The Typology of Alcohol Use Disorder: Latent Class Analyses of a Population-Based Swedish Sample

Abstract

Objective:

The purpose of this study was to clarify the clinical heterogeneity of alcohol use disorder (AUD) in a national Swedish sample of affected individuals using latent class analysis.

Method:

Using a Swedish population-based sample of AUD cases ascertained from medical, criminal, and pharmacy registries (n = 217,071), we applied latent class analysis to sex, externalizing and internalizing syndromes before first registration, and age at first registration. The resulting types were evaluated against 15 diverse validators and degree of resemblance in relative pairs concordant for AUD.

Results:

A three-class solution was preferred by fit indices. The three classes were as follows: type 1 (male preponderant, externalizing; 32%), type 2 (minimal prior psychopathology; 46%), and type 3 (mixed-sex internalizing; 23%). Repeated split-half analyses revealed the statistical stability of these solutions. Meaningful differences emerged between the classes on many validators. Type 1 had the greatest family disruption, lowest educational levels, most AUD registrations, highest rates of criminal registration, and highest genetic risk for externalizing disorders and AUD. Type 2 had the least social dysfunction. Type 3 had the highest educational attainment, genetic liability to internalizing disorders, and proportion of women. All types significantly aggregated in affected pairs of relatives.

Conclusions:

Meaningful and reproducible subtypes of AUD, consistent with prior typological results, can be obtained from national registry–based samples. Using a range of external validators and patterns of familial aggregation, our results suggest that our three-class solution captured a meaningful proportion of the clinical heterogeneity of AUD.

The heterogeneity of alcoholism (hereafter termed alcohol use disorder [AUD]) in terms of clinical presentation, pattern of comorbidity, course, and outcome has long been noted. As a result, typologies of alcoholism have been proposed and discussed as early as the mid-19th century (Babor, 1996; Babor & Dolinsky, 1988; Del Boca & Hesselbrock, 1996; Leggio et al., 2009) and have remained an active area of research into the modern era. Babor (1996), in a thoughtful review of the extensive literature on AUD typology, suggests that most binary typologies of AUD describe an Apollonian subtype—characterized by a “later onset, a slower disease course, fewer complications, less psychological impairment, and a better prognosis” (Babor,) and a Dionysian subtype “characterized by early onset, more severe symptomatology, greater psychological vulnerability, and more personality disturbance” (p. 13). In his review, a number of the typologies noted that the Apollonian subtype was more common in women and more often related to mood symptoms, whereas the Dionysian subtype occurred more frequently in men and was more often related to antisocial behaviors (Babor, 1996). Both of the two most influential modern AUD typologies by Cloninger (Cloninger et al., 1981, 1996) and Babor himself (Babor et al., 1992) produced subtypes broadly consistent with the Apollonian–Dionysian schema. Since Babor's review, studies of AUD typology have continued, including three-class (Hauser & Rybakowski, 1997), four-class (Del Boca & Hesselbrock, 1996; Windle & Scheidt, 2004), and five-class solutions (Moss et al., 2007).

In this report, we add to this literature on empirical approaches to the clinical heterogeneity of AUD by examining, using latent class analysis (LCA; Collins & Lanza, 2009; McCutcheon, 1987), a large-scale population-based sample of cases of AUD ascertained through national medical, criminal, and pharmacologic registries in Sweden. We use a very simple model, subdividing our AUD cohort by LCA using four variables: sex, age at first AUD registration, and the presence/number of internalizing and externalizing syndromes before first AUD registration.

We then evaluate the resulting classes in two ways, first by testing their performance on 15 diverse potential validators. Second, given the strong evidence for familial/genetic factors in the etiology of AUD (Cotton, 1979; Verhulst et al., 2015), we examine whether the proposed subtypes of AUD aggregate in sibling, half-sibling, and cousin pairs concordant for AUD.

Method

We analyzed information on individuals from Swedish population-based registers with national coverage. The registers were linked using each person's unique identification number replaced by a serial number to preserve confidentiality. We secured ethical approval from the Regional Ethical Review Board in Lund (No. 2008/409, 2012/795, and 2016/679). The database for the LCA was created by selecting all individuals born in Sweden from 1950 to 1980 and registered with AUD at some point between 1972 and 2018 (N = 217,074). In the database, we included registrations for externalizing disorders (EDs; drug use disorder [DUD] and criminal behavior [CB]), registrations for internalizing disorders (IDs; major depression [MD] and anxiety disorder [AD]), age at first registration for AUD, and sex of the individual. For a definition of all variables, see Appendix Table 1. (A supplemental appendix appears as an online-only addendum to this article on the journal's website.) For both IDs and EDs, we created a three-level variable. For IDs, the three levels were (a) no registration for MD or AD, (b) registration for MD only or AD only, and (c) registration for both MD and AD. For EDs, the three levels were (a) no registration for DUD or CB, (b) registration for only DUD or only CB, and (c) registration for both DUD and CB. For age at first registration, we standardized (M = 0, SD = 1) the distribution for each year of birth and then divided the 31 combined distributions into three groups (<25th percentile, 25th–75th percentile, >75th percentile).

We used LCA to identify homogeneous AUD classes based on ID registrations, ED registrations, age at AUD registration, and sex. The number of latent classes indicated by the observed variables was determined by comparing model fit statistics between nested models. Improvement in model fit was indicated by smaller values of the Bayesian information criterion (BIC; Vrieze, 2012), Akaike's information criterion (AIC; Akaike, 1987), the sample size–adjusted BIC (ABIC; Sclove, 1987), by significant improvement in fit assessed by the Lo–Mendell–Rubin (LMR) test (Nylund et al., 2007), and by entropy values close to 1.0. However, as the number of classes is influenced by the number of observed variables, both empirical (improved model fit) and theoretical (model interpretability) aspects were considered. Given our large sample size, we were concerned that we were well powered to identify classes too small to be useful in clinical or research work. We set an arbitrary limit that we would only consider solutions where all identified classes had a prevalence in our AUD cohort of 5% or more. In addition, we split our sample into two random halves three times and compared the solutions across the two samples. We then assigned individual subjects to specific classes based on the likelihood of their particular response profile.

Thereafter, we included 15 external validators (year of birth, years of education, family genetic risk score [FGRS] for AUD, FGRS for ID, FGRS for ED, type of AUD registration [criminal, medical, and pharmacy], number of AUD registrations, early retirement, unemployment, recipient of social welfare, neighborhood deprivation, marital status at first AUD registration, and residing with both parents at age 15) in the database. We included different kinds of measures of the AUD-associated outcomes—number of registrations and several indices of social and economic functioning— with the goal of capturing the diversity in the impact of the subject's alcohol abuse on their life course.

See Appendix Table 2 for definitions of these validators, including the FGRS, which, briefly, integrates data on risk for specific disorders or groups of disorders in first- through fourth-degree relatives, weighted by age, sex, and genetic relationship with the proband. This method has been used in several recent studies from our group (Kendler et al., 2021a, 2021b, 2021c). We used the most likely class membership from the LCAs to calculate the share of variance in the validators accounted for by class membership. In an additional attempt to validate the LCAs, using the Swedish Multi-Generation Registry, we included all possible full-sibling, half-sibling, and cousin pairs among our sample concordant for AUD registration. Tetrachoric correlations were used to estimate the similarity between relatives based on class membership. Finally, using the entire Swedish population born from 1950 to 1980, we investigated, using tetrachoric correlations, how our AUD subtypes differed in their ability to predict any case of AUD in their full siblings. We used Mplus Version 7.31 software for the LCA (Muthén & Muthén, 2012) and SAS Version 9.4 software for all other analyses (SAS Institute Inc., Cary, NC).

Compliance with ethical standards

The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008. We secured ethical approval for this study from the Regional Ethical Review Board of Lund University (No. 2008/409).

Informed consent

As approved by Swedish ethical authorities, informed consent was not obtained from individual participants included in this study.

Results

Descriptive findings

Our final sample size of AUD cases equaled 217,071, of whom 72% were men. The majority of cases had neither an ED (59%) nor an ID (83%) before their first AUD registration. Further descriptive statistics on this sample are seen in Table 1, including that their mean FGRS for AUD was 0.55 SD above the population mean, whereas the elevation in genetic risk for EDs equaled 0.41 SD and the parallel value for IDs was considerably lower at 0.18. Nearly three quarters of the sample had presented for medical care with an AUD-related diagnosis, and one third had an alcohol-related criminal registration or received a prescription for drug treatment for AUD. Less than 60% of the sample was residing with both biological parents at age 15.

Table 1.

Descriptive statistics of the sample

Variable	n, %, or M (SD)	Missing
N cases of alcohol use disorder (AUD)	217,071
Men	72.3%
Externalizing disorders (ED) before AUD onset
No externalizing disorders	58.8%
Drug abuse or criminal behavior	35.1%
Drug abuse and criminal behavior	6.2%
Internalizing disorders (ID) before AUD onset
No internalizing disorders	83.2%
Major depression or anxiety disorder	11.8%
Major depression and anxiety disorder	5.0%
Age at AUD onset
<25th percentile	25%
25th to 75th percentile	50%
>75th percentile	25%
Descriptors
Year of birth, M (SD)	1963 (8.6)
FGRS AUD, M (SD)a	0.55 (1.4)	8,695
FGRS ED, M(SD)a	0.41 (1.2)	8,698
FGRS ID, M(SD)a	0.18 (1.1)	8,689
Years of education, M (SD)	11.4 (2.4)	4,381
Early retirement (Population: 9.8%)	26.2%
Unemployment (Population: 50.4%)	59.9%
Social welfare (Population: 26.0%)	59.5%
Neighborhood deprivation (1 year before AUD registration)a	0.18 (1.5)	62,095
Married at AUD registration	19.3%
Medical complication	7.8%
Residing with both parents at age 15	59.7%
Mean AUD registrations at geographical area at age 15	0.49 (0.9)	516
Type of AUD registration
Criminal	34.6%
Medical	74.3%
Prescription	35.3%

Notes: FGRS = family genetic risk score; ED = externalizing disorder; ID = internalizing disorder.

^a Standardized to population M = 0, SD = 1.

Latent class analysis

We obtained one- to six-class LCA solutions, as we had insufficient degrees of freedom to identify a seven-class solution (Table 2). The AIC, BIC, and ABIC kept improving with more classes, as is common in large, well-powered samples. The entropy score was considerably higher for the three- versus the four-class solution, and according to the LMR, the four-class solution did not provide an improvement in fit over the three-class solution. Furthermore, both the five- and six-class solutions included a class with a prevalence lower than we wanted to consider (4.6% and 2.6%, respectively). Therefore, we focused further analyses on the three-class solution.

Table 2.

Fit indices of our latent class analysis models

No. of classes	AIC	BIC	ABIC	Entropy	LMR (p value)
1	1,321,522	1,321,594	1,321,572
2	1,277,451	1,277,605	1,277,557	.588	<.0001
3	1,269,134	1,269,371	1,269,297	.807	<.0001
4	1,264,264	1,264,583	1,264,485	.586	1.0000
5	1,263,244	1,263,645	1,263,521	.721	<.0001
6	1,262,736	1,263,220	1,263,070	.719	<.0001

Notes: No. = number; AIC = Akaike's information criterion; BIC = Bayesian information criterion; ABIC = sample size–adjusted BIC; LMR = Lo–Mendell–Rubin test.

Description of the three-class model

The frequencies of the three classes in our first model were, respectively, 32%, 46%, and 23%. In all three classes, the mean posterior probabilities for each case assigned to that class were high, ranging from 92% to 95%. The entropy scores of our four classification variables reflected the discriminatory power of each variable. The distribution was quite uneven, with ED having by far the highest (0.50), followed by ID (0.30) and then rather low values for both age-at-onset (AAO; 0.15) and sex (0.08). As seen in Figure 1, class 1 was distinguished from both other classes on three of four of our variables: highest rates of men, highest rates of ED, and earliest AAO. We call such cases type 1 AUD. Class 3 could also be easily distinguished from the other two classes on three variables: highest rate of women, highest rate of IDs, and latest AAO. Members of this class will be called type 3 AUD. By contrast, class 2—the most common class—was best characterized by very low rates of both EDs and IDs before AUD onset. It was between classes 1 and 3 for both the percentage of men and AAO, and its members will be called type 2 AUD.

Figure 1.

Results of the three-class latent class analysis of alcohol use disorder (AUD). The y-axis equals the percent of the three classes that have the specific characteristic given in the x-axis, respectively, left to right: (a) % male, (b) % with the three possible distributions of externalizing disorders (ED) (none, drug use disorder [DUD] or criminal behavior [CB], or both DUD and CB), (c) % with the three possible distributions of internalizing disorders (ID) (none, major depression [MD] or anxiety disorder [AD], or both MD and AD), and (d) age at onset (assessed by age at first registration) divided into three groups: <25th percentile, in between the 25th and 75th percentile, and >75th percentile.

Reproducibility of the latent class solutions

To address the statistical stability of our latent class solutions, we performed, with our total sample, a random split half and then analyzed each half by LCA following the same procedure we used on the total sample. We did this three times. As seen in the Appendix Tables 3–5, in all three replicates we reassuringly produced solutions quite similar to those found in the entire sample.

Performance on validators

Table 3 summarizes the performance of our three-class solution for our 15 validators, the performance of each being assessed by the percent of interclass variance they explain (i.e., r²). We will focus on the five best performing validators. The strongest validator was criminal AUD registration, which cleanly separated all three of our classes from each other, with highest scores in class 1, intermediate in class 2, and lowest in class 3. The second strongest validator was “being on social welfare prior to AUD registration,” in which class 1 had the highest endorsement, followed by class 3 and then class 2. In years of education, class 1 stood out as having by far the lowest average educational attainment. Genetic risk for ED was the fourth most potent validator with by far the highest mean score for Class 1, followed by class 3 and then class 2. Finally, receipt of a prescription drug for treatment of AUD was seen in much higher proportions of cases from classes 1 and 2.

Table 3.

Validators for the three-class solution

Variable	Class 1	Class 2	Class 3	r 2
Year of birth, M	1963	1962	1965	1.54%
Years of education, M	10.7	11.7	12.0	4.54%
Genetic risk score AUD, M	0.78	0.42	0.45	1.33%
Genetic risk score ID, M	0.22	0.10	0.36	0.80%
Genetic risk score ED, M	0.74	0.21	0.34	3.78%
Type of AUD registration
Criminal, %	51%	32%	10%	8.84%
Medical, %	69%	76%	80%	0.88%
Prescription, %	29%	32%	55%	3.76%
No. of registrations, M	0.19	−0.08	−0.08	1.66%
Early retirement, %	31%	21%	31%	1.25%
Unemployment, %	61%	57%	64%	0.28%
Social welfare, %	76%	48%	59%	6.50%
Neighborhood deprivation, M	0.43	0.01	0.22	1.54%
Married at AUD registration, %	11%	22%	27%	2.39%
Residing with both parents at age 15, %	49.9%	66.0%	61.0%	2.16%

Notes: AUD = alcohol use disorder; ID = internalizing disorder; ED = externalizing disorder; no. = number.

Validation by familial aggregation of our classes

Before examining the specific results, we examined the correlation for AUD itself in these pairs in the entire population, which was +0.29, +0.12, and +0.10 in full siblings, half-siblings, and cousins, respectively. Table 4 depicts the familial aggregation, as assessed by a tetrachoric correlation, of classes of AUD in full-sibling (n = 41,122), half-sibling (n = 18,420), and first-cousin (n = 46,992) pairs concordant for AUD from our three-class models.

Table 4.

Familial aggregation, as assessed by a tetrachoric correlation (± SE), of classes of alcohol use disorder (AUD) in siblings, half-siblings, and cousins from the three-class model

Outcome	Predictor	Siblings	Half-siblings	Cousins
N pairs		41,122	18,420	46,992
Type 1	Type 1	0.24 (0.01)	0.07 (0.01)	0.09 (0.01)
	Type 2	-0.18 (0.01)	-0.07 (0.01)	-0.07 (0.01)
	Type 3	-0.11 (0.01)	-0.02 (0.01)	-0.04 (0.01)
Type 2	Type 2	0.18 (0.01)	0.09 (0.01)	0.08 (0.01)
	Type 3	0.01 (0.01)	-0.02 (0.01)	-0.02 (0.01)
Type 3	Type 3	0.16 (0.01)	0.06 (0.02)	0.08 (0.01)
Total population born 1950–1980
AUD	AUD	0.29 (0.00)	0.12 (0.00)	0.10 (0.00)
AUD	Type 1	0.30 (0.00)	0.13 (0.00)	0.11 (0.00)
	Type 2	0.20 (0.00)	0.07 (0.00)	0.06 (0.00)
	Type 3	0.17 (0.00)	0.07 (0.00)	0.06 (0.00)

Notes: Bold indicates within-class correlations.

Within-class correlations in Table 4 are bolded. Without exception, within-class resemblance exceeds cross-class resemblance in all relative pairs. We also examined, in the entire sample, whether our AUD subtypes differed in their ability to predict all cases of AUD in their relatives (bottom of Table 4). Type 1 AUD predicted the risk for AUD in full and half-siblings and cousins considerably more strongly than types 2 or 3, which had broadly similar predictive power.

Discussion

In these analyses, we sought to determine if a meaningful, valid, and reproducible typology of AUD would emerge from the application of LCA to a large, population-based sample of cases ascertained by contact with the medical system or criminal authorities in Sweden. Our approach was to use a small group of variables for primary classification: sex, age at first registration, and a history of IDs and EDs.

From our fit indices, we obtained a plausible best-fit three-class model and demonstrated, by a series of split-half analyses, that these results were highly reproducible. Our first class, which we termed type 1 AUD, accounted for 32% of our sample and was nearly 90% male, had the earliest onset of any type, and only contained individuals with at least one prior externalizing syndrome. Type 2 was our largest class, constituting nearly half of our sample, and was characterized by minimal prior internalizing or externalizing psychopathology, an intermediate AAO, and a sex distribution almost identical to that of the entire AUD cohort (~70% male). Type 3, our smallest class, containing 23% of the sample, had a nearly equal number of men and women, the latest AAO, and by far the highest proportion of individuals with prior IDs.

We sought to validate our typology by exploring its performance on a diverse group of potential validators and examining the familial transmission of the resulting classes within pairs of relatives concordant for AUD registrations. To avoid large age differences in the pairs of relatives, we examined only relationships of the same generation: full siblings, half-siblings, and cousins.

Type 1 AUD stood out on a range of validators, including (a) lowest educational level, (b) highest rates of broken families of origin, (c) highest genetic risk scores for EDs and AUD, (d) highest rates of criminal registration, (e) largest number of AUD registrations, (f) highest rate of receipt of social welfare, and (g) residence in the most deprived environments. Type 2 AUD also had a distinctive set of validators that all reflected the least social dysfunction among the types as evidenced by (a) the lowest rates of early retirement, unemployment, and receipt of welfare; (b) the lowest rates of community social deprivation; and (c) the lowest rates of broken families of origin. As expected, it also had lower genetic risks for EDs and IDs. Type 3 AUD stood out with a different set of validators showing the (a) highest education attainment, (b) highest genetic liability to IDs, (c) highest proportion of medical and pharmacy registrations, and (d) tied with type 2 for the lowest number of AUD registrations.

Our second approach to validation, consistent with a long history of using family data to validate psychiatric disorders (Robins & Guze, 1970) and prior studies of AUD typology (Kendler et al., 1998), explored the degree to which our individual classes aggregated within relatives (here full siblings, half-siblings, and cousins) concordant for AUD registration. Our results were striking. All three of our classes significantly aggregated in each of our groups of relatives and all of the cross-correlations were negative. Furthermore, all the correlations in siblings were substantially higher than those observed in half-siblings and cousins, suggesting that our findings might reflect genetic effects.

Relationship of our findings with prior AUD typologies

Our type 1 AUD bore a close resemblance to the Dionysian subtype extracted by Babor from his detailed review of prior AUD typologies (Babor, 1996), whereas our class 3 was quite similar to his Apollonian subtype. By contrast, class 2—the largest in our sample—did not have a clear precedent in Babor's review, being best characterized by mild symptomatology and clinical course. It is also useful to compare our results more closely with two influential modern AUD typologies. Cloninger's type I and type II alcoholism was originally derived from 151 Swedish adoptees (from a total cohort of 862) with registration for alcohol abuse (Cloninger et al., 1981, 1996). His type II alcoholism had four features (highly male preponderance, early AAO frequent criminal background, and high rates of recurrence) that closely resemble our type 1 AUD. His type I alcoholism shared a number of features with our type 3: mixture of men and women, later AAO, and high harm avoidance—which reflects a liability to internalizing symptoms and disorders.

Babor's typology was based on 321 alcohol patients recruited from residential treatment centers (Babor et al., 1992) and also contained two categories. His type A alcoholics had at least some features that resemble our type 3, particularly later onset and fewer related social consequences. His type B had more features of resemblance to our type 1, including higher levels of childhood and familial risk factors, earlier onset, polydrug use, and more serious consequences. As pointed out by Babor (1996), the subtypes in both of these binary typologies reflect themes across a range of prior studies dating back to the 19th century. They are also reflective of more modern work suggesting one subform sometimes termed “antisocial alcoholism” (Lappalainen et al., 1998) or “externalizers” (Del Boca & Hesselbrock, 1996) versus another form of illness that results from an “internalizing pathway” to AUD (Hussong et al., 2011), sometimes called “internalizers” (Del Boca & Hesselbrock, 1996) or “negative affect alcoholism” (Babor, 1996).

Our type 2 AUD had fewer parallels in the prior literature but bears some similarity to cluster 1—the largest single class—derived from an LCA of data from the National Epidemiologic Survey on Alcohol and Related Conditions that was also characterized by having quite low rates of internalizing and externalizing psychopathology (Moss et al., 2007).

Our findings can also be usefully compared to two population-wide Swedish studies that used LCA. The first used male twins registered with the Swedish temperance board born between 1902 and 1949 and, therefore, did not overlap with our cohort (Kendler et al., 1998). Although five classes were identified, we will focus here on only the two common classes: “single-cause registration–drunk” and “early onset–multiple cause registration.” The early onset group had high numbers of registrations and criminal registrations, a high risk for imprisonment, especially high novelty seeking, and a quite high risk for AUD was seen in the co-twins. All of these features are reminiscent of our type 3 AUD. The single-cause registration group had an intermediate AAO, low recurrence rate, low rates of criminal associations, and low levels of novelty seeking. This class had some features of our types 1 and 2. Of note, consistent with our findings, both of these types strongly aggregated within twin pairs concordant for temperance board registration.

The second study was conducted on Swedish individuals with an AUD registration and born between 1960 and 1990 (Long et al., 2019). The four classification variables, differing substantially from those used here, were the sources of AUD ascertainment, as follows: prescription, crime, inpatient, and outpatient/primary health care. A four-class solution fitted best. Class 4—called crime—closely resembled our type 1, with male preponderance and high registration for EDs. Class 1—called outpatient/prescription—bore some resemblance to our type 3, with the highest proportion of women, high rates of psychiatric disorders, and lowest rates of crime. The other two classes did not map clearly onto types in our analyses.

Limitations

These results should be interpreted in the context of at least four potentially important limitations. First, these results are dependent on the method of the diagnosis of AUD through Swedish registries. Although such administrative data have important advantages (e.g., no refusals or reporting biases), they cannot be expected to identify cases in a manner similar to interview-based assessments. Our subjects with AUD are on average probably more severe than those meeting criteria for AUD according to the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5; American Psychiatric Association, 2013), at interview, although the lifetime prevalence of alcohol dependence in nearby Norway is only moderately higher than the estimates obtained in Sweden (Kringlen et al., 2001). However, the validity of our AUD definition is supported by the high rates of concordance observed across our ascertainment methods (the M [SD] odds ratio is 24.5 [11.7]) (Kendler et al., 2018) and by the pattern of resemblance in relatives that we see here and previously (Kendler et al., 2015, 2016), which is similar to that found in studies based on personal interviews (Heath et al., 1997; Prescott & Kendler, 1999).

Second, LCA is only one of a large number of potential approaches to classifying subjects into relatively homogeneous subgroups. It assumes “local independence”—that subjects are distributed at random once subgrouping is associated for—an assumption unlikely to be precisely true. Furthermore, LCA is distinct from other traditional clustering approaches as it offers a model-based approach (Collins & Lanza, 2009). The choice of the cluster criterion therefore relies on rigorous statistical tests. In addition, it provides more information to judge the overall quality of the model. As a result, LCA may be somewhat more objective than most other clustering approaches.

Third, the choice of a best latent class solution is not always an unambiguous one. In our case, the five-class solution also seemed viable. It had the second best entropy score, and its smallest class—at 4.6%—was only slightly below our arbitrary cutoff of 5%. We therefore examined the five-class model. As seen in Appendix Table 6, this solution left our class 2 virtually unchanged but split classes 1 and 3 into two classes each. In subsequent analyses, we could see that the differences were modest in these new subclasses. The added information did not compensate for the substantial additional complexity, justifying our original decision to focus on the three-class solution.

Fourth, to simplify our LCA, we combined MD and AD into one variable—internalizing—as we did for CB and DUD to constitute EDs. To explore the impact of this decision, we repeated our analyses with the four separate syndromes: MD, AD, CB, and DUD. As outlined in Appendix Tables 7–9, the best-fit model in these new analyses also contained three classes, which closely resembled those found in our original analyses. Furthermore, class membership assignments in the two analyses were extremely similar, with agreements ranging from 94.5% to 100% (Appendix Table 10). We conclude that the results presented here were not substantially dependent on our decision to combine our two IDs and two EDs into one variable each.

Conclusions

Although it is easy to claim that AUD is a heterogeneous syndrome, it is more challenging to first demonstrate that heterogeneity using rigorous statistical methods and second to present validated evidence of meaningful subtypes. We have attempted this using the epidemiologic resources of a range of national Swedish registries relying on a simple set of phenotypes used for LCA. We obtained a clear three-class solution, which showed a large class characterized by no prominent prior psychopathology, an intermediate-sized class that was predominantly male and early onset in which all members had prior externalizing syndromes, and a still smaller class with equal numbers of men and women, later AAO, and most members with only prior IDs. These three classes differed meaningfully from each other across a wide range of potential validators. Furthermore, all three of our classes robustly aggregated in pairs of relatives, both of whom had AUD—indicating that each of the three subtypes had at least partially independent familial risk factors.

Our findings, especially our first and third classes, were broadly consistent with a large prior literature on AUD typology. That these classes have been detected across a wide range of samples and methods provides important support for their underlying validity and utility. Our study provides further important evidence that for both clinical and research applications, the heterogeneous syndrome of AUD can be meaningfully subdivided into valid subtypes, the use of which might aid future studies into etiology and treatment. Further research and treatment studies for AUD should collect the data needed to subtype their sample and then conduct analyses on those subtypes to determine if meaningful differences emerge.