Ethics statement

All analyses were approved by the Regional Ethical Review Board of Lund University (no. 2008/409 and later amendments).

Sample

We used several Swedish nationwide registers, linked using the unique 10-digit identification number assigned at birth or immigration to all Swedish residents. The identification number was replaced by a serial number to ensure anonymity. The registry resources are described in S1 Methods, and included the population register, multi-generation register, multiple health care registers, and the Longitudinal Integrated Database for Health Insurance and Labour Market Statistics (LISA), which included information on income and education. To maximize registry coverage based on preliminary analyses, we included females and males, born between 1950 and 1970, and residing in Sweden at age 40, without a prior AMC registration. Due to missing data on education, N = 916 females and N = 845 males from the cohort were excluded; due to missing data on marital status, N = 15 females and N = 11 males were excluded.

Measures

Our outcome variable, AMC, was defined from Swedish medical registers by the following ICD codes, described further in S1 Methods: ICD8: 571.0, ICD9: 357F, 425F, 535D, 571A, 571B, 571C, 571D; and ICD10: E24.4, G31.2, G62.1, G72.1, I42.6, K29.2, K700, K701, K702, K703, K704, K709, K85.2, K86.0, and O35.4. Individuals with AMC registrations prior to age 40 were excluded.

The primary independent variables of interest were education and familial income. Income was assessed at age 40 and categorized based on the income quartiles for the working-age population in Sweden (ages 20 to 65), although analyses included both employed and unemployed individuals. Educational attainment was categorized into low (compulsory school only), mid (upper secondary school), and high (university level), based on the subject’s highest level of school completion. These variables allow us to jointly account for, and distinguish the impact of, indices of individual-level SEP (educational level) and access to resources at the family level (familial income). An individual could have low educational attainment, and/or no personal income, but still have access to resources through a spouse’s income.

Sociodemographic covariates included year of birth, marital status, and region of origin. Marital status was assessed at age 40 and obtained from the total population register. The categories were: unmarried, married, divorced, or widowed. Region of origin was defined by country of birth and categorized into: Finland, Western Countries, Eastern Europe, Latin America and the Caribbean, Middle East and Northern Africa, Africa (excluding Northern Africa), and Asia (excluding Middle East) and Oceania. Finland was examined separately because they represent one of the largest immigrant groups in Sweden and because they have higher AUD rates.

From Swedish medical registers, AUD was defined by the following ICD codes: ICD8: 291, 303; ICD9: 305A, 291, 303; and ICD 10: ICD-10 codes: F10.1, F10.2, F10.3, F10.4, F10.5, F10.6, F10.7, F10.8, and F10.9. Additionally, AUD registrations were identified using the Prescribed Drug Register by retrieving “Drugs used in alcohol dependence” (Anatomical Therapeutic Chemical [ATC] Classification System code N07BB); this includes disulfiram (N07BB01), acamprosate (N07BB03), naltrexone (N07BB04), or nalmefene (N07BB05). Finally, we identified AUD cases as those convicted for or suspected of at least 2 alcohol-related crimes according to law 1951:649, paragraph 4 and 4A and law 1994:1009, Chapter 20, paragraphs 4 and 5 from the Swedish Crime Register, and code 3005 and 3201 in the Suspicion register. Most of these crime types include driving or operating a boat while intoxicated. Among criminal AUD registrations, 99.7% were related to drunk driving; however, only 18.3% of AUD registrations are detected using only in the criminal registries.

Internalizing disorder (ID) was defined using the ICD codes for major depression (MD); ICD-9: 296B, 298A, 300E; and ICD-10: F32 and F33, and anxiety disorder (AD); ICD-10: F41. Externalizing disorder (ED) was defined as criminal behavior including white collar crime, property crime, or violent crime as described in detail elsewhere [18]; or drug use disorder (DUD). DUD was identified from medical registers using codes from ICD-9: 292, 304, and 305C - 305I; ICD-10: F11-F16 and from the Prescribed Drug Register if a person had retrieved (on average) more than 4 defined daily doses a day for 12 months of a drug to treat drug dependence (excluding those suffering from cancer), defined by the following ATC codes: N02A, N05C, N05BA, or N07BC. We further defined DUD from the Suspicion Register by codes 3070, 5010, 5011, and 5012; and the Crime Register if convicted by law 1968:64, paragraph 1, point 6 or and law 1951:649, paragraph 4, subsection 2 and paragraph 4A, subsection 2.

Finally, aggregate genetic liability for AUD, which was used in a series of secondary analyses, was operationalized using the family genetic risk score for AUD (FGRS_AUD), which assesses genetic risk by an examination of first through fifth degree relatives, correcting for cohabitation effects which could make relatives more similar in their AUD risk. This measure is a relative measure which is operationalized as a continuous, standardized indicator of aggregate genetic risk, so that 1 unit corresponds to 1 standard deviation in the population. The details of this measure have been described in detail elsewhere [19]; FGRS have been validated in subsequent research [20].

Statistical analyses

These analyses were not prospectively described in an analysis plan. S1 Methods provides information on the timing of analyses. Cox proportional hazard models were utilized to estimate the time to AMC, censoring at end of follow-up (December 2018), death, or emigration. Models were stratified by sex to facilitate direct comparisons and enhance data transparency [21].

We conducted a series of preliminary analyses to decide whether time-varying coefficients were necessary for the socioeconomic predictors. This was investigated by including a linear interaction between follow-up time and education or income. If the interaction was significant and the modification of the magnitude of the hazard ratio (HR) was potentially meaningful, we elected to allow for a time-dependent coefficient in the final models.

We then pursued model building to estimate the association between education and/or income with AMC and investigated whether observed associations were robust to the inclusion of ID, ED, and AUD, which were included as time-varying variables. These models included birth year, marital status, and region of birth as covariates. In Models 1A and 1B, we tested the association between education or income, respectively, with AMC, otherwise including only the sociodemographic covariates. In Model 2, we jointly estimated these associations, including both education and income. In Model 3, we added registrations for ID and ED, as these could act as important mediators between education/income and AMC. Finally, in Model 4, to provide insight into whether education/income is associated with risk for AMC above and beyond variance that could be accounted for by the association between socioeconomic indicators and AUD, we also included AUD as a time-varying variable.

In a secondary set of analyses, we included genetic risk of AUD, operationalized as FGRS_AUD, to understand how genetic liability to AUD might influence the associations with the investigated risk factors. FGRS_AUD is estimated with greater precision for Swedish born subjects with 2 Swedish-born parents, as these individuals have more family members whose data is available in the Swedish registries. Therefore, this series of analyses included only Swedish-born individuals with 2 Swedish-born parents. We ran the same models as above, excluding region of birth. These models are referred to as Model S1A, S1B, S2, S3, and S4, respectively.

We ran an additional model, Model S5, including an interaction between FGRS_AUD and education and between FGRS_AUD and income, to test whether genetic liability to AUD moderated the associations between education/income and AMC. To be able to interpret the results of the interactions from these multiplicative models on an additive scale, which is often of greater relevance in a public health context [22], we assessed additivity by estimating the relative excess risk due to interaction (RERI) and synergy index (S) [23].

This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (S1 STROBE Checklist).

Descriptive statistics

Tables 1 and 2 provide descriptions of the female and male cohorts, respectively, in the sample of N = 2,359,338 individuals. In the female cohort, N = 4,253 individuals had an AMC registration, corresponding to 0.37% of the sample. Among males, N = 11,183 (0.93%) had an AMC registration. For both sexes, the distribution of sociodemographic variables, including the primary predictors of interest, varied significantly across the groups without and with AMC. For example, individuals with AMC were less likely to be married than their counterparts without AMC (females: 42.04% versus 56.42%; males: 32.59% versus 50.75%). In addition, among individuals residing in Sweden who were born elsewhere, AMC registrations were disproportionately common only among those from Finland: 2.81% of females and 2.06% of males in the cohort were Finnish, but Finns represented 6.51% of females and 5.54% of males with AMC. In contrast, immigrants from Asia, Africa, and the Middle East had disproportionately few AMC registrations. Incidences of AMC for each variable, per 10,000 person years, are provided in the S1 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive information on the female cohort, born 1950–1970, with observation beginning at age 40.

https://doi.org/10.1371/journal.pmed.1004359.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Descriptive information on the male cohort, born 1950–1970, with observation beginning at age 40.

https://doi.org/10.1371/journal.pmed.1004359.t002

The polychoric correlation between income quartile and education level for females was 0.185 (SE-0.001) and for males was 0.198 (SE = 0.001), suggesting that while these socioeconomic measures are related, they also provide nonoverlapping information. The S2 and S3 Tables provide cross-tabulations for females and males, respectively, of income quartile and education level.

Crude models

In crude models unadjusted for any covariates, females with lower educational attainment were at higher risk of AMC (low versus high: HR = 4.48, 95% CIs = 3.66, 5.47, Chi-square p < 0.001; mid versus high: HR = 1.98, 95% CIs = 1.66, 2.34, p < 0.001); the same was true for males, at lower effect sizes (low versus high: HR = 2.27, 95% CIs = 2.01, 2.55, p < 0.001; mid versus high: HR = 1.46, 95% CIs = 1.32, 2.61, p < 0.001). Likewise, females in lower income quartiles had higher risk of AMC (income quartile 1 versus 4: HR = 11.52 [9.30, 14.27], p < 0.001; income quartile 2 versus 4: HR = 3.90 [3.12, 4.87], p < 0.001; income quartile 3 versus 4: HR = 1.60 [1.26, 2.03], p = 0.001). Results for males were comparable (income quartile 1 versus 4: HR = 11.95 [10.41, 13.72], p < 0.001; income quartile 2 versus 4: HR = 3.32 [2.86, 3.86], p < 0.001; income quartile 3 versus 4: HR = 1.63 [1.39, 1.91], p < 0.001).

Multivariable models

We next estimated the association between educational attainment (in Model 1A) or income at age 40 (in Model 1B) with risk of AMC, adjusting only for sociodemographic covariates. We provide snapshots of the HRs for education and income at 4 time points: time 0 (age 40), after 5 years of observation (age 45), after 10 years of observation (age 50), and after 15 years (age 55 or older). As shown in the Fig 1, in which estimates from Models 1A and 1B are superimposed for ease of illustration, individuals with lower levels of education or income at age 40 were at substantially increased risk for AMC. Complete results, including for all covariates, are provided in S4 and S5 Tables.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. HRs between education level and/or income with risk of AMCs, across time, from Models 1A/1B–Model 4.

Models 1A and 1B include education level or income, respectively, as the primary predictors of interest, and are adjusted for birth year, marital status, and region of origin. Model 2 includes both education and income level with the aforementioned covariates. Model 3 further adjusts for internalizing and externalizing registrations. Model 4 further adjusts for AUD registrations. Results for females are presented in the top panel and for males in the bottom panel. The horizontal dashed gray line at an HR of 1 represents the null hypothesis. AMC, alcohol-related medical condition; AUD, alcohol use disorder; HR, hazard ratio.

https://doi.org/10.1371/journal.pmed.1004359.g001

Among females (Fig 1 and S4 Table), having the lowest level of education was associated with a nearly 5-fold increase in risk of AMC in Model 1A (HR = 4.67; 95% CI 3.82, 5.72; p < 0.001). Relative to individuals with the highest level of education, those with mid- or low-level education at age 40 were still at increased risk of AMC 15 years later (HRs = 1.68 [95% CIs 1.56, 1.81], p < 0.001; and 2.84 [95% CIs 2.60, 3.11), p < 0.001, respectively). HRs for the 3 lower quartiles of income were also elevated (Model 1B; Fig 1 and S5 Table). The magnitude of effect declined across observation time, most notably for the lowest quartile (from HR = 11.26 at observation onset to HR = 3.86, 15 years later), and was more pronounced for income than for education. We observed a general dose-response pattern for levels of education/income and for time elapsed since the start of observation; i.e., the strongest effects were observed at the start of observation, and at the lowest education and income levels.

HRs were lower, but still consistently >1, for males (Fig 1 and S4 Table). In males, being in the lowest education category (Model 1A) was associated with a 2-fold increase in risk of AMC (HR = 2.04, 95% CIs 1.81, 2.30; p < 0.001) at the start of observation. Over time, these estimates declined, but remained above 1 (HRs = 1.26 to 1.57 for those with mid-level and low-level education). Similar to females, the effect size for being in the lowest income quartile was high: HR = 10.21 (95% CIs 8.86, 11.77; p < 0.001; Model 1B; Fig 1 and S5 Table). Individuals in the 2 middle quartiles were at lower, but still elevated, risk of AMC. Also consistent with females, the HRs for education were more stable over time than those for income.

In Model 2 (S6 Table), education and income were jointly included as predictors. The HRs were attenuated, but only slightly, indicating that these measures of socioeconomic position account for independent components of AMC risk. We next adjusted for psychiatric illness using time-varying covariates, in Model 3 (S7 Table). Among females, HRs for ID and ED were associated with an approximately 2.6- to 3.7-fold increase in AMC risk. We observed a slight attenuation in the effect sizes of education and income from Model 2. The lowest income quartile remained at high risk of AMC at the start of observation (HR = 7.80, 95% CIs 6.24, 9.75; p < 0.001), declining to HR = 2.88, 15 years later. A similar pattern was observed for males, among whom HRs were approximately 2.1 to 3.1 for ID and ED. For males, having a mid- or low-level education was associated with a 1.21- to 1.63-fold increase in AMC risk. Being in the lowest income quartile was still strongly associated with AMC (HR = 7.76, 95% CIs 6.73, 8.95; p < 0.001 at start of observation), though the effects were attenuated as time progressed and being in the third quartile 15 years after the onset of observation was not significantly associated with higher AMC risk.

Finally, in Model 4, we adjusted for AUD as a time-varying covariate. AUD was prominently associated with AMC risk: the associated HRs were HR = 34.00 (95% CIs 31.73, 36.42; p < 0.001) for females and HR = 10.71 (95% CIs 10.28, 11.15; p < 0.001) for males (S8 Table). For both sexes, the inclusion of AUD led to lower HRs for education and income relative to previous models. However, with the exception of the estimate for individuals in the third income quartile, 15 years after the start of observation, all estimates remained significantly >1. We continued to observe a dose-response pattern for income quartiles and as a function of time elapsed since onset of observation in most cases, the exception being that the effect of having mid-level education at age 40 was relatively stable across the observation period. This pattern was clearer among males.

Secondary analysis

We pursued a secondary set of analyses that corresponded to the original models, replacing region of origin with FGRS_AUD, as described in the Methods. In Models S1A/S1B through Model S4, the effect of FGRS_AUD ranged from HR = 1.17 to 1.39 for females and HR = 1.21 to 1.36 for males (S9–S13 Tables) per standard deviation increase. The HRs for education level and income did not differ substantially from the corresponding HRs from the original Models 1 to 4. We added a model with an interaction term between education/income and FGRS_AUD, Model S5 to evaluate the possibility that level of aggregate genetic liability to AUD leads to differential susceptibility to the risky effects of lower education or income. The main effects of all predictors and covariates are reported in S14 Table, and we report the RERI and S (estimates of interactions on the additive scale) in S15 Table. For females, there was no evidence that FGRS_AUD moderated the association between education level and AMC risk. However, RERI and S estimates were consistent with an interaction between FGRS_AUD and the first and second income quartiles, such that individuals at higher genetic liability to AUD were more susceptible to the adverse impact of lower income on risk for AMC. Similarly, in males, we observed no interaction between FGRS_AUD and education level, but the association between income and AMC was significantly moderated by FGRS_AUD.