Genetic and Environmental Factors Underlying Parallel Changes in Body Mass Index and Alcohol Consumption: A 36-Year Longitudinal Study of Adult Twins

Obesity and heavy alcohol consumption are major public health issues worldwide due to their high and increasing prevalence in recent decades [1, 2]. Body mass index (BMI) is a widely used indicator for measuring obesity and is associated with a large number of diseases such as cardiovascular diseases, several cancers, and mental health [3], reinforcing the value of studying the genetics of obesity and weight gain. The genetics of obesity is well established through twin and family studies [4], and genome-wide association studies (GWAS) [5, 6], allowing for more in-depth approaches, such as precision nutrition [7] and nutrigenomic [8] studies.

Alcohol consumption, despite its high caloric content [9], influences weight gain in a way that remains unclear [10] due notably to nutritional disorders induced by alcohol use [11]. The relationship between obesity and alcohol-related behaviors hinges on plural and complex patterns incorporating dimensions of mental health [3, 12], behavioral sciences [13], and molecular biology [14]. Although alcohol consumption and obesity seem to be positively associated according to a recent meta-analysis [15], the absence of significant pooled results within cohort studies has been reported, showing important discrepancies between studies, even in cross-sectional designs [15]. Despite the existence of such inconsistencies, heavy alcohol consumption and obesity constitute a public health burden in that their combined association causally increases the risk of developing high-prevalence diseases, such as hypertension [16] or liver disease [17], reinforcing the importance of disentangling how obesity and alcohol consumption are associated.

Association studies involving BMI polygenic risk scores (BMI-PRS) and alcohol consumption do not yet appear to have been reported in the literature. Associations between PRS-BMI and energy intake have, however, been investigated but have not led to significant results [18]. Two other studies investigated associations between PRS-BMI and lifestyle factors [19, 20], without including alcohol consumption as such in the analyses. Despite the lack of evidence at the polygenic score level in the literature, a weak negative genetic correlation (rG = −0.05) between BMI and drinks per day was recently reported in a very large multi-cohort genome-wide association study [21].

Alcohol use is widely known to have a genetic component [22, 23] and substantial interindividual differences, such as sex differences [24, 26]. Large epidemiological studies have also found differences and inconsistencies between men and women in the association between drinking at baseline and weight gain [27, 30]. While sex differences may be reflected in the magnitude of associations between weight change and alcohol consumption, as in the study by Downer et al. [27], other studies show significant results in men but not in women [30]. Additional sex-based inconsistencies have been observed in a study with longitudinal measures of alcohol use, where alcohol consumption significantly correlated with changes in BMI in men but not in women [31]. Another study based on more than 19,000 women reported less weight gain in initially normal-weight women who drank light to moderate amounts of alcohol compared with nondrinkers, despite the caloric intake of alcohol. These findings encourage a sex-specific approach to the study of the longitudinal relationship between BMI and alcohol consumption.

Previous longitudinal twin studies have independently described BMI and alcohol consumption trajectories in Finnish adults [22, 32] using so-called latent growth curve modeling (LGCM). However, these studies have not allowed the investigation of potential genetic or environmental correlations between BMI and alcohol consumption trajectories. The aims of the current study were therefore (1) to study the relationship between BMI and alcohol consumption trajectories, (2) to highlight the genetic and environmental components underlying changes in BMI and alcohol consumption, and (3) to investigate possible genetic and environmental correlations between BMI and alcohol consumption trajectories. We derived indicators of BMI and alcohol consumption trajectories using LGCM from a cohort of twins followed over 36 years. Genetic and environmental influences associated with changes in BMI and alcohol consumption and their mutual associations were then estimated using multivariate twin modeling. Further, we analyzed sex differences in these associations.

Data from the older Finnish Twin Cohort (FTC) were used in the current study. The FTC is a population-based twin cohort [33] which provides valuable resources for the study of complex phenotypes, including obesity and alcohol use [34, 35]. The study participants were monozygotic (MZ) and same-sex dizygotic (DZ) twins born before 1958 [36], for which questionnaire data were first collected in 1975. FTC participants responded to up to four waves of questionnaires from surveys conducted in 1975, 1981, 1990, and 2011. The data included measures of weight, height, and alcohol consumption. Self-reported height and weight measures were found to be reliable when they were examined from a subsample of participants who completed the 2011 questionnaire: correlations between measured and self-reported BMI were 0.95 for both men and women [37].

A total of 24,384 adult same-sex twin participants had data from at least one survey, and they entered the data preprocessing phase (Fig. 1). After applying exclusion criteria, 4,461 participants constituted the final sample (58% females). The male subsample included 483 complete same-sex twin pairs (34% MZ pairs), and the female subsample included 803 complete same-sex twin pairs (40% MZ pairs). The pre-processing steps required to obtain the final sample are contextualized in Figure 1 and detailed as follows:

1. 1.

   BMI measures for each wave were calculated using the associated participant height and weight measures. Height measurements identified as missing were supplemented with height measurements from other questionnaires when the information was available. Three participants with inconsistent BMI values were excluded from the study. Alcohol measures, expressed in grams per month [23], were log-transformed as well as BMI measures to remove the skewness of the original variable (BMI skewness ranged 0.62–1.07 in men and 1.07–2.47 in women before log-transformation; 0.25–0.48 and 0.48–1.11, respectively, after) [22, 32]. To correct for any zeros that may be present in the alcohol consumption variable, f:t → log(1+t) was used as log-transformation for the alcohol consumption measures (alcohol consumption skewness ranged 2.26–3.61 in men and 4.22–4.99 in women before log-transformation; −1.93 to −1.54 and −1.51 to −1.21, respectively, after). Of the initial participants, only those whose BMI and alcohol measures were complete on all four questionnaires were retained. This pre-selection led to a first sample of 4,663 participants (Fig. 1).

2. 2.

   The use of covariates for additional exploratory purposes was considered 38‒40 (online suppl. Table S1, S2; for all online suppl. material, see www.karger.com/doi/10.1159/000529835), and led to a reduction of the sample. Among these covariates, a variable indicating smoking status was not recorded for 202 participants. Despite the absence of these covariates in our study, participants with missing smoking data were excluded to compare the phenotypic correlations obtained in the main study and those produced in the supplementary sensitivity analysis (online suppl. Table S3). This exclusion criterion led to a final sample of 4,461 participants.

Overall, the average weight and alcohol consumption tended to increase over time in men and women. Alcohol consumption in men ranged from 2.29-fold to 2.56-fold more than women across the 4 questionnaires. The average BMI increased over the period 1975–2011 from 22.5 (standard deviation: 2.5) kg/m² to 26.6 (3.9) in men and from 20.8 (2.5) to 25.7 (4.4) in women. The proportion of participants with overweight (BMI >25 kg/m²) increased over the 4 follow-up timepoints from 13.6% to 63.2% in men and from 5.8% to 50.2% in women. The proportion of participants with obesity (BMI >30 kg/m²) increased from 0.9% to 15.9% and from 0.6% to 15.2% in men and women during follow-up, respectively. The distributions of logarithmic BMI, logarithmic alcohol consumption, age, and untransformed raw values across questionnaires are described in Table 1. Multivariate twin modeling was performed only on the subset of complete same-sex pairs.

Structural Equation Modeling (SEM) has become a common set of methods used in the analysis of behavioral phenotypes [41, 42]. We used LGCM, an SEM-based method, to represent repeated-measures data as a function of time. The linear LGCM derives two latent variables, hereafter called growth factors, to characterize within-person change over time. The growth factors consist of an intercept and a slope that can be interpreted as a rate of change. The means of the growth factors therefore describe the average trajectory at the population level, while their (co-)variances provide key information about the magnitude of inter-individual differences.

The LGCM, in the study of parallel processes, allows (1) intercepts and slopes of traits to freely covary or (2) estimation of growth factors to be constrained from other growth factors. In our multivariate study, the slopes and intercepts derived from the four measures of BMI and alcohol consumption were freely derived. That is, they were free to correlate without influencing each other during the model-fitting phase (Fig. 2). This modeling was sex-specific to correct for structural differences in alcohol consumption between the sexes that are well-established in most populations [22, 24, 26]. Nullity tests of the means and variances of the growth factors were performed to describe the trajectory of each trait, for each sex. Phenotypic correlations between growth factors were also calculated, and growth values were used in twin modeling (Fig. 1). A p value lower than 0.05 was considered sufficient to reject the null hypothesis of nullity of the mean, variance, and correlation. The modeling was conducted using the R package lavaan [43], version 0.6-10.

The value of using twin cohorts to understand the genetic and environmental components underlying complex phenotypes is well established [33]. One of the applications of twin study designs is quantitative genetic modeling, based on the comparison of MZ and DZ twins. MZ twins share 100% of their DNA, while DZ twins share on average 50% of their segregating genes. This model sources on the differences in phenotypic intra-class correlations between MZ and DZ twins: these are assumed to reflect genetic and environmental components.

Classical twin modeling decomposes the total trait variance into genetic and environmental components. If the DZ twin intra-class correlation (rDZ) is less than half of the MZ twin intra-class correlation (rMZ), the genetic component can be subdivided into an additive (A) and non-additive (D) component, reflecting the summed allelic effects across genes and the genetic dominance, respectively. If rDZ is greater than half of rMZ, a shared (C) environmental component reflecting shared environmental effects (e.g., such as family common environment and other effects that make siblings similar) can be estimated along with the A component. Further, in all models, a non-shared (E) environmental component (i.e., unique to each sibling) will be estimated, reflecting individual-level environmental factors as well as measurement error. Since the D and C components cannot be estimated simultaneously, the combinations of classical twin modeling in a univariate context where a genetic component is considered certain are mainly summarized in the ACE and ADE models, which can be compared downstream to the simplified AE model.

We calculated the intraclass correlations of growth factors between DZ and MZ twins to determine the most appropriate C or D component for model fitting (Table 2). The pattern of correlation for the alcohol consumption intercept suggested the estimation of the C component rather than the D component: the ratio of DZ to MZ intraclass correlations was substantially greater than 50% in women and men. The set of intra-class correlation estimates with their associated 95% confidence intervals is provided in Table 2.

The heritability (h²) is defined, in this classical configuration, as the share of the phenotypic variance attributed to additive genetic variance. The coefficients e² and c² (or d²) are defined as above and provide information on the part of the variance explained by the individual’s unique environment and the environment shared by the co-twins (or genetic dominance, respectively). In a multivariate context, the covariances between traits can be decomposed into genetic and environmental parts using the Cholesky decomposition and can provide genetic and environmental correlations. The path coefficients associated with the respective components A, E, and C or D thus provide information on the cross influence of genetic and environmental components on the set of growth factors.

We investigated the genetic and environmental influences of the growth factors of BMI and alcohol consumption in a multivariate setting. Twin modeling was performed using the complete same-sex twin pairs derived from the LGCM with their growth values (Fig. 1). Because of heterogeneous intra-class correlation between MZ and DZ twins within the four latent variables (Table 2), several models were tested. An unconstrained saturated model estimating all parameters was first performed to build models based on A, E and C or D components to ensure that the twin modeling assumptions were met. These assumptions include the equality of means and variances between co-twins and between zygosity types [44]; we therefore tested these assumptions for each sex. The equality of environments between MZ and DZ twins was also assumed. Evaluation of the ACE and ADE models was investigated by comparing them to the AE model through the Akaike information criterion (AIC), which was preferred to the −2 log-likelihood measure (−2LL) [45]. Statistical power to detect genetic and environmental components was evaluated [46, 47] and is presented in the supplementary material (online suppl. document, section S2). Confidence intervals for the parameters (h², e², c², and d²) and for correlations between genetic components and between environmental components were calculated with a 95% confidence margin. The genetic twin modeling was conducted using the R package OpenMx [48], version 2.20.3.

An ACE and ADE model were fitted for men and women by decomposing the variation of each growth factor into A, C or D, and E variance components. These two models, hereafter referred to as full ACE and full ADE, did not decrease the model fit significantly and can thus be regarded as more parsimonious as compared to the saturated model in both men and women (p value <0.05, likelihood ratio test). These two models were then compared with the AE model, deriving only an additive genetic component and a non-shared environmental component. The AE model outperformed the full ADE and full ACE models in terms of AIC, for modeling of men (Table 3). The full ACE and full ADE models showed better AIC performance than the AE model (full ACE: −20161.51 vs. AE: −20160.96; full ADE: −20159.51 vs. AE: −20158.96) for women but not for men. To remove the inflection effect of the shared environmental component estimate for all growth factors on the performance measures, a partial ACE model deriving only a shared environmental component at the alcohol intercept was constructed because of (1) the intraclass correlation ratio of this growth factor for both sexes and (2) the significant c² component estimate in the full ACE model (Table 3). This partial model increased the difference in AIC with the AE model compared to the differences observed with the full ACE model. The partial ACE model also outperformed the ADE model when compared to the saturated model (partial ACE: AIC = −12448.52; ADE: AIC = −12438.55). This model being the most parsimonious, it was kept as the final model.

Phenotypic correlation analysis was performed using growth values characterizing trait trajectories derived from the LGCM (Fig. 1). The trajectory associated with log alcohol consumption showed a general trend, among both men and women, toward a generalized increase over the period 1975–2011. This rate was 0.007 log-units per year from baseline for women and 0.011 log-units per year for men from baseline. The population-wide trend in BMI increase was 0.004 log BMI units per year in men and 0.006 log BMI units per year in women (Table 1). Additional information on log BMI trajectories, log alcohol consumption trajectories, and growth factor distributions can be found in the online supplementary document.

Associations between growth factors, i.e., correlations or inverse correlations, were observed. The correlation between intercept and slope of alcohol consumption was moderate in men (r = −0.55, 95% CI: −0.58, −0.51) and somewhat weaker in women (r = −0.33, 95% CI: −0.37, −0.30). The relationship between BMI intercept and BMI slope was positive in women (r = 0.08, 95% CI: 0.04, 0.12) but negative in men (r = −0.10, CI: −0.14, −0.05). Alcohol and BMI slopes were significantly correlated and estimated at 0.16 (95% CI: 0.11, 0.20) in men but were not significant in women (r = 0.02, 95% CI: −0.01, 0.06). Phenotypic correlations between growth factors were therefore found to be of different magnitude, sign, or significance depending on sex. The set of correlations is shown in Table 2.

A sensitivity analysis was performed to compare the phenotypic correlations of the study with those obtained by adjusting log BMI and log alcohol consumption variables for covariates such as smoking status, marital status, or education level in addition to age. This comparative analysis is shown in the supplementary document (online suppl. Fig. S1).

A total of 483 same-sex pairs of male twins and 803 same-sex pairs of female twins were used to decompose the variance and covariance of the LGCM-derived growth factors (Fig. 1). The heritability of the BMI intercept was estimated to be h² = 79% (95% CI: 74, 83) in men and h² = 77% (95% CI: 73, 81) in women, suggesting no sex-related difference in the heritability estimate of the BMI intercept. The share of heritability within alcohol intercepts was estimated at h² = 49% (95% CI: 32, 67) in men and h² = 45% (95% CI: 29, 62) in women. The common environmental variance component for the alcohol consumption intercept was estimated at c² = 19% (95% CI: 5, 33) in men and c² = 18% (95% CI: 4, 32) in women. All the estimates of the full ACE, full ADE, partial ACE, and AE models can be found in Table 3. No sex differences in the heritability of the BMI intercept (log-likelihood ratio test: p = 0.51) and the alcohol consumption intercept (p = 0.17) were observed.

Men, for whom the relationship between alcohol consumption intercept and slope was significantly stronger than in women (Table 3), were conditioned by a hereditary component for alcohol consumption slope estimated at h² = 45% (95% CI: 34, 54), whereas it was h² = 31% (95% CI: 22, 38) in women. On the other hand, the non-shared environment component for BMI change was e² = 48% (95% CI: 39, 58) in men and e² = 37% (95% CI: 37, 50) in women. A model constraining heritability estimates of the slope of alcohol consumption to be equal in men and women significantly decreased the model’s fitting performance (log-likelihood ratio test: p = 0.03), suggesting sex differences in the heritability of changes in alcohol consumption. The heritability of BMI change was estimated to be h² = 52% (95% CI: 42, 61) in men and h² = 57% (95% CI: 50, 63) in women, with no evidence for sex differences in heritability (p = 0.39).

A positive genetic correlation was observed between the BMI intercept and the BMI slope in women (rG = 0.24, 95% CI: 0.14, 0.33) but not in men (rG = −0.08, 95% CI: −0.19, 0.05). The additive genetic components of the intercept and slope of alcohol consumption were found to be significantly correlated for women (rG = −0.37, 95% CI: −0.49, −0.24) and men (rG = −0.82, 95% CI: −1.00, −0.65), but with different magnitudes (Fig. 3). A similar genetic correlation between BMI intercept and alcohol consumption slope was observed in men (rG = −0.17, 95% CI: −0.29, −0.04) and women (rG = −0.18, 95% CI: −0.31, −0.06). These results support the existence of complex genetic architectures influencing (1) alcohol consumption and changes in alcohol consumption, (2) BMI with changes in BMI, and (3) BMI intercept with alcohol-related behavior over time. No other genetic correlations between BMI and alcohol consumption growth factors reached significance.

Correlations between non-shared environmental components were also widely observed (Fig. 3). The non-shared environmental components of BMI intercepts and slopes were negatively correlated in men (rE = −0.22, 95% CI: −0.34, −0.09) and women (rE = −0.15, 95% CI: −0.24, −0.05). Growth factors for alcohol consumption also covaried via their environmental components in both men (rE = −0.32, 95% CI: −0.43, −0.20) and women (rE = −0.26, 95% CI: −0.35, −0.17). In addition to environmental correlations between growth factors of the same trait, correlations between environmental components of BMI and alcohol consumption were observed. The non-shared environmental components of alcohol consumption intercept and BMI slope were negatively correlated in women (rE = −0.11, 95% CI: −0.20, −0.01). BMI change was also positively associated with alcohol consumption change in men via their non-shared environmental components (rE = 0.18, 95% CI: 0.06, 0.30). All correlation estimates between environmental components are presented in Figure 3.

The objectives of our study were, from a sex-specific perspective, (1) to study the relationship between BMI and alcohol consumption trajectories, (2) to highlight the genetic and environmental components underlying changes in weight and alcohol consumption, and (3) to investigate possible genetic and environmental correlations underlying the parallel BMI and alcohol consumption trajectories. Ultimately, our study revealed substantial sex differences in the relationship between BMI change and alcohol use change. Sex-related differences were primarily observed in longitudinal BMI or alcohol consumption changes, sharing common sex-specific environments with baseline. At the univariate level, a significant sex difference in the heritability of change in alcohol consumption was found, but no evidence of such sex differences in other growth factors was observed.

The heritability of baseline BMI (i.e., intercept) remained consistent with traditional BMI estimations [49] and was similar to those measured in the three waves of 1975, 1981, and 1990 on the same cohort (h² = 80% in men, h² = 82% in women) [32]. Estimates of the genetic and environmental components of the baseline alcohol consumption (intercept) were also consistent with the literature [22, 50]. Thus, these heritability estimates of BMI and alcohol consumption baseline levels were similar in men and women and echoed those of the literature.

The phenotypic correlation between baseline alcohol consumption and change in BMI was found to be weak but significantly negative in women. This result is consistent with large-scale studies of non-overweight women with shorter follow-ups of 7 years and 12.9 years [28, 29], as these studies also showed a negative correlation between baseline alcohol consumption and weight gain in women. In our study, the non-shared environmental component of the baseline alcohol consumption and change in BMI were also significantly correlated in women (rE = −0.11), which provides additional evidence for a potentially causal relationship, which is difficult to establish in standard epidemiological analyses. In very large genome-wide analyses of alcohol consumption [21], the genetic correlation with body mass index was weak (rG = −0.05), leaving plenty of variation to be accounted for by either unmeasured shared genetic effects or causal environmental influences. Being cross-sectional data, the large GWAS does not permit teasing out the possible direction of causality or whether it is bidirectional.

Phenotypically, parallel changes in weight and alcohol consumption were significantly correlated in men, but not in women. This sex-specific relationship, although modestly positive in men, is consistent with what was observed by French et al. [30]. The subtle positive intensity of this relationship in men was also reported in Downer et al. [27]. We found a significant correlation between the non-shared environmental components of the slopes of BMI and alcohol consumption in men, demonstrating the existence of a common environmental structure influencing weight gain and change in alcohol use in the same direction. This is consistent with a causal association, where increased alcohol consumption would lead to higher BMI (or vice versa). The trajectories in alcohol consumption and BMI, being consistent with the literature, reinforce both the need to appreciate the sex-induced complexity in the relationship between changes in weight and alcohol consumption, and the confidence that can be brought to the reproducibility of our twin model estimates. Such understanding is also needed in the design of randomized clinical trials, either of weight loss or of reducing alcohol consumption. A recent pilot study [51] showed that moderate alcohol use attenuated the beneficial effects of weight loss in premenopausal women with obesity, indicating the need to consider alcohol in obesity-related clinical trials. Our analyses suggest that both genetic and environmental factors may play a role and should be taken into account, for example, through the use of polygenic genetic risk scores. Such scores can now be reliably obtained from very large GWAS studies [21]. Incorporating genetic information may help to understand why some participants benefit more from interventions than others, and could lead to more targeted interventions based on genetics.

Twin and family analyses of longitudinal alcohol use and misuse indicate substantial genetic stability across time [52], but molecular genetic analyses are few and of relatively small size [53, 54], and BMI has rarely been accounted for in such analyses. Although we found no genetic correlation between baseline BMI and alcohol consumption, a genetic correlation associating baseline BMI with a change in alcohol consumption was observed in both men and women. Such a genetic correlation seems to have not yet been observed in the literature. This genetic relationship, although weakly negative, extends insights gained in cross-sectional designs [14, 21] to longitudinal settings. The nature of this genetic relationship remains undetermined, but could possibly be related to smoking, which is known to be causally associated with obesity [55, 57] and alcohol consumption [14, 21]. Expanded research on this topic could further investigate the genetic relationship between BMI, weight change, and longitudinal drinking behavior. Similarly, the negative correlation observed between baseline BMI and change in alcohol consumption was also reflected in a correlation between the non-shared environmental components in men, in the same direction, indicating the likely existence of environmental factors influencing both traits. Again, cigarette smoking could explain some of this environmental correlation. However, model adjustment with cigarette smoking (online suppl. document) did not reduce the phenotypic correlation between baseline BMI and change in alcohol consumption to zero, meaning that cigarette smoking, alone, could explain only part of the observed correlations.

A major limitation of our study lies in the statistical power in multivariate twin modeling. This lack of power mainly prevented (1) determining more precise sex differences, (2) obtaining narrower confidence intervals for heritability estimates, and (3) identifying common environments or genetics between growth factors. This last point applies particularly to men, for whom the difference in the bounds of the confidence intervals was sometimes substantial. As such, the significant phenotypic correlation between BMI and alcohol consumption intercepts in men is neither explained by a genetic correlation nor by a correlation between environmental components. The nature of this relationship thus remains unclear in our study, despite some evidence of genetic structure in the literature [21]. Further investigation, incorporating a larger sample size, may better capture the essence of this phenotypic relationship. However, comparable datasets are rare.

Other limitations related to the sample design can also be mentioned. The first concerns sample selection, for which the exclusion of individuals who did not respond to all 4 questionnaires could induce biases in the estimates of BMI trajectories due to deaths occurring before 2011. Our estimates of mean log-BMI over the first 3 questionnaires, however, are consistent with those described by Hjelmborg et al. [32] using a similar approach on the first 3 questionnaires of the same cohort (N = 10,556 twins). Deaths between 1990 and 2011, which prevented the collection of the 4th questionnaire from deceased individuals and led to their rejection from our study, seem unlikely to have biased the BMI trajectory estimates. Further, BMI remains a common indicator of obesity but is not systematically associated with physical activity, especially in non-obese individuals [58]. The inclusion of data on physical activity in the coupled analysis of BMI and alcohol consumption could provide major insights, particularly as physical activity was found to be positively associated with moderate alcohol consumption [59]. Overall multivariate and longitudinal analyses of changes in all major risk factors would be ideal but challenging to compile and analyse.

Our study drew strength from the 36-year follow-up time in the Finnish Twin Cohort study. Modeling the trajectory of alcohol consumption using growth factors allowed us to capture the effects associated with Finnish policies on alcohol consumption, such as those initiated in the second half of the 1980s [60], also visible in Table 1, and those in the early 2000s [61]. These relatively abrupt changes in alcohol consumption provided an opportunity to detect associations between changes in alcohol consumption and weight gain at both population and individual levels. However, accordingly, linear modeling of the alcohol consumption trajectory remains questionable and simplistic in such a context. The limited number of measures of alcohol consumption to four is a constraint to quadratic modeling, although it may provide interesting results at the population level. The use of quadratic or cross-lagged modeling [62] to draw nonlinear conclusions is therefore a credible avenue to further explore the longitudinal relationship between BMI and alcohol consumption, as might be the use of direction of causation modeling to infer the causal nature of this relationship [63].

In sum, changes in alcohol consumption and BMI were conditioned by both genetic and environmental factors, with often different intensity according to sex. In particular, the longitudinal relationship between BMI and alcohol consumption was rooted in a significant positive correlation between associated environmental components in men. BMI change in women was also negatively associated with baseline alcohol consumption, as their respective non-shared environmental components were significantly correlated. A negative genetic correlation between baseline BMI and change in alcohol consumption was observed, however, in both men and women. Since changes in weight and alcohol consumption are behavioral traits, large disparities between cohorts can be anticipated, and more studies are needed to provide an accurate and generalized relational picture. Other stratifications in addition to sex may be considered, since the correlations between indicators of obesity and alcohol consumption differ according to whether alcohol consumption is heavy or moderate [10, 27]. A funnel approach could, in this sense, allow a better understanding of the mechanisms associating weight change with alcohol consumption.

The authors gratefully acknowledge Dr. Alyce Whipp of the University of Helsinki Language Service for language revision of this paper.

FTC participants responded to up to four waves of questionnaires from surveys conducted in 1975, 1981, 1990, and 2011. For the 2011 survey, ethical approval was obtained from the Ethics Committee of the Helsinki and Uusimaa Hospital District (ID 01/2011). As for all earlier questionnaire surveys, informed consent to participate was not directly obtained but inferred by completion of the questionnaires. The study purpose was explained to the participants, and all participants were aware they could withdraw from the study at any time without any consequence to themselves.

The authors have no conflicts of interest to declare.

G.D. has received funding from the University of Helsinki, Faculty of Medicine, Doctoral School of Population Health (DOCPOP), PhD program. A.L. was supported by Grant #308698 and J.K. by Grants #312073 and #336823 from the Academy of Finland.

The authors’ contributions to the completion of this study are divided as follows: J.K. and G.D. conceptualized the study; G.D. conducted the study and performed the formal statistical analyses with support from K.S. and J.K. in twin modeling; G.D., K.S., A.L., and J.K. participated in the interpretation of the results and their relation to the literature; G.D. drafted the first version of the manuscript; all authors participated in the revision and writing of the final version of the manuscript and approved it.

The FTC data is not publicly available due to the restrictions of informed consent. However, the FTC data is available through the Institute for Molecular Medicine Finland (FIMM) Data Access Committee (DAC) (fimmdac@helsinki.fi) for authorized researchers who have IRB/ethics approval and an institutionally approved study plan. To ensure the protection of privacy and compliance with national data protection legislation, a data use/transfer agreement is needed, the content and specific clauses of which will depend on the nature of the requested data. A preprint version of this paper was posted on medRxiv [64] on April 10, 2022.