Alcohol exposure definitions

In the baseline questionnaire, participants were asked “Are you a non-drinker/teetotaller now?” and “At present, about how many alcoholic drinks do you have each week” for four types of alcohol: Beer, cider or lager (pints); wine (glasses), sherry or fortified wines (glasses) and spirits (singles). Current non-drinkers were defined as those who answered “yes” to being a non-drinker now and did not report consuming beer, wine or spirits at present. Similarly, current drinkers were defined as answering “no” to the question or report drinking at present.

Participants were also asked “Have you ever drunk alcohol in the past?” and two similar questions relating to consumption of the four alcohol types when aged 20 and aged 30. Former drinkers were defined as current non-drinkers who answered “yes” to ever drinking alcohol or reported consuming alcohol aged 20 or 30. Lifelong abstainers were defined as participants who are neither current drinkers nor former drinkers.

Current units and past units were calculated from the questionnaire responses with one unit equal to a half pint of beer, one glass of wine or fortified wine or a single measure of spirits. The capacity of a glass was not specified, but assumed to be 125ml for wine and 50ml for fortified wines. An additional category “occasional”, representing consumption of less than one drink per week, contributed half a unit when ticked for an alcohol type. Heavy current drinkers are defined as participants with >35 current units per week while heavy former drinkers are defined as participants with >35 past units per week. Those with current units greater than zero were divided into four categories: (0,7], (7,14], (14,21] and >21 units per week. Past alcohol consumption is defined as the higher of units reported consumed aged 20 and aged 30.

Other covariates

Participants were also asked details of their current job and their partner's current job. Occupational social class was defined according to the Registrar General's classification [25]. Non-manual occupations were represented by codes 1 (professional), 2 (managerial and technical), 3.1 (non-manual skilled) occupations, while manual occupations were represented by codes 3.2 (manual skilled), 4 (partly skilled), and 5 (unskilled) occupations. Partner's social class was used where available for women and former occupation used where no current occupation was reported [26].

Educational attainment was established using the question “Do you have any of the following qualifications?” followed by a list of common UK qualifications. Participants were categorised according to the highest qualification attained in four groups: those with no formal qualifications; those with formal qualifications usually associated with a school age between 16 ('O' level or equivalent) or 18 years ('A' level or equivalent); and those with degree level qualifications.

Smoking status was derived from two questions each of which could be answered as yes or no: “Have you ever smoked as much as one cigarette a day for as long as a year?” and for those who answered “yes” to the first question “Do you smoke cigarettes now?”

Participants were asked: “Has a doctor ever told you that you have any of the following?” followed by a list of conditions including diabetes, heart attack, stroke and cancer. Personal history of disease was defined by “yes” responses to these four conditions. Trained nurses measured height and weight according to standard protocols at the health examination. Body Mass Index (BMI) was calculated as weight in kilogrammes divided by height squared in square metres.

Ascertainment of hospital usage and mortality through record linkage

Between 1999 and 2009, cohort participants were also linked to hospital records held by the East Norfolk Primary Health Care Trust using their unique National Health Service number. The database contained hospital episode statistics (HES) coded using the International Classification of Disease (ICD), revision 10, for all Norfolk residents wherever they were treated, including hospitals in other areas in the UK. Linking the EPIC-Norfolk cohort to HES records enables a well defined population denominator to explore future hospital usage patterns. Details of the linkage and outcome variables have been previously reported [22].

Time in hospital over the ten year period was calculated using admission and discharge dates from HES. The sum (in days) of one plus (discharge date minus admission date) was used in order that day cases (admission and discharge on the same day) were considered as well as bed days (overnight stays). The number of hospital admissions was also determined from the HES data with contiguous admissions counted as a single admission. Three dichotomous outcome categories were then calculated: 'Any hospital admissions' and '7 or more admissions' using total admissions and '>20 hospital days' using total bed days and day cases. In addition to total hospital usage for any reason, we also explored usage related to conditions that have been associated with alcohol in systematic literature reviews [8,27].

All participants were followed up for mortality by cause by flagging for death at the UK Office for National Statistics (ONS), and trained nosologists coded death certificates using the ICD, revisions 9 and 10.

Statistical analysis

For the analyses presented here, we excluded 625 men and women from the baseline cohort who died before 1999 and excluded 1274 for whom alcohol intake was not known or inconsistent leaving 23,740 individuals. Men and women were examined separately recognising the different alcohol consumption patterns and conditions between the sexes. Logistic regression models were used to examine associations between alcohol intake and hospital usage outcome categories for total admissions, and in exploratory analyses for various diagnostic codes. The terms “beneficial” and “detrimental” used in S3 and S4 Tables were defined by Rehm and colleagues in their systematic reviews of disease burden [8,20] and approximated by the lists of ICD version 10 codes shown. Logistic regression was used rather than survival analysis since the outcomes under examination are the total number of admissions and total bed days and day cases occurring over a fixed period of ten years. The numbers of individuals with missing values for covariates were: 51 BMI, 180 smoking status, 466 social class. Logistic regression was also used to examine the risk of death in alcohol drinkers compared to non-alcohol drinkers over the period under examination. Three sensitivity analyses were conducted: using the random forest non-parametric algorithm for multiple imputation; using the value of 1.8 units per glass of wine instead of 1 unit per glass; admissions limited to those after March 2004. All analyses were performed using the R statistical language (R Foundation for Statistical Computing, Vienna, Austria version 3.4.0 with packages knitr, Gmisc, missForest) and Stata statistical software version 14 (Stata Corporation, College Station, Texas, USA).

Confounding

The frequency and pattern of alcohol use is strongly related to age, sex, education, social class, obesity, and prevalent ill health, all of which are also related to hospital use so confounding is a major issue. However, multivariable regression models adjusting for all these variables hardly changed the findings. In addition, we stratified by main confounders (Table 5) as well as excluding those with known prevalent heart disease, cancer and diabetes, and the results remained consistent in the subgroups. However, measurement of covariates might not be sufficiently accurate to ensure adequate adjustment and we cannot exclude the possibility of residual confounding with known or other unknown factors associated with both alcohol intake and hospital usage, which could either attenuate or strengthen the associations.

Bias

Differential follow-up might have occurred if participants had chosen to use private hospitals instead of NHS hospitals and the alcohol consumption of those participants differed from the study population. Participants in higher social class groups might be higher alcohol consumers and also use private healthcare not recorded in NHS hospital statistics. If this occurred it might attenuate some of the inverse associations observed. However, private health care use is minimal in Norfolk and these results do reflect the use of National Health Service hospitals which is the predominant health care system.

Similarly, differential misclassification in hospital use may be explained by early death rates. Participants who died early from alcohol attributable diseases may have lower hospital usage over the period under examination having not used hospital services for the entire period. This is unlikely as over this time period the risk of death was in fact lower in alcohol drinkers compared to non-alcohol drinkers hazard ratio (OR) 0.67 (95% CI 0.57–0.80) for men and OR 0.66 (95% CI 0.57–0.76) for women. A sensitivity analysis excluding hospital admissions prior to 2004 showed attenuated associations for men which might indicate that prevalent illness could lead to a reduction or cessation in alcohol consumption although this was not apparent in women.

Under-reporting of consumption in this study is likely given the known problems in capturing alcohol intake by questionnaire. Self-reported alcohol consumption in surveys suggest much lower consumption than estimates based on alcohol sales data [29–31]. In the 1998 Australian National Drug Strategy Household Survey, reported intake accounted for only 46.5% of known alcohol sales for the preceding 12-month period. When asked to estimate average consumption, there is a tendency to report a figure closer to median than mean consumption with heavy drinking episodes disregarded. There is also a tendency for past alcohol consumption to be remembered less well than more recent consumption. Questions relating to past consumption are insufficiently sensitive to determine periods of abstaining, binge drinking, patterns of consumption or heavy use. Nevertheless, random measurement errors or systematic underreporting of heavy alcohol consumption would only attenuate the findings observed.

Those who enrol in studies, typically in middle age, represent healthy survivors while those worst affected by alcohol misuse may be less likely to participate. Participants who drink moderately may not be representative of moderate drinkers of similar age in the general population due to differing consumption patterns over the life course [32,33]. It has also been suggested that while high levels of alcohol consumption are associated with harm in all socioeconomic groups, there appears to be a disproportionate level of harm for individuals with low socioeconomic status [34,35]. A meta-analyses that controlled for quality-related study characteristics found that moderate drinking had no net mortality benefit compared with lifetime abstention or occasional drinking [36]. However, in a large study of linked electronic UK health records using recruitment at general practice rather than individual level, moderate drinking was associated with a lower risk of several cardiovascular diseases [37].

Inclusion of former drinkers in the current non-drinkers reference group

The choice of reference group in describing our results may influence interpretation. Non-drinkers comprise heterogeneous subgroups with different characteristics. Former drinkers may have stopped consuming alcohol because of illness, irrespective of whether their illness was caused by drinking. They have been reported to have increased risk for cardiovascular mortality compared to long-term abstainers, a phenomenon described as the “sick-quitter” hypothesis [11]. Lifelong abstainers, ostensibly an ideal reference group having no exposure, may have characteristics that are unusual in the general population [38–40]. Lifelong teetotalism is rare in men (less than 2% of those in the current study) and the reasons for abstaining such as cultural or religious beliefs, may introduce other biases obscuring the results. It has also been noted that there are substantial inconsistencies in self-reports of lifetime abstention. Others have suggested moderate drinkers with no previous history of heavy drinking as a reference group since that is the most commonly observed behaviour and forms the largest group [12,41]. The consumption of alcohol in middle aged men and women tends to decline with age with the largest decline in heavy drinkers but with a reduction across all intake categories.

We opted to use both current non-drinkers and lifelong abstainers as reference groups in the main analyses presented. In the context of hospital usage our objective was to examine the burden on hospital services of cohort participants in relation to current alcohol use rather than pathological processes that may be involved in alcohol and disease associations. To this extent participants' previous history of alcohol consumption was less relevant than the more pragmatic question of their use of services given their current drinking status. However, we have also presented analyses using the alternative reference group of lifetime abstainer in order to explore better this issue. Estimates are less stable given the very small proportion of men who were lifetime non-drinkers. These analyses suggest that in men, the highest hospital usage was observed in former drinkers but with current alcohol drinkers also having higher hospital usage than lifelong abstainers. However, findings in women were not materially different, irrespective of whether lifelong abstainers or former drinkers were used as a reference group. There was no evidence of the “sick quitter” effect found in women affecting the risk of hospitalisation observed in current drinkers.

Findings in context

These results are somewhat unexpected in the light of current beliefs about alcohol intake in the general population and hospital usage. The many diseases related to the high consumption of alcohol would lead us to expect a positive association between hospital usage and alcohol intake. Mortality rates for liver disease have increased four-fold since 1970 with liver disease the third most common cause of premature death in the UK [9]. Obesity related diseases also have a profound impact on hospital services and since alcohol's energy density is second only to fat, a positive association might be expected. However, cardiovascular disease is a predominant reason for hospital admissions, and an inverse association between alcohol intake and cardiovascular disease has been reported in many epidemiological studies [10,11,14,42]. While causality has not been established, plausible biological mechanisms such as the reduction of plaque deposit in arteries, the reduction of blood clot formation and the dissolving of blood clots[18], have supported the reported beneficial associations for ischaemic heart disease (IHD) and diabetes at moderate levels of alcohol intake. Hospitalisations might reflect the balance between positive and negative health effects of alcohol consumption in a particular study population. Most studies based on hospital cases without a population denominator are unable to assess the potential impact of moderate alcohol consumption if associated with lower hospital use.

Strengths of the study

Most studies of hospital use only have data on patients who are hospitalised, that is, cases without denominators so are unable to assess overall risk associated with alcohol consumption in the general population. We were able to examine hospital usage over a defined time period in a clearly defined community based population using a prospective cohort design. Use of record linkage with routinely collected hospital admissions data means that ascertainment is virtually complete as use of private healthcare in Norfolk at this time period was minimal. We have previously reported that age, BMI and smoking status predict future hospital use in this cohort over a ten year period of follow-up [22]. Loss to follow-up is small (approximately 2%) as few study participants have moved away from the area they were recruited.

Study participants are very well characterised and we were able to take into account many potentially confounding variables documented to relate to hospital usage in this population as well as prevalent ill health. Income was not measured in EPIC-Norfolk. However, in the UK national health system, income is not a major determinant of hospital admissions, and education and occupational social class are stronger sociodemographic indicators in this respect than income. EPIC-Norfolk is homogeneous with respect to race and ethnicity with 99% describing themselves as white. The assessment by study participants of their alcohol intake in their 20s and 30s enabled us to differentiate between current non-drinkers and lifelong abstainers.

The measurement of past alcohol consumption allows the separation of non-drinkers into former drinkers and lifelong abstainers.

Limitations in generalisability

Potential selection biases may limit the interpretation of the data since participants were recruited in middle-age and represent survivors who may over-represent resilient and less risky drinkers. Since very few cohort participants reported heavy drinking, a limitation of the study is the inability to examine any possibly deleterious effect of very high consumption. While we did not observe a higher risk of admissions even with the highest alcohol intake categories when comparing current non-drinkers to current drinkers, there were very few people in this study population with very high alcohol consumption levels. Hence the generalisability of these findings to other populations where there are substantially more heavy drinkers may be limited. The use of current non-drinker as reference category must also be considered alongside any interpretation of these results as evidence that the consumption of alcohol may be beneficial but we had very few men who were lifelong abstainers in this cohort.

By using total hospital usage, we were able to assess hospital admissions not just for conditions for which alcohol might increase risk, but also the possible lower service use if alcohol at moderate intake levels were to have the postulated cardioprotective effects. The results presented here reflect hospital usage in a middle aged and older age group and thus we are not able to comment on associations in younger people where binge drinking resulting in acute alcohol poisoning, road traffic and other accidents are a major problem. Nevertheless, older people are by far the greatest users of hospital services and in this older cohort, which was similar to UK national samples in many respects, there was no evidence that current alcohol intake was associated with a higher level of hospital use.