DSM-IV alcohol dependence: a categorical or dimensional phenotype?

Psychological Medicine, 2006, 36, 1695 1705. f 2006 Cambridge University Press doi:10.1017/s0033291706009068 First published online 12 October 2006 Printed in the United Kingdom DSM-IV alcohol dependence: a categorical or dimensional phenotype? DEBORAH S. HASIN 1,2, XINHUA LIU 3, DONALD ALDERSON 2 AND BRIDGET F. GRANT 4 * 1 Department of Epidemiology, Mailman School of Public Health, and Department of Psychiatry, College of Physicians and Surgeons, Columbia University, NY, USA; 2 New York State Psychiatric Institute, New York, NY, USA; 3 Department of Biostatistics, Mailman School of Public Health, Columbia University, New York, NY, USA; 4 Laboratory of Epidemiology and Biometry, Division of Intramural Research, National Institute on Alcohol Abuse and Alcoholism, Bethesda, MD, USA ABSTRACT Background. Etiologic research on complex disorders including alcohol dependence requires informative phenotypes. Information is lost when categorical variables represent inherently dimensional conditions. We investigated the validity of DSM-IV alcohol dependence as a dimensional phenotype by examining evidence for linearity and thresholds in associations with validating variables. Method. Current drinkers in the National Longitudinal Alcohol Epidemiologic Survey (NLAES) (n=18 352) and National Epidemiologic Survey of Alcohol and Related Conditions (NESARC) (n=20 836) were analyzed. Validating variables included family alcoholism, early-onset drinking, and alcohol treatment. Logistic or Poisson regression modeled the relationships between the validating variables and dependence in categorical, dimensional or hybrid forms, with severity defined as number of current DSM-IV alcohol-dependence criteria. Wald tests assessed differences between models. Results. No evidence was found for boundaries between categories. Instead, the association of alcohol dependence with the validating variables generally increased in linear fashion as the number of alcohol-dependence criteria increased. For NLAES models of family alcoholism, early-onset drinking and treatment, the lines had zero intercepts, with slopes of 0. 18, 0. 27, 0. 70, respectively. For NESARC models of family history and early-onset drinking, the zero intercept lines had slopes of 0. 20, 0. 33, and 0. 77, respectively. Wald tests indicated that models representing alcohol dependence as a dimensional linear predictor best described the association between dependence criteria and the validating variables. Conclusions. The sample sizes allowed strong tests. Diagnoses are necessary for clinical decisionmaking, but a dimensional alcohol-dependence indicator should provide more information for research purposes. INTRODUCTION Research on complex traits such as alcohol use disorders involves multiple genetic and environmental risks. The success of such studies is likely to be limited by the quality of the phenotypes (Crabbe, 2002). This awareness has focused attention on the possibility that clinical diagnoses may not be informative or specific enough for some research needs, as has been found in * Address for correspondence: Bridget F. Grant, Ph.D., Laboratory of Epidemiology and Biometry, Room 3077, Division of Intramural Clinical and Biological Research, National Institute on Alcohol Abuse and Alcoholism, National Institutes of Health, M.S. 9304, 5635 Fishers Lane, Bethesda, MD 20892-9304, USA. (Email: bgrant@willco.niaaa.nih.gov) The views and opinions expressed in this report are those of the authors and should not be construed to represent the views of any of the sponsoring organizations, agencies, or the U.S. government. 1695

1696 D. S. Hasin et al. other areas of medicine (e.g. epilepsy) (Greenberg et al. 1988, 2000). Genetic studies generally use categorical diagnoses of DSM-III-R or DSM-IV alcohol dependence or dependence/ abuse as phenotypes. However, some investigators have suggested (Meyer, 2001, 2003; Hasin et al. 2003; Gunzerath & Goldman, 2003) or simply assumed (Edwards, 1986) that alcohol dependence should be studied as a dimensional condition. When the measure of an inherently dimensional condition is dichotomized, information is lost. Such information may be critical when statistical power is limited, as it often is in studies of gene gene or gene environment interaction. If DSM-IV alcohol dependence in categorical form is psychometrically sound (i.e. reliable and valid) but dichotomizes an inherently dimensional condition, then converting its elements to a dimensional measure should produce a more informative phenotype. However, the validity of such a dimensional model of dependence should be tested empirically. The theoretical basis of the DSM-IV alcohol use disorders (Rounsaville et al. 1986), the Alcohol Dependence Syndrome (ADS) of Edwards and Gross (Edwards & Gross, 1976), provides a context for this work. The ADS was described as a complex physiological and psychological process impairing the ability to control one s drinking. Importantly, the ADS was described as dimensional rather than categorical. The ADS was considered one axis of alcohol problems that was distinct from consequences of heavy drinking such as injuries, and social or legal problems related to drinking. This theoretical conceptualization was supported by latent variable analysis of national data (Muthen et al. 1993; Harford & Muthen, 2001) showing that abuse and dependence constitute two separate factors, each with a dimensional aspect. This bi-axial (Edwards, 1986) distinction between dependence and its consequences guided the separation of dependence and abuse criteria in DSM-III-R and DSM-IV (Rounsaville et al. 1986) and between dependence and hazardous use in ICD-10 (WHO, 1992). Due to the clinical needs of DSM-III-R, DSM-IV and ICD-10, dependence was dichotomized at a threshold of three or more criteria. When DSM-IV went to press, relatively little was known about the reliability and validity of this formulation of alcohol dependence. A substantial body of research on the reliability and validity of alcohol dependence has accumulated since then, as reviewed in detail in Hasin et al. (2006). This material largely pertains to alcohol dependence as a categorical diagnosis. Most of the studies were conducted in general population samples, avoiding numerous biases. The reliability and validity of the DSM-III-R or DSM-IV alcohol-dependence diagnosis was strongly and consistently supported. Therefore, when considering further analyses to identify a more informative phenotype for research on alcohol use disorders, we focused on dependence. Since the theoretical basis of the dependence diagnosis described a dimensional condition, (Edwards, 1986) dichotomizing such a condition may lose critical information. Because of the need for research phenotypes to be maximally informative, we investigated whether empirical evidence supported the validity of DSM-IV alcohol dependence as a dimensional condition. This work was based on two assumptions described previously in investigation of whether diagnosis carves nature at its joints (Kendell & Brockington, 1980, Kendler & Gardner, 1998). Assumption 1. If a condition is inherently categorical, boundaries between the categories (e.g. cases and non-cases) should be clear. Discontinuities should appear at those boundaries, for example, in the level of risk factors or consequences. Such boundaries or discontinuities should not be evident if a condition is inherently dimensional. Assumption 2. Within categories of an inherently categorical condition (e.g. present or absent), members should be homogeneous with regard to risk factors or consequences. Regular gradation within categories (e.g. strength of association with risk factors) should not appear (e.g. the lack of slopes below and above the threshold in Fig. 1). Because patient samples such as treated alcoholics may represent a truncated severity range, it is important to address the issue across the full range of severity in representative samples. We examined the applicability of these two assumptions to alcohol dependence among current drinkers from a large national sample.

DSM-IV alcohol dependence 1697 Log odds ratio for Tx and onset, log proportion for family history 7 6 5 4 3 2 1 0 0 1 2 3 4 5 6 7 Criterion count FIG. 1. NLAES model results, Model 1 (M 7Dum ) and Model 2 (M Dimensional ) (Model 1 results shown in symbols; Model 2 by lines)., Family history; &, onset of drinking before age 15 years;, treatment in last 12 months. METHOD Samples The 1992 National Longitudinal Alcohol Epidemiologic Survey (NLAES) was a face-to-face interview survey of a nationally representative U.S. sample aged o18 years (n=42 862) (Grant, 1997; Hasin & Grant, 2002). The response rate was 90. 0% and Blacks and young adults were oversampled. To adjust for non-response and selection probability, the sample was weighted and adjusted to reflect the U.S. population from the 1990 Decennial Census in terms of age, race, sex, and ethnicity. All current drinkers, defined as those drinking o12 drinks in the previous year were included in this report (n=18 352). The mean age was 39. 8, 87. 5% were white and 60. 3% were male. The National Epidemiologic Survey of Alcohol and Related Conditions (NESARC) is a nationally representative United States sample of 43 093 civilian non-institutionalized participants aged o18 years. Details of the sampling frame are described elsewhere (Grant et al. 2003 b). Like the NLAES, the research protocol, including informed consent procedures, received full ethical review and approval from the U.S. Census Bureau and U.S. Office of Management and Budget. Young adults, Hispanics, and African-Americans were oversampled. The overall response rate was 81%. To adjust for non-response and selection probability, the sample was weighted and adjusted to reflect the U.S. population from the 2000 Decennial Census in terms of age, race, sex, and ethnicity. All current drinkers, using the same definition as the NLAES, were included in this report (n=20 836). The mean age was 42. 2, 76. 0% were white and 57. 3% were male. Measures The Alcohol Use Disorder and Associated Disabilities Interview Schedule (AUDADIS; Grant et al. 1995; Grant et al. 2003a), is a structured diagnostic interview designed for trained non-clinician interviewers. The AUDADIS showed excellent psychometric properties in U.S. and international reliability and validity studies of DSM-IV alcohol dependence, including psychiatrist reappraisal (Grant et al. 1995, 2003a; Chatterji et al. 1997; Cottler et al. 1997; Hasin et al. 1997a c; Canino et al. 1999; Hasin & Grant, 2004). To operationalize DSM-IV alcohol dependence, the AUDADIS includes detailed questions covering repeated instances of the seven criteria: (1) tolerance; (2) withdrawal or relief/avoidance of withdrawal; (3) persistent desire/unsuccessful attempts to reduce drinking; (4) time drinking or recovering from its effects; (5) giving up/reducing occupational/social/recreational activities to drink; (6) impaired control; and (7) continued drinking despite physical or psychological problems. A dimensional scale using the current (last 12 months) dependence criteria (range 0 7) has excellent test retest reliability (ICC=0. 75; Grant et al. 1995). We used three validating variables to examine evidence for boundaries and homogeneity. (1) Family history of alcoholism, represented by the proportion of first-degree biological relatives with a history of drinking problems (physical, emotional, interpersonal, occupational, or legal problems with alcohol, or a great deal of time drinking or getting over its effects), reported as greater than zero by 35. 9% of the subjects in the NLAES, and 28. 9% in the NESARC. (2) Early onset of drinking (before age 15, excluding tastes or sips), a strong predictor of DSM-IV alcohol dependence (Grant et al. 2001), reported by 7. 5% of the subjects in the NLAES, and 7. 9% in the NESARC. (3) Treatment for alcohol problems (e.g. AA, physician, specialized services in the last 12 months), reported by 1. 6% of the subjects in the NLAES, and 1. 5% in the NESARC. The widely differing nature and distributions of the three validating

1698 D. S. Hasin et al. variables provided contrasting tests for the assumptions addressed. Statistical analysis To take account of the complex sampling design of the NLAES and NESARC, SUDAAN was used throughout to conduct weighted analyses. To describe the patterns in the association between number of alcohol-dependence criteria and the family history validating variable, we applied Poisson models for count outcomes, which can be extended to outcomes in the form of a proportion (EY)/N, where Y is the count of affected relatives and N is the number of relatives in each family. With a log link function between the proportion and the linear combination of predictors, the model parameters associated with predictors can be interpreted as ratios of proportions in log scale. To describe the patterns in the association between number of alcoholdependence criteria and early onset of drinking and treatment in last 12 months, we applied logistic regression analysis for binary outcomes. Early-onset drinking was coded as prior to age 15 versus o15 years. The model parameters associated with these predictors are interpreted as log odds ratios. Age, gender and race were included in all models as control variables. We began with a basic model (Model 1) with ten predictors (M 7Dum ): the three control variables and seven non-ordered dummy variables to represent the seven levels of severity (1 7) of the alcohol-dependence criteria. Persons with 0 alcohol-dependence criteria constituted the reference group. Persons with one dependence criterion were compared to the reference group, as were persons with two dependence criteria, etc., up to persons with seven dependence criteria. Dummy variables were used at this stage because we did not know in advance if a trend would be found. The seven parameters (b 1, b 2,, b 7 ) associated with the dummy variables indicate the effect of the number of alcohol-dependence criteria. Consistently increasing regression coefficients for dummy variables that can be indicated by a line (slope) would be consistent with an underlying dimensional variable. This result would suggest a model with a single parameter representing the slope (indicating linearity), which would be more parsimonious. Model 1 also provides a basis for comparison with several other models using different forms of alcohol dependence as the predictor, such as forms involving a boundary. To test whether a linear trend represented the effect of criteria count on an outcome, we used Model 2 (M Dimensional ) to test for a dimensional trend in criteria count. Instead of seven dummy variables, one predictor was used to represent dependence criteria as a single dimensional measure (i.e. 0 7 criteria) in which the associated parameter b could be used to indicate a trend in the effect of criterion count. In addition to Model 1 (M 7Dum ) and Model 2 (M Dimensional ), we used three other models with varying functional forms to describe specific patterns in the relationship between the number of symptoms and the validating variables. Models 3 and 4 were plausible alternative or hybrid models involving an element of category and an element of continuity, while Model 5 represented the current DSM-IV binary approach. Model 3 (M 2trends ) described two distinct trends with discontinuity at the DSM-IV threshold of >3 criteria by using 3 variables: a variable contrasting 0 2 criteria versus o3, a dimensional variable for 0 2 criteria, and a dimensional variable for 3 7 criteria. Model 4 (M Threshtrend ) used a variable contrasting 0 2 criteria versus o3 and a single dimensional variable for 3 7 criteria. The lack of a dimensional variable for 0 2 criteria illustrated the possibility of homogeneity (lack of slope) within this category, but not above the threshold of 3 criteria, the diagnostic threshold. Model 5 (M DSM-IV ) consisted of a single variable contrasting 0 2 criteria versus o3, most directly portraying a dichotomous diagnosis with homogeneity (zero slope) within the two categories defined by the threshold. To compare the models incorporating the weights reflecting the complex sample design, the Wald test was used to test hypotheses on the model parameters. The Wald test statistic is a squared distance between two vectors of estimated effects in the two nested models, following a x 2 distribution with the degrees of freedom defined as the difference in the number of parameters in the two nested models. While differences in the fit of nested models are often compared using the likelihood ratio test, this test cannot be used with weighted data. The Wald test is used to test differences in

DSM-IV alcohol dependence 1699 patterns of associations. With large samples such as the NLAES and NESARC, the likelihood ratio test and Wald test are usually equivalent in testing the hypotheses on model parameters for the pattern in the association between outcome and predictors if no sampling weights applied. We first used the Wald test to determine if the set of dummy variables in Model 1 (M 7Dum ) had an effect on the outcome variable, controlling for demographic variables. The null hypothesis on the parameters of Model 1 is (b 1, b 2,, b 7 )=(0, 0,, 0) for no association between the alcohol-dependence criteria count and the validating variable. With Model 1 (free of a constrained pattern) as the basis of comparison, we choose from a set of models to describe the pattern of effects. We first used the Wald test for a difference in the pattern of associations between symptom count and the outcome variable described with Model 1 and Model 2, i.e. comparing (b 1, b 2,,b 7 ). In this case, the Wald test was used to determine whether alternative parameterizations of the association between the validating variable and the number of dependence criteria produce significantly different estimates between Model 1 and Model 2, with a single linear predictor. Little or no difference would indicate support for Model 2, as it is most parsimonious in terms of number of parameters. The Wald test was also used to explore differences in patterns of associations between the number of dependence criteria and the validating variables described by Model 1 and the alternative models (Models 3, 4 and 5). We considered that Model 2 fit the pattern in Model 1 better than Model 5 (for example) if the difference between Models 1 and 2 were small (non-significant), while the difference between Models 1 and 5 were not small (e.g. significantly difference). For all tests, the level of statistical significance was set at 0. 05. RESULTS The null hypothesis for Model 1 was rejected. The variables representing the seven levels of alcohol-dependence criteria counts were significantly related to all three validating variables in Model 1 in both the NLAES and the NESARC (p<0. 0001). Log odds ratio for Tx and onset, log proportion for family history 7 6 5 4 3 2 1 0 0 1 2 3 4 5 6 7 Criterion count FIG. 2. NESARC model results, Model 1 (M 7Dum ) versus Model 2 (M Dimensional ) (Model 1 results shown in symbols; Model 2 by lines)., Family history; &, onset of drinking before age 15 years;, treatment in last 12 months. Family history as a validating variable The seven log proportions obtained from Model 1(M 7Dum ) were plotted to show the pattern in the relationship between family history and number of DSM-IV alcohol-dependence criteria in the NLAES (Fig. 1) and the NESARC (Fig. 2). Results in the two samples were very similar. For both surveys, the relationship increased in strength as the number of criteria met increased, with no visual evidence of a cut-off at the DSM-IV threshold (3 criteria) or anywhere else. The relationship of the dependence criteria count to family history (log proportion of alcoholic relatives) is described with a zero intercept line with slope of 0. 18 (S.E.=0. 01) for the NLAES and a zero intercept line with slope of 0. 20 (S.E.=0. 01) for the NESARC. Model 1 (M 7Dum ) and Model 2 (M Dimensional ) did not differ significantly in the NLAES (x 2 = 6. 6, df=6, p=0. 36; Table 1) or the NESARC (x 2 =11. 2, df=6, p=0. 08; Table 2). Thus, the models described the pattern of associations similarly. In this case, Model 2 (M Dimensional ) is the preferred model because it has fewer parameters. The pattern of associations in Model 3 (M 2trends ) also did not differ significantly from Model 1 (M 7Dum ) in the NLAES (p=0. 35). However, since it also contains more parameters than Model 2, Model 2 remains the preferred model. Also, a Wald test on the difference between the two slope parameters in Model 3 indicated no difference in slope (p=0. 25), making this model redundant with Model 2. The patterns of associations in NLAES Models 4

1700 D. S. Hasin et al. Table 1. Comparison of models representing alcohol dependence in different forms: NLAES (n=18 352) Validating variables Model 1 estimate (M 7Dum ) Model 2 (M Dimensional ) Model 3 (M 2trends ) Model 4 (M Threshtrend ) Model 5 Estimate a (M DSM-IV ) % Positive family history alcohol problems Log P(K)/P(0) 1 0. 20 (0. 04) 0. 18 (0. 01) 0. 18 (0. 02) 0. 00 0. 00 2 0. 35 (0. 05) 0. 36 (0. 02) 0. 37 (0. 04) 0. 00 0. 00 3 0. 59 (0. 06) 0. 54 (0. 03) 0. 61 (0. 14) 0. 51 (0. 13) 0. 66 (0. 04) 4 0. 77 (0. 10) 0. 72 (0. 04) 0. 75 (0. 19) 0. 65 (0. 19) 0. 66 (0. 04) 5 0. 92 (0. 10) 0. 90 (0. 06) 0. 89 (0. 22) 0. 80 (0. 22) 0. 66 (0. 04) 6 1. 19 (0. 10) 1. 08 (0. 07) 1. 04 (0. 24) 0. 94 (0. 24) 0. 66 (0. 04) 7 1. 01 (0. 15) 1. 27 (0. 08) 1. 18 (0. 27) 1. 08 (0. 27) 0. 66 (0. 04) 0. 36 0. 35 <0. 01 <0. 01 Age of onset <15 years Log [Odds(K)/Odds (0)] 1 0. 29 (0. 09) 0. 27 (0. 02) 0. 34 (0. 05) 0. 00 0. 00 2 0. 69 (0. 10) 0. 54 (0. 05) 0. 67 (0. 08) 0. 00 0. 00 3 0. 80 (0. 14) 0. 81 (0. 07) 0. 84 (0. 28) 0. 66 (0. 29) 0. 92 (0. 10) 4 1. 06 (0. 21) 1. 09 (0. 10) 1. 09 (0. 41) 0. 91 (0. 41) 0. 92 (0. 10) 5 1. 48 (0. 20) 1. 36 (0. 12) 1. 34 (0. 46) 1. 16 (0. 46) 0. 92 (0. 10) 6 1. 62 (0. 28) 1. 63 (0. 14) 1. 59 (0. 52) 1. 41 (0. 52) 0. 92 (0. 10) 7 1. 64 (0. 33) 1. 90 (0. 17) 1. 84 (0. 58) 1. 66 (0. 58) 0. 92 (0. 10) 0. 74 0. 87 <0. 01 <0. 01 Treatment Log [Odds(K)/Odds (0)] 1 0. 77 (0. 32) 0. 70 (0. 04) 0. 82 (0. 14) 0. 00 0. 00 2 1. 63 (0. 28) 1. 39 (0. 08) 1. 63 (0. 29) 0. 00 0. 00 3 2. 40 (0. 31) 2. 09 (0. 11) 2. 46 (0. 39) 1. 91 (0. 34) 2. 67 (0. 18) 4 3. 14 (0. 29) 2. 78 (0. 15) 3. 04 (0. 48) 2. 50 (0. 45) 2. 67 (0. 18) 5 3. 51 (0. 37) 3. 48 (0. 19) 3. 63 (0. 52) 3. 09 (0. 50) 2. 67 (0. 18) 6 4. 32 (0. 35) 4. 18 (0. 23) 4. 22 (0. 57) 3. 68 (0. 55) 2. 67 (0. 18) 7 4. 73 (0. 38) 4. 87 (0. 27) 4. 81 (0. 62) 4. 27 (0. 61) 2. 67 (0. 18) 0. 53 0. 98 <0. 01 <0. 01 a Estimate derived from coefficient on variable representing the number of symptoms. (M Threshtrend ) and 5 (M DSM-IV ) were significantly different from Model 1 (Table 1). This indicates that Models 4 and 5 did not reflect the pattern of associations clearly indicated by the trend in magnitude of the regression coefficients in Model 1. In the NESARC, Model 3(M 2trends ) also did not differ from Model 1 (Table 2), and a Wald test for the difference between the two slope parameters in Model 3 also indicated no difference in slope (p=0. 14). The Wald tests in the NESARC for Models 4 (M Threshtrend )and5(m DSM-IV ) were significant, indicating that the pattern of associations in Models 4 and 5 failed to reflect the increasing trend indicated in the regression coefficients found in Model 1. Early drinking onset as validating variable As shown in Figs 1 and 2, no visual evidence of a threshold was found for the relationship of early onset of drinking and number of DSM-IV dependence criteria in the NLAES or NESARC. The log odds ratio of the association between the alcohol-dependence criteria count and early onset of drinking is described with a zero intercept line with slope 0. 27 (S.E.=0. 02) for the NLAES and a zero intercept line with slope 0. 33 (S.E.=0. 02) for the NESARC. Model 1 (M 7Dum ) and Model 2 (M Dimensional ) did not differ significantly in the NLAES (x 2 = 3. 6, df=6, p=0. 73; Table 1) or the NESARC (x 2 =5. 5, df=6, p=0. 48; Table 2). Thus, the

DSM-IV alcohol dependence 1701 Table 2. Comparison of models representing alcohol dependence in different forms: NESARC (n=20 836) Validating variables Model 1 estimate (M 7Dum ) Model 2 (M Dimensional ) Model 3 (M 2trends ) Model 4 (M Threshtrend ) Model 5 Estimate a (M DSM-IV ) % positive family history of alcohol problems Log P(K)/P(0) 1 0. 28 (0. 06) 0. 20 (0. 01) 0. 26 (0. 03) 0. 00 (0. 00) 0. 00 (0. 00) 2 0. 51 (0. 06) 0. 40 (0. 02) 0. 52 (0. 06) 0. 00 (0. 00) 0. 00 (0. 00) 3 0. 64 (0. 06) 0. 59 (0. 02) 0. 59 (0. 12) 0. 41 (0. 12) 0. 71 (0. 03) 4 0. 74 (0. 06) 0. 79 (0. 03) 0. 80 (0. 14) 0. 62 (0. 13) 0. 71 (0. 03) 5 0. 98 (0. 07) 0. 99 (0. 04) 1. 01 (0. 15) 0. 82 (0. 15) 0. 71 (0. 03) 6 1. 30 (0. 08) 1. 19 (0. 05) 1. 22 (0. 17) 1. 03 (0. 17) 0. 71 (0. 03) 7 1. 39 (0. 08) 1. 39 (0. 06) 1. 42 (0. 18) 1. 24 (0. 18) 0. 71 (0. 03) 0. 08 0. 53 <0. 01 <0. 01 Age of onset <15 years Log [Odds(K)/Odds (0)] 1 0. 40 (0. 10) 0. 33 (0. 02) 0. 40 (0. 06) 0. 00 (0. 00) 0. 00 2 0. 80 (0. 13) 0. 65 (0. 03) 0. 80 (0. 13) 0. 00 (0. 00) 0. 00 3 0. 94 (0. 12) 0. 98 (0. 05) 0. 90 (0. 22) 0. 60 (0. 21) 1. 16 (0. 07) 4 1. 23 (0. 12) 1. 31 (0. 06) 1. 27 (0. 24) 0. 97 (0. 23) 1. 16 (0. 07) 5 1. 62 (0. 14) 1. 64 (0. 08) 1. 64 (0. 26) 1. 34 (0. 26) 1. 16 (0. 07) 6 2. 01 (0. 14) 1. 96 (0. 09) 2. 01 (0. 29) 1. 71 (0. 28) 1. 16 (0. 07) 7 2. 43 (0. 15) 2. 29 (0. 11) 2. 38 (0. 32) 2. 08 (0. 31) 1. 16 (0. 07) 0. 48 0. 99 <0. 01 <0. 01 Treatment Log [Odds(K)/Odds (0)] 1 2. 16 (0. 58) 0. 77 (0. 04) 0. 96 (0. 18) 0. 00 (0. 00) 0. 00 (0. 00) 2 2. 27 (0. 59) 1. 55 (0. 08) 1. 91 (0. 35) 0. 00 (0. 00) 0. 00 (0. 00) 3 2. 97 (0. 56) 2. 32 (0. 11) 2. 32 (0. 52) 1. 43 (0. 43) 2. 99 (0. 20) 4 3. 83 (0. 55) 3. 09 (0. 15) 3. 13 (0. 54) 2. 23 (0. 46) 2. 99 (0. 20) 5 4. 51 (0. 54) 3. 86 (0. 19) 3. 94 (0. 57) 3. 04 (0. 50) 2. 99 (0. 20) 6 5. 12 (0. 54) 4. 64 (0. 23) 4. 74 (0. 61) 3. 84 (0. 54) 2. 99 (0. 20) 7 6. 29 (0. 53) 5. 41 (0. 26) 5. 55 (0. 65) 4. 65 (0. 58) 2. 99 (0. 20) 0. 08 0. 07 <0. 01 <0. 01 a Estimate derived from coefficient on variable representing the number of symptoms. models described the pattern of associations similarly and Model 2 is the preferred model for both surveys because it has fewer parameters. The pattern of associations in Model 3 (M 2trends ) did not differ significantly from Model 1 for the NLAES (p=0. 87) or the NESARC (p=0. 99). A Wald test of the difference between the two slope parameters in Model 3 indicated no difference in slope for NLAES (p=0. 33) or NESARC (p=0. 72). Therefore, Model 2 remained the preferred model. For Models 4 (M Threshtrend ) and 5 (M DSM-IV ), Wald tests showed that the patterns of associations were significantly different from Model 1 in the NLAES and NESARC (Tables 1 and 2), indicating that the pattern of associations in Models 4 and 5 failed to reflect the increasing trend indicated in the regression coefficients found in Model 1. Treatment as a validating variable For the NLAES data, the plot of the relationship between the number of DSM-IV alcoholdependence criteria and the outcome resembled those for the other validating variables (Fig. 1), with no visual evidence of a threshold. The log odds ratio of the relationship between the number of DSM-IV dependence criteria and treatment is described with a zero intercept line with slope of 0. 70 (S.E.=0. 04). In the NESARC, a zero intercept line could be with slope of 0. 77 (S.E.=0. 04). Model 1 (M 7Dum ) and Model 2 (M Dimensional ) did not differ significantly in the NLAES

1702 D. S. Hasin et al. (x 2 =5. 1, df=6, p=0. 53), so Model 2 was preferred due to parsimony. Model 3 (M 2trends ) produced results similar to Model 1, but was not preferred because it has more parameters than Model 2 and a test of the difference between the two slope parameters in Model 3 indicated again no difference in slopes (p=0. 17), making this model redundant with Model 2. Models 4 (M Threshtrend ) and 5 (M DSM-IV ) were significantly different from Model 1, failing to capture the trend in the magnitude of the regression coefficients in Model 1. In the NESARC data, Model 2 (M Dimensional ) did not differ significantly from Model 1 (M 7Dum )(x 2 =11. 3, df=6, p=0. 08). Model 3 also did not differ significantly from Model 1 and again, a Wald test of the difference in the two slopes revealed no significant difference (p=0. 45). The functional forms represented by Models 4 and 5 failed to portray the pattern represented in Model 1 accurately (Table 2); all produced p values of 0. 01 or lower when compared to Model 1 (M 7Dum ). DISCUSSION Since publication of DSM-IV, much research has accumulated on the reliability and validity of alcohol dependence. The need for a more informative alcohol-dependence phenotype led us to empirical investigation of alcohol dependence as a dimensional condition. Graphic representations as well as tests of the relationships between increasing severity of dependence and three validation variables (family history, early drinking onset, and treatment) failed to indicate a clear boundary along the severity continuum of DSM-IV alcohol dependence. Instead, gradations were found within categories superimposed on the continuum. These results were confirmed by tests of differences in patterns of associations between alcohol dependence in different forms and the three validating variables. The tests had ample power to detect slight differences due to the large samples. A group of models representing various hypothetical relationships between alcohol-dependence severity and the three validation variables were investigated. Of these, the model with a single parameter, M Dimensional, representing alcohol dependence as a dimensional (in fact, linear) condition from 0 7 criteria parsimoniously described the pattern shown by Model 1 estimates without assumptions on the relationship. In contrast, Model 5, representing DSM-IV alcohol dependence as a binary variable with homogeneous risk for the validation variables within the two classes (above and below the threshold), clearly differed from Model 1, failing to reflect the increasing trend indicated in the regression coefficients found in Model 1. The same was true for two hybrid models that we tested that included boundaries as well as the possibility for order within some aspect of the variable definition. Two molecular genetics studies illustrate the utility of the dimensional measure of DSM-IV alcohol dependence compared to a binary diagnosis. Both studies focused on polymorphism in an alcohol dehydrogenase gene, ADH1B. In one study, the sample was large and the allele of interest was rare (Heath et al. 2001). In the other study, the allele of interest was more common but the sample was much smaller (Hasin et al. 2002). In both studies, the relationships between DSM-IV alcohol-dependence diagnoses and ADH1B polymorphisms were not significant. The predicted relationships became significant only when dependence as a dimensional variable (range 0 7) was used. These results suggest that the dimensional dependence measure provided more information. Moreover, national data on 4339 individuals meeting DSM-III-R criteria for alcohol dependence (Grant, 1993) showed associations between DSM-III-R alcohol-dependence severity and family history of alcoholism, early onset of drinking, and treatment (Hasin & Glick, 1992). This study did not include the statistical tests described above, include sub-threshold cases or address the issue of boundaries in the manner we did. However, earlier results were consistent with the present findings. Further, latent class analysis of alcohol-dependence symptoms in relatives of treated alcoholics (Bucholz et al. 1996) showed that 4 6 classes adequately fit the data, with classes differentiated primarily by severity. This is also consistent with our present results, although applying latent classes derived in one dataset to use in different datasets can be a complex procedure. The above findings are consistent with the original concept of alcohol dependence as a dimensional condition (Edwards & Gross, 1976;

DSM-IV alcohol dependence 1703 Edwards, 1986). The results are important, as they are the first to graphically illustrate and statistically test issues of boundaries and gradation in alcohol dependence as defined by DSM-IV, a common method of defining alcohol dependence in psychiatry and other disciplines. Furthermore, we used a highly reliable measure and two very large national samples. The representative nature of the samples eliminates many potential sources of bias in the results. The three validation variables were distinct in nature but produced consistent findings. Thus, this study builds on previous work and helps move discussion about DSM-IV alcohol dependence as a dimensional condition out of the realm of expert opinion and into one of firm empirical support. Because statistical methods now allow analysis of quantitative traits in genetic linkage and association studies (Almasy & Blangero, 1998; Elston & Buxbaum, 2000; Zhu et al. 2002), these issues are pertinent to analysis of existing as well as future data. An advantage of alcohol dependence as a dimensional phenotype based on the seven DSM-IV criteria is that it can be applied to existing datasets that include measures of the seven DSM-IV alcoholdependence criteria without requiring new, costly data collection. The scale used in this study could more accurately be characterized as ordinal rather than dimensional, as it is a sequence of whole numbers. However, while the term, ordinal is accurate in a strict sense, it may not be recognized easily by readers whose expertise lies outside quantitative areas. The DSM-V work group convened to address these issues across diagnostic categories uses the term, dimensional. Since ready recognition of terminology across disciplines is important in nosological work, we used the term dimensional to refer to the graded nature of dependence examined in this paper. To investigate whether the association between validating variables and DSM-IV alcohol-dependence criteria could be represented by a single dimensional trend for criteria count or by another pattern, we used logistic regression models with binary validation variables as outcomes for two of the three validating variables (early age at onset of drinking and treatment). Our purpose was not to investigate the best form of the validating variables, a task outside the scope of the present report, but the best form of dependence, for which purpose the two binary outcomes were wellsuited. The remaining variable, family history, worked well as a quantified variable (as density, or a proportion of relatives with alcohol disorders) so family history was used in this form. The existence of a clear biological distinction between a problem drinker and an individual with alcohol addiction who has permanently lost control over drinking is not supported by our data, but if such a distinction could be found, it would fill an important gap in the knowledge. The applicability of these results to dependence on other substances is unknown but should be tested. This research does not address categorical subtypes based on psychiatric co-morbidity. We did not subject the four DSM-IV alcohol abuse criteria to the same type of model-fitting analysis, although this could be the topic of a separate analysis and report later. We did not include criteria from DSM-III-R or ICD-10 that were not included in DSM-IV because criteria from each nomenclature should be analyzed separately, reflecting the way they are used. Such analyses would be useful to conduct in the future. Separate analyses for subjects sub-grouped by demographic or clinical variable such as gender, age, or race would also be useful to present later as sample sizes allow, and would provide valuable additional information on DSM-IV alcohol dependence as a dimensional measure in the different groups. Earlier, examination of major depressive disorder as a dimensional condition (Kendler & Gardner, 1998) produced similar results: little evidence for boundaries between categories and if boundaries were imposed, graded levels within categories. Such findings do not alter the importance of categorical diagnoses for clinical and policy purposes. The categorical approaches will remain important for a number of purposes, such as communication between clinicians, inclusion criteria in clinical trials, and for some etiologic studies. However, understanding etiology and treatment of these clinical categories, research definitions yielding more informative phenotypes may be necessary, as has been the case for other medical disorders. Phenotypes such as the dimensional measure validated above retain a link to DSM through

1704 D. S. Hasin et al. shared criteria. The dimensional form of DSM-IV alcohol dependence does not require development of new assessments and datasets, another important practical advantage. Such research definitions would promote consistency in measurement across studies while addressing the needs of research that can differ from the needs of clinical practice. ACKNOWLEDGMENTS Support is acknowledged from NIAAA grants AA K05 AA014223 and R01AA8159 (D.H.), and contributing support from the New York State Psychiatric Institute. The National Longitudinal Alcohol Epidemiologic Survey is sponsored and funded by the National Institute on Alcohol Abuse and Alcoholism. This research paper, in part, was supported by the Intramural Research Program of the National Institutes of Health, National Institute on Alcohol Abuse and Alcoholism. The authors thank Valerie Richmond, M.A. for editorial assistance and manuscript preparation. DECLARATION OF INTEREST None. REFERENCES Almasy, L. & Blangero, J. (1998). Multipoint quantitative-trait linkage analysis in general pedigrees. American Journal of Human Genetics 62, 1198 1211. Bucholz, K. K., Heath, A. C., Reich, T., Hesselbrock, V., Kramer, J. R., Nurnberger Jr., J. I. & Schuckit, M. A. (1996). Can we subtype alcoholism? A latent class analysis of data from relatives of alcoholics in a multicenter family study of alcoholism. Alcoholism: Clinical and Experimental Research 20, 1462 1471. Canino, G. J., Bravo, M., Ramı rez, R., Febo, V., Ferna ndez, R., Hasin, D. & Grant, B. F. (1999). The Spanish AUDADIS: reliability and concordance with clinical diagnoses in a Hispanic population. Journal of Studies on Alcohol 60, 790 799. Chatterji, S., Saunders, J., Vrasti, R., Grant, B. F., Hasin, D. & Mager, D. (1997). Reliability of the alcohol and drug modules of the AUDADIS Alcohol/Drug Revised (AUDADIS-ADR): an international comparison. Drug and Alcohol Dependence 47, 171 185. Cottler, L. B., Grant, B. F., Blaine, J., Mavreas, V., Pull, C. B., Hasin, D., Compton, W. M., Rubio-Stipec, M. & Mager, D. (1997). Concordance of DSM-IV alcohol and drug use disorder criteria and diagnoses as measured by AUDADIS-ADR, CIDI and SCAN. Drug and Alcohol Dependence 47, 195 205. Crabbe, J. C. (2002). Alcohol and genetics: new models. American Journal of Medical Genetics 114, 969 974. Edwards, G. (1986). The Alcohol Dependence Syndrome: a concept as stimulus to enquiry. British Journal of Addiction 81, 171 183. Edwards, G. & Gross, M. (1976). Alcohol dependence: provisional description of a clinical syndrome. British Medical Journal 1, 1058 1061. Elston, R. C. & Buxbaum, S. (2000). Haseman and Elston revisited. Genetic Epidemiology 19, 1 17. Grant, B. (1997). Prevalence and correlates of alcohol use and DSM-IV alcohol dependence in the United States: results of the National Longitudinal Alcohol Epidemiological Survey. Journal of Studies on Alcohol 5, 464 473. Grant, B. F. (1993). The relationship between ethanol intake and DSM-III-R alcohol dependence: results of a national survey. Journal of Substance Abuse 5, 257 267. Grant, B. F., Dawson, D. A., Stinson, F. S., Chou, S. P., Kay, W. & Pickering, R. P. (2003a). The AUDADIS-IV: reliability of alcohol consumption, tobacco use, family history of depression and psychiatric diagnostic modules in a general population sample. Drug and Alcohol Dependence 71, 7 16. Grant, B. F., Harford, T. C., Dawson, D. A., Chou, P. S. & Pickering, R. (1995). The Alcohol Use Disorder and Associated Disabilities Schedule (AUDADIS): reliability of alcohol and drug modules in a general population sample. Drug and Alcohol Dependence 39, 37 44. Grant, B. F., Moore, T. C., Shepard, J. & Kaplan, K. (2003b). Source and Accuracy Statement, l Wave 1 National Epidemiologic Survey on Alcohol and Related Conditions (NESARC) (http://niaaa. census.gov/pdfs/source_and_accuracy_statement.pdf). Accessed 26 September 2006. Grant, B. F., Stinson, F. S. & Harford, T. C. (2001). Age at onset of alcohol use and DSM-IV alcohol abuse and dependence: a 12-year follow-up. Journal of Substance Abuse 13, 493 504. Greenberg, D. A., Delgado-Escueta, A. V., Maldonado, H. & Widelitz, H. (1988). Segregation analysis of Juvenile Myoclonic Epilepsy. Genetic Epidemiology 5, 81 94. Greenberg, D. A., Durner, M., Keddache, M., Shinnar, S. S., Resor, S. R., Moshe, S. L., Rosenbaum, D., Cohen, J., Ballaban-Gil, K., Tomasini, L., Zhou, G., Klotz, I. & Dicker, E. (2000). Reproduceability and complications in gene searches: linkage, heterogeneity, association, and inheritance in juvenile myoclonic epilepsy. American Journal of Human Genetics 66, 508 516. Gunzerath, L. & Goldman, D. (2003). GrE: A NIAAA workshop on gene-environment interactions. Alcoholism: Clinical and Experimental Research 27, 540 562. Harford, T. C. & Muthen, B. O. (2001). The dimensionality of alcohol abuse and dependence: a multivariate analysis of DSM-IV symptom issues in the National Longitudinal Survey of Youth. Drug and Alcohol Dependence 62, 150 157. Hasin, D., Aharonovich, E., Liu, X., Maman, Z., Matseoane, K., Carr, L. G. & Li, T-K. (2002). Alcohol dependence symptoms and Alcohol Dehydrogenase 2 Polymorphism: Israeli Ashkenazis, Sephardics and recent Russian immigrants. Alcoholism: Clinical and Experimental Research 26, 1315 1321. Hasin, D., Carpenter, K. M., McCloud, S., Smith, M. & Grant, B. F (1997a). The AUDADIS: reliability of alcohol and drug modules in a clinical sample. Drug and Alcohol Dependence 44, 133 141. Hasin, D. & Glick, H. (1992). Severity of alcohol dependence: results from a national survey. British Journal of Addiction 87, 1725 1730. Hasin, D. S. & Grant, B. F. (2002). Major depression in 6,050 former drinkers: association with past alcohol dependence. Archives of General Psychiatry 59, 794 800. Hasin, D. S. & Grant, B. F. (2004). The co-occurrence of DSM-IV alcohol abuse in DSM-IV alcohol dependence: NESARC results on heterogeneity that differs by population subgroup. Archives of General Psychiatry 61, 891 896. Hasin, D., Hatzenbuehler, M. L., Keyes, K. & Ogburn, E. (2006). Substance use disorders: Diagnostic and Statistical Manual of Mental Disorders, fourth edition (DSM-IV) and International Classification of Diseases, tenth edition (ICD-10). Addiction (Suppl. 1) 101, 59 75.

DSM-IV alcohol dependence 1705 Hasin, D., Schuckit, M. A., Martin, C. S., Grant, B. F., Bucholz, K. K. & Helzer, J. E. (2003). The validity of DSM-IV alcohol dependence: what do we know, what do we need to know? Alcoholism: Clinical and Experimental Research 27, 244 252. Hasin, D., Van Rossem, R., McCloud, S. & Endicott, J (1997b). Alcohol dependence and abuse diagnoses: validity in community sample heavy drinkers. Alcoholism: Clinical and Experimental Research 21, 213 219. Hasin, D., Van Rossem, R., McCloud, S. & Endicott, J (1997c). Differentiating DSM-IV alcohol dependence and abuse by course: community heavy drinkers. Journal of Substance Abuse 9, 135. Heath, A. C., Whitfield, J. B., Madden, P. A. F., Bucholz, K. K., Dinwiddie, S. H., Slutske, W. S., Bierut, L. J., Statham, D. B. & Martin, N. G. (2001). Towards a molecular epidemiology of alcohol dependence: analysing the interplay of genetic and environmental risk factors. British Journal of Psychiatry 40, s33 s40. Kendell, R. E. & Brockington, I. F. (1980). The identification of disease entities and the relationship between schizophrenic and affective psychoses. British Journal of Psychiatry 137, 324 331. Kendler, K. S. & Gardner, C. O. (1998). Boundaries of major depression: an evaluation of DSM-IV criteria. American Journal of Psychiatry 155, 172 177. Meyer, R. (2003). The Disease Concept in Addiction, 2003. College on Problems of Drug Dependence. Symposium, Annual Meeting, 17 June 2003. Meyer, R. E. (2001). Finding paradigms for the future of alcoholism research: an interdisciplinary perspective. Alcoholism: Clinical and Experimental Research 25, 1393 1406. Muthen, B. O., Hasin, D. & Wisnicki, K. S. (1993). Factor analysis of ICD-10 symptom items in the 1988 National Health Interview Survey. Addiction 88, 1071 1077. Rounsaville, B. J., Spitzer, R. L. & Williams, J. (1986). Proposed changes in DSM-III substance use disorders: description and rationale. American Journal of Psychiatry 143, 463 468. WHO (1992). International Classification of Diseases (10th edn). World Health Organization: Geneva. Zhu, X., Zhang, S., Zhao, H. & Cooper, R. S. (2002). Association mapping, using a mixture model for complex traits. Genetic Epidemiology 23, 181 196.