Abstract
Background
Alcohol-related health inequalities remain a major public health challenge in England, with those from more disadvantaged socioeconomic backgrounds experiencing the greatest burden of harm despite consuming similar or lower levels of alcohol compared to those from more advantaged backgrounds. Yet, most research in this area has relied on single or composite measures of socioeconomic status (SES) that do not capture the overlapping dimensions of advantage and disadvantage that shape people’s lives and can be difficult for policymakers to interpret. We used a person-centred approach to examine how differing latent SES profiles relate to alcohol consumption before and during the COVID-19 pandemic.
Methods
We analysed data from 8,204 adults in 2019 and 5,880 adults in 2021 from the Health Survey for England. A latent class analysis was conducted on seven indicators of SES in 2019 (income, education, occupational grade, housing tenure, benefit receipt, car ownership and employment status), and six in 2021 as occupational grade was not collected that year. Multinomial logistic regression, adjusting for age, sex, ethnicity and marital status, examined associations between latent SES classes and four alcohol consumption risk categories (non-drinker, low-risk, increasing-risk and high-risk).
Results
Analysis revealed five latent classes in 2019 and four in 2021, each representing different constellations of social and economic conditions. Across both years, similar latent classes were identified and drinking patterns across classes were consistent despite the disrupting effects of the COVID-19 pandemic. Compared to non-drinkers in the Professional/Employed Homeowners class, (Skilled) Low-Income Renters and Retired Homeowners had far lower odds of drinking at all risk levels, while Professional/Employed Private Renters had odds of increasing and high-risk drinking similar to the reference group.
Conclusions
The identification of similar latent SES classes in 2019 and 2021 supports the utility of latent class analysis for capturing the multidimensional nature of SES over time and strengthens the case for future research to employ person-centred approaches when examining the relationship between alcohol consumption and SES. Capturing these constellations of SES indicators through latent class analysis may provide a stronger evidence base for designing targeted interventions and assessing the equity impacts of population-level alcohol control policies.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12889-026-26651-7.
Keywords: Socioeconomic status, Alcohol consumption, Health inequalities, Health behaviours, Alcohol harm paradox, Latent class analysis, Targeted interventions, Population-level interventions, Alcohol control policy, Health survey for England
Introduction
In England, alcohol is the leading risk factor for premature death and ill-health among individuals aged 15 to 49 [1]. National survey data indicate that alcohol consumption patterns shifted substantially during the COVID-19 pandemic [2]. Evidence shows that some individuals reduced or stopped drinking altogether, while others, especially those already drinking at harmful levels, reported significant increases in both frequency and quantity [2–6]. Consequently, the proportion of adults drinking at levels associated with long-term health risks increased, a trend that has continued in the years following the pandemic [2, 7]. This rise in harmful consumption has led to measurable population-level consequences. Every year since 2019 has recorded a new high in alcohol-specific deaths, with the rate increasing by over 20% between 2019 and 2021 and by a further 22.2% between 2021 and 2023 [8]. Recent estimates suggest that the sustained population-level rise in drinking above the recommended guidelines will lead to a substantial rise in alcohol-related health harms over the next decade, with the largest impact seen amongst the most disadvantaged groups across England [4, 9, 10].
Alcohol-specific mortality in the most deprived areas of England is more than double that in the least deprived, and the gap in potential years of working life lost is similarly large [8]. Although individuals from more affluent socioeconomic groups on average consume equal or greater quantities of alcohol compared to those from more disadvantaged groups, the burden of alcohol-attributable harm disproportionately affects individuals in lower socioeconomic positions [11–13]. This well-documented phenomenon, referred to as the alcohol harm paradox, underscores a complex and insufficiently understood relationship between patterns of alcohol consumption, socioeconomic position, and related health outcomes [14–17].
The alcohol harm paradox has been consistently observed in studies across the United Kingdom, as well as in other high-income countries such as Australia and across Europe [12, 18–23]. A growing body of literature has attempted to explain the mechanisms underlying this paradox [14, 17, 24]. A recent systematic review identified over 40 hypothesised mechanisms, many of which remain under-explored or lack empirical support [14]. Among the most frequently cited explanations are differential patterns of alcohol use, as well as a higher prevalence of other co-occurring health risk behaviours, such as smoking or poor diet, amongst more disadvantaged groups [11]. Research attempting to elucidate the mechanisms that underlie the alcohol harm paradox may be hindered by the complex nature of socioeconomic status (SES). While SES is a multifaceted construct comprised of several overlapping components [25, 26], many studies in this area adopt single indicators such as income, education, or occupational class to represent SES [12, 27, 28], or alternatively, use composite indices that are conceptually opaque and difficult for policymakers to translate into practical interventions [21, 27, 29]. Previous research studies have found a linear relationship between composite measures of SES and alcohol-related harms, whereas single SES indicators often display differing and, in many cases, non-linear relationships with alcohol consumption and harm [15, 27].
In practice, socioeconomic disadvantage is rarely experienced as a simple continuum [30]. Individuals may occupy multiple, overlapping positions that reflect both privilege and disadvantage. For example, holding a university degree while being unemployed, or owning one’s home but receiving a low income. Thus, composite measures of SES may fail to account for how these overlapping aspects of socioeconomic disadvantage combine to shape health behaviours and outcomes in non-additive ways. The assumption that SES can be treated as a uniform gradient likely overlooks the complexity and diversity of lived experiences, thereby obscuring the mechanisms through which social position influences alcohol-related harm [31].
In response to these limitations, recent research has used person-centred approaches that can capture a substantial degree of heterogeneity in the overlap between different indicators of SES [30, 31]. One such approach, latent class analysis (LCA), has received growing attention in clinical and public health research due to its ability to identify unobserved (latent) subgroups within a population based on shared response patterns across multiple indicators [32, 33]. Previous studies have used LCA to cluster individuals based on a range of indicators, including lifestyle behaviors (e.g., smoking, diet, physical activity), patterns of healthcare utilization, and psychosocial risk factors [34–37]. In the context of the association between SES and alcohol consumption, instead of examining the relationships between single indicators of SES, or computing composite SES measures, a person-centred approach such as LCA could enable the identification of distinct socioeconomic profiles that more accurately reflect the heterogeneity of social positioning.
Previous work using similar analytical approaches at the local level has demonstrated that distinct socioeconomic groups, defined by distinct combinations of socioeconomic indicators, may exhibit elevated risks of harmful or dependent drinking, even when they share few observable characteristics [31]. While these findings point to the utility of using LCA to identify socioeconomic groups, the utility of this person-centred approach at a national level has yet to be explored. The present study aims to address these gaps by investigating how multiple dimensions of socioeconomic position coalesce to shape alcohol use patterns in England.
In England, public health and clinical guidance commonly categorises alcohol consumption according to the Chief Medical Officer’s recommendations into non-drinkers, low-risk (up to 14 units/week), increasing-risk (15–35 units/week for women, 15–50 units/week for men), and higher-risk (over 35 units/week for women, over 50 units/week for men), providing a well-established framework for assessing the health risks associated with different levels of weekly alcohol consumption [38].
In light of the significant population-level changes in patterns of alcohol consumption observed following the COVID-19 pandemic [2, 7] the current study utilises two independent cross-sectional samples from the Health Survey for England (HSE) conducted in 2019 (prior to the COVID-19 pandemic) and 2021(during the COVID-19 pandemic). In 2021, a third national lockdown was introduced in England with strict stay-at-home rules, school closures, and limits on social contact [39]. Although restrictions were gradually eased over the rest of the year, ongoing measures such as limits on social gatherings, mandatory face coverings in certain settings, and widespread testing requirements continued to disrupt people’s social lives throughout 2021. The analysis employs LCA to identify distinct socioeconomic groups, which serve as a more nuanced proxy for social position than traditional SES measures, allowing for the examination of how intersecting aspects of disadvantage relate to differing levels of alcohol use. While the two samples do not track the same individuals over time, they offer nationally representative snapshots prior to (2019), and during the second year (2021) of restrictions imposed in response to the COVID-19 pandemic, allowing for a comparative examination, at the population level, of the association between differing socioeconomic profiles and alcohol use across two timepoints.
Methods
Study design and data sources
This cross-sectional study utilized publicly available data from the Health Survey for England (HSE) 2019 and 2021, the most recent archived datasets in the series. The HSE, commissioned by NHS England, is an annual, nationally representative survey that monitors health trends, disease prevalence, and lifestyle behaviours in England [40]. It aims to inform healthcare policy, service planning, and enable the evaluation of public health initiatives. Participants provided consent for their data to be made available for secondary data analysis through the UK Data Archive: https://www.data-archive.ac.uk/. Ethical approval for the current study was granted by the University of Southampton ethics committee (ERGO: 98296). Both surveys employed a multi-stage, stratified random sampling method to ensure representation of individuals living in private households [41, 42].
In the 2019 HSE sample, data collection occurred through face-to-face interviews from January to December 2019, with 9,612 addresses selected across 534 postcode sectors [41]. The original dataset included N = 10,299 participants (adults aged ≥ 16yrs and children aged 0–15yrs). Only adults were included in the analysis, resulting in a final sample of n = 8,204 adults for the 2019 sample.
Due to the COVID-19 pandemic, data collection for the 2021 HSE sample shifted from face-to-face interviews to remote telephone and video interviews [42]. Data collection spanned from January 2021 to March 2022, with 12,798 addresses selected across 711 postcode sectors. The original dataset included N = 5,880 adults and N = 1,240 children. Only adults were retained for analysis, yielding a final sample of n = 5,880 adults for the 2021 sample.
Socioeconomic indicators
The socioeconomic indicators used in the current study were based on previous research utilising LCA to identify distinct socioeconomic groups [30, 31]. These indicators included housing tenure, occupational grade, employment status, educational level, household income, car ownership, and receipt of means-tested benefits. Housing tenure was categorised into four groups: rent-free, social housing, private rented, and mortgage. Occupational grade was classified according to the National Statistics Socio-economic Classification (NS-SEC) into four categories: no occupational grade assigned, routine and manual occupations, intermediate occupations, and managerial and professional occupations. Employment status was divided into three groups: inactive (including unemployed individuals, those who were sick or disabled, students, and carers), retired, and economically active (defined as part-time or full-time employment). Educational level was assessed based on the highest qualification obtained and grouped into three categories: no qualifications, below degree level, and degree-level. Equivalised household income was divided into five quintiles: lowest quintile (£14,956 or less), second lowest quintile (£14,957–£23,443), middle quintile (£23,444–£35,540), second highest quintile (£35,541–£56,000), and highest quintile (greater than £56,000). Car ownership was included as a binary variable, indicating whether the respondent had access to a car. Receipt of means-tested benefits was also included as a binary variable, indicating whether participants were currently receiving any such benefits.
Covariates
Demographic variables, including age (categorized as 16–24yrs, 25–34yrs, 35–44yrs, 45–54yrs, 55–64yrs, 65 + years), gender (male or female), ethnicity (White, Black, Asian, or Mixed/Other), and marital status (single, married/cohabiting, or separated/divorced/widowed), were included as covariates in the model. These variables were selected based on prior research identifying them as key confounders in the relationship between SES and alcohol use [31, 43].
Distal outcome: Weekly alcohol consumption
Participants self-reported the amount of alcohol in units (10 ml or 8 g of pure alcohol) consumed on a typical day over the last 12 months, which was subsequently used to compute the average level of weekly alcohol consumption. The volume of weekly alcohol consumption was grouped into four categories, in line with the Chief Medical Officer’s (CMO) guidelines for lower-risk drinking [38]. These categories included: non-drinker (lifetime abstinence or no alcohol consumption in the last 12 months), lower risk (up to 14 units per week for both men and women), increasing risk (above 14 and up to 50 units per week for men, and above 14 and up to 35 units for women), and higher risk (above 50 units per week for men and above 35 units for women).
Analysis
In this study, latent class analysis (LCA) was employed to identify distinct socioeconomic subgroups within two nationally representative datasets: the Health Survey for England (HSE) 2019 (referred to as the 2019 Model) and the HSE 2021 (referred to as the 2021 Model). As occupational grade was not collected in the 2021 sample, it was not included as an indicator of latent class membership in the 2021 model. LCA is a person-centred approach that classifies individuals into unobserved (latent) subgroups based on their response patterns across multiple categorical indicators [32]. Unlike variable-centred approaches, which examine linear relationships between socioeconomic variables, LCA captures heterogeneity by grouping individuals with similar socioeconomic profiles into discrete latent classes. This approach is particularly advantageous for understanding the relationship between SES and alcohol use, as it allows for the identification of population subgroups that may differ meaningfully in their experiences and health-related behaviours. Furthermore, compared to traditional clustering methods such as k-means clustering, LCA provides a probabilistic framework that accounts for measurement error and uncertainty in class assignment, yielding more reliable and interpretable results [44].
To determine the optimal number of latent classes, a series of models ranging from one to six classes were estimated separately for the 2019 model and the 2021 model. The class enumeration process involved systematically increasing the number of latent classes from 1 to 6 and evaluating model fit using multiple statistical criteria. The selection of the best-fitting model was informed by both statistical and theoretical considerations. Several fit indices were examined, including the Bayesian Information Criterion (BIC), the Sample-Size Adjusted BIC (SABIC), the Akaike Information Criterion (AIC), the Consistent AIC (CAIC), and the Approximate Weight of Evidence (AWE), with lower values indicating better model [44–47]. Additionally, entropy values were assessed, with values closer to 1, suggesting clearer class separation and more precise classification of individuals [48]. The Lo–Mendell–Rubin adjusted likelihood ratio test (LMR-LRT) was also considered, as it compares models with k and k-1 classes to determine whether adding an additional class significantly improves model fit [49].
Model interpretability was a key consideration in selecting the final class solution, ensuring that the identified classes represented meaningful socioeconomic subgroups. The class enumeration process ensured that the best log-likelihood value was replicated across multiple runs, reducing the likelihood of the model settling on a local solution [50]. To account for missing data on the latent class indicators, the analysis employed Maximum Likelihood Robust (MLR) estimation. [51, 52]. This method allowed all available data to be included in the LCA without resorting to listwise deletion, which can lead to biased results and reduced statistical power.
The current analysis employed the manual three-step method for including covariates and distal outcomes in the latent class model described by Nylund-Gibson et al. (2019) to examine associations between latent class membership and weekly alcohol consumption outcomes while adjusting for the covariates [53]. In step one, a latent class model was estimated using the SES indicators to identify subgroups of individuals. In step two, classification errors were retained by fixing the measurement parameters to those obtained in step one. In step three, the distal outcome (weekly alcohol consumption) and covariates (sex, marital status, ethnicity, age) were included and regressed on latent class membership. Unlike the one-step method, which estimates all parameters simultaneously and risks altering class definitions in response to external variables, the manual 3-step method preserves the latent class solution and produces unbiased estimates by explicitly modelling classification error, thereby enhancing both validity of inferences drawn about the association between latent class membership and weekly alcohol consumption [54].
All analyses adjusted for age, gender, ethnicity, and marital status, to account for potential confounding effects. The analysis was conducted using Mplus version 8.1 [55], and results are reported in accordance with best practices for LCA and regression modelling [53, 56]. All analyses accounted for the complex survey design using pre-calculated survey weights available in the HSE 2019 and HSE 2021 datasets, which adjusted for selection probabilities, non-response bias, and population demographics. Detailed information on the computation of survey weights can be found elsewhere [41, 42].
Sensitivity analyses
Continuous weekly alcohol units
As a sensitivity analysis we re-estimated the models using continuous weekly alcohol units as the distal outcome using the three-step method detailed above. Wald tests of equality were computed to examine whether the class-specific mean estimates of weekly alcohol consumption were significantly different across the latent SES classes. Right skewed weekly alcohol units were square root transformed to approximate normality. Age, gender, ethnicity, and marital status, were adjusted for to account for potential confounding effects.
2019 Model without occupational grade
Occupational grade was included in the main 2019 latent class model because prior research indicates it is an important socioeconomic indicator in relation to weekly alcohol use [57]. However, as this indicator was not available in the 2021 data, we performed a sensitivity analysis in which the 2019 model was re-estimated without occupational grade. Age, gender, ethnicity, and marital status were adjusted for using the three-step method previously described.
Supplementary analysis
Associations between individual SES predictors and alcohol use risk
We also examined associations between the individual SES indicators used in class formation (household income, educational level, occupational grade, employment status, housing tenure, means-tested benefit receipt, and car ownership) and alcohol consumption using separate regression models adjusting for age, ethnicity, sex and marital status.
Results
Overall, across the two models, N = 14,084 adults were included in the analysis (2019 model, N = 8,204; 2021 model, N = 5,880). Table 1 presents the overall sample characteristics for both models. The distribution of differing levels of alcohol use risk across the socioeconomic indicators of latent class membership can be found in the supplementary materials (see Additional File 1: Tables S1 and S2).
Table 1.
Overall sample characteristics
| 2019 model | 2021 model | ||
|---|---|---|---|
| N(%) | N(%) | ||
| Gender | Male | 4011 (48.9%) | 2892 (49.2%) |
| Female | 4192 (51.1%) | 2987 (50.8%) | |
| Age | 16–24 | 1081 (13.2%) | 748 (12.7%) |
| 25–34 | 1349 (16.5%) | 986 (16.8%) | |
| 35–44 | 1309 (16.0%) | 939 (16.0%) | |
| 45–54 | 1396 (17.0%) | 978 (16.6%) | |
| 55–64 | 1241 (15.1%) | 910 (15.5%) | |
| 65+ | 1824 (22.2%) | 1316 (22.4%) | |
| Ethnic Group | White | 6817 (83.4%) | 4964 (85.0%) |
| Black | 267 (3.3%) | 213 (3.7%) | |
| Asian | 848 (10.4%) | 500 (8.6%) | |
| Mixed/Other | 242 (3.0%) | 161 (2.8%) | |
| Marital Status | Single | 1954 (23.8%) | 1347 (22.9%) |
| Married/Cohabiting | 5138 (62.6%) | 3760 (64.0%) | |
| Divorced/Separated/Widowed | 1109 (13.5%) | 770 (13.1%) |
Note. 2019 model (N = 8,204) reflects data from the HSE 2019 dataset. 2021 model (N = 5880) reflects data from the HSE 2021 dataset. Data from both models have been weighted to account for non-response and the complex survey design
Class enumeration
Goodness-of-fit statistics supported the selection of a five-class solution for the 2019 model and a four-class solution for the 2021 model (Table 2). In the 2019 model, the model fit statistics continued to decline through all tested solutions. Entropy was highest for the five-class model, and the LMR-LRT indicated a statistically significant improvement from four to five classes, with no significant gain beyond that. The five-class solution was the most meaningful overall and the latent classes identified were distinct from one another. Similarly, for the 2021 model, the model fit statistics decreased with each additional class. The four-class model showed the highest entropy and an acceptable class prevalence, while the LMR-LRT confirmed a significant improvement from three to four classes only. The four-class solution was deemed the most appropriate model based on interpretability, and the relative similarity of the latent classes to those obtained in the 2019 model.
Table 2.
Model fit statistics for the 2019 and 2021 models
| Classes | LL | AIC | BIC | SABIC | CAIC | AWE | LMR-LRT | Entropy | Smallest Class |
|---|---|---|---|---|---|---|---|---|---|
| 2019 Model | |||||||||
| 1 | -51159.34 | 102350.68 | 102462.87 | 102412.03 | 102397.30 | 102405.30 | 100% | ||
| 2 | -47600.57 | 95267.13 | 95498.54 | 95393.67 | 95363.29 | 95379.79 | < 0.001 | 0.73 | 35.50% |
| 3 | -46494.53 | 93089.06 | 93439.68 | 93280.78 | 93234.76 | 93259.76 | < 0.001 | 0.71 | 23.30% |
| 4 | -46003.38 | 92140.76 | 92610.59 | 92397.67 | 92336.00 | 92369.50 | < 0.001 | 0.79 | 15.70% |
| 5 | -45737.82 | 91643.65 | 92232.69 | 91965.75 | 91888.42 | 91930.42 | 0.01 | 0.80 | 7.40% |
| 6 | -45528.88 | 91259.76 | 91968.01 | 91647.06 | 91554.08 | 91604.58 | 0.62 | 0.77 | 7.50% |
| 2021 Model | |||||||||
| 1 | -27557.22 | 55140.44 | 55227.27 | 55185.96 | 55176.44 | 55182.94 | |||
| 2 | -25931.23 | 51916.45 | 52096.79 | 52011.00 | 51991.23 | 52004.73 | < 0.001 | 0.74 | 23.40% |
| 3 | -25451.08 | 50984.17 | 51258.02 | 51127.73 | 51097.71 | 51118.21 | < 0.001 | 0.71 | 18.50% |
| 4 | -25188.88 | 50487.76 | 50855.12 | 50680.35 | 50640.08 | 50667.58 | < 0.001 | 0.75 | 14.20% |
| 5 | -25062.62 | 50263.24 | 50724.11 | 50504.85 | 50454.33 | 50488.83 | 0.34 | 0.73 | 7.20% |
| 6 | -24957.06 | 50080.11 | 50634.50 | 50370.75 | 50309.97 | 50351.47 | 0.63 | 0.68 | 7.90% |
Note. The 2019 model (N = 8,204) reflects data from the HSE 2019 dataset. The 2021 model (N = 5,880) reflects data from the HSE 2021 dataset. LL Loglikelihood, AIC Akaike Information Criterion, BIC Bayesian Information Criterion, SABIC Sample-size Adjusted Bayesian Information Criterion, CAIC Consistent Akaike Information Criterion, AWE Approximate Weight of Evidence, LMR-LRT Lo-Mendell-Rubin Adjusted Likelihood Ratio Test. Values in bold reflect the model fit statistics for the selected class solutions
Conditional dependence
Latent Class Analysis (LCA) assumes conditional independence, meaning that any association between socioeconomic indicators should be fully accounted for by the latent class structure, therefore, the indicators should be uncorrelated within each latent class. Violations of this assumption can result in biased parameter estimates and poor model fit [58, 59]. Potential conditional dependencies were identified by inspecting the bivariate residual (BVR) associations using the Pearson test statistic, provided in the ‘TECH-10’ output in Mplus. Additional details on how violations of the assumption of conditional dependence were handled can be found in the supplementary materials (See Additional File 1: Supplement 1 and Additional File 1: Table S3).
Class descriptions
See Table 3 for conditional item response probabilities for the 2019 and 2021 models.
Table 3.
Conditional item response probabilities
| 2019 model (N = 8,204) |
2021 model (N = 5,880) |
||||||||
|---|---|---|---|---|---|---|---|---|---|
| LC1 | LC2 | LC3 | LC4 | LC5 | LC1 | LC2 | LC3 | LC4 | |
| Housing Tenure | |||||||||
| Rent Free | 0.01 | 0.01 | 0.01 | 0.01 | 0.01 | 0.02 | 0.01 | 0.01 | 0.01 |
| Social Housing | 0.56 | 0.05 | 0.03 | 0.01 | 0.12 | 0.53 | 0.06 | 0.21 | 0.00 |
| Private Rented | 0.30 | 0.95 | 0.02 | 0.00 | 0.19 | 0.27 | 0.02 | 0.78 | 0.00 |
| Mortgage | 0.12 | 0.00 | 0.95 | 0.98 | 0.68 | 0.18 | 0.91 | 0.00 | 0.99 |
| Occupational Grade | |||||||||
| No Occupational Grade Assigned | 0.11 | 0.01 | 0.01 | 0.01 | 0.10 | - | - | - | - |
| Routine and manual occupations | 0.67 | 0.07 | 0.46 | 0.00 | 0.48 | - | - | - | - |
| Intermediate occupations | 0.14 | 0.15 | 0.32 | 0.16 | 0.31 | - | - | - | - |
| Managerial and professional occupations | 0.08 | 0.77 | 0.21 | 0.83 | 0.10 | - | - | - | - |
| Employment Status | |||||||||
| Inactive | 0.47 | 0.07 | 0.04 | 0.03 | 0.02 | 0.41 | 0.06 | 0.09 | 0.03 |
| Retired | 0.19 | 0.00 | 0.96 | 0.19 | 0.00 | 0.17 | 0.83 | 0.01 | 0.11 |
| Employed (PT/FT) | 0.34 | 0.93 | 0.00 | 0.78 | 0.98 | 0.42 | 0.11 | 0.90 | 0.86 |
| Educational Level | |||||||||
| No Qualification | 0.40 | 0.01 | 0.40 | 0.01 | 0.17 | 0.31 | 0.32 | 0.09 | 0.05 |
| Below Degree | 0.53 | 0.17 | 0.56 | 0.32 | 0.71 | 0.57 | 0.53 | 0.45 | 0.49 |
| Degree | 0.07 | 0.83 | 0.05 | 0.67 | 0.12 | 0.13 | 0.15 | 0.46 | 0.46 |
| Equivalized Household Income | |||||||||
| £0 - £14,956 | 0.61 | 0.06 | 0.22 | 0.03 | 0.12 | 0.63 | 0.30 | 0.13 | 0.04 |
| £14,957 - £23,443 | 0.29 | 0.14 | 0.36 | 0.06 | 0.22 | 0.33 | 0.40 | 0.19 | 0.15 |
| £23,444 - £35,540 | 0.08 | 0.14 | 0.28 | 0.15 | 0.30 | 0.04 | 0.23 | 0.28 | 0.20 |
| £35,541 - £56,000 | 0.02 | 0.33 | 0.09 | 0.34 | 0.26 | 0.00 | 0.05 | 0.30 | 0.31 |
| £56,000+ | 0.01 | 0.33 | 0.04 | 0.43 | 0.11 | 0.01 | 0.02 | 0.10 | 0.29 |
| Car Ownership | |||||||||
| No | 0.56 | 0.36 | 0.15 | 0.05 | 0.04 | 0.42 | 0.14 | 0.10 | 0.03 |
| Yes | 0.44 | 0.65 | 0.85 | 0.95 | 0.96 | 0.58 | 0.86 | 0.90 | 0.97 |
| Means Tested Benefits | |||||||||
| No | 0.30 | 0.93 | 0.92 | 0.95 | 0.84 | 0.24 | 0.94 | 0.73 | 0.97 |
| Yes | 0.70 | 0.07 | 0.08 | 0.05 | 0.16 | 0.76 | 0.07 | 0.28 | 0.03 |
Note. LC = Latent Class. The 2019 model reflects data from the HSE 2019 dataset. The 2021 model reflects data from the HSE 2021 dataset. Data from both models have been weighted to account for non-response and the complex survey design
A vignette with descriptions of the latent classes retained for both final models is presented in Table 4. A description of the socio-economic and demographic characteristics for models 1 and 2 are presented in the supplementary materials (Additional File 1: Tables S4 and S5).
Table 4.
Vignette describing the latent classes from the 2019 and 2021 models
| 2019 model | 2021 model | ||
|---|---|---|---|
| LC1: ‘Skilled Low Income Renters’ (N = 1520, 18.5%) | LC1: ‘Low Income Renters’ (N = 850, 14.4%) | ||
| Majority in either social housing (56%) or privately renting (30%) | Majority in either social housing (53%) or privately renting (27%) | ||
| Routine and manual occupations (67%) | Mixed employment status, with moderated levels of employment (42%) and economic inactivity (41%) | ||
| Low household income, concentrated in the lowest (61%) and second lowest (29%) quintiles | Low educational attainment, majority either no qualifications (31%) or educated below degree level (57%) | ||
| Low educational attainment, majority either no qualifications (40%) or educated below degree level (53%) | Low household income, concentrated in the lowest (63%) and second lowest (33%) quintiles | ||
| Majority not in employment (47%) with sizable minority employed (34%) | Moderate levels of access to a car (58%) | ||
| The majority reported having no access to a car (56%) | High receipt of means-tested benefits (76%) | ||
| High receipt of means-tested benefits (70%) | |||
| LC2: ‘Professional Private Renters‘(N = 651, 7.9%) | LC2: ‘Retired Homeowners’ (N = 1020, 17.3%) | ||
| Private Renters (95%) | Homeowners (91%) | ||
| Managerial and professional occupations (77%) | Retired (83%) | ||
| High household income, concentrated in the highest (33%) and second-highest quintiles (33%) | Low educational attainment, with the majority either having no qualifications (32%) or educated below degree level (53%) | ||
| High educational attainment, majority holding a degree qualification (83%) | Low household income, concentrated in the second lowest (40%) and lowest quintiles (30%) | ||
| High level of employment (93%) | High Car ownership (86%) | ||
| Car ownership was moderate (65%) | Low receipt means-tested benefits (6%) | ||
| Low receipt of means-tested benefits (7%) | |||
| LC3: ‘Retired Homeowners’ (N = 1274, 15.5%) | LC3: ‘Employed Private Renters’ (N = 912, 15.5%) | ||
| Homeowners (95%) | Private renters (78%) | ||
| Retired (96%) | Mixed educational attainment, majority either holding a degree (46%) or educated below degree level (45%) | ||
| Mixed occupational grade. Majority in routine and manual (46%) or intermediate occupations (32%) | High level of employment (90%) | ||
| Low educational attainment, with the majority either having no qualifications (40%) or educated below degree level (56%) | Mixed household income, concentrated in the middle (28%) second highest (30%) and highest (10%) quintiles | ||
| Low household income, concentrated in the second lowest (36%) and middle quintiles (28%). | High car ownership (90%) | ||
| High car ownership (85%) | Significant minority in receipt of means-tested benefits (28%) | ||
| Low receipt means-tested benefits (8%) | |||
| LC4: ‘Professional Homeowners’ (N = 2142, 26.11%) | LC4: ‘Employed Homeowners’ (N = 3099, 52.7%) | ||
| Homeowners (98%) | Homeowners (99%) | ||
| Managerial and professional occupations (83%) | High level of employment (86%) | ||
| High household income, concentrated in the highest (43%) and second-highest quintiles (34%) | Mixed educational attainment, majority either holding a degree qualification (46%) or educated below degree level (49%) | ||
| High educational attainment, with the majority possessing a degree qualification (67%) | High household income, concentrated in the highest (29%) and second-highest quintiles (31%) | ||
| High level of employment (78%) | High car ownership (97%) | ||
| High car ownership (95%) | Low receipt of means-tested benefits (3%) | ||
| Low receipt of means-tested benefits (5%) | |||
| LC5: ‘Skilled Homeowners’ (N = 2617, 31.9%) | |||
| Homeowners (68%) | |||
| Mixed occupational grade, with majority in routine and manual (48%) and intermediate occupations (31%) | |||
| Moderate household income, spread across the middle (30%) and second-highest quintiles (26%) | |||
| Moderate educational attainment, mostly below degree level (71%) | |||
| High levels of employment (98%) | |||
| High car ownership (96%) | |||
| Low receipt means-tested benefits (16%) | |||
Note. LC Latent Class. The 2019 model reflects data from the HSE 2019 dataset (N = 8,204). The 2021 model reflects data from the HSE 2021 dataset (N = 5,880). Data from both models have been weighted to account for non-response and the complex survey design
Multinomial logistic regression results
For the 2019 model, the Professional Homeowners class served as the reference class as this class exhibited the fewest indicators of socioeconomic disadvantage overall. For the 2021 model, the Employed Homeowners class served as the reference class as this group was the most comparable to the Professional Homeowners class in the 2019 model. In both models, non-drinker served as the reference category. The predicted probabilities for each latent class and each category of alcohol use are presented visually in Fig. 1. All estimates presented for both models have been adjusted for by gender, age, marital status and ethnicity.
Fig. 1.
Predicted probabilities of alcohol use risk by latent class. Note. Panel a illustrates the predicted probabilities of alcohol use risk by latent classes in the 2019 model which reflects data from the HSE 2019 dataset (N = 8,204). Panel b illustrates the predicted probabilities of alcohol use risk by latent classes in the 2021 model which reflects data from the HSE 2021 dataset (N = 5,880). Data from both models have been weighted to account for non-response and the complex survey design
Across both models, distinct class differences in alcohol use patterns emerged (See Fig. 1). Skilled Low Income Renters (2019 model) and Low Income Renters (2021 model) consistently had the highest likelihood of being non-drinkers and the lowest likelihood of engaging in increased- or high-risk alcohol use overall. Retired Homeowners (both models) also showed relatively high rates of abstention and low engagement in higher-risk drinking. In contrast, Professional Private Renters and Professional Homeowners (2019 model), as well as Employed Private Renters and Employed Homeowners (2021 model), were the most likely to exceed the threshold for low-risk alcohol consumption. Skilled Homeowners (2019 model) tended to fall between these groups, with lower levels of abstention than retired groups and moderately elevated levels of increased- and high-risk drinking.
2019 Model: 2019 health survey for England
The results of the multinomial regression for the 2019 model are presented in Table 5. Latent class membership was significantly associated with alcohol use risk category in the 2019 model, as indicated by the overall Wald test (p < .001).
Table 5.
Multinomial logistic regression of the association between latent class membership and alcohol use (2019 model)
| Probabilities | N | RRR | SE | Lower 95% CI |
Upper 95% CI |
p value | |
|---|---|---|---|---|---|---|---|
| Latent Class 1: Skilled Low Income Renters | |||||||
| Non Drinker | 0.33 | 498 (34%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Low Risk | 0.53 | 734 (50.1%) | 0.43 | 0.05 | 0.34 | 0.53 | < 0.001 |
| Increased Risk | 0.10 | 164 (11.2%) | 0.14 | 0.03 | 0.09 | 0.19 | < 0.001 |
| High Risk | 0.04 | 68 (4.7%) | 0.31 | 0.08 | 0.15 | 0.46 | < 0.001 |
| Latent Class 2: Professional Private Renters | |||||||
| Non Drinker | 0.07 | 82 (12.8%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Low Risk | 0.66 | 401 (62.8%) | 1.11 | 0.08 | 0.95 | 1.27 | 0.201 |
| Increased Risk | 0.23 | 134 (21%) | 1.19 | 0.28 | 0.65 | 1.72 | 0.476 |
| High Risk | 0.04 | 22 (3.5%) | 1.43 | 0.60 | 0.26 | 2.61 | 0.396 |
| Latent Class 3: Retired Homeowners | |||||||
| Non Drinker | 0.20 | 248 (19.8%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Low Risk | 0.63 | 738 (59.1%) | 0.66 | 0.07 | 0.53 | 0.79 | < 0.001 |
| Increased Risk | 0.14 | 223 (17.8%) | 0.30 | 0.06 | 0.18 | 0.42 | < 0.001 |
| High Risk | 0.02 | 41 (3.3%) | 0.27 | 0.09 | 0.08 | 0.45 | < 0.001 |
| Latent Class 4: Professional Homeowners [REF] | |||||||
| Non Drinker | 0.09 | 248 (11.7%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Low Risk | 0.64 | 1231 (58.2%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Increased Risk | 0.24 | 545 (25.8%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| High Risk | 0.04 | 91 (4.3%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Latent Class 5: Skilled Homeowners | |||||||
| Non Drinker | 0.18 | 520 (20.8%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Low Risk | 0.63 | 1454 (58.2%) | 0.71 | 0.06 | 0.59 | 0.82 | < 0.001 |
| Increased Risk | 0.15 | 428 (17.1%) | 0.36 | 0.06 | 0.24 | 0.47 | < 0.001 |
| High Risk | 0.03 | 95 (3.8%) | 0.40 | 0.10 | 0.20 | 0.60 | < 0.001 |
Note. The 2019 model represents data from the HSE 2019 dataset (N = 8,204). Analysis adjusted for age, sex, ethnicity and marital status with 95% CIs
RRR Relative Risk Ratio, SE Standard Error, [REF] refers to the reference category (Non-Drinker)
Compared to Professional Homeowners, individuals in the Skilled Low Income Renters class were substantially less likely to report alcohol use across all levels of risk. Specifically, members of this class were 57% less likely to be low-risk drinkers (RRR = 0.43, 95% CI: 0.34–0.53, p < .001), 86% less likely to be increased-risk drinkers (RRR = 0.14, 95% CI: 0.09–0.19, p < .001), and 69% less likely to be classified as high-risk drinkers (RRR = 0.31, 95% CI: 0.15–0.46, p < .001) relative to being non-drinkers.
A similar pattern was observed for the Retired Homeowners class. Compared to the Professional Homeowners, individuals in this group were 34% less likely to be low-risk drinkers (RRR = 0.66, 95% CI: 0.53–0.79, p < .001), 70% less likely to be increased-risk drinkers (RRR = 0.30, 95% CI: 0.18–0.42, p < .001), and 74% less likely to be high-risk drinkers (RRR = 0.27, 95% CI: 0.08–0.45, p < .001), again indicating a higher prevalence of abstention in this class.
The Skilled Homeowners class also exhibited a consistently lower likelihood of alcohol use at all levels when compared with the reference class. Members of this group were 29% less likely to be low-risk drinkers (RRR = 0.71, 95% CI: 0.59–0.82, p < .001), 64% less likely to be increased-risk drinkers (RRR = 0.36, 95% CI: 0.24–0.47, p < .001), and 60% less likely to be high-risk drinkers (RRR = 0.40, 95% CI: 0.20–0.60, p < .001), relative to non-drinkers.
In contrast, the Professional Private Renters class did not differ significantly from the Professional Homeowners in any of the alcohol use categories. The relative risk ratios for high-risk, increased-risk, and low-risk drinking were all close to 1.0 and were non-significant (all ps > 0.20), suggesting similar patterns of alcohol use in these two socioeconomically advantaged groups.
2021 Model: 2021 health survey for England
The results from 2021 model showed similar associations between latent class membership and alcohol use risk category (See Table 6). Latent class membership was significantly associated with alcohol use risk category in the 2021 model, Wald test (p < .001).
Table 6.
Multinomial logistic regression of the association between latent class membership and alcohol use (2021 model)
| Probabilities | N | RRR | SE | Lower 95% CI |
Upper 95% CI |
P value | |
|---|---|---|---|---|---|---|---|
| Latent Class 1: Low Income Renters | |||||||
| Non Drinker | 0.38 | 313 (38.7%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Low Risk | 0.52 | 402 (49.7%) | 0.52 | 0.06 | 0.41 | 0.64 | < 0.001 |
| Increased Risk | 0.07 | 64 (7.9%) | 0.13 | 0.04 | 0.05 | 0.22 | < 0.001 |
| High Risk | 0.03 | 29 (3.6%) | 0.38 | 0.13 | 0.12 | 0.64 | 0.006 |
| Latent Class 2: Retired Homeowners | |||||||
| Non Drinker | 0.28 | 198 (20%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Low Risk | 0.60 | 593 (59.8%) | 0.69 | 0.09 | 0.52 | 0.87 | 0.001 |
| Increased Risk | 0.11 | 170 (17.2%) | 0.29 | 0.08 | 0.13 | 0.44 | < 0.001 |
| High Risk | 0.02 | 30 (3.0%) | 0.29 | 0.20 | -0.09 | 0.67 | 0.064 |
| Latent Class 3: Employed Private Renters | |||||||
| Non Drinker | 0.14 | 188 (21.4%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Low Risk | 0.62 | 504 (57.5%) | 0.98 | 0.08 | 0.83 | 1.12 | 0.740 |
| Increased Risk | 0.18 | 135 (15.4%) | 0.85 | 0.19 | 0.48 | 1.22 | 0.443 |
| High Risk | 0.06 | 50 (5.7%) | 1.82 | 0.68 | 0.49 | 3.14 | 0.113 |
| Latent Class 4: Employed Homeowners | |||||||
| Non Drinker | 0.13 | 517 (17.3%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Low Risk | 0.63 | 1746 (58.2%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| Increased Risk | 0.21 | 640 (21.3%) | 1.00 | 0.00 | 1.00 | 1.00 | |
| High Risk | 0.03 | 95 (3.2%) | 1.00 | 0.00 | 1.00 | 1.00 | |
Note. The 2021 model represents data from the HSE 2021 dataset (N = 5,880). Analysis adjusted for age, sex, ethnicity and marital status with 95% CIs
RRR Relative Risk Ratio, SE Standard Error
Compared with the reference class, the Low Income Renters class had a markedly reduced likelihood of alcohol use at all risk levels. Members of this class were 48% less likely to be low-risk drinkers (RRR = 0.52, 95% CI: 0.41–0.64, p < .001), 87% less likely to be increased-risk drinkers (RRR = 0.13, 95% CI: 0.05–0.22, p < .001), and 62% less likely to be high-risk drinkers (RRR = 0.38, 95% CI: 0.12–0.64, p = .01), when compared to non-drinkers.
The Retired Homeowners class also showed significantly lower probabilities of alcohol use across all levels of risk compared to Employed Homeowners. Individuals in this group were 31% less likely to be low-risk drinkers (RRR = 0.69, 95% CI: 0.52–0.87, p = .001) and 71% less likely to be increased-risk drinkers (RRR = 0.29, 95% CI: 0.13–0.44, p < .001) relative to being non-drinkers. Although the relative risk of high-risk drinking in this class was also 71% lower (RRR = 0.29, 95% CI: -0.09–0.67), in contrast to 2019 model, it was not statistically significant (p = .06) in the 2021 model.
The Employed Private Renters class did not significantly differ from the Employed Homeowners in any drinking category. The relative risk of high-risk drinking in this group was higher in point estimate (RRR = 1.82), but, as with the 2019 model comparisons between the Professional Private Renters and the Professional Homeowners, the result was not statistically significant (p = .11). Nor were there any significant differences observed for increased-risk or low-risk drinking in this class (both ps > 0.40).
Sensitivity analyses
Differences in weekly alcohol units across latent SES classes
Weekly alcohol units were back-transformed for interpretation purposes. In 2019, the adjusted equality test of means across latent classes revealed a significant overall difference, χ²(4) = 86.16, p < .001. Pairwise comparisons using Professional Homeowners as the reference class showed that Skilled Low Income Renters (M = 4.28, 95% CI [3.67, 4.93]; χ²(1) = 65.02, p < .001), Retired Homeowners (M = 4.62, 95% CI [3.74, 5.56]; χ²(1) = 30.74, p < .001), and Skilled Homeowners (M = 5.24, 95% CI [4.72, 5.79]); χ²(1) = 43.71, p < .001) consumed significantly less alcohol compared to Professional Homeowners (M = 8.30, 95% CI [7.60, 9.02]). In contrast, no significant difference was observed between Professional Private Renters (M = 8.52, 95% CI [6.91, 10.29]) and Professional Homeowners, χ²(1) = 0.06, p = .81.
In 2021, the adjusted equality test of means across classes revealed a significant overall difference, χ²(3) = 50.63, p < .001. Post-hoc comparisons using Employed Homeowners as the reference class showed that Low Income Renters (M = 3.02, 95% CI [2.28, 3.83]; χ²(1) = 47.37, p < .001) and Retired Homeowners (M = 3.49, 95% CI [2.24, 4.93]; χ²(1) = 11.97, p < .001) had significantly lower mean values than Employed Homeowners (M = 6.81, 95% CI [6.21, 7.44]). In contrast, no significant difference was observed between Employed Private Renters (M = 7.38, 95% CI [5.39, 9.62]; χ²(1) = 0.28, p = .60) and Employed Homeowners.
2019 Model without occupational grade
The results from the sensitivity analysis where occupational grade was omitted from the 2019 model are presented in the Supplement (Supplementary Tables S6-S8). Goodness-of-fit measures favoured a four-class solution for the 2019 model that excluded occupational grade (Supplementary Table S6). The four-class model showed the highest entropy, and the LMR-LRT suggested a statistically significant improvement when moving from three to four classes, with no additional benefit observed beyond four classes. Overall, the four-class solution was the most interpretable, and the latent classes identified were clearly distinct from each other. As the identified classes were similar to those identified in the 2021 model (see Supplementary Table S7 for conditional item response probabilities), the same naming conventions were used for the latent SES classes. As can be seen in the Supplement (Supplementary Table S8), the associations between the latent SES classes and alcohol use risk categories were consistent with those observed in the 2021 model.
Supplementary analysis
The results from the supplementary analysis where individual SES predictors predicted the likelihood of being classified into low-risk, increasing-risk, or high-risk drinking categories (relative to being a non-drinker), are presented in the Supplement (Supplementary Tables S9 and S10). All analyses adjusted for age, sex, ethnicity, and marital status.
Association between individual SES predictors and alcohol consumption (2019)
The full results are presented in the Supplement (Supplementary Table S9). In 2019, living in social housing was associated with a substantially lower likelihood of drinking at any risk level. Compared with homeowners, individuals in social housing were less likely to be low-risk drinkers (54%), increasing-risk drinkers (72%), or high-risk drinkers (54%, ps < 0.001), relative to being non-drinkers. Those living rent-free or in private rented accommodation did not differ significantly from homeowners across drinking categories.
Relative to those in managerial or professional occupations, individuals with no occupational grade were less likely to be low-risk (86%), increasing-risk (95%), or high-risk drinkers (87%, ps < 0.001) relative to being non-drinkers. Similarly, those in routine or manual occupations were less likely to be low-risk (56%), increasing-risk (71%), or high-risk drinkers (57%, ps < 0.001). Individuals in intermediate occupations were also less likely to be low-risk (33%) or increasing-risk drinkers (39%, ps < 0.001), although the association with high-risk drinking was not statistically significant (p = .07). Compared with those in employment, economically inactive individuals were less likely to be low-risk (56%), increasing-risk (70%), or high-risk drinkers (38%, ps ≤ 0.04), relative to being non-drinkers. Retired individuals were likewise less likely than those in employment to be low-risk (41%), increasing-risk (35%), or high-risk drinkers (58%, ps ≤ 0.01), relative to being non-drinkers.
Compared with degree holders, individuals with no qualifications were less likely to be low-risk (73%), increasing-risk (85%), or high-risk drinkers (70%, ps < 0.001), relative to being non-drinkers. Those with below-degree qualifications were also less likely to be low-risk (39%) and increasing-risk drinkers (49%, ps < 0.001) compared to being non-drinkers. Differences in high-risk drinking were not significant (p = .30). Relative to those earning £56,000 or more, individuals earning £0–£14,956 were substantially less likely to be low-risk (71%), increasing-risk (87%), or high-risk drinkers (72%, ps < 0.001) relative to being non-drinkers. Similar associations were observed for those earning £14,957–£23,443 (ps ≤ 0.001) and £23,444–£35,540 (59%, ps ≤ 0.01). In contrast, those earning £35,541–£56,000 were less likely to be increasing-risk drinkers (47%, p = .002) but did not differ from the highest-income group in the likelihood of being low- or high-risk drinkers (ps = 0.23–0.14) relative to non-drinkers.
Individuals with access to a car were more likely to be low-risk (44%, p = .001) or increasing-risk drinkers (74%, p < .001), with no difference in high-risk drinking (p = .96), relative to being non-drinkers. Finally, recipients of means-tested benefits were significantly less likely to be low-risk (60%), increasing-risk (75%), or high-risk drinkers (65%, ps < 0.001), relative to being non-drinkers.
Association between individual SES predictors and alcohol consumption (2021)
Patterns observed in 2019 largely persisted in 2021 (see Supplementary Table S10). Relative to homeowners, individuals living in social housing were again significantly less likely to be low-risk (62%), increasing-risk (83%), or high-risk drinkers (63%, ps < 0.001), relative to being non-drinkers. Those living rent-free were also less likely to be low-risk drinkers (76%, p = .002). Private renters did not differ from homeowners in low- or increasing-risk drinking (ps > 0.05) but were significantly more likely to be high-risk drinkers (85%, p = .04) relative to being non-drinkers. Compared with employed individuals, those who were economically inactive were less likely to be low-risk (53%, p < .001) and increasing-risk drinkers (55%, p = .02), with no difference in high-risk drinking (p = .44) relative to being non-drinkers. Retired individuals were less likely to be increasing-risk drinkers compared to being non-drinkers (35%, p = .01) but did not differ significantly in low-risk (p = .05) or high-risk drinking (p = .69).
Relative to degree holders, individuals with no qualifications were less likely to be low-risk (62%), increasing-risk (84%), or high-risk drinkers (62%, ps ≤ 0.01) relative to being non-drinkers. Those with below-degree qualifications were modestly less likely to be low-risk (19%, p = .046) and increasing-risk drinkers (40%, p < .001), with no difference observed for high-risk drinking (p = .90). Compared with the highest-income group, individuals earning £0–£14,956 were significantly less likely to be low-risk (65%), increasing-risk (82%), or high-risk drinkers (67%, ps ≤ 0.001). Those earning £14,957–£23,443 and £23,444–£35,540 were also less likely to drink at any risk level (ps < 0.05). Individuals earning £35,541–£56,000 were less likely to be increasing-risk (44%, p = .02) and high-risk drinkers (51%, p = .03) but did not differ in the likelihood of being low-risk drinkers relative to non-drinkers (p = .06).
Individuals with a car were more likely to be low-risk drinkers (61%, p = .001) and more than twice as likely to be increasing-risk drinkers (p < .001) relative to being non-drinkers, with no difference in high-risk drinking (p = .19). Finally, recipients of means-tested benefits were consistently less likely to be low-risk (60%), increasing-risk (70%), or high-risk drinkers (65%, ps < 0.001), relative to being non-drinkers.
Discussion
This study employed a person-centred approach (LCA) to identify distinct socioeconomic groups in nationally representative samples from England in 2019 and 2021 and examined their associations with alcohol use. Due to the lack of information on occupational grade in the 2021 model, the two models differed slightly in how the latent classes were defined. The 2019 model allowed finer differentiation of a Professional Homeowners class versus a Skilled Homeowners class. Conversely, the 2021 model yielded one less latent class than the 2019 model, likely due to the inability to separate individuals based on their occupational grade. The authors propose that this may have resulted in the Employed Homeowners class in the 2021 model, which reflects a combination of individuals with a range of differing occupational grades. Regardless, these technical differences did not appreciably change the association with drinking outcomes. Our sensitivity analysis, in which the 2019 model was re-estimated without occupational grade, produced latent classes and associations with alcohol consumption that were similar to those observed in the 2021 model. This reproducibility lends confidence that the identified latent classes capture meaningful and enduring profiles of social disadvantage and advantage as they relate to alcohol use.
Across both models, distinct SES classes exhibited markedly different patterns of alcohol consumption. A consistent pattern emerged, with all SES classes, except Professional Private Renters in the 2019 model and Employed Private Renters in the 2021 model, exhibiting a significantly lower likelihood of reporting alcohol use at any risk level compared to the relatively more advantaged reference classes. Notably, in both models, there was no significant difference between these Professional/Employed Private Renter classes and the Professional/Employed Homeowner reference classes at any level of alcohol use, indicating these groups share similar patterns of alcohol use despite the difference in housing tenure. While this finding appears to contradict previous research which suggests that housing tenure is a robust predictor of alcohol use [57], the authors propose that the contribution of housing tenure may become less salient when multiple other overlapping aspects of SES are accounted for.
Importantly, the findings were consistent across both years, despite the disruptive effects of the COVID-19 pandemic, further confirming the stability of the relationship between SES and alcohol use that has been consistently observed in the literature [17, 18, 23, 57]. Crucially, the most disadvantaged groups, Skilled Low Income Renters in the 2019 model, and Low Income Renters in the 2021 model, were significantly less likely to consume alcohol at any level of risk and were disproportionately abstinent. The persistence of lower alcohol use among disadvantaged groups lends partial support to the alcohol harm paradox wherein individuals from less advantaged backgrounds report similar or lower levels of alcohol consumption relative to individuals from more advantaged backgrounds. However, without including measures of alcohol-related harm the current study cannot determine whether these groups also experience disproportionately higher levels of alcohol-related harm. As alcohol-related harm was not measured in the HSE, future research using LCA to conceptualize SES is warranted to investigate the relationship between latent SES classes and alcohol-related harm.
In both the 2019 and 2021 models, the sensitivity analysis in which weekly alcohol consumption was modelled as a continuous distal outcome produced a pattern that was highly consistent with the main analysis. Specifically, in the 2019 model, the three classes that exhibited the lowest likelihood of reporting alcohol use at any risk level in the main analysis (Skilled Low-Income Renters, Retired Homeowners, and Skilled Homeowners), also showed significantly lower levels of mean weekly alcohol consumption relative to the Professional Homeowners class in the continuous model. Moreover, in both analyses, the Professional Private Renters did not differ from Professional Homeowners. Similarly, in the 2021 model, the two employed classes (Employed Private Renters and Employed Homeowners) exhibited the highest mean weekly alcohol consumption, while the Low-Income Renters and Retired Homeowners classes showed significantly lower levels of consumption relative to the Employed Homeowners class. Consistent with the main analysis, the Employed Private Renters did not significantly differ from Employed Homeowners. Taken together, these consistent findings support the robustness of our results, as similar patterns of association with alcohol consumption were observed regardless of how alcohol consumption was operationalized. However, we note that using categorical alcohol-consumption measures based on well-established thresholds offers clear advantages over continuous weekly units. High variability in weekly alcohol units within latent SES classes obscured meaningful differences in the proportion of individuals drinking above recommended guidelines. Although sensitivity analyses with continuous measures confirmed the robustness of our findings, alcohol-use risk categories provide a more informative framework for understanding how at-risk drinking is distributed across latent SES classes and offer more actionable insight for designing interventions tailored to different SES profiles.
Our findings highlight the advantages of a person-centred approach to conceptualising SES compared with the conventional view of SES as a linear or ordinal construct, as the LCA-derived classes did not align neatly along a single continuum of advantage to disadvantage. For example, groups such as Professional Private Renters or Skilled Homeowners combined elements of both privilege and insecurity, such as high educational attainment and household income but no homeownership, or homeownership but lower levels of educational attainment and household income. Supplementary analyses using individual SES indicators further demonstrated that the associations between SES and alcohol consumption are complex and often non-linear. Individual SES indicators showed broadly similar patterns, as more disadvantaged socioeconomic circumstances were associated with a higher likelihood of being a non-drinker relative to drinking at any risk level. However, these single SES measures do not capture how multiple dimensions of SES co-occur and interact to shape lived experiences. For example, across both the 2019 and 2021 samples, having access to a car was associated with a higher likelihood of drinking at any risk level relative to being a non-drinker. However, despite similarly high levels of car ownership, the latent SES classes displayed distinct patterns of alcohol consumption, showing that the relationship between any single SES indicator and drinking depends on the broader socioeconomic context. In conjunction, our findings further underscore that individuals’ socioeconomic positions are shaped by complex intersections across multiple indicators of SES that are not adequately captured by single SES indicators.
These findings build on earlier work by Boniface et al., who similarly employed LCA to explore the association between SES and higher risk drinking and found that certain socioeconomically disadvantaged classes, most notably Economically Inactive Renters and Economically Inactive Homeowners, had higher odds of harmful or dependent drinking relative to the most advantaged group [31]. This appears to contrast with our results, where the most disadvantaged classes were significantly less likely to report alcohol consumption at any level of risk. However, the divergence likely stems from differences in population and outcome measures as Boniface et al. [31] used the Alcohol Use Disorders Identification Test (AUDIT), which captures not only consumption but also patterns of harm and dependence, whereas our outcome was limited to volume-based risk thresholds (i.e. low-, increasing-, and high-risk consumption based on weekly units). Additionally, as the Boniface et al. [31] sample comprised individuals from two boroughs in South London, the findings likely reflect demographic and socioeconomic conditions specific to that area and may not be generalizable to the wider population captured in the nationally representative samples used in the present study.
It is plausible that, while most individuals in disadvantaged groups drink little or not at all, a small minority within these groups may engage in hazardous or dependent drinking, which would be more readily captured using the AUDIT. This interpretation is supported by Lewer et al., who investigated the relationship between SES and very high levels of alcohol consumption [28] They found that individuals in lower SES groups were less likely to exceed moderate drinking guidelines but were significantly more likely to engage in very high levels of alcohol consumption (e.g. >110 units per week for men). This suggests that harm may be driven by a relatively small subset of individuals in deprived groups who drink at extreme levels, a nuance that cannot be captured by the standard consumption thresholds used in our analysis.
Beard et al. similarly found that lower socioeconomic groups reported lower levels of alcohol consumption compared to higher SES groups, which closely mirrors our results [57]. However, this study used the full AUDIT, allowing them to distinguish between consumption, dependence, and harm. Akin to the current study, they found that individuals from more disadvantaged groups drank less overall, however, their use of the full AUDIT enabled them to show that these individuals also scored significantly higher on the harm and dependence subscales, indicating a greater likelihood of negative outcomes when they did drink. In conjunction with the current study, these findings provide complementary support for the alcohol harm paradox: the most socioeconomically disadvantaged groups may be the least likely to drink, but those who do may experience disproportionate harm. The convergence of these findings across differing methodological approaches strengthens the evidence base for the paradox and underscores the need for both consumption- and harm-based outcome measures in future studies using LCA to investigate the relationship between SES and alcohol-related health inequalities.
The value of examining multiple SES indicators concurrently is reinforced by other studies that have adopted intersectional or person-centred analytical strategies. For instance, Goodwin et al. used LCA to define classes combining SES, ethnicity, and migration status, and found six distinct groups with differing mental health outcomes [30]. In another example, Demant et al. analysed intersecting social identities (ethnic minority, sexual minority, gender) and found that some intersections unexpectedly conferred protective effects on alcohol use, as individuals who were both ethnic and sexual minorities had lower odds of heavy drinking than those with one minority status [60] Their results underscore that risk is not simply additive across differing dimensions of disadvantage. While our study focused on structural indicators of SES rather than identity-based categories, intersectional approaches, whether rooted in identity or socioeconomic positioning, are essential for revealing the complexity of mechanisms underlying the relationship between SES, alcohol use and adverse alcohol related health outcomes.
A major strength of this study is the application of LCA to construct empirically derived SES groupings. This approach offers several advantages over conventional SES measures. First, it allows for the identification of distinct socioeconomic profiles that reflect real-world experiences of advantage and disadvantage. Second, the resulting classes are intuitively interpretable, enabling researchers and policymakers to better understand the social positioning of different groups and their associations with alcohol use. Third, the use of two independent, nationally representative samples enhances the robustness of the socioeconomic groups identified in the current study. The consistency across pre- and during-COVID periods adds further weight to the argument that LCA comprises a useful tool for identifying distinct socioeconomic groups with differing patterns of alcohol use at a national level.
Despite its strengths, the study has important limitations. Most notably, alcohol use was measured using weekly consumption thresholds, which do not capture patterns such as binge drinking or quantify alcohol-related harm. As a result, the lack of AUDIT data in the Health Survey for England restricts our ability to directly compare findings with studies that use harm-based outcomes. Additionally, the absence of occupational grade in the 2021 dataset limited our ability to capture certain dimensions of SES and reduced comparability with the 2019 model. Relatedly, there remains little clarity on the salience of the different latent class indicators used in the current study for identifying meaningful latent SES groups. For instance, while housing tenure is a strong marker of social position in the many parts of the UK, where homeownership is closely tied to financial security and intergenerational wealth, in more urban areas where renting is the norm, its utility may be far more limited. Future research could systematically test alternative combinations of indicators or employ sensitivity analyses to establish which indicators consistently shape class structure, thereby strengthening the reliability and interpretability of using LCA as a person-centred approach to capturing meaningful and replicable latent SES groups. Moreover, the current study did not find evidence of a population-level rise in increased and higher risk drinking from 2019 to 2021 despite prior evidence demonstrating a population level increase throughout 2021 [2, 7]. This may, in part be due to the transition to remote data collection from face-face interviews due to the COVID-19 restrictions that were still in place during 2021. Although HSE 2021 data were weighted to adjust for non-response, there was a much lower household response rate in 2021 compared with 2019 (32% vs. 60%) which raises concerns about comparability across the years. Evidence from the HSE feasibility study showed that respondents were disproportionately older, from white backgrounds, and from more affluent areas, and such differences persisted after weighting [61]. This suggests that weighting cannot fully adjust for all systematic differences between respondents and non-respondents. Such biases may affect the observed associations between latent SES groups and weekly alcohol consumption, and caution is therefore needed when comparing results across survey years. Finally, as with any cross-sectional design, we cannot establish causality. Longitudinal studies are needed to determine whether movement into or out of particular SES classes influences an individual’s alcohol consumption over time. The observed associations may also be influenced by participant selection, as previous research indicates that heavy drinkers are less likely to participate in health research [62]. In addition, alcohol intake was self-reported and therefore vulnerable to underestimation, and such reporting bias may vary by SES due to several factors, such as differences in societal norms around drinking, or perceived social desirability [17]. Additionally, potential confounding factors such as psychosocial stressors, cultural attitudes toward alcohol, neighborhood alcohol availability, or underlying health conditions may influence drinking behavior differentially across the latent SES groups.
Future work should utilise the AUDIT and assess socio-economic differences in alcohol dependence and related harms, such as hospitalizations and alcohol-related illnesses, which were not available in the HSE, to enhance our understanding of how different socioeconomic groups experience alcohol-related harms and to inform public health strategies for addressing alcohol related inequalities. Furthermore, expanding LCA models to include dimensions such as race or gender, could allow for a more comprehensive intersectional understanding of alcohol-related inequalities. Moreover, future research should also examine the SES–alcohol relationship in subsequent years, as throughout 2021, there was some form of restrictions and lockdowns in England, and at times, these restrictions differed substantially from one locality to the next [39]. This lack of consistency throughout 2021 in exposure to COVID-19 restrictions and lockdowns, warrants future research when the immediate effects of the pandemic on everyday life were less apparent. Given that the 2021 Health Survey for England (HSE) dataset was the most recent available at the time of this study, future research should investigate whether the socioeconomic latent classes identified here, along with their associations with alcohol consumption, persist in the post-pandemic period.
Implications and conclusions
This research is situated within a broader imperative to inform both targeted and population-level interventions. Public health interventions that target vaguely defined “low SES” populations may lack precision and effectiveness if they do not account for the heterogeneity within these groups. Hence, the use of person-centred methods, like LCA, to identify which constellations of socioeconomic factors are linked to higher or lower alcohol consumption, and alcohol-related harm, could provide a more granular understanding of the alcohol harm paradox and help target interventions more precisely.
The alcohol harm paradox is a multifaceted public health issue, with previous evidence suggesting that differences in overall consumption explain only 5–15% of the disparity in related harms [12]. Complementary theories point to drinking patterns, broader social determinants of health, and the commercial determinants of health as non-mutually exclusive explanations for why more disadvantaged socioeconomic groups, who were more likely to abstain or drink within the low-risk drinking guidelines, nevertheless experience disproportionate alcohol-related harms compared to more advantaged socioeconomic groups [63].
Targeted interventions, such as public health messaging aimed at reducing alcohol consumption amongst individuals from more deprived SES groups, often focus on individual behaviour change without addressing the underlying drivers of harmful alcohol consumption [64]. Such approaches are unlikely to effectively address alcohol-related health inequalities unless they also tackle the social and economic conditions that shape alcohol use behaviour [14, 63]. The Social Determinants of Health (SDH) framework suggests that patterns of alcohol consumption and related health outcomes are not only shaped by individual behaviour, but also by the broader structural, material, and social conditions that people are born, live and work in [65, 66]. Factors such as housing, educational attainment, employment status, benefit receipt and income interact to influence drinking patterns and contribute to alcohol-related health inequalities [16, 17]. These overlapping disadvantages likely heighten vulnerability to initiating/maintaining harmful patterns of drinking, particularly where alcohol is used to cope with chronic stress or marginalisation [64, 67]. Consequently, tailoring interventions to the specific social and material conditions associated with each latent SES class may enable a closer match between the underlying drivers of alcohol use and the form of support offered, thereby enhancing the effectiveness of targeted interventions for reducing alcohol consumption.
For instance, for individuals in the (Skilled) Low Income Renters latent classes, which were characterised by low income, low educational attainment, moderate levels of unemployment, high benefit receipt, and limited car access, targeted interventions could be made more holistic by integrating brief alcohol interventions (BAIs) within broader social support mechanisms. For instance, promoting and delivering brief interventions for reducing alcohol consumption within job centres, housing associations, and/or community hubs, where individuals may already be accessing services, could improve accessibility and reduce stigma. Moreover, evidence suggests that incorporating Individual Placement and Support (IPS) to address unemployment, alongside referrals to mental health, housing, and financial advice services, could facilitate integrated support to address the multiple co-occurring disadvantages that drive increasing and higher risk drinking in this group [65, 68]. Neighbourhood health centres introduced in the Labour government’s 10-year health plan could serve as effective settings for delivering BAIs in conjunction with support for the material and social conditions that serve to maintain increasing and higher risk levels of drinking [69]. By integrating services such as mental health care, financial advice, housing and employment support, they offer a potential avenue for delivering holistic support that addresses both alcohol consumption and the underlying social and economic conditions that drive disproportionately higher rates of alcohol-related harm in deprived communities.
Conversely, less resource-intensive targeted interventions may be suitable for groups experiencing fewer material disadvantages but exhibiting the highest levels of increasing and higher-risk drinking, such as the Professional/Employed Homeowner/Private Renters classes. Given that individuals in these classes were predominantly employed, workplace-based screening and BAIs may be an effective strategy, as evidence has illustrated their efficacy in reducing alcohol consumption among non-dependent drinkers [70]. Integrating the Drink Less app into such interventions may further enhance effectiveness. Developed using evidence-based behavior change techniques [71], evidence shows that the Drink Less app is effective for supporting individuals to reduce their alcohol consumption [72]. Although modelling by Angus et al. suggests that referral to the app through primary care could yield substantial public health benefits [73], previous research indicates that this approach may not adequately reach more socioeconomic advantaged groups who are less likely to receive a BAI from their GP [74]. Integrating Drink Less into workplace-based interventions may therefore represent a more effective and cost-efficient means of reaching increasing and higher-risk drinkers in these subgroups.
These examples serve to highlight the potential value of using LCA to guide intervention design by identifying specific socioeconomic subgroups and informing the most effective settings and methods for delivering interventions that maximise reach within each subgroup. Future research could build on our findings by using the identified latent SES classes to inform the design of tailored, context-specific interventions. By aligning interventions with the specific material and social contexts of the latent SES groups, these targeted interventions may be better positioned to reduce alcohol-related harm and address underlying inequalities.
Beyond informing targeted interventions, the latent classes identified in this study offer a valuable tool for assessing the equity of population-level alcohol policies. These interventions, which aim to reduce alcohol-related harm across entire populations by altering the broader social, physical, and economic contexts in which alcohol consumption occurs, include measures such as alcohol pricing policies, marketing restrictions, and mass media campaigns [75, 76]. Although population-level policies target the general population, their impacts may differ significantly depending on an individual’s socioeconomic context, as evidence suggests that alcohol control policies can have varying effects across social strata [77].
For instance, minimum unit pricing (MUP), which disproportionately affects cheaper, high-strength alcoholic beverages commonly consumed by heavy drinkers in lower socioeconomic groups, has been associated with reductions in alcohol purchases, particularly among low-income households with high alcohol consumption [78–80]. This suggests that MUP may be especially impactful for members of the (Skilled) Low Income Renters classes, who face high levels of unemployment and financial vulnerability, making them more sensitive to price-based deterrents. Conversely, individuals in more advantaged classes, such as Professional/Employed Homeowners and Professional Private Renters, may be insulated from the effects of price increases due to their greater financial security.
However, findings regarding the impact of MUP on other socioeconomic indicators, such as occupational grade, remain mixed [77]. Since previous research has often relied on single proxies or composite SES measures [74, 78, 79, 81], it remains unclear how the diverse social, economic, and material conditions individuals experience interacts to shape responses to population-level interventions. Further research is needed to examine how alcohol-control policies like MUP differentially influence patterns of alcohol consumption across the socioeconomic latent classes identified in this study. As these classes reflect multiple, overlapping dimensions of SES, using LCA offers a method of evaluating policy impacts that more accurately reflects the complex nature of socioeconomic disadvantage in the real-world.
Finally, in addition to the influence of the social determinants of health, commercial determinants of health also play a crucial role in shaping alcohol-related health outcomes and contribute significantly to alcohol-related health inequalities [82]. Commercial determinants of health, defined as “the systems, practices, and pathways through which commercial actors influence health and equity in pursuit of profit”, encompass not only the production, pricing, and marketing of unhealthy commodities such as alcohol, but also the lobbying of policymakers, the funding of research to shape evidence, and the framing public debates to delay or weaken regulation [83–85]. Alcohol industry narratives often highlight health inequalities across socioeconomic groups to argue that the relationship between alcohol use and health outcomes is too complex to disentangle. Instead, the industry often places large emphasis on individual behaviours while downplaying the role of their own practices in shaping consumption and harm [82]. Yet, evidence from England challenges these industry narratives as alcohol outlets are disproportionately concentrated in disadvantaged areas [86, 87], and areas that have a higher density of alcohol outlets experience higher rates of alcohol-related hospital admissions [88]. Additionally, research suggests that there are around six times more outdoor advertisements for unhealthy commodities, such as alcohol, in the most deprived, compared to the least deprived, areas of the country [89, 90]. Future research could build on the current findings by linking the identified latent SES subgroups to data on outlet density and exposure to alcohol marketing or price promotions, to facilitate a more nuanced understanding of how corporate strategies influence patterns of alcohol consumption across differing SES groups. This person-centred approach may help clarify whether the latent SES groups at highest risk of harm also encounter the most intensive commercial targeting, potentially helping to challenge industry narratives through providing evidence that alcohol-related inequalities are shaped not only by broader social and material conditions but also by the actions of the alcohol industry.
Conclusion
In conclusion, this study demonstrates the utility of LCA in capturing the multidimensional nature of SES and its relationship to alcohol consumption. By moving beyond single indicators and composite measures of SES, LCA may enable more context-sensitive design of targeted interventions and support a more nuanced evaluation of the equity of population-level alcohol control policies. Ultimately, person-centred approaches could provide a robust foundation for future research into the underlying drivers of the alcohol harm paradox, as well as for designing and evaluating public health efforts that not only reduce harmful patterns of alcohol consumption but also begin to address the persistent socioeconomic inequalities in alcohol-related harms seen across England.
Supplementary Information
Acknowledgements
Not applicable.
Abbreviations
- AIC
Akaike information criterion
- AUDIT
Alcohol use disorders identification test
- AWE
Approximate weight of evidence
- BAI
Brief alcohol interventions
- BIC
Bayesian information criterion
- BVR
Bivariate residual
- CAIC
Consistent akaike information criterion
- HSE
Health survey for England
- IPS
Individual placement and support
- LC
Latent class
- LCA
Latent class analysis
- LL
Loglikelihood
- LMR-LRT
Lo-Mendell-Rubin adjusted Likelihood Ratio Test
- MUP
Minimum unit pricing
- SABIC
Sample-size adjusted bayesian information criterion
- SDH
Social determinants of health
Authors’ contributions
J.K., S.B. and K.S. contributed to the conceptualization of the research design. J.K. contributed to the formal analysis, interpretation and writing of the original draft. J.K., K.S., S.B. and V.B. contributed to the methodology. K.S., S.B. and V.B. supervised the work. J.K., K.S., S.B. and V.B. contributed to reviewing and editing the final manuscript.
Funding
The funder had no role in study design, data collection, analysis, interpretation, or report writing.
Data availability
The datasets supporting the conclusions of this article are available in the UK Data Service Archive, subject to their end user license agreement: NatCen Social Research, University College London, Department of Epidemiology and Public Health. Health Survey for England. [data series]. 8th Release. UK Data Service, 2024 [Accessed 17 Oct 2024]. Available from: DOI: http://doi.org/10.5255/UKDA-Series-2000021.
Declarations
Ethics approval and consent to participate
The current study used secondary data from the Health Survey for England (HSE) 2019 and HSE 2021. Access to the anonymised dataset was obtained through the UK Data Archive. Participants provided informed consent before participation in the study and for their data to be made publicly available for secondary analysis. This study obtained ethics approval from the University of Southampton ethnics committee (ERGO: 98296).
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Brauer M, Roth GA, Aravkin AY, Zheng P, Abate KH, Abate YH, et al. Global burden and strength of evidence for 88 risk factors in 204 countries and 811 subnational locations, 1990–2021: a systematic analysis for the global burden of disease study 2021. Lancet. 2024;403(10440):2162–203. 10.1016/S0140-6736(24)00933-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Public Health England. Monitoring alcohol consumption and harm during the COVID-19 pandemic. 2021. https://www.gov.uk/government/publications/alcohol-consumption-and-harm-during-the-covid-19-pandemic. Accessed 1 Apr 2025.
- 3.Anderson P, O’Donnell A, Llopis EJ, Kaner E. The COVID-19 alcohol paradox: British household purchases during 2020 compared with 2015–2019. PLoS ONE. 2022;17(1):e0261609–e. 10.1371/JOURNAL.PONE.0261609. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Boniface S, Card-Gowers J, Martin A, Retat L, Webber L. The COVID hangover Addressing long-term health impacts of changes in alcohol consumption during the pandemic. 2022. https://www.ias.org.uk/wp-content/uploads/2022/07/The-COVID-Hangover-report-July-2022.pdf. Accessed 25 Apr 2025.
- 5.Institute of Alcohol Studies. Alcohol consumption during the COVID-19 lockdown: Summary of emerging evidence from the UK. 2020. https://www.ias.org.uk/wp-content/uploads/2020/06/sb28062020.pdf. Accessed 24 May 2025.
- 6.Institute of Alcohol Studies. Alcohol consumption during the COVID-19 pandemic in the UK: Second IAS briefing. 2020. https://www.ias.org.uk/wp-content/uploads/2020/10/sb29102020.pdf. Accessed 24 May 2025.
- 7.Alcohol Toolkit Study. Monthly Tracking KPI - Graphs - Alcohol in England. 2025. https://www.alcoholinengland.info/graphs/monthly-tracking-kpi. Accessed 3 Mar 2025.
- 8.Department of Health & Social Care. Local Alcohol Profiles | Fingertips. 2025. https://fingertips.phe.org.uk/profile/local-alcohol-profiles. Accessed 12 Mar 2025.
- 9.Angus C, Henney M, Pryce R. Modelling the impact of changes in alcohol consumption during the COVID-19 pandemic on future alcohol-related harm in England. Sheffield Alcohol Research Group; 2022. https://orda.shef.ac.uk/articles/report/Modelling_the_impact_of_changes_in_alcohol_consumption_during_the_COVID-19_pandemic_on_future_alcohol-related_harm_in_England/19597249?file=36419202. Accessed 11 Aug 2024.
- 10.Angus C, Henney M, Pryce R. Modelling the longer-term health and health inequality impacts of changes in alcohol consumption during the COVID-19 pandemic in England. J Public Health. 2024;46(2):286–93. 10.1093/PUBMED/FDAE010. [DOI] [PubMed] [Google Scholar]
- 11.Bellis MA, Hughes K, Nicholls J, Sheron N, Gilmore I, Jones L. The alcohol harm paradox: using a National survey to explore how alcohol May disproportionately impact health in deprived individuals. BMC Public Health. 2016;16(1):1–10. 10.1186/s12889-016-2766-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Jones L, Bates G, McCoy E, Bellis MA. Relationship between alcohol-attributable disease and socioeconomic status, and the role of alcohol consumption in this relationship: A systematic review and meta-analysis. BMC Public Health. 2015;15(1). 10.1186/s12889-015-1720-7. [DOI] [PMC free article] [PubMed]
- 13.Jones L, McCoy E, Bates G, Bellis M, Sumnall H. Understanding the alcohol harm paradox in order to focus the development of interventions. 2015. https://s3.eu-west-2.amazonaws.com/files.alcoholchange.org.uk/documents/FinalReport_0122.pdf. Accessed 25 Mar 2025.
- 14.Boyd J, Sexton O, Angus C, Meier P, Purshouse RC, Holmes J. Causal mechanisms proposed for the alcohol harm paradox-a systematic review. Addiction. 2022;117(1):33–56. 10.1111/ADD.15567. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Collins SE. Associations between socioeconomic factors and alcohol outcomes. Alcohol Research: Curr Reviews. 2016;38(1):83–94. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Institute of Alcohol Studies. Alcohol and Health Inequalities. 2020. https://www.ias.org.uk/wp-content/uploads/2020/12/Alcohol-and-health-inequalities.pdf. Accessed 24 May 2025.
- 17.Probst C, Kilian C, Sanchez S, Lange S, Rehm J. The role of alcohol use and drinking patterns in socioeconomic inequalities in mortality: a systematic review. Lancet Public Health. 2020;5(6):e324–32. 10.1016/S2468-2667(20)30052-9. [DOI] [PubMed] [Google Scholar]
- 18.Katikireddi SV, Whitley E, Lewsey J, Gray L, Leyland AH. Socioeconomic status as an effect modifier of alcohol consumption and harm: analysis of linked cohort data. Lancet Public Health. 2017;2(6):e267–76. 10.1016/S2468-2667(17)30078-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Livingston M. Socioeconomic differences in alcohol-related risk-taking behaviours. Drug Alcohol Rev. 2014;33(6):588–95. 10.1111/DAR.12202. [DOI] [PubMed] [Google Scholar]
- 20.Paljärvi T, Suominen S, Car J, Koskenvuo M. Socioeconomic disadvantage and indicators of risky Alcohol-drinking patterns. Alcohol Alcohol. 2013;48(2):207–14. 10.1093/ALCALC/AGS129. [DOI] [PubMed] [Google Scholar]
- 21.Sadler S, Angus C, Gavens L, Gillespie D, Holmes J, Hamilton J, et al. Understanding the alcohol harm paradox: an analysis of sex- and condition-specific hospital admissions by socio-economic group for alcohol-associated conditions in England. Addiction (Abingdon England). 2017;112(5):808–17. 10.1111/ADD.13726. [DOI] [PubMed] [Google Scholar]
- 22.Van Oers JAM, Bongers IMB, Van De Goor LAM, Garretsen HFL. Alcohol consumption, alcohol-related problems, problem drinking, and socioeconomic status. Alcohol Alcohol. 1999;34(1):78–88. 10.1093/ALCALC/34.1.78. [DOI] [PubMed] [Google Scholar]
- 23.Giskes K, Turrell G, Bentley R, Kavanagh A. Individual and household-level socioeconomic position is associated with harmful alcohol consumption behaviours among adults. Aust N Z J Public Health. 2011;35(3):270–7. 10.1111/J.1753-6405.2011.00683.X. [DOI] [PubMed] [Google Scholar]
- 24.Boyd J, Bambra C, Purshouse RC, Holmes J. Beyond behaviour: How health inequality theory can enhance our understanding of the ‘alcohol-harm paradox’. International Journal of Environmental Research and Public Health: MDPI; 2021. Accessed. [DOI] [PMC free article] [PubMed]
- 25.Shavers VL. Measurement of socioeconomic status in health disparities research. J Natl Med Assoc. 2007;99(9):1013. [PMC free article] [PubMed] [Google Scholar]
- 26.Wilkinson R, Marmot M. Social determinants of health: the solid facts. 2nd ed. Copenhagen: World Health Organization. Regional Office for Europe; 2003. p. 9289013710. [Google Scholar]
- 27.Beard E, Brown J, West R, Angus C, Brennan A, Holmes J, et al. Deconstructing the alcohol harm paradox: A population based survey of adults in England. PLoS ONE. 2016;11(9):e0160666–e. 10.1371/JOURNAL.PONE.0160666. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Lewer D, Meier P, Beard E, Boniface S, Kaner E. Unravelling the alcohol harm paradox: A population-based study of social gradients across very heavy drinking thresholds. BMC Public Health. 2016;16(1):1–11. 10.1186/s12889-016-3265-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Beard E, Brown J, West R, Holmes J, Kaner E, Meier P, et al. Characterising the alcohol harm paradox: a population-based survey of adults in England. Lancet. 2015;386:S23–S. 10.1016/s0140-6736(15)00861-2. [Google Scholar]
- 30.Goodwin L, Gazard B, Aschan L, MacCrimmon S, Hotopf M, Hatch SL. Taking an intersectional approach to define latent classes of socioeconomic status, ethnicity and migration status for psychiatric epidemiological research. Epidemiol Psychiatric Sci. 2018;27(6):589–600. 10.1017/S2045796017000142. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Boniface S, Lewer D, Hatch SL, Goodwin L. Associations between interrelated dimensions of socio-economic status, higher risk drinking and mental health in South East london: A cross-sectional study. PLoS ONE. 2020;15(2). 10.1371/JOURNAL.PONE.0229093. [DOI] [PMC free article] [PubMed]
- 32.Collins LM, Lanza ST. Latent Class and Latent Transition Analysis: With Applications in the Social, Behavioral, and Health Sciences. Latent Class and Latent Transition Analysis: With Applications in the Social, Behavioral, and Health Sciences. 2010:1-295. 10.1002/9780470567333
- 33.Sinha P, Calfee CS, Delucchi KL. Practitioner’s guide to latent class analysis: methodological considerations and common pitfalls. Crit Care Med. 2021;49(1):e63–e. 10.1097/CCM.0000000000004710. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.Grant RW, McCloskey J, Hatfield M, Uratsu C, Ralston JD, Bayliss E, et al. Use of latent class analysis and k-Means clustering to identify complex patient profiles. JAMA Netw Open. 2020;3(12):e2029068–e. 10.1001/JAMANETWORKOPEN.2020.29068. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Roberts E, Hotopf M, Strang J, Marsden J, White M, Eastwood B, et al. The hospital admission profile of people presenting to specialist addiction services with problematic use of alcohol or opioids: A National retrospective cohort study in England. Lancet Reg Health – Europe. 2021;3. 10.1016/j.lanepe.2021.100036. [DOI] [PMC free article] [PubMed]
- 36.Suhag A, Webb TL, Holmes J. Longitudinal clustering of health behaviours and their association with Multimorbidity in older adults in england: A latent class analysis. PLoS ONE. 2024;19(1):e0297422–e. 10.1371/JOURNAL.PONE.0297422. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Visser E, Oudsten BLD, Gosens T, Lodder P, De Vries J. Psychological risk factors that characterize the trajectories of quality of life after a physical trauma: a longitudinal study using latent class analysis. Qual Life Res. 2021;30(5):1317–35. 10.1007/S11136-020-02740-X. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Department of Health. UK Chief Medical Officers’ Low Risk Drinking Guidelines. 2016. https://assets.publishing.service.gov.uk/media/5a80b7ed40f0b623026951db/UK_CMOs__report.pdf. Accessed 19 Aug 2024.
- 39.Institute for Government. Timeline of UK government coronavirus lockdowns and restrictions. 2022. https://www.instituteforgovernment.org.uk/data-visualisation/timeline-coronavirus-lockdowns. Accessed 25 Jan 2025.
- 40.NatCen Social Research UCL, Department of Epidemiology and Public Health, Health Behaviour Unit. Health Survey for England. [data series]. UK Data Service, 2023. Available from: 10.5255/UKDA-Series-2000021. Accessed 1 Apr 2025.
- 41.NatCen Social Research. Health Survey for England 2019: User Guide. University College London, Department of Epidemiology and Public Health., 2020. https://doc.ukdataservice.ac.uk/doc/8860/mrdoc/pdf/8860_hse_2019_dataset_documentation.pdf. Accessed 25 Jan 2025.
- 42.NatCen Social Research. Health Survey for England 2021: User Guide. University College London, Department of Epidemiology and Public Health., 2022. https://doc.ukdataservice.ac.uk/doc/9319/mrdoc/pdf/9319_hse_2021_user_guide_v1.pdf. Accessed 25 Jan 2025.
- 43.Penã S, Mäkelä P, Härkänen T, Heliövaara M, Gunnar T, Männistö S, et al. Measurement error as an explanation for the alcohol harm paradox: analysis of eight cohort studies. Int J Epidemiol. 2021;49(6):1836–46. 10.1093/IJE/DYAA113. [DOI] [PubMed] [Google Scholar]
- 44.Nylund-Gibson K, Choi AY. Ten frequently asked questions about latent class analysis. Translational Issues Psychol Sci. 2018;4(4):440–61. 10.1037/TPS0000176. [Google Scholar]
- 45.Bozdogan H. Model selection and akaike’s information criterion (AIC): the general theory and its analytical extensions. Psychometrika. 1987;52(3):345–70. 10.1007/BF02294361. [Google Scholar]
- 46.Schwarz G. Estimating the dimension of a model. Ann Statist. 1978;6(2):461–4. 10.1214/AOS/1176344136. [Google Scholar]
- 47.Sclove SL. Application of model-selection criteria to some problems in multivariate analysis. Psychometrika. 1987;52(3):333–43. 10.1007/BF02294360. [Google Scholar]
- 48.Celeux G, Soromenho G. An entropy criterion for assessing the number of clusters in a mixture model. J Classif. 1996;13(2):195–212. 10.1007/BF01246098. [Google Scholar]
- 49.Lo Y, Mendell NR, Rubin DB. Testing the number of components in a normal mixture. Biometrika. 2001;88(3):767–78. 10.1093/BIOMET/88.3.767. [Google Scholar]
- 50.Hipp JR, Bauer DJ. Local solutions in the Estimation of growth mixture models. Psychol Methods. 2006;11(1):36–53. 10.1037/1082-989X.11.1.36. [DOI] [PubMed] [Google Scholar]
- 51.Enders CK. Applied missing data analysis. New York, NY, US: Guilford Press; 2010. 978-1-60623-639-0 (Hardcover). [Google Scholar]
- 52.Enders CK. Applied missing data analysis. 2 ed. Guilford Press; 2022. p. 9781462549863.
- 53.Nylund-Gibson K, Grimm RP, Masyn KE. Prediction from latent classes: A demonstration of different approaches to include distal outcomes in mixture models. Struct Equation Modeling: Multidisciplinary J. 2019;26(6):967–85. 10.1080/10705511.2019.1590146. [Google Scholar]
- 54.Asparouhov T, Muthén B. Auxiliary variables in mixture modeling: Three-Step approaches using Mplus. Struct Equation Modeling: Multidisciplinary J. 2014;21(3):329–41. 10.1080/10705511.2014.915181. [Google Scholar]
- 55.Muthén LK, Muthén BO. Mplus User’s Guide. Eighth Edition. Los Angeles; 2017. https://www.statmodel.com/download/usersguide/MplusUserGuideVer_8.pdf. Accessed 1 Apr 2025.
- 56.Weller BE, Bowen NK, Faubert SJ. Latent class analysis: A guide to best practice. J Black Psychol. 2020;46(4):287–311. 10.1177/0095798420930932. [Google Scholar]
- 57.Beard E, Brown J, West R, Kaner E, Meier P, Michie S. Associations between socio-economic factors and alcohol consumption: A population survey of adults in England. PLoS ONE. 2019;14(2):e0209442. 10.1371/journal.pone.0209442. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 58.Lee J, Jung K, Park J. Detecting conditional dependence using flexible bayesian latent class analysis. Front Psychol. 2020;11:482868. 10.3389/fpsyg.2020.01987. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 59.Visser M, Depaoli S. A guide to detecting and modeling local dependence in latent class analysis models. Struct Equation Modeling: Multidisciplinary J. 2022;29(6):971–82. 10.1080/10705511.2022.2033622. [Google Scholar]
- 60.Demant D, Oviedo-Trespalacios O, Carroll JA, Ferris JA, Maier L, Barratt MJ, et al. Do people with intersecting identities report more high-risk alcohol use and lifetime substance use? Int J Public Health. 2018;63(5):621–30. 10.1007/S00038-018-1095-5. [DOI] [PubMed] [Google Scholar]
- 61.NHS England Digital. The Health survey for England 2020/2021 Feasibility study. 2021. https://digital.nhs.uk/data-and-information/areas-of-interest/public-health/health-survey-for-england---health-social-care-and-lifestyles/feasibility-study-2020---2021? Accessed 25 Apr 2025.
- 62.Lemmens PHHM, Tan ES, Knibbe RA. Bias due to Non-response in a Dutch survey on alcohol consumption. Br J Addict. 1988;83(9):1069–77. 10.1111/j.1360-0443.1988.tb00534.x. [DOI] [PubMed] [Google Scholar]
- 63.Bloomfield K. Understanding the alcohol-harm paradox: what next? Lancet Public Health. 2020;5(6):e300–1. 10.1016/S2468-2667(20)30119-5. [DOI] [PubMed] [Google Scholar]
- 64.Roche A, Kostadinov V, Fischer J, Nicholas R, O’Rourke K, Pidd K, et al. Addressing inequities in alcohol consumption and related harms. Health Promot Int. 2015;30(suppl2):ii20–35. 10.1093/HEAPRO/DAV030. [DOI] [PubMed] [Google Scholar]
- 65.Marmot M, Allen J, Boyce T, Goldblatt P, Morrison J. Health equity in England: The Marmot Review 10 years on. London; 2020. Accessed 1 Apr 2025.
- 66.World Health Organization. A conceptual framework for action on the social determinants of health. Geneva: World Health Organization. 2010. https://iris.who.int/handle/10665/44489. Accessed 26 May 2025.
- 67.Heim D, Monk RL, Qureshi AW. An examination of the extent to which drinking motives and problem alcohol consumption vary as a function of deprivation, gender and age. Drug Alcohol Rev. 2021;40(5):817–25. 10.1111/dar.13221. [DOI] [PubMed] [Google Scholar]
- 68.Marsden J, Anders P, Shaw C, Amasiatu C, Collate W, Eastwood B, et al. Superiority and cost-effectiveness of individual placement and support versus standard employment support for people with alcohol and drug dependence: a pragmatic, parallel-group, open-label, multicentre, randomised, controlled, phase 3 trial. eClinicalMedicine. 2024;68:102400. 10.1016/j.eclinm.2023.102400. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 69.Department of Health & Social Care. Fit for the future: 10 Year Health Plan for England. London. 2025. Accessed 1 Aug 2025.
- 70.Yuvaraj K, Eliyas SK, Gokul S, Manikandanesan S. Effectiveness of workplace intervention for reducing alcohol consumption: a systematic review and Meta-Analysis. Alcohol Alcohol. 2019;54(3):264–71. 10.1093/ALCALC/AGZ024. [DOI] [PubMed] [Google Scholar]
- 71.Garnett C, Crane D, West R, Brown J, Michie S. The development of drink less: an alcohol reduction smartphone app for excessive drinkers. Translational Behav Med. 2019;9(2):296–307. 10.1093/TBM/IBY043. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 72.Oldham M, Beard E, Loebenberg G, Dinu L, Angus C, Burton R, et al. Effectiveness of a smartphone app (Drink Less) versus usual digital care for reducing alcohol consumption among increasing-and-higher-risk adult drinkers in the UK: a two-arm, parallel-group, double-blind, randomised controlled trial. eClinicalMedicine. 2024;70. 10.1016/j.eclinm.2024.102534. [DOI] [PMC free article] [PubMed]
- 73.Angus C, Oldham M, Burton R, Dina LM, Field M, Hickman M, et al. Modeling the potential health, health Economic, and health inequality impact of a Large-Scale rollout of the drink less app in England. Value Health. 2025;28(2):215–23. 10.1016/j.jval.2024.11.007. [DOI] [PubMed] [Google Scholar]
- 74.Angus C, Brown J, Beard E, Gillespie D, Buykx P, Kaner EFS, et al. Socioeconomic inequalities in the delivery of brief interventions for smoking and excessive drinking: findings from a cross-sectional household survey in England. BMJ Open. 2019;9(4):e023448. 10.1136/BMJOPEN-2018-023448. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 75.Babor TF, Casswell S, Graham K, Huckle T, Livingston M, Österberg E, et al. Alcohol: no ordinary commodity: research and public policy. Oxford University Press; 2022. p. 9780192844484.
- 76.World Health Organization. Tackling NCDs. ‘best buys’ and other recommended interventions for the prevention and control of noncommunicable diseases. Geneva: World Health Organization; 2017. https://iris.who.int/handle/10665/259232. Accessed 1 Mar 2025.
- 77.Kilian C, Lemp JM, Llamosas-Falcón L, Carr T, Ye Y, Kerr WC, et al. Reducing alcohol use through alcohol control policies in the general population and population subgroups: a systematic review and meta-analysis. EClinicalMedicine. 2023;59. 10.1016/J.ECLINM.2023.101996. [DOI] [PMC free article] [PubMed]
- 78.Giles L, Mackay D, Richardson E, Lewsey J, Robinson M, Beeston C. Evaluating the impact of minimum unit pricing (MUP) on alcohol sales after 3 years of implementation in scotland: A controlled interrupted time-series study. Addiction. 2024;119(8):1378–86. 10.1111/ADD.16492. [DOI] [PubMed] [Google Scholar]
- 79.O’Donnell A, Anderson P, Jané-Llopis E, Manthey J, Kaner E, Rehm J. Immediate impact of minimum unit pricing on alcohol purchases in scotland: controlled interrupted time series analysis for 2015-18. BMJ. 2019;366:l5274–l. 10.1136/BMJ.L5274. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 80.Public Health Scotland. Evaluating the impact of minimum unit pricing for alcohol in Scotland: Final report. 2023. https://publichealthscotland.scot/publications/evaluating-the-impact-of-minimum-unit-pricing-for-alcohol-in-scotland-a-synthesis-of-the-evidence/. Accessed 1 Apr 2025.
- 81.Anderson P, O’Donnell A, Kaner E, Llopis EJ, Manthey J, Rehm J. Impact of minimum unit pricing on alcohol purchases in Scotland and wales: controlled interrupted time series analyses. Lancet Public Health. 2021;6(8):e557–65. 10.1016/S2468-2667(21)00052-9. [DOI] [PubMed] [Google Scholar]
- 82.Maani N, Collin J, Friel S, Gilmore AB, McCambridge J, Robertson L, et al. The need for a conceptual Understanding of the macro and meso commercial determinants of health inequalities. Eur J Pub Health. 2021;31(4):674–5. 10.1093/EURPUB/CKAB048. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 83.Mialon M. An overview of the commercial determinants of health. Globalization Health. 2020;16(1):74. 10.1186/S12992-020-00607-X. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 84.World Health Organization. Global Action Plan for the Prevention and Control of Noncommunicable Diseases 2013–2020. 2013. https://iris.who.int/bitstream/handle/10665/94384/9789241506236_eng.pdf?sequence=1. Accessed 1 Mar 2025.
- 85.World Health Organization. Commercial determinants of noncommunicable diseases in the WHO European region. Copenhagen: World Health Organization. Regional Office for Europe; 2024. p. 9789289061162. [Google Scholar]
- 86.Angus C, Holmes J, Maheswara R, Gree MA, Meie P, Brennan A. Mapping patterns and trends in the Spatial availability of alcohol using Low-Level geographic data: A case study in England 2003–2013. Int J Environ Res Public Health. 2017;14(4):406. 10.3390/IJERPH14040406. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 87.Shortt NK, Tisch C, Pearce J, Mitchell R, Richardson EA, Hill S, et al. A cross-sectional analysis of the relationship between tobacco and alcohol outlet density and neighbourhood deprivation. BMC Public Health. 2015;15(1):1–9. 10.1186/S12889-015-2321-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 88.Maheswaran R, Green MA, Strong M, Brindley P, Angus C, Holmes J. Alcohol outlet density and alcohol related hospital admissions in england: a National small-area level ecological study. Addiction. 2018;113(11):2051–9. 10.1111/ADD.14285. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 89.Adfree Cities. UNAVOIDABLE IMPACT How outdoor advertising placement relates to health and wealth inequalities. 2024. https://adfreecities.org.uk/wp-content/uploads/2024/03/Unavoidable-Impact-Advertising-Inequality-Full-Report-Adfree-Cities-4th-March-2024-WEB.pdf. Accessed 21 Aug 2025.
- 90.Institute of Alcohol Studies. Outdoor Alcohol Advertising by Area of Deprivation. London. 2024. Accessed 25 Apr 2025.
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets supporting the conclusions of this article are available in the UK Data Service Archive, subject to their end user license agreement: NatCen Social Research, University College London, Department of Epidemiology and Public Health. Health Survey for England. [data series]. 8th Release. UK Data Service, 2024 [Accessed 17 Oct 2024]. Available from: DOI: http://doi.org/10.5255/UKDA-Series-2000021.