were significant (Supplemental Table 3).The genetic correlation between the life satisfaction factor and the alcohol factor was small and negative but significant (r g : − 0.07, p < 0.05).
Three factor model/model C. The three-factor model (Model C; Supplemental Table 4) fit adequately (χ 2 = 762.45,AIC = 812.45,CFI = 0.90, SRMR = 0.11; Table 2).All of the loadings on these factors were significant.There was a small, negative association between AUD and the life satisfaction factor (r g : − 0.17, p < 0.001).The correlation between the alcohol use and life satisfaction was non-significant (r g : 0.00, NS), but the association between AUD and alcohol use was large, positive, and significant (r g : 0.72, p < 0.001).
Bifactor model/model D. To test whether the associations between the life satisfaction factor, alcohol use, and AUD would shift when accounting for the common variation shared across all alcohol-related items, we fit a Bifactor model (Model D; See Fig. 1 and Supplemental Table 5) in which the alcohol items loaded onto 3 factors: a common factor, a residual alcohol use factor and a residual AUD factor.The correlations across these factors were fixed to zero.Model D also estimated the correlations between the life satisfaction factor and each of the three alcohol factors.Model D fit the data well (χ 2 = 553.54,AIC = 615.54,CFI = 0.93, SRMR = 0.09; Table 2) and provided the best fit across the considered models.The genetic correlation between the life satisfaction factor and the common alcohol factor was negative (r g : − 0.17, p < 0.001).The correlation between the life satisfaction factor and the unique AUD factor was not significant (r g − 0.03, NS).Finally, the correlation between the life satisfaction factor and the unique alcohol use factor was positive (r g : 0.22, p < 0.001).

Discussion
We employed gSEM to examine multivariate associations between alcohol use, AUD, and life satisfaction.This study builds upon our previous work demonstrating both improved model fit after partitioning genetic variance into separate alcohol use and AUD factors 26 and differential genetic associations across these factors and PTSD 22 .
Based on this prior work, we sought to explore potential mechanisms that may contribute to these differences in the genetic architecture between PTSD and alcohol use versus PTSD and more problematic forms of alcohol use, such as AUD.Given the positive genetic correlations between PTSD and AUD compared to the negative or non-significant genetic correlations between PTSD and alcohol use 22 , we hypothesized that the alcohol use and AUD genetic factors may represent drinking patterns influenced by distinct reasons for drinking.As such, we hypothesized that genetic correlations between unique alcohol use and life satisfaction and unique AUD and life satisfaction would be in opposing directions (i.e., positive and negative, respectively).

Best fitting model and its implications.
A Bifactor model partialling out the shared variance across all alcohol-related items was the best fitting model.Therefore, associations between factors in that model are the focus of our interpretation and discussion.Consistent with our hypotheses, we found evidence of a positive genetic correlation between unique alcohol use and life satisfaction.This is consistent with phenotypic research indicating that alcohol use is positively associated with general well-being among low-risk (but not high-risk) drinkers 40 and that low-risk drinkers report higher levels of general well-being compared to individuals who abstain from alcohol 41,42 .Previous research demonstrating that socially-motivated factors may mediate the relationship between happiness and alcohol use 43 points to the possibility that individuals who drink alcohol at higher, but not necessarily problematic, levels tend to do so out of a desire to be social and that this may, in turn, be associated with higher general life satisfaction.Previous work demonstrating that social motives for drinking are the most commonly reported reasons for drinking 44,45 and have been shown to be positively associated with www.nature.com/scientificreports/frequency and quantity of use, but not to heavy drinking or alcohol-related problems 46 may lend further support for this theory.Another possible explanation for the positive genetic association between alcohol use and life satisfaction could be the "healthy volunteer" effect, such that large population-based cohorts like those used in the current gSEM analyses tend to be overall healthier and of higher socioeconomic status than the broader population.As such, it is possible that our results are subject to selection bias.This aligns with previous work demonstrating that individuals of higher socioeconomic status tend to endorse more frequent alcohol use in the absence of non-problematic alcohol use 18,47 .
Revisiting hypotheses about correlations among factors.Our hypothesis that the genetic association between unique AUD and life satisfaction would be negative was not supported, as we did not find evidence for any association between two genetic factors.Our hypothesis was formulated based on prior work demonstrating positive genetic associations between AUD/AD and PTSD in across both behavioral genetic 48 and molecular genetic 22,26,49 research designs.We had thought that the AUD factor might represent the propensity to drink for reasons more consistent with negatively valenced intentions, such as to avoid aversive emotional states, specifically in the context of PTSD 50 .However, prior work suggests that psychopathology and life satisfaction are distinct, albeit related, phenotypes [51][52][53] and, therefore, it is possible our findings reflect a similar genotypic distinction.The null association between life satisfaction and the alcohol factor unique to disordered use suggests that genetic risk for problematic/pathological alcohol use is not associated with genetic risk for life dissatisfaction.This would be consistent with some phenotypic work suggesting that psychopathology and psychological well-being (e.g., life satisfaction, self-realization, social well-being) are not two ends of a single dimension 52,53 .
We also tested the genetic association between the common factor reflecting shared variance across alcohol use and AUD with life satisfaction.Contrary to our hypothesis that this common factor would show no significant association with life satisfaction as was the case with PTSD in prior work; Ref. 26 , the common factor was negatively associated with life satisfaction in the present study.One possible explanation for this negative association could be that the genetic signal that is shared across all alcohol use indicators, including measures of alcohol use in the general population, is generally indexing genetic propensities for heavier drinking.This may then indicate patterns of drinking associated with, or motivated by, overall life dissatisfaction.Conversely, it is possible that drinking at high levels leads to greater life dissatisfaction across the domains captured by the indicators of the life satisfaction factor in the present study (e.g., work, family, friends, finances).There is phenotypic support for the bidirectional nature of these two constructs.Specifically, findings from a large (n = 14,083), 15-year longitudinal study of healthy Finnish twins 54 showed that life dissatisfaction and adverse alcohol use (including binge drinking, passing out, high use) reciprocally influenced each other over time, and that the magnitude of this relationship increased with heavier alcohol use.
The lack of evidence of a genetic association between AUD and life satisfaction, as well as the negative genetic association between the common alcohol-related factor and life satisfaction were unexpected in light of previous findings presented by our group.Specifically, using gSEM techniques, we reported that AUD is negatively genetically correlated with PTSD and that there was no evidence of a genetic association between PTSD and the common factor representing shared genetic variance across AUD and use 26 .The discrepancy between these findings and those from the present study highlight a few important points.First, in our prior work, the PTSD factor was capturing what is common between PTSD case/control status and PTSD Re-experiencing symptoms (i.e., the two indicators); that is, the factor is primarily measuring intrusive symptoms such as nightmares, flashbacks, and repetitive, distressing images, typically related to the experienced traumatic event.In contrast, our life satisfaction factor was capturing what is common among indicators of family, friend, work, and financial satisfaction-so some general contentment with all these areas of life.In phenotypic analyses, more intrusive trauma-related (PTSD) symptoms tend to be associated with thought disorders/conditions such as mania psychosis 55 , while higher satisfaction tends to be associated with more happiness, well-being, and less neuroticism, loneliness, and depressive symptoms 33 .Second, and relatedly, it seems important to apply gSEM to questions involving positive constructs, such as life satisfaction and other positively valenced constructs associated with psychological well-being, in addition to psychiatric disorders.Doing so not only provides biological support for the notion that life satisfaction and psychopathology are not opposite ends of the same dimension, but also allows for the identification of novel patterns between previously studied phenotypes and other constructs relevant to, but distinct from, psychopathological functioning.As such, incorporation of positively valenced constructs into gSEM provides clinically relevant insight into the potential biological impact of specific treatment intervention and prevention approaches (e.g., interventions to increase life satisfaction).
Limitations.The present findings should be interpreted in light of several limitations.First, the summary statistics that were available for our multiple phenotypes of interest, particularly when collated across various datasets, were limited to individuals of European Ancestry.This is problematic for a number of reasons, including the fact that initial work from our group demonstrated that zero order genetic correlations between PTSD and alcohol-related constructs differed across European Ancestry and admixed populations 22 .Exploring these associations among more diverse populations will become more feasible as more summary statistics including individuals from other ancestral groups become available.Additionally, while not inclusive of all individuals/ populations, particularly outside of European Ancestry, the datasets from which gSEM analyses were conducted in general are large and representative, or are from consortia of numerous studies, adding to the generalizability of the findings.Second, the available summary statistics for alcohol-related phenotypes precluded analyses split by sex, which is problematic given known sex differences in molecular genetic associations between alcoholrelated phenotypes and other constructs, such as PTSD 49 , as well as known sex differences in the prevalence of alcohol-related phenotypes 56 , motives for drinking 46 , and relationships between life satisfaction and alcohol www.nature.com/scientificreports/use 57 .Third, due to modeling issues, the 23andMe AUDIT-T item was excluded from each of the models.This choice was made in large part to statistical reasons; however, for substantive reasons as well, it makes sense that this item was not included.It was important for us to examine the genetic architecture of alcohol use and AUDrelated phenotypes separately, and the AUDIT-T scale included both consumption and problems items.Thus, it does not make clear sense on which factor(s) that item would load.Additionally, due to other modeling issues, the PGC AD factor was excluded from the unique AUD factor.These modifications resulted in a Bifactor model that differed from that which was run in our earlier 26 .Finally, these findings should be interpreted in the context of the authors not having pre-registered study hypotheses.
Take-home points and implications.Despite these limitations, this study advances the field by applying a novel methodology (i.e., gSEM) to a question relating to a positive psychological construct, which is rare in the field of psychiatric genetics, and is of high importance, given the potential to increase our understanding surrounding the etiology of alcohol phenotypes and prevention of problematic alcohol use.This study also adds to a growing body of work suggesting that the genetic architecture of alcohol use and AUD are distinct and that these distinctions extend to external correlates ranging from psychopathology (e.g., PTSD) and psychological well-being (e.g., life satisfaction).This work also provides some initial evidence for shared etiology across life satisfaction and typical alcohol use, and life dissatisfaction and heavy alcohol use.Although our findings reflect association and not causality, they provide some foundation on which longitudinal, causal models might build.Specifically, work testing whether intervening on life dissatisfaction reduces heavy alcohol use specifically, may be a beneficial next step in this line of research.

Table 1 .
Descriptive information about phenotypes planned to be included.PGC-SUD psychiatrics genomics consortium substance use disorder workgroup, MVP million veteran program, UKB United Kingdom biobank, GSCAN GWAS & sequencing consortium of alcohol and nicotine use.*AUDIT-T from 23andMe not included in final models.

Table 2 .
Fit indices for Models.