Determinants of harmful drinking in old age: a synopsis of the literature

A recent literature review of alcohol use and alcohol-use disorders among older adults in India reports significant statistical associations between alcohol consumption in old age and higher educational status, younger age, better health status, lower chronic morbidity, employment status, socioeconomic status, auditory/locomotor impairment and asthma.18

Some of these variables (eg, age, income, education and gender) have been consistently found to be associated with risk of harmful drinking and alcohol abuseⁱ in old age across the literature at least since the late 1980s,22 although some studies are based on such small samples23 that low statistical power may have affected the SEs and, consequently, the significance of the results. A study based on a sample large enough for statistical purposes found that age is not a significant predictor.24 Two recent papers also report significant findings among alcohol consumption and these four variables, as well as with other covariates: significant associations have been found between alcohol consumption and education, age, subjective health status and gender among the older population in Sweden;25 in turn, the following list of characteristics were reported to be associated with unhealthy drinking among older people in Belgium: age, gender, social contacts, education, health status and socioeconomic status.26

Similarly, the literature on transitions has identified some recurring associations, such as declining risk with advanced age.27 Using data from the USA,28 it was found that increasing alcohol consumption among the older population was more likely among the more affluent, better educated, whites, males, unmarried, less religious and those in better self-reported health. Furthermore, a longitudinal study with US data on people aged 53 and 64 years found that changes in drinking categories were associated with gender, health, education, adolescent IQ, income, lifetime history of alcohol-related problems, religious service attendance, depression, debt and changes in employment.29 A study into changes in alcohol consumption among people aged 55–76 in the USA found that they were associated with gender, ethnicity, smoking, use of substances to reduce tension and social connectedness (and whether their contacts approved of heavy drinking or not).30 More recently, a 10-year study of alcohol use transitions among men aged between 50 and 65 in the USA reported that the different trajectories of risk were associated with age, education, smoking, binge drinking, depression, pain and self-reported health.31

Some culturally relevant aspects have also been identified: for example, a study on change in drinking consumption among older people in Japan found that apart from gender, education, age and depression, a related factor is employment status: consumption tends to drop significantly with retirement, which the authors ascribed to the role of alcohol consumption in Japanese work culture.32 This paper used data from England where, along with the rest of the UK, drinking patterns are shifting towards a Nordic-European pattern, ‘characterised by non-daily drinking, irregular heavy and very heavy drinking episodes (such as during weekends and at festivities) and a higher level of acceptance of drunkenness in public’ (p. 11, ref. 33). Therefore, consideration needs to be given to its results as applicable to a particular cultural milieu with evolving drinking patterns.

Table 1 presents a summary of the main findings in the literature.

Data

This paper is based on data from the English Longitudinal Study of Ageing (ELSA), a longitudinal multidisciplinary data study from a representative sample of men and women aged 50 years and over living in private households in England.55 To study the determinants of harmful drinking, we used the latest wave (wave 5) of ELSA corresponding to the years 2010–2011 (sample size: 9251) and, for the transitions analysis, the last two waves (wave 4 corresponds to the years 2008–2009).

We defined risk of harmful drinking following the guidelines set out by the National Institute for Health and Care Excellence (NICE). NICE has defined the following levels of risk of harmful drinking:

• Lower risk drinking: ≤21 units per week (adult men) or ≤14 units per week (adult women).

• Increasing-risk drinking: 22≤50 units per week (adult men) or 15≤35 units per week (adult women).

• Higher risk drinking: >50 alcohol units per week (adult men) or >35 units per week (adult women).

This definition is part of the classification adopted by NICE to set guidelines and public health guidance in England. See also.56 It is consistent with a 1995 report by the Department of Health on ‘sensible’ drinking,57 which had set daily but not weekly guidelines. Although it is not without problems, see58 ,59 for a discussion. Several hundred estimation measures and instruments, primarily designed as screening tools, are available (the most widely used include the Severity of Alcohol Dependence Questionnaire; the Alcohol Use Disorders Identification Test, the CAGE test, the Michigan Alcoholism Screening Test, the FAST alcohol screening test, etc).60 Furthermore, we carried out a sensitivity analysis and found that relatively minor changes in the cut-off points for higher risk drinking altered the significance of our regression results (see online supplementary annex 1).ⁱⁱ

Two additional considerations regarding the definitions of risk by NICE merit a comment here: the use of a weekly, instead of daily, guideline may mask episodic heavy drinking—a non-negligible and increasing issue among middle-aged people in England61 ,62—and the lack of specific guidelines for older people (eg, for people aged 65 or older) is problematic in the light of evidence on physiological and metabolic changes in alcohol absorption and effects associated with individual ageing,63 which led the Royal College of Psychiatrists to recommend reduced thresholds for older people.64

However, we have used this definition of risk given that this paper is based on English data and that one of the objectives of the paper is to inform and influence policymakers and practitioners in England who have to follow NICE guidelines.

The literature on alcohol abuse in old age distinguishes between people who consume low quantities of alcohol and abstainers.65–67 Our data do not allow us to make such a distinction, but we have separated those respondents who said they had not drunk any alcohol at all during the previous week from those who did. Therefore, in this paper, we have included a fourth risk category which we have termed the ‘abstainers’, but we emphasise that the individuals in this category may not necessarily be teetotallers but consistent infrequent drinkers,31 and therefore there may be some overlap between this group and the lower risk drinkers. Nevertheless, since in our models we focused on higher risk drinkers, this distinction between abstainers and lower risk drinkers is inconsequential for our results.

The literature has also introduced a distinction between early-onset and late-onset cases,68–70 as well as that between having a history of problem drinking or not.28 ,52 However, the data do not allow us to distinguish between early and late onsets or about individual lifetime alcohol consumption patterns, so the paper considers current abuse risk and changes across risk categories in relation to covariates measured at baseline with a disregard of alcohol ingestion patterns at early ages.

Finally, the literature that focuses on the relation between stressful life events and changes in alcohol consumption in old age distinguishes between events with short-term effects and those with long-term implications.71 In order to discern between short-term and long-term effects of life events, we would need data that covered longer than 2 years as in this paper; furthermore, we would also need data from years before the events took place to account for simultaneous causality and unobserved individual endogeneity;72 hence, we are not going to investigate whether this distinction holds for our sample.

ELSA contains questions about weekly consumption of spirits, wine and beer. In order to use NICE guidelines, we used the conversion tables in the ‘alcohol unit calculator’ produced by the National Health Service (NHS) to help find out how many units there are in the most popular glass sizes and measures in the country.ⁱⁱⁱ According to this tool, 1 glass of wine is equivalent to 2.1 units, 1 pint of beer to 2.8 units and 1 measure of spirits to 1 unit.

Following the suggestions by one of the reviewers, we estimated the effect on our results of changing the conversion ratios for wine and beer. Thus, we calculated consumption according to the conversion measures introduced in 2007 to the General Lifestyle Survey (GLS):^iv one pint of a normal strength beer is equivalent to 2 units, a 175 mL glass of wine is equivalent to 2 units and a 250 mL glass of wine is equivalent to 3 units. Furthermore, we also applied the conversion tables in the drinkaware website (http://www.drinkaware.co.uk): 1 glass of wine, equivalent to 3 units, and 1 pint of beer, equal to 3 units.

Table 2, the starting point for the transitions analysis, presents the data on risk for those respondents surveyed in both waves. It also presents the transitions using the two other alternative conversion measures;–the widest variations are found using the drinkaware definitions.

We included the following variables in our models:

Age

Income: equivalised household net income per week (truncated at 1 penny). The equivalisation was performed by the OECD equivalence scale. Income is imputed net of taxes, and there may be records with negative incomes due to, for example, self-employment losses. We set any negative incomes to one penny, but not exactly to zero as in,73 to allow the logarithmic transformation.

Education: we categorised this variable into four levels:

• No Qualifications

• NVQ1/CSE other grade equivalent

• NVQ2/NVQ3/GCE A/GCE O Level or equivalent and Foreign

• NVQ4/NVQ5/Degree or Higher education below degree

Smoking: number of cigarettes smoked per day

Physical activity: we categorised this variable into four levels:

• Sedentary

• Low

• Moderate

• High

Depression: whether the respondent felt depressed much of the time during the past week

Loneliness: four-item scale based on the 20-item Revised UCLA loneliness scale,74

• How often do you feel you lack companionship?

• How often do you feel isolated from others?

• How often do you feel left out?

• How often do you feel in tune with the people around you?

(For the first three questions, categorised as lonely if responded ‘some of the time’ or ‘often’; for the last question, if responded ‘hardly ever’ or ‘never’).

Self-reported health: categorised into Poor, Fair, Good, Very Good, Excellent

Ethnicity: whether respondent is white (=0) or not (=1)

Gender: Female=0; Male=1

Marital Status: three categories:

• Single, legally separated, divorced

• Married, first and only marriage, remarried, civil partner

• Widowed

Caring responsibilities: whether respondent looked after anyone in the past week (=1) or not (=0)

Children in household: total number of children inside the household

Religion: how important religion is in the respondent’s daily life. Four levels:

• Very important

• Somewhat important

• Not very important

• Not at all important

We excluded this variable from the transitions section, because we detected potential errors in the categorisation of the questions about religion in wave 4.

Economic Activity: Three categories:

• Employed

• Inactive (including unemployed, due to the low number of cases)

• Retired (including semiretired)

Social detachment: if detached in three or more of the following four domains. Source:75

• Civic participation: Not a member of a political party, a trade union, an environmental group, a tenants’ or neighbourhood group or neighbourhood watch, a church or religious group, or a charitable association; and did not do voluntary work at least once in the past year.

• Leisure activities: Not a member of an education, arts or music group or evening class, a social club, a sports club, gym or exercise class, or of another organisation, club or society.

• Cultural engagement: neither went to a cinema, an art gallery, a museum, a theatre or a concert nor to an opera performance at least once in the past year.

• Social networks: Do not have any friends, children or other immediate family or if they have friends, children or other immediate family, have contact with all of them (meeting, phoning or writing) less than once a week.

Healthy diet: whether respondent consumes five or more portions of vegetables (excluding potatoes) or fruit a day

Table 3 presents the descriptive statistics for each variable for wave 5.

For wave 5, table 4 presents the distribution of the sample by risk category and gender and figure 1 presents the distribution of consumption of alcohol units (excluding abstainers) by gender. We notice the largest gender discrepancy among abstainers: 64% of abstainers are women; besides, 50% of women are abstainers (against 34% of men).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Distribution of respondents by number of alcohol units consumed.

Both table 4 and figure 1 suggest that we should run separate models for women and men, further confirmed by a χ² test of independence applied to table 4: there is a significant statistical association between risk and gender (χ²=259.07, df=3, p value=0).

For the transition analysis, we used a balanced panel, that is, only respondents in both waves were included. This could in principle introduce bias if dropping out of the survey was to any extent contingent on drinking levels. Therefore, we checked for a significant association between attrition and consumption levels^v. To this purpose, we fitted an OLS regression with dropping out as the outcome variable on age (and age squared), gender, smoking, health status and alcohol consumption units. Alcohol consumption units are not significant. Further, we compared the statistical distributions of alcohol consumption units among respondents in wave 4 who took part in the survey in wave 5 and those who did not, for men and women separately, by means of the consistent univariate entropy density equality test bootstrapped with 100 replications.^vi Again, we failed to find an association between alcohol consumption in wave 4 and attrition in wave 5.^vii Therefore, our results would not suffer from ‘sick quitter’ bias. It is worth noting that studies of attrition by alcohol consumption patterns report inconsistent findings across waves. For example,77 the odds for dropping out at one wave was higher among respondents with high-alcohol intake, but that abstainers had increased odds for dropping out at other waves. Furthermore, a literature review78 failed to find any association between attrition and alcohol consumption levels.

Statistical methods

To identify variables associated with the probability of being or not being in the higher risk category, we used a logistic regression model. Given the small proportion of respondents in this category, we applied the Firth correction to account for any bias introduced by zero inflation.

We included age squared as a term in the regression models to account for any non-linearity in the association between age and risk.

Heavy drinking in old age is associated with short-term mortality;12 ,15 ,79–83 hence, the sample could suffer from an over-representation of moderate drinkers at higher ages as heavy drinkers might have died earlier. Therefore, we checked for this possible selection bias: given that this is a cross-sectional study and that the data do not include any information about past drinking behaviour, we ran models with age truncated at 70. We compared the results with those from the full sample but failed to find any significant differences, so preliminarily we can rule out the presence of selection bias in our sample. Comparing logistic models fitted on different groups or samples is not straightforward, and there is no one accepted method. We applied the heterogeneous choice model procedure,84 for which we fitted heteroscedastic binary response models via maximum likelihood and ran likelihood tests to compare both models (results can be requested from the author).

With regard to the transition analysis, we carried out a Markov chain model assuming continuous and homogeneous time fitted by maximum likelihood, with the same number of regressors as in the previous models (except religion, as indicated above). We used the package ‘msm’ in R.85

Markov chains are stochastic processes according to which, for each respondent, the risk of being in any one particular state in one wave depends solely on the state in which the person was in the previous wave, plus a number of characteristics observed also in the previous wave.86 ,87

We computed discrepancy scores88 for consumption of wine, spirits and beer and for the number of drinking days, for men and women separately, as the difference between units in waves 4 and 5. We failed to find any systematic or consistent patterns and the correlation coefficients between the discrepancy scores of each type of alcoholic drink and the number of drinking days were all very low (max=0.33, for consumption of beer and drinking days for women). Therefore, we can rule out inconsistency in the answers.

In this paper, we ran models for two and four states. The two-state models fitted transitions between being and not being in the higher risk category; the four-state models allowed for transitions across any of the four risk categories. However, owing to sample limitations, the four-state models could not be computed: with four states, none of them irreversible, there are 16 possible transitions, which proved too many for our samples. Consequently, we only report the results for the two-state models.

Our two-state models, shown separately for women and men, estimate the probability of a respondent becoming a higher risk alcohol drinker in wave 5, given that they were not a higher risk alcohol drinker in wave 4, and the probability of becoming a not at-higher risk alcohol drinker in wave 5, given they were a higher risk alcohol drinker in wave 4—conditional on a number of personal characteristics such as age, education attainment, marital status, etc—as measured in wave 4.