Sociodemographic characteristics associated with alcohol consumption and alcohol-related consequences, a latent class analysis. The WIRUS screening study

Background For alcohol, the association with socioeconomic status (SES) is different than for other public health challenges – the associations are complex, and heterogeneous between socioeconomic groups. Specifically, the relationship between alcohol consumption per se and adverse health consequences seems to be different across SES. This observation is called the ‘alcohol harm paradox’. Aims This study aims to describe different patterns of alcohol use and potential problems. Next, the associations between sub-groups characterized by different patterns of alcohol use and potential problems, and age, gender, educational level, percentage of employment, occupational level and income is analysed. Methods Employing data from the ongoing cross-sectional WIRUS-study, N=4,318 participants were included in the present study. Individual response patterns of the ten-item Alcohol Use Disorders Identification Test (AUDIT) were analysed and latent class analysis (LCA) was used to identify latent groups. Next, the associations between the classes identified in the best fitting LCA-model and sociodemographic factors was analysed and presented. Results We found support for 3 classes based on the response patterns on AUDIT. Class 1 was characterised by low-level alcohol consumption and very low probability of negative alcohol-related consequences (items 4-10 on AUDIT) related to their alcohol consumption. Class 2 was characterised by a higher level of consumption, but despite this, class 2 also had a relatively low probability of reporting negative alcohol-related consequences. Class 3, however, was characterised by high levels of alcohol consumption, and a high probability of reporting negative consequences of their consumption. The classes identified were systematically differentially associated with the included measures of SES. Conclusions Our findings highlight the interconnectedness of alcohol consumption and alcohol-related consequences. Further, our related the identified classes and yields further insights into to intricate relationship between various socioeconomic factors, alcohol use patterns and related negative consequences.


Introduction
Alcohol use seems to be unlike other public health challenges when it comes the relationship with socioeconomic status (SES).Lower socioeconomic status is usually related to higher levels of disadvantageous health-related behaviour, such as less exercise, unhealthy diet and smoking [1], resulting in negative health consequences for the individual.For alcohol, the associations are complex, and heterogeneous between socioeconomic groups [2].Several studies indicate that individuals with higher socioeconomic status consume more alcohol compared to groups in lower social strata [2].There are exceptions to this pattern, however, and some find this association only among females and in particular countries [3,4].In a comparative study of 15 countries, it was for example reported that whereas heavy drinking was related to higher education among women in most countries, it was associated with lower education among men in specific countries, including Norway [4].Studies also find patterns of more bingeing and higher scores on the Alcohol Use Disorders Identification Test (AUDIT; [5,6]) among men with lower SES [3,4].
The complexity involved is evident as we observe the risks for negative consequences of alcohol consumption which appear higher among those with lower socioeconomic status compared to those with higher socioeconomic status despite similar consumption levels.This means that the relationship between alcohol consumption per se and adverse health consequences seems to be different across the socioeconomic spectrum, dubbed the 'alcohol harm paradox'.For each alcohol unit consumed, alcohol-related harm are greater among people with lower socioeconomic status [7].The mechanisms behind the 'alcohol harm paradox' is not very well understood [3,8].Broadly, putative mechanisms related to socioeconomic status can be divided into 'differential exposure' and 'differential vulnerability' [9]; i.e. a situation where different groups have different exposures to a specific factor and a situation where the effect of a given exposure varies between groups, respectively.In relation to alcohol-harm, the most obvious a-priori candidate for differential exposure is the level of alcohol consumption.However, as highlighted in the 'alcohol harm paradox', a common finding is that alcohol consumption per se is lower in groups with lower socioeconomic status compared to groups in higher ends of the socioeconomic spectrum, despite higher levels of alcohol-related harm in the former group.For instance, a systematic review and meta-analysis on socioeconomic differences in alcohol-attributable mortality concluded that differences in alcohol exposure does not fully explain the observed association between lower socioeconomic status and alcohol-attributable mortality [10].
Several studies have modified this general finding somewhat, and called for a more detailed investigation into alcohol consumption patterns [8,[11][12][13][14].This may include investigating differential proportions at the extreme ends of the alcohol-consumption spectrum, amount consumed during one drinking session ('binge drinking'/'heavy episodic drinking'), density of drinking occasions (e.g.weekend heavy drinking), and type of beverage being consumed.As a case in point, although the mean level of consumption in low SES groups is less than for higher SES groups, Lewer and colleagues (2016) found that lower SES groups also were more likely to drink at extreme levels compared to other socioeconomic groups [8].This observation may in part help to explain the paradox.Still, Katikireddi and colleagues (2017) found that low socioeconomic status was consistently associated with raised alcohol-attributable harm (hospital admissions or death) compared to higher socioeconomic status despite controlling for weekly consumption and binge drinking [15].Other studies have qualified this finding and uncovered specific associations between alcohol-attributable conditions and socioeconomic status, where low -relative to high -socioeconomic status is related to some cancers, stroke, hypertension and liver disease [16].It has also been suggested that it is not consumption per se that is related to increased morbidity and mortality among those with low socioeconomic status, but the clustering of adverse behaviours such as smoking, being overweight, poor diet and lack of exercise [17].In summary, to understand the underpinnings of the paradox there is a need for more studies investigating different aspects of alcohol consumption patterns and consequences of alcohol consumption and the relationship with socioeconomic status.Specifically, there is a need for studies that are able to distinguish groups based on their consumption patterns, the related negative consequences of these patterns and to determine the socioeconomic characteristics that defines those groups.

Aims
The overall aim is to identify sub-groups of alcohol consumption and potential alcohol-related problems and investigate how these sub-groups relate to sociodemographic factors, including indicators of SES.This will be accomplished by a) investigating the individual response patterns of AUDIT and identify latent groups using latent class analyses (LCA).Next, b) the associations between the classes identified in the best fitting LCA-model and sociodemographic factors will be analysed and presented.The sociodemographic factors included are age, gender, educational level, percentage of employment, occupational level and income.

Design and setting
The present study has a cross-sectional design, and is part of the ongoing Norwegian national WIRUS project ("Workplace Interventions preventing Risky Use of alcohol and Sick leave").Data for the current study was obtained from the alcohol screening component of WIRUS project.More details and other results from the WIRUS can be found elsewhere [18][19][20][21][22][23].

Participants and data collection
In the WIRUS screening study, 20 large companies in Norway were recruited.These private (n = 8) and public sector (n = 12) companies employed approximately 18,000 persons.Individual-level criteria for being included in the screening study were: a) aged 16-72, b) status as an employee i.e., salaried person), in a medium or large company (c) employed in a company served by one of the participating Occupational Health Service units, regardless of work division or geographical region, d) basic understanding of the Norwegian language, and e) provided written informed consent to participate.Included companies provided email addresses for all their employees.Employees (n = 18,000) were invited to participate by receiving a web-based questionnaire.A total of 5,136 employees responded on the questionnaire (28.5 %), while 4,311 (24.0 %) responded on relevant items for this study and thus constitute the final study sample.

AUDIT
A translated Norwegian full version of the AUDIT was used in the present study, consisting of 10 items measuring different aspects of alcohol habits during the last 12 months and potential negative consequences of these alcohol habits [5,6].A recent confirmatory factor analysis of AUDIT based on the WIRUS-study concluded that AUDIT consists of one factor, and that there were no indications of differences in the factor structure or metric across gender [21].All 10 items were used as individual indicators in the latent class analysis.

Covariates
Gender, age and educational level was recorded based on self-report information.Age was divided into five groups for the purposes of this study: 18-29, 30-39, 40-49, 50-59, and ≥60years.
Educational level was recorded as a four-level variable, discriminating between primary/lower secondary, upper secondary, university/college ≤4 years, and university/college >4 years.

Percentage of employment
The participants could indicate the current percentage of employment.In the present study percentage of employment was divided into those with less than full-time employment (<100%) and those with full-time or more.

Occupational level
The participants could indicate their occupational level according to four categories: 1) Employee, 2) Middle management, 3) Upper management and 4) Other (e.g.substitute, apprentice).Four percent indicated 'other' as occupational level and were excluded from this study due to the lack of specificity of the content.The three former categories were retained and used in the present study.

Income
The yearly household income was asked for in an open-ended question.In the present study, income was used as quintiles to mitigate potential problems regarding distribution and misclassification.A little over five percent (N = 228) had missing information about yearly family income.

Identification of number of classes
Latent class analysis (LCA) was used to identify subsets of participants who shared a similar response pattern on the ten items of AUDIT.LCA can be understood as a person-centred approach that estimates the number of latent classes that can be discerned based on the total pool of individual response patterns.The approach is probabilistic rather than deterministic, and when deciding on the number classes to retain statistic criteria, parsimony and meaningfulness of the classes should be collectively considered [24].The following statistical criteria were used to decide on the number of classes to retain: Consistent Akaike information criterion (AIC), Bayesian information criterion (BIC) and adjusted BIC (aBIC) [25], were lower values indicate better model fit.Also, we used entropy to assess the quality of classification (ranging from 0 to 1 with higher values indicating better discrimination between classes), as well as the likelihood-ratio between the different models.The LCA was done iteratively, beginning with one class (i.e.similar response patterns across all participants), and increasing the number of classes up to 5.After identification of the number of classes that best fitted the data, the response patterns across AUDIT items was presented for each of the classes.As the AUDIT items differ in number of response options, the response patterns were also presented as binary variables differentiating between scoring 0 and more than zero to ease interpretation.

Association with covariates
The association between covariates and the latent classes was done using two different steps.First, the association between most likely class membership and covariates was estimated using traditional multiple logistic regression models with most probable manifest class membership as the dependent variable.Second, the association was estimated using the 3-step approach recommended by Vermunt [26] and available in Mplus as the 'R3STEP'-procedure [27].The latter procedure yields estimates from multiple logistic regression models while also taking into account the probabilistic nature of the LCA approach.The associations were estimated for each covariate separately.Only the odds ratios with p-values from the R3STEP-procedure is presented, as both approaches yielded very similar results.R [28] and the poLCA package [29] was used for the initial LCA, while inclusion of covariates using the 'R3STEP'-procedure was done in the Mplus statistical package version 8 [30].

Ethical considerations
The study was approved by the Regional Committee for Medical and Health Research in Norway (REK) (approval no.2014/647).Participants provided written informed consent upon participation.

Results
Among the eligible participants, 67% were female, and the mean age for the eligible participants was 45.0 (standard deviation 11.6) years.Only one in forty reported primary/lower secondary education, 24% reported upper secondary, 34% reported ≤4 years of university/college education, while 40% reported more than four years of university/college education.Almost 80% reported full-time employment or more, the most common occupational level was employee (81%), and 17% were middle management.The median reported household income was 918,000, with an interquartile range of 530,000.

Number of classes
Compared with a 1-class solution, there was a substantial decrease in the values related to model fit for a 2-class solution (table 1).There was a further substantial decrease when moving from a 2-to a 3-class solution, while the entropy increased from 0.769 to 0.825.When estimating more than 3 classes, however, the decrease in values related to model fit was comparatively small, and a drop in entropy was also observed.Based on this, and after visual inspection of the meaningfulness of classes 2, 3 and 4, it was decided to retain the 3-class solution.

Characteristics of retained classes
For class 1 there was a relatively high probability (81.2%) for scoring more than 0 points on item 1, and a rather low probability (17.9%) for scoring more than 0 on item 2 (for more details see figure 1 and table 2).For the rest of the items, the probability of scoring more than 0 was negligible (between 0.1 and 1.6%).The mean sum score for class 1 was 1.6 (median 1, interquartile range 1-2).
All the members of class 3 scored more than 0 points on item 1 (100%) and had a very high probability of scoring more than 0 points on item 2 (87.6%) and item 3 (99.1%).Furthermore, there was over 50% probability of scoring more than 0 on item 5, 7 and 8 (range 50.5%-68.5%),while the probability for scoring more than 0 in item 4 was 45.2%.The probability for scoring more than 0 on the rest of the items (items 6, 9 and 10) was moderate ranging between 14.4% and 19.8%.The mean sum score for class 3 was 9.4 (median 9, interquartile range 7-10).Based on the response patterns of the classes, they were characterised as: 'Class 1: Low-level consumption, no negative consequences' (38.2%), 'Class 2: Moderate level consumption, almost no negative consequences (47.2%)', and 'Class 3: Higher-level consumption, prone to negative consequences (14.6%)'.

Class belongingness and covariates
When comparing class 1 with class 2, there was an increased odds of belonging to class 1 with increasing age, being female, having higher education, while there was a decreased odds of having a higher occupational level (i.e.managerial role), an employment size of 100% or more and higher income.In the adjusted model the same pattern of associations were observed but the association with employment percentage was no longer statistically significant (see figure 2 for crude proportion across indicators of socioeconomic status and table 3).
When comparing class 1 with class 3, there was an increased odds of belonging to class 1 with increasing age, being female, having higher education and higher income, while there was a decreased odds of having an employment size of 100% or more.There was no difference between class 1 and 3 with regards to occupational level.In the adjusted model a similar pattern was observed, but the difference in income was no longer statistically significant (table 3).
When comparing class 2 with class 3, there was an increased odds of belonging to class 2 with increasing age, being female, having higher education and higher income.There was no difference between class 2 and 3 with regards to occupational level or employment percentage.In the adjusted model, the associations were similarly patterned, but education was no longer significant.
Overall, there were only small differences between the crude association estimates and the adjusted ones, and four associations were rendered non-significant (table 3).
Upon investigating the distribution of socioeconomic status across the three classes, some notable differences are worth mentioning (figure 3): Those belonging to class 3 report a higher proportion of only primary education (3.8%) and a lower proportion of higher education (33.9%) compared to the two other classes (figure 2).On the other hand, class 1 reported a substantially higher proportion of higher education (44.5%) compared to the remainder.Furthermore, over 30% of those belonging to class 3 reported a family income in the lowest quintile, compared to around 22% in the other classes.

Main findings
In the present study, we found support for 3 substantive classes based on the response patterns on AUDIT in a cohort of Norwegian workers.Class 1 could adequately be described in terms of low-level alcohol consumption both with regards to frequency and intensity/'binge drinking', as well as reporting a very low probability of negative consequences (as measured by items 4-10 on AUDIT) related to their alcohol consumption.Class 2 was characterised by a higher level of consumption both in terms of frequency and intensity, but despite this, class 2 also had a relatively low probability of reporting any negative consequences related to their alcohol consumption.The last class, however, was characterised by even higher levels of frequency and intensity of consumption, as well as a rather high probability of reporting negative consequences of their consumption.In relation to the included covariates, important differences were observed across classes.For age, gender and education, class 1 was characterised by older age, a higher proportion of females and higher educational attainment, and class 3 by younger age, more males and lower educational attainment.
Class 2 fell somewhere in between these two classes with regards to these factors.For income, class 2 was characterised by higher income compared to the other classes, and class 3 were characterised by the lowest level of income.Comparing work-related factors, the differences were less overarching, and occupational level only differed between class 1 and class 2, where the latter class was characterised by higher occupational level.For percentage of employment, class 1 was more likely to report less 100% occupation compared to the remaining classes.In the adjusted models, four of the associations were rendered non-significant, but only small changes to the point estimates were observed indicating little confounding.No other differences were observed between classes for the included covariates.

Interpretation of findings
The three different classes we identified makes intuitively sense with regards to the relationship between alcohol consumption patterns (AUDIT items 1-3) and self-reported alcohol-related consequences (AUDIT items 4-10).Our results also yields indirect support for the conventional cutpoint of 8 for alcohol-related problems as the central tendency scores (mean and median) of the class described as 'higher-level consumption, prone to negative consequences' (class 3) was close to the suggested cut-point while the other classes cental tendency scores were well below [5].The general distribution of age, gender and education across classes is in line with previous findings from the WIRUS-study where cut-points and the sum score of AUDIT was used [20].With regards to findings from other studies, we found that income and occupational level is independently differentially associated with the retained classes in mostly expected ways [3,12,31].However, we also found that those with lower occupational level (compared to class 2), less income (compared to class 2) and less than 100% employment (compared to class 2 and 3) are more likely to belong to the low risk class (class 1), even though they are more likely to report higher education in this class (compared to class 2 and 3).This could be due to gender differences in the association between education and occupational level and income (i.e.weaker association for women), and not a reflection of actual disparities between education and the other socioeconomic indicators in relation to alcohol.This notion was, however, not reflected in our findings as there was only small changes from unadjusted to adjusted estimates, despite the tendency for socioeconomic factors to cluster and co-vary.Residual confounding is however always an issue and it is possible that inclusion of other unmeasured covariates would have yielded greater evidence for confounding.Some previous studies have adjusted for drinking patterns when investigating the association between SES and alcohol-related consequences [12,15].Our findings indicate that individual drinking patterns is intrinsically related to self-reported negative consequences, as the only identified class (class 3) with a high probability of reporting negative consequences of their alcohol habits also was the class characterised by substantially higher volume of consumption (quantity and frequency, item 1 and 2) as well as frequency of binge drinking (item 3).This suggests that the level of alcohol consumption and consumption patterns, and the consequences cannot be understood separately from each other.Class 3 also differed on key demographic and socioeconomic indicators from the other classes in meaningful ways-characterised by lower age, more males, lower education and lower income.Taken together, the clear relationship between alcohol consumption patterns (AUDIT items 1-3), self-reported alcohol-related consequences (AUDIT items 4-10) and socioeconomic status suggest Third, the WIRUS-study does not include questions regarding other life-style or health factors, such as smoking, diet, physical activity, general or mental health.Inclusion of such factors would have yielded more information regarding the relationship between the identified classes and measure of socioeconomic status.Fourth, despite the inclusion of several measures of socioeconomic status, even more measures would increase the value of our findings.This includes for instance area-or neighbourhood-based deprivation, home/car ownership as socioeconomic status is a complex phenomenon with many facets.

Implications
Our findings highlight a need for differentiation of various aspects of alcohol-related harm and behaviour, and a differential association of the separate factors that are often used to gauge SES.Specifically, our findings indicate that being young, male, having low educational attainment and low income were associated with particular exposure to both high levels of alcohol consumption and alcohol-related harm.Such knowledge may have practical implications for alcohol-preventive efforts within the frame of workplace interventions.For instance, companies that to a large extent employ individuals associated with the sociodemographic characteristics of 'higher-level consumption, prone to negative consequences' (class 3), should make alcohol-preventive efforts an overall priority.

Future direction: the alcohol harm paradox
As mentioned in the introduction, both 'differential exposure' and 'differential vulnerability' is relevant when trying to understand mechanisms underlying social inequalities in health.Regarding alcohol harm and differential vulnerability, several aspects may be relevant.First, it is plausible that other life-style factors, such as smoking, unhealthy diet and less physical activity, increases the vulnerability for exposure to alcohol.A recent study concluded for instance that combinations of lifestyle factors such as smoking, excessive alcohol consumption, poor diet and low physical activity is associated with an excess risk for poor health in socioeconomically deprived populations [1].Despite such findings, Katikreddi and colleagues (2017) reported that smoking and obesity could not help explained the increased alcohol-attributable harm observed among lower versus higher socioeconomic groups [15].Second, other factors such as chronic stress [32], stressful life-events [33,34], stigma [35], mental health problems [36] and poor health in general [37] may put individuals in lower socioeconomic status groups at a higher risk for alcohol harm despite similar consumption patterns.Third, a more limited access to high-level quality health care [38,39] or less efficient application of available health information ('health literacy') [40,41] may also result in excess exacerbation of symptoms or conditions attributable to alcohol in groups with lower socioeconomic status compared to groups higher on the socioeconomic ladder.The abovementioned vulnerabilityfactors may modify the relationship between alcohol exposure and alcohol-related harm in complex ways.We are not aware of studies which have specifically investigated this future research need to further address differential exposure and differential vulnerability as a potential mechanism.In that respect an evidence review, Roche and colleagues highlighted some of the challenges with respect to giving an overview of social inequality in alcohol-related health [2]: The use of different measures of socioeconomic status as well as alcohol use across studies Different determinants can interact with each other in a multitude of ways The different components of socioeconomic status can act as mediators for each other Disadvantaged groups may be encumbered by several risk factors, which in turn can interact and modify each other Navigating these abovementioned factors in future research will provide further knowledge about the true relationship between alcohol consumption patterns, alcohol harm and socioeconomic status.

Conclusions
We found evidence for three latent classes based on the participants' response pattern on AUDIT.The class with reporting the lowest levels of alcohol consumption, also reported the lowest level of negative alcohol-related consequences.For the class with the highest levels of alcohol consumption, also reported the highest level of negative consequences.This highlight the interconnectedness of alcohol consumption and alcohol-related consequences, and also suggest that efforts to disentangle is very challenging.Our finding that the classes identified were systematically differentially associated with the included measures of socioeconomic status yields further insights into to intricate relationship between various socioeconomic factors, alcohol use patterns and related negative consequences.

Figures Figure 1
Figures

Figure 2 Distribution
Figure 2 Distribution of indicators of socioeconomic status across classes.Crude proportions based on most probable class belongingness.

Table 1 :
Comparison of model fit from 1 to 5 classes

Table 2 :
Probability of scoring more than 0 on AUDIT items across retained classes.

Table 3 :
Comparison of class belongingness across covariates.<0.05; ** <0.01; *** <0.001 a N=173 deleted observations due to missing specific information regarding occupational level.b N=228 deleted observations due to missing information regarding family income.c Model include age, gender, education, occupational level, employment percentage and income in same multinomial regression (N=3,925).d Model include age, gender, education, occupational level, employment percentage, income and company in same multinomial regression (N=3,925). *