Liu I, Blacker DL, Xu R, Fitzmaurice G, Lyons MJ, Tsuang MT. Genetic and Environmental Contributions to the Development of AlcoholDependence in Male Twins. Arch Gen Psychiatry. 2004;61(9):897-903. doi:10.1001/archpsyc.61.9.897
Information on the heritability of the development of alcohol dependence
could provide a better understanding of the importance of genetic components
in disease transition.
To examine the genetic and nongenetic contributions to the age at onset
of regular alcohol use, the age at diagnosis of alcohol dependence, and the
transition from regular alcohol use to alcohol dependence.
Classic twin study.
This study included 3372 twin pairs of known zygosity from the Vietnam
Era Twin Registry. The diagnosis of DSM-III-R–defined
alcohol dependence and related information were obtained through telephone-administered
interviews conducted in 1992.
Main Outcome Measures
Standardized proportions due to genetic vs nongenetic factors of the
total variation in twin resemblance on the age at onset of regular alcohol
use, the age at meeting criteria for a diagnosis of alcohol dependence, and
the transition period from regular alcohol use to a diagnosis of alcohol dependence.
Genetic influence accounted for 49% of the variation in the age at diagnosis
of alcohol dependence. After adjusting for co-occurring psychiatric diseases,
additive genetic factors still explained more than 37% of the variance in
age at onset of alcohol dependence and at least 25% of the variance in the
transition period between regular drinking and the diagnosis of alcohol dependence.
Additionally, after grouping participants as early and late regular users
of alcohol, the genetic effects on the transition period for early regular
users were statistically significantly greater than those for late regular
Our results demonstrate a substantial heritable basis for alcohol dependence
according to its developmental sequence, including age at onset of regular
use, age at diagnosis, and the transition period between regular use and diagnosis.
Alcohol use disorder has long been considered an important issue inpublic health and medicine because of its high prevalence and devastatingeffect on individuals, families, and society.1,2 Twinstudies3- 5 haverevealed a substantial genetic influence on liability for alcohol dependencein registry data and clinically ascertained samples. However, still unclearis the extent to which genetic factors affect the developmental course ofalcohol dependence. The developmental sequence for alcohol dependence mightbe conceptualized as a series of steps or phases. One step represents theage when an individual becomes a regular alcohol user; a second step wouldbe the age when a regular user becomes an abuser. Genetic factors might contributeto each phase to a different extent.6,7
Although knowledge on the role of genetic factors in the early or latedevelopmental stage is of great importance for prevention and early intervention,few studies have focused on this issue. Jennison and Johnson8 conducteda study to relate the development of alcohol dependence in adult childrento a positive family history of alcohol dependence. Obot et al9 foundthat children with parental alcohol problems tended to initiate use of alcoholat an earlier age than others. Even fewer twin studies have examined the geneticcontribution to the development of alcohol problems because of limited statisticalstrategies. One study by Stallings et al10 exploredthe genetic impact on the ages at onset of regular drinking and smoking, andlinear regression analysis was applied to elderly affected twin pairs withthe assumption that all of the subjects in the sample had passed the at-riskperiod for developing these 2 behaviors. However, this approach is difficultto extend to middle-aged twins or studies that concentrate on diseases ratherthan behaviors, because outcomes of interest could occur after data collection.In this situation, we need an analysis in which time factors are taken intoconsideration to appropriately handle censored twin data. Given recent remarkableprogress on analytical techniques, we are able to use survival analysis techniquesthat incorporate random effects to handle these types of data and then toobtain valid estimates of parameters of interest.11- 13
The primary objective of this study is to apply a proportional hazardsregression model with random effects to estimate genetic and nongenetic effects.The study uses middle-aged male twin pairs from the Vietnam Era Twin (VET)Registry. Even though genetic contribution to determine the similarity ofalcohol dependence in twins has been confirmed, the heritability of the developmentof alcohol dependence has yet to be determined. With eyes toward preventionand early intervention, we seek to better understand the role of genetic andnongenetic factors in determining the information on age at onset of regularuse of alcohol, age at diagnosis of alcohol dependence, and the transitionperiod from regular alcohol use to a diagnosis of alcohol dependence.
The members of the VET Registry are 7375 male-male twin pairs born between1939 and 1957. Both twins served in the military during the Vietnam War era(1965-1975). In 1992, trained interviewers from the Institute for Survey Researchat Temple University invited 5150 twin pairs from the registry to participatein telephone-administrated interviews based on the DiagnosticInterview Schedule Version III Revised (DIS-III-R).14 Among them, 8169 individuals gaveverbal informed consent and were successfully interviewed. The interviewsincluded questions about several aspects of alcohol use, such as the frequency,amount, age at onset of each symptom, how long each symptom lasted, and whatmood or perception changes associated with alcohol drinking the twins experienced.
The eligible twins in this study were the 3372 pairs with known zygosity.Zygosity had been previously established based on responses to a questionnaireand limited blood group typing.15- 17 Intotal there were 1874 monozygotic (MZ) twin pairs and 1498 dizygotic (DZ)twin pairs. The mean age of this sample at interview in 1992 was 41.95 years(SE, 2.75 years; range, 33-52 years). The racial distribution was 93.7% non-Hispanicwhite, 5.9% African American, 0.3% Hispanic, and more than 0.01% other. Ofthe sample, 88.4% reported that they were "ever regular drinkers."
The age at onset of regular alcohol use, age at diagnosis of alcoholdependence, and the period from the age at onset of regular alcohol use tothe onset of diagnosis of alcohol dependence are the 3 primary outcomes inthis study. One question in the interview was, "How old were you when youfirst had any wine, beer, or other alcohol at least once a month for 6 monthsor more?" We took the response to this question as the definition for age at onset of regular alcohol use. People who were notregular drinkers were considered as having censored observations.Individuals who were not regular drinkers were by definition precluded fromdeveloping alcohol dependence.
Using standard DIS-III-R computer algorithms,we derived individual symptoms and the diagnosis of alcohol dependence basedon DSM-III-R.18 Weordered the ages at onset for the 9 DSM-III-R–defineddiagnostic symptoms for individuals who were regular drinkers. Because theoccurrence of at least 3 of the symptoms is the first criterion for makinga diagnosis of alcohol dependence and each of the 3 symptoms must last 1 monthto make a diagnosis, the age at onset was defined as the age at onset of thethird criterion symptom. Subjects free of the diagnosis at interview werecoded as censored at the age at interview.
The reliability and validity of the diagnosis of alcohol dependencein the VET Registry cohort were examined in a 1998 study.19 Researchersinvited 146 individuals of 5000 twin pairs to have an additional interviewand obtained detailed information regarding the diagnosis of alcohol dependence.The test-retest reliability of the 2 interviews, which were approximately466 days apart, was assessed with the κ coefficient of 0.61 for alcoholdependence. The long interval between the 2 interviews may be partly responsiblefor the low reliability in this study. Criterion validity was assessed throughcomparison with the clinical diagnosis for 89 discharged patients with alcoholismin a Veterans Affairs medical center; the sensitivity of telephone-administratedinterviews in this study was 96%.
For individuals who were regular drinkers and ultimately became alcoholdependent, we derived a transition period in years by subtracting the ageat regular use from the age at diagnosis. Individuals who were regular usersbut who did not become alcohol dependent were censored fortransition period. For this outcome measure, we also stratified thesample based on the mean age at onset of regular use to compare genetic effectsin those younger than 17 years with those 17 years and older.
The 1992 DIS-III-R interview also includedquestions that allowed us to define several other psychiatric disorders basedon the DSM-III-R. The diagnoses include nicotinedependence, antisocial personality disorder, depression, and anxiety disorder.The relationship among these variables, genetic factors, and alcohol dependencehas been established in the literature.20- 22 Weregarded the variables as potential confounders and handled them as binarydata. For example, we coded subjects as having presence of antisocial personalitydisorder if they had experienced the diagnosis before the interview regardlessof age at onset.
Initial analysis of the data involved the prevalence of 4 co-occurringpsychiatric disorders and the mean ages at onset of regular alcohol use andmeeting a diagnosis of alcohol dependence (Table 1 and Table 2).We then conducted a biometrical genetic survival analysis of the 3 alcoholdependence development measures. Based on the difference in the genetic relationshipin MZ and DZ pairs (MZ pairs share 100% of their genes and DZ pairs shareon average 50% of their genes), the method attempts to decompose the phenotypicvariance in a trait (such as the alcohol developmental measures) into 3 components,known as ACE: additive genetics (A), common environment (C), and unique environment(E). For regular drinking, we assessed all 3772 twin pairs, whereas the 2740twin pairs with the experience of regular drinking were further analyzed toinvestigate the age at onset of alcohol dependence and the transition periodbetween regular alcohol use and alcohol dependence. Before including co-occurringpsychiatric diseases in statistical models, 2 different analytic strategieswere used: structural equation models (SEMs) and random-effects models (REMs).We then took comorbid conditions into consideration and treated the age atonset of regular drinking as an effect modifier in subsequent analyses usingthe REMs we proposed. Each of these approaches is discussed herein.
The SEMs have been widely used in twin studies to partition the covariancebetween phenotypes into genetic and environmental parts.23 However,the application of SEMs to censored data can be complex, because survivaldata are neither binary nor continuous and individuals can develop an event after the time of data collection. The most convenientmethod is to apply the multiple cut points to censored twin outcomes and constructseveral contingency tables for the pairwise twin data.24,25 Then the polychoric correlations for MZ and DZ twins from 2 sets of contingencytables can be compared.
We implemented the SEM approach by classifying twins into 6 age categories:younger than 25 years, 25 to 29 years, 30 to 34 years, 35 to 39 years, 40to 44 years, and 45 years or older. Using these categories, we cross-classifiedeach of the age at onset phenotypes and calculated polychoric correlations.For twins 45 years or older, ages at onset for both twins were cross-tabulatedinto a 6 × 6 table in which censored data were placed exclusively inthe last column or the last row—the last age category. Then, we createda 5 × 5 table for MZ twins aged 40 to 44 years. The procedure was repeatedas subjects aged 30 to 34 years (the youngest age group in our sample) wereplaced. The same technique for forming contingency tables was then appliedto DZ twins.
Next, one summary polychoric correlation coefficient for each zygositywas attained, and the magnitude of genetic contribution was assessed in thebest-fitting model based on the principle of parsimony. The model with ACEcomponents was considered the full model, and it was then compared with otherreduced models. The analyses were implemented using the computer analysissoftware package Mx.26
The REMs, in which the random effects are used to explain the heterogeneityin survival time, have become popular for modeling the dependence among theobservations of correlated survival data.27 Theirapplication has been proposed for solving variable ages at onset among familymembers in human genetics28,29 andrecurrent depressive episodes in a population-based registry.30
We used a Cox proportional hazards model with random effects to analyzecensored time to event data in twins. Here, random-effects terms were addedto describe the relationship within a twin pair. In addition to 2 shared randomeffects for MZ and DZ twin pairs, individual random effects were used to representcomplicated dependence in twin data. Thus, there is 1 shared random effectand 2 individual random effects for MZ and DZ twin pairs, respectively. Theformulation of this model is as follows:
λij(t) = λ0(t)exp(βXij + bjZij), i = 1, 2; j = 1, 2, . . . , n
where λij(t) denotes the hazardfunction at time t for the ithmember of the jth twin pair, and λ0(t) is called the baseline hazard function. Xijdenotes a covariate vector for the fixed effects, β,for an individual, and Zij, usually asubset of Xij, denotes a covariate vectorfor the random effects, bj, for observationsin the jth cluster. If there are no covariates, Zi is simply a vector of 1 and 0s.
Here, bj is a vector of 6 randomeffects (b0, b1, b2, b3, b4, b5) for the jth twin pair as described. b0, b1,And b2 are 3 random effects for the MZpairs. The random effects terms are b0 + b1 and b0 + b2 for the 2 twins, respectively. Here, b0 represents the overall mean for the MZ twins,and b1 and b2 represent the effect specific to the twins in an MZ pair. Similarly,for the DZ pairs, the random-effects terms are b3 + b4 and b3 + b5.
The components of bj (b0, b1, b2, b3, b4, b5) are assumedto follow independent normal distributions with means equal to zero and with6 variance parameters (σ2b)to summarize their distributions. Then these 6 parameters are reparameterizedto be 3 major parameters of interest in twin studies: the part of the totalvariance due to A, C, or E effects (σ2A, σ2C, σ2E). For example, within MZ twinpairs, the variance of b0, the sharedpart for 2 twins, equals the sum of A effects and C effects (σ2b0 = σ2A + σ2C).Within DZ twin pairs, the variance of b3 equalsthe sum of the C effects and half of the A effects (σ2b3 =σ2C +[σ2A/2]). In contrast,the dissimilarity of 2 twins within a DZ twin pair can come from either Efactors or the difference of A factors between 2 twins (σ2b4 = σ2b5 = σ2E +[σ2A/2]). The followingequation shows the relationship between the 6 independent random effects and3 major parameters in twin studies:
To assess the interaction between age at onset of regular use and Aor C effects, we added 6 more random-effect terms (b6, b7, b8, b9, b10, b11) to the model. One set ofparameters, σ2A1, σ2C1, and σ2E1, was estimated from the first 6 random effectsto generate the within-twin dependence among early regular users, whereasa second set of parameters (σ2A2, σ2C2, and σ2E2) was generated based on thesecond set of 6 random effects for late regular users. In addition to fittingcore REMs, we added significant comorbid psychiatric disorders as covariatesfor fixed effects in the REMs to obtain adjusted estimates of parameters.
We used an expectation-maximization algorithm, which treats random effectsas missing data, to obtain maximum likelihood estimates.13,31 Tomake the results more interpretable, we used an equivalent linear transformationmodel to convert the parameter estimate for σ2E32 and calculatedstandardized proportions of the total variance in similarity of time-scaledoutcomes from A, C, and E. All of the analyses used an algorithm developedin C programming language.
The mean age at meeting a diagnosis of DSM-III-R–defined alcohol dependence in the VET Registry cohort is 25.3 years(SE, 6.4 years). Table 3 presentsthe parameter estimates for the standardized proportion of variation attributableto ACE components for age at onset of regular alcohol use and age at onsetof alcohol dependence. For age at onset of regular alcohol use, genetic effectsare similarly estimated in both types of models with the amount of nearlyone third of variation for age at onset of regular use. In contrast, the resultsfrom the SEM are different from those from the REM on the C and E parts. Morethan 30% of the total variance for onset of regular drinking could be attributedto C factors in the SEM, whereas C effects were estimated to affect only 13%of the total variation based on the REM. In the REM for age at onset of diagnosis,the expectation-maximization algorithms converge, with σ2C approaching 0, and the results of all estimatesapproximated those from the SEM. Almost half of the variation in latent liabilityfor age at onset could be explained by A factors, whereas the remaining partof the total variation arose from an E origin unshared by 2 members of a twinpair. As for the transition period from regular alcohol use and onset of diagnosis,SEMs are difficult to handle with bivariate censored data with age-incongruentoutcomes like this. According to the REM, A factors could account for a substantial(42%) portion of the total variance in liability.
Based on Cox proportional hazards models, univariate analysis showedthat all covariates were significant in predicting the outcomes of interestas the dependency within twin pairs was taken into consideration. We retained4 co-occurring psychiatric disorders for fixed effects in further analysesand presented estimates for the age at onset of diagnosis and the transitionperiod in Table 4. According tothe 95% confidence intervals, all adjusted estimates of fixed effects remainedstatistically significant in predicting the hazard rate. For age at onset,people with antisocial personality disorder tended to have a 3.8 (exp[1.34])–foldincrease in the hazard rate for developing alcohol dependence at any specificage compared with people without antisocial personality disorder, with theadjustment of other co-occurring diseases and the dependence within a twinpair. Controlling for other co-occurring conditions, subjects with a lifetimehistory of depressive disorder at interview were more likely to develop alcoholdependence at any specific age than people without depression based on a positiveestimate of the fixed effect (0.58).
We calculated the standardized proportion of the total variance in failuretime similarity owing to each component here. In these full models, 37% ofthe total variation in age at meeting a diagnosis can be attributed to geneticfactors, which exert less influence than those in the core model given in Table 3. For transition period, the discrepancyin estimates of interest between core and full models becomes more evidentwith the increment of C from 0% to 4.24% and the decrement of A from 42% tojust 25%.
To further examine the genetic contribution to the transition periodbased on age at onset of regular alcohol use, we fitted the models with interactionbetween random effects and age at onset of regular use, and the results arepresented in Table 5. Four co-occurringconditions significantly predicted the transition period. Irrespective ofthe adjustment of covariates for fixed effects, the C estimate converged to0 during parameter estimation for early regular users, whereas it accountedfor more than 10% of the variability in late regular users. The A variationis consistent in both core and full models to have significantly greater effectson early users than on late users. In the comparison of results in core andfull models, A factors made a decreased contribution by controlling for 4psychiatric disorders, and only 9.9% of variation in the transition time couldbe explained by A factors among late regular users.
Our results demonstrate the heritability of the development course ofalcohol dependence. The A factors influence when an individual becomes a regulardrinker, how fast an individual progresses from regular drinking to meetinga diagnosis of alcohol dependence, and when an individual develops alcoholdependence. To our knowledge, this study is the first twin study to addressthe genetic influence on different phases in the development of alcohol dependence.We could benefit from the present study by having a better understanding regardingthe genetic contribution over time. However, our findings seem to be comparableto results from previous studies that focused on the heritable basis for theoccurrence of alcohol dependence.
Approximately 40% to 60% of the variation in risk of alcohol abuse anddependence has been identified in the literature as being due to genetic influence.5 For age at onset of regular alcohol use, C factorshave effects in determining twin resemblance, yet A effects were found tocontribute approximately one third of the total variation. On the other hand,almost half of the variation in determining age at onset could be explainedby A factors. Such a finding must be interpreted cautiously, because, as hasbeen pointed out by Lyons et al,21 the estimateof genetic contribution could vary with the reliability of the report. Thus,unreliable reports on an earlier event, such as age at onset of regular use,may lead to underestimation of the genetic contribution. Nevertheless, itis reasonable that a shared familial environmental influence may contributemore to a temporally early event than to a late one. Despite the possibilityof unreliability of reports, our findings still provide evidence of a strongheritable basis for age at onset of regular alcohol use. As for the transitionperiod, the study shows that once an individual became a regular alcohol user,the length of the progression to alcohol dependence is still substantiallyaffected by underlying genetic factors.
The issue of confounding is not straightforward in twin studies. In1999, MacGregor33 cited one example about smokingas a confounding factor for genetic influence on the occurrence of smoking-relateddisease. The genetic component was incorrectly regarded as an important factorin determining the disease because smoking was ignored. It has long been debatedwhether there is a common vulnerability for alcohol dependence among peoplewith co-occurring diseases or causally related pathways from these diseases.34,35 If a common etiologic basis is adopted,all these diseases would be regarded as manifestations of a general vulnerability.Otherwise, parallel pathways from genetic liabilities to co-occurring diseases,which predispose to the development of alcohol dependence, have to be established.However, the distinction between these 2 hypotheses is barely verified ina retrospective study with an adult sample.36
To address this concern, we fitted both core and full models with theadjustment of 4 co-occurring conditions to data from the VET Registry. Notsurprisingly, the adjusted estimates of genetic effects from the full modelsare smaller, especially for the transition period (from 42% to 25%). As mentioned,it is not yet clear which model would give us the best information. However,our findings support the hypothesis that the development of alcohol dependenceis substantially heritable, and the adjusted genetic estimates could be deemedthe lower boundary of the true effect. In this case, we could claim that Afactors account for more than 37% of the variance for age at meeting a diagnosisof alcohol dependence and at least 25% of the variance for the transitionperiod.
Age at first drink has been found to be highly associated with the occurrenceof alcohol dependence.36,37 Inour study, age at onset of regular alcohol use acted as an effect modifierin the development of alcohol dependence. Through fitting the models withinteraction, we obtained significantly different genetic effects on the lengthof the transition period. Of the people who had become regular users, earlyregular users seem to be influenced by genetic factors in the developmentof alcohol dependence more than late users do. In addition, common environmentsshared by 2 members of a twin pair play a moderate role in determining thelength of the progression to alcohol dependence in late regular users. However,we should cautiously interpret the results because, as is the case with theassociation between alcohol dependence and associated psychiatric diseases,different hypotheses for the relationship between age at first use and therisk of alcohol dependence have been proposed.36- 38
As for the comparison of SEMs and REMs, our findings show that the estimatesof the variance of random effects are not always similar. Using SEMs we wouldlose important information by transforming censored data into an ordinal arrangement.In addition, SEMs have limitations when they are extended to handle censoreddata with age-incongruent outcomes such as the transition period in our study.In performing SEMs, we cannot place all twins with the censored transitionperiod in the last row or column of a contingency table as usual because bothtwins in a twin pair might have censored outcomes with different magnitudes.In addition, individual covariates or time-varying covariates would be hardto handle in SEMs, whereas even interactive terms can be flexibly examinedin REMs. To deal with correlated censored data, we suggest the use of appropriatemodels based on survival analysis such as REMs. It is important to have futureresearch on the application of REMs, especially model diagnostics concerningthe effect of violating the assumption of proportional hazards or the influence,because a great deal of nonaffected individuals exist in data.
There are limitations in our study that should be considered. First,the study results cannot be generalized to women. The role of genetic inheritancein the occurrence of alcohol use disorder has been found to differ by sexin previous twin studies.39 More studies ofwomen will be needed to confirm the importance of genetic components in diseasetransition. Second, the equal environment assumption and the lack of assortativemating are 2 critical assumptions to the validity of twin studies.23 If either of these assumptions is incorrect, thegenetic influence will be biased. However, one study40 conductedto test equal environment assumption for the VET Registry cohort showed nosignificant violations. Third, the data collection of the VET Registry reliedon retrospective reports of age at onset of each symptom, which resulted inan inability to avoid recall bias. In addition, subjects with alcohol dependencewould tend to have recall problems regarding age at onset of any alcoholicsymptom. However, recall errors for both zygotic types would operate in thesame direction and are likely to be of similar magnitude. The recall errorwould be nondifferential, and the bias of our study results would be towardthe null.
Notwithstanding these limitations, our registry-based study has enoughpower to detect the significance of genetic effects of interest because ofthe large sample size, and, moreover, sampling bias, which is a major concernin studies using either volunteers or hospital-based samples, can be avoided.Although it is true that this exclusively male-male twin sample limits thegeneralizability of our study, by restricting sex, we can prevent confoundingeffects by sex and then explore the pure relation of genetic factors on diseaseprogression.
Our findings confirm a strong heritable basis for the progression fromregular drinking to alcohol dependence in men in addition to the risk of alcoholdependence. However, some research questions surrounding the nature of therelationship between this disease and associated psychiatric diseases remainunanswered.
Correspondence: I-Chao Liu, MD, DSc, Department of Psychiatry, CardinalTien Hospital, No. 5 Alley 20 Ln 790, Chung-hsiao East Road, Section 5, Taipei,Taiwan 110 (email@example.com).
Submitted for publication June 19, 2003; final revision received March15, 2004; accepted March 15, 2004.
The US Department of Veterans Affairs has provided financial supportfor the development and maintenance of the VET Registry.
Numerous organizations have provided invaluable assistance in the conductof this study, including Department of Defense (Washington, DC); NationalPersonnel Records Center, National Archives and Records Administration (StLouis, Mo); the Internal Revenue Service (Washington); National Opinion ResearchCenter (Chicago, Ill); National Research Council, National Academy of Sciences(Washington); and the Institute for Survey Research, Temple University (Philadelphia,Pa). Most important, we gratefully acknowledge the continued cooperation andparticipation of the members of the VET Registry and their families. Withouttheir contribution this research would not have been possible.