Data

Hospital Episode Statistics (HES) for hospital admissions in England for the 11-year period 2002/03 to 2013/14 (financial year 1st April to 31st March) were supplied by NHS Digital. HES data are routinely collected administrative data that record any hospital activity. Information extracted included the personal characteristics of each patient (age, sex, location of residence), along with up to 20 diagnoses. Data comprised individual ‘episodes’, each of which is a continuous period of care under a single consultant doctor during an admission.

We followed the approach taken by Public Health England, an executive agency of the Department of Health responsible for health promotion and protection, for identifying alcohol-related admissions [25]. Firstly, the data set was cleaned to address some known issues with HES. Only episodes that were finished, and were ‘ordinary’ (non-elective admissions or an elective admission expected to remain in hospital overnight), day case (elective admissions not requiring an overnight stay) or maternity admissions were included. We removed admissions with an age at the beginning of admission outside the range of 0 to 120, or where the sex was not recorded as male or female. We only considered admissions with an English postcode of residence so that our national estimates remained consistent with our estimates by level of deprivation, since the deprivation data were assigned using a patient’s residential address. This produced a small undercount of estimates as some admissions with unknown address or no fixed abode were alcohol-related (see Note 7 in Additional file 1: Table S1), but was necessary to allow for consistency in the data used throughout our results. The percentage of episodes which remained following cleaning over the study period was 89.72% (see Additional file 1: Table S1 for more details).

Alcohol-related conditions were weighted using Public Health England’s age group and sex specific population attributable fractions (PAFs) [25, 26]. PAFs represent the proportion of cases (i.e. hospital admissions) at the population level that might be attributed to an exposure (i.e. alcohol consumption). PAFs were age- and sex-specific, and were adjusted for alcohol consumption [26]. Negative PAFs, which suggest a protective effect of alcohol, were excluded. We identified alcohol-related admissions based on Public Health England’s ‘narrow’ measure. Our approach is defined in the paragraph below.

Each HES episode can contain up to 20 diagnoses, with the first diagnosis being the primary diagnosis (diagnoses were coded using ICD-10 throughout the study period). The narrow measure was calculated based on the primary diagnosis of an admission, with external conditions taken from secondary diagnostic positions 2 to 14 (external conditions, which are environmental causes of injury occurring outside the body, do not feature as the primary diagnosis). During the time period studied, the number of secondary diagnostic positions was expanded from 14 to 20, but we only considered up to 14 so that there was consistency throughout the period. Where there were multiple alcohol-related conditions recorded within an episode, we classified the admission using the condition with the largest PAF. If there were two conditions with the same PAF, we used the one from the lowest diagnostic position (i.e. closest to diagnosis position 1). An admission can comprise more than one episode of care if patients are transferred from the care of one consultant to another. However, 86.7% of all episodes only contained a single episode. We considered only the first episode from each admission.

We used four outcome measure categories of alcohol-related harm based on previous research [25–27]. These were (i) acute conditions wholly attributable to alcohol consumption, (ii) chronic conditions wholly attributable to alcohol consumption, (iii) acute conditions partially attributable to alcohol consumption, and (iv) chronic conditions partially attributable to alcohol consumption. Additional file 2: Table S2 presents which conditions were found in each category (based on Public Health England’s guidance). For acute conditions, we considered only emergency admissions. For chronic conditions, we combined emergency and non-emergency admissions.

In addition, we also examined four specific conditions which were wholly attributable to alcohol consumption, two acute and two chronic. These were the most common specific conditions within the acute and chronic wholly attributable to alcohol categories. The two acute conditions were (i) ‘Acute Intoxication subcategory of Mental and Behavioural Disorders due to use of Alcohol’ (ICD-10 code F10.0), and (ii) ‘Intentional self-poisoning due to alcohol’ (ICD-10 code X65). The two chronic conditions were (i) ‘All other Mental and Behavioural Disorders due to use of Alcohol’ (ICD-10 code F10.1-F10.9), and ‘Alcoholic Liver Disease’ (ICD-10 code K70).

We also calculated ‘non-alcohol’ related admissions to understand how our other measures compared to general trends in admissions. Non-alcohol related admissions were identified as any admission that contained a condition that had no known association to alcohol as defined using the population attributable fractions used previously (i.e. PAF = 0). We chose to exclude any condition with a partial association.

Counts of admissions were calculated for each outcome measure by Lower Super Output Area (LSOA). LSOAs are census areas created to disseminate administrative data and contain an average population size of approximately 1500 people. This allowed analysis by area-based socioeconomic deprivation. Annual mid-year Office for National Statistics (ONS) population estimates for LSOAs by five year age band and sex were used to calculate rates. We also attached the English Indices of Deprivation (IMD). IMD is a multi-dimensional neighbourhood-level indicator of socioeconomic deprivation for LSOAs [28]. We used the ‘income deprivation’ domain to avoid circularity issues since the overall index contains a health domain. Each measure of IMD was divided into categories based on quintiles of its level of income deprivation. IMD scores were assigned to HES records based on the closest year for which IMD scores were available (2004, 2007, and 2010). The 2015 index was not available at the time of analysis.

Statistical analysis

We calculated the crude rate of admissions at the national level and age specific rates for five year age bands. The results were further disaggregated by quintile of deprivation, and we then calculated the directly standardised rate of admission for each quintile category to account for differing age structures between categories (age-specific rates were used for analysing trends by age). 95% confidence intervals were calculated for estimates using Byar’s methodology [25]. All analyses were stratified by sex. We visualised our estimates using line plots and Lexis surfaces. Lexis surfaces allow for the simple visualisation of complex patterns through the plotting of admission rates by age band and year together.

Key results

Our study presents population-level trends in hospital admissions attributable to alcohol by age, sex and level of socioeconomic deprivation. Total admissions attributable to alcohol increased from 201,398 in 2002/03 to 303,716 in 2013/14. While these increases have occurred during a period of overall increases in non-alcohol attributable admissions, the relative increase in alcohol-attributable admissions was larger than increases in overall, emergency and non-emergency non-alcohol attributable admissions. The largest relative increase in admission type was for acute admissions wholly attributable to alcohol which doubled. Admission rates were consistently higher for males compared to females across all of our outcome measures other than ‘Intentional self-poisoning due to alcohol’ which has more than doubled for both males and females over our study period.

Chronic admissions wholly attributable to alcohol were concentrated in middle age, whereas for acute admissions wholly attributable to alcohol there was a clear bimodal distribution (although admission rates in adolescents/young adults declined over the period). Both acute and chronic partially attributable admissions were higher in the oldest ages. However, the concentrated in middle aged males resulted that in 2013/14 men aged between 35 and 54 consisted 27.4% of the male population but accounted for 34% of all male admissions attributable to alcohol, 44.5% of all male acute admissions wholly attributable to alcohol and 56.4% of all male chronic admissions wholly attributable to alcohol.

Finally, we observe a social gradient with admission rates higher in the most deprived areas compared to the least deprived areas. Social inequalities were wider for males compared to females, and they became wider over the study period particularly for acute conditions wholly attributable to alcohol. ‘Intentional self-poisoning due to alcohol’ was the only condition where admission rates were higher for females across each deprivation category.

Limitations

While HES is routinely collected data on all hospital admissions across England, there are multiple issues with the data which limit the interpretation of our observations. We do not include A&E attendances in our data (only if an individual was admitted) which may lead to an undercount of the harms associated with alcohol. There may be individuals who do not seek medical care and if over the period these individuals have changed their behaviours through seeking care it may affect our observed trends. The trends we observe may be partly explained by changes over time in the quality of HES. Changes in coding practice over time could partly influence our observed trends. In 2003/04, ‘Payment by Results’ (PbR) was introduced whereby hospitals were paid for fully reporting all treatments provided to patients [29, 30]. Increasing admission rates may partly reflect such financial incentives which capture more information which previously was missed. While we only use the primary diagnosis for most alcohol-related conditions, PbR has been associated with increasing depth of secondary diagnoses which may influence trends for external conditions. These changes have also occurred alongside rising levels of overall admissions which could influence our trends ([29]; also see Additional file 1: Table S1). However, increases in alcohol-attributable admissions were larger than increases in admissions for non-alcohol-attributable admissions (including split by emergency and non-emergency admissions).

Missing data was also an issue particularly as it could not always be ascertained that it was not missing randomly (i.e. see Additional file 1: Table S1 Note 7) which may introduce bias into our results (although the level of bias is likely to be small). There have also been increasing usage of ‘bucket codes’, which are diagnoses whereby the cause is unknown and therefore represent missing data, for emergency admissions over the same period (Additional file 3: Figure S1). The use of bucket codes may produce an undercount of admission rates. Diagnostic accuracy is also important with Burns et al. [30] estimating that 80% of primary diagnoses are correctly coded.

The partially attributable conditions used in calculating our measures are based on PAFs that assume a causal relationship between alcohol and each condition [26]. Whilst PAFs are useful for estimating population level patterns, we do not know which individual admissions were due to alcohol. Incorrectly attributing non-alcohol associated admissions to our partially-attributable estimates may introduce error into our results. This is evident in our age-specific analyses for partially attributable conditions that suggest greater alcohol-related harm in the elderly, which is mostly driven by the higher rates of some partially-attributable conditions at old age despite their small PAFs (e.g. falls). The issue is also problematic when analysing trends since factors unrelated to alcohol may be driving changes over time [31]. We therefore suggest interpreting our results for conditions partially attributable to alcohol cautiously.

We used the narrow measure of alcohol-related admissions, which only consider the primary diagnosis to ascertain admissions for non-external conditions [25]. While this approach may underestimate alcohol-related admissions, particularly wholly attributable conditions found in the secondary diagnostic positions that are not evident in the primary diagnosis, the measure is less sensitive to changes in coding practice over time (e.g. increased usage of secondary diagnoses over the period). We also did not investigate trends in case severity of admissions, which may again lead to an underestimate of alcohol-related admissions.

Whilst we focussed on hospital admissions as our measure of alcohol-related harm, our estimates included multiple repeat admissions of individuals which may have overestimated the scale of the problem if we are interested in understanding individual-level patterns. However, admissions represent the actual demand for hospitals and therefore are important for understanding the total pressure on health services.

Interpretation

Increasing alcohol-attributable admissions have been noted for England both overall [19] and for specific conditions [13, 14, 20, 21]. Our study builds on this small evidence-base by presenting a detailed investigation of how trends in alcohol-attributable admissions vary by sex, age and socioeconomic deprivation. Our results also corroborate with experiences in other countries [32–35].

The increasing burden of alcohol-attributable admissions placed on the NHS, combined with increasing pressures on scarce healthcare resources; suggest the need for increased focus on preventive measures. Our results suggest that the optimum intervention may differ by population sub-group, depending on age, sex and level of deprivation. The scale of the problem combined with the high costs involved in treating alcohol-attributable harms (as well as the wider costs to society) suggest that population-level preventive measures could be cost effective [10].

Recent declines in total alcohol consumption experienced at the national level since 2005 have not translated into downward trends in alcohol-related hospital admissions [5, 10]. Similar diverging trends of consumption and hospital admissions have been reported elsewhere [31]. There are multiple explanations for this. There may be a time lag effect and the effects of declining consumption have yet to result in fewer admissions. This would be a stronger explanation for trends in chronic conditions compared to acute. Whilst declining consumption appears to be driven by declines amongst younger adults, we only see some evidence for this in admissions due to ‘Acute Intoxication subcategory of Mental and Behavioural Disorders due to use of Alcohol’ [5]. It may also be that under-reporting bias in consumption is increasing over time [36]. Obtaining reliable self-reported survey information on alcohol consumption is difficult. Finally, it may be that consumption is declining faster in those individuals at lower risk of admission, and in those at high risk of admission consumption may be increasing [10]. Trends in alcohol-related mortality rates have, however, followed consumption trends more closely [37]. Understanding the disparity between morbidity and mortality trends, and how they relate to trends in consumption will be an important direction for future research.

While admission rates were higher for males for most conditions and ages, we found higher rates of ‘Intentional self-poisoning due to alcohol’ in females, particularly younger females. This finding follows increases in overall levels in self-poisoning and self-harm amongst females in England [38, 39]. Increased ‘felt’ pressures on younger women may be leading to greater self-harm through alcohol [40–42]. Our figures may also underestimate the scale of the issue since not all individuals who self-harm will be admitted to hospital [43]. However, it also follows trends that self-harm cases are now more likely to be admitted to hospital [44]. Whilst alcoholic liver disease has received much attention in the literature [10, 13, 20, 21, 45], there is clear need for greater investigation of the wider health-related harms attributable to alcohol.

Our results demonstrate wide social inequalities in alcohol-attributable admissions, supporting evidence from elsewhere [4, 8, 9, 11, 12]. There is a clear social gradient in admissions for each outcome measure, and inequalities were widest for middle aged males. The social gradient exists despite similar levels of consumption in deprived and affluent areas; termed the ‘alcohol harm paradox’. Bellis and colleagues offer two possible explanations for the paradox [11]. Firstly, individuals from deprived areas tend to engage in multiple risky behaviours (e.g. smoking or unhealthy diets), which interact with alcohol consumption to result in a greater risk of hospitalisation. Secondly, individuals in deprived areas engage in different drinking behaviours (e.g. binge drinking) despite similar overall levels of consumption. Siegler and colleagues also suggest that social inequalities in harms may be influenced by ‘social drift’, whereby heavy drinkers move down the social gradient due to the harms of their consumption, although there is less evidence for this [8]. Given that many of these alcohol-attributable admissions are preventable, this represents a key policy area to reduce overall social inequalities in health.