Procedure and participants

As part of a comprehensive research project, patients with chronic AUD receiving inpatient treatment were involved from the Department of Psychiatry, Faculty of Medicine, University of Szeged. The inclusion criteria for participation in this study were the following: having met the criteria of DSM-5 diagnosis of AUD, completing at least primary education, surpassing the level of intellectual disability (IQ above 70). Participants were excluded if they had past history of comorbid substance use disorder (except caffeine and/or nicotine), neurodegenerative or neurological disorders or psychosis spectrum disorders. For a detailed description of exclusion and inclusion criteria and patient enrolment, see Kovács et al. [3,18]. Data of 102 patients with AUD were analyzed (Mean age: 45.68, SD: 10.35, 71.2% men, age of onset: regular alcohol consumption: 24.61 SD: 10.09, Education% (primary: 7.8%; secondary 70.9%; higher 21.4%)). The study was conducted in accordance with the Declaration of Helsinki and was approved by the Human Investigation Review Board, University of Szeged (ethical approval number: 49/B-53/2016KK). Prior to enrolment, every patient signed an informed consent form.

Measures

Data were collected for all patients after the withdrawal syndrome has subsided. Addiction Severity Index semi-structured clinical interview was applied to reveal the following (AUD-related) demographic data: age, gender, education, age of onset of regular alcohol use, family history of AUD, and number of previous inpatient treatments [19]. The severity of AUD was assessed with the 20-item self-administered questionnaire, the Severity of Alcohol Dependence Questionnaire, which measures the psycho-biological aspects of alcohol withdrawal [20]. The following self-measurement scales were applied to operationalize externalizing characteristics: The 29-item Buss-Perry Aggression Questionnaire (BPAQ) was used to reveal four facets of physical and verbal aggression, hostility and anger [21,22]. The 21-item Barratt Impulsivity Scale (BIS) was calculated to assess different components of impulsiveness, such as cognitive impulsivity, behavioral impulsivity and impatience & restlessness [23,24]. The individual level of Novelty-seeking was measured with the Temperament and Character Inventory–Revised (TCI-R), which is one of the most widely used tool developed to measure personality traits assessing four temperament and three character dimensions, from which novelty seeking is a temperament dimension [25,26]. The anxio-depressive characteristics were measured by the total scores of the 21-item Beck Depression Inventory, which evaluated the severity of depressive symptoms [27]; and the 40-item Spielberger State–Trait Anxiety Inventory [28,29] for measuring the presence of state and trait anxiety.

Statistical analysis

To test the hypothesis whether externalizing characteristics have a group-forming role in AUD, latent profile analysis (LPA) was conducted. This method can allow to identify empirically-based groups with distinct profiles of externalizing characteristics. That is, LPA can contribute to having more precise classification models than other, more arbitrary classification methods (e.g., splitting participants into categories based on median, and/or quartiles). The present analytical approach was in line with previous latent class analytic studies which examined the link between problematic alcohol use and psychopathological symptoms [30].

The estimated models contained three continuous and observed indicator variables: aggression, novelty seeking and impulsivity. All variables were standardized to assist interpretation. These constructs were measured by the BPAQ total score, by the novelty seeking total score of the TCI-R and by the BIS total score, respectively. Previous studies also calculated total scores based on these questionnaires to assess these externalizing constructs [3,31]. Moreover, the use of the total scale scores was also warranted due to the relatively low sample size (instead of the use of the more specific subscale scores). Independence of the indicator variables was suggested by significant, positive and moderately strong correlations between aggression, novelty seeking and impulsivity (r = 0.34–0.47; S1 Table). Previous theoretical and empirical findings also supported the distinguishability of these constructs (see: Introduction).

LPA models with increasing number of latent classes were estimated starting with the most parsimonious model with one latent class, and were compared in terms of various model fit indices. Lower levels of the Akaike and the Bayesian Information Criteria (AIC and BIC) and the sample size-adjusted Bayesian Information Criteria (SSA-BIC), and higher level of classification accuracy based on the measure of Entropy indicated more optimal model fitness. Decision regarding the number of classes to be retained was primarily based on the Lo-Mendel-Rubin adjusted likelihood ratio test (LMR-LRT): it tested whether a given model with k number of latent classes offered a more optimal and parsimonious solution in contrast with a model with k-1 number of latent classes. The maximum likelihood robust to non-normality (MLR) estimation method was used to perform LPA.

To test the study hypothesis whether AUD patients with more expressed externalizing symptoms show more severe anxio-depressive symptoms, the latent classes were compared in terms of trait and state anxiety and depressive symptoms. Furthermore, the identified latent classes were also compared in terms of socio-demographic and alcohol misuse related variables. For the continuous validating variables (i.e., age, age of onset of regular alcohol consumption, number of previous inpatient treatments, alcohol dependence severity, depressive symptom severity, state and trait anxiety), the 3-step Block-Croon-Hagenaars (BCH) method was used to perform bivariate comparisons between the latent classes [32]. Chi-square statistic (χ²) was calculated as a test statistic related to the BCH method [32], and Cohen’s d represented effect size. For the comparisons in terms of categorical validating variables (i.e., gender, level of education, family history of AUD), Pearson’s Chi-square statistic (χ²) was calculated as a test statistic with Fisher’s exact test for significance testing. Effect sizes were represented by Phi correlation estimates (φ).

Furthermore, to control for potential confounding effects, a multivariate logistic regression model was also constructed by using the R3Step method [33]. This allowed to compare the latent classes in terms of anxio-depressive symptoms while controlling for socio-demographic and alcohol misuse related variables. As there were very high correlations between depressive symptoms, trait and state anxiety (r = 0.69–0.77; S1 Table), a composite score of anxio-depressive symptoms was constructed by using principal component analysis (component weights = 0.87–0.92; explained variance = 82.06%). The latter step was necessary to avoid issues related to multicollinearity in the multivariate model. Other validating variables did not present high correlations with each other (S1 Table) [34].

Analyses were carried out by using Mplus 8.0 [35] and IBM SPSS Statistics 26.0 software [36].

Model selection

Models with one to four latent classes were estimated and assessed (see Table 1). Based on the indices of the AIC, the SSA-BIC and the Entropy, the four-class model provided the most optimal model fit. However, the BIC and the LMR-LRT indices suggested that the two-class solution showed the most optimal classification. The LMR-LRT showed non-significant result for both the three-class and the four-class models, which indicated that it might not be reasonable to include an additional latent class over two groups.

Due to these conflicting findings, it was not possible to determine unequivocally the best fitting model. Overall, due to multiple considerations, the two-class model was selected as the most optimal classification solution and retained for further analyses. First, previous studies suggested that results from the LMR-LRT can indicate the number of classes to be retained more precisely compared with the other fit indices [37]. Second, in the three-class and the four-class models only one and two participants were assigned to some of the latent classes. Thus, LPA was performed again without these three potential outlier cases. The model fit indices suggested similar pattern than in the full sample (S2 Table). The indices of the AIC, the SSA-BIC and the Entropy indicated that the four-class model presents the most optimal model fit. Alternatively, the BIC and the LMR-LRT suggested that the two-class solution should be retained. The latent classes of the two-class model presented similar characteristics to the full sample, thus it was possible to replicate these classes even after the exclusion of the potential outlier cases. However, in the cases of the three-class and the four-class models, the best loglikelihood values were not replicated, therefore these solutions might not have been trustworthy. Therefore, the two-class model was considered as the most reliable solution out of the competing models. Finally, it was also an aim to select a more parsimonious solution (i.e., a model with fewer latent classes) as it was not possible to determine unequivocally the best fitting model.

Profile characteristics of the latent classes

To examine profile characteristics, class-based mean scores of the indicator variables were considered. Fig 1 displays the profile characteristics of the latent classes. Class 1 (‘Moderately low externalizing characteristics’) presented moderately low rates of novelty seeking, aggression and impulsivity. In the case of Class 2 (‘Moderately high externalizing characteristics’), moderately high levels of novelty seeking, aggression and impulsivity were shown. The average latent class probabilities for the most likely latent class memberships were 0.91 and 0.91, respectively.

Associations of the latent classes with anxio-depressive symptoms and demographic variables

Results of the latent class comparisons are presented in Table 2. Members of Class 2 showed significantly higher levels of severity of alcohol dependence, depressive symptoms, and levels of trait and state anxiety compared to Class 1 with strong effect sizes in each case. Class 2 was also characterized by significantly lower age of onset of regular alcohol drinking (with moderate effect size) and by higher rates of lack of vocational or high-school graduation and family history of AUD (with small effect sizes).

Next, binary logistic regression was performed to control for potential confounding effects in the comparisons. Table 3 presents predictive effects on the membership of the ‘Moderately high externalizing characteristics’ latent class. Due to the small sample size and the limited statistical power to detect significant effects, only those variables were included in the multivariate model that showed significant differences between the two classes (Table 2). Higher levels of alcohol dependence severity and anxio-depressive symptoms had significant and positive predictive effects on the membership of the ‘Moderately high externalizing characteristics’ latent class.

Statements of ethics

The study was conducted in accordance with the Declaration of Helsinki and was approved by the Human Investigation Review Board, University of Szeged (ethical approval number: 49/B-53/2016KK). Prior to enrolment, every patient signed an informed consent form.