RESEARCH REPORT doi:10.1111/add.

13396

Understanding trends in Australian alcohol

consumption—an age–period–cohort model

Michael Livingston1,2, Jonas Raninen3,4, Tim Slade1, Wendy Swift1, Belinda Lloyd5,6 & Paul Dietze7,8
National Drug and Alcohol Research Centre, University of New South Wales, Sydney, Australia,1 Centre for Alcohol Policy Research, La Trobe University, Melbourne,
Australia,2 CAN (Swedish Council for Information on Alcohol and Other Drugs), Stockholm, Sweden,3 Karolinska Institutet, Department of Clinical Neuroscience,
Stockholm, Sweden,4 Monash University, Melbourne, Australia,5 Turning Point, Melbourne, Australia,6 The Burnet Institute, Melbourne, Australia7 and School of public
Health and Preventive Medicine, Monash University, Melbourne, Australia8

ABSTRACT

Aims To decompose Australian trends in alcohol consumption into their age, period (survey year) and cohort
(birth year/generation) components. In particular, we aimed to test whether recent declines in overall consumption have
been influenced by reductions in drinking among recently born cohorts. Design Seven cross-sectional waves of the
Australian National Drug Strategy Household Survey (1995–2013). Age, period and cohort effects were estimated using
a linear and logistic cross-classified random-effects models (CCREMs). Setting Australia Participants A total of
124 440 Australians (69 193 females and 55 257 males), aged 14–79 years. Measurements Whether or not respon-
dents consumed alcohol in the 12 months prior to the survey and, for those who did, the estimated volume of pure alcohol
consumed, derived using standard quantity–frequency survey questions. Findings Controlling for age and period effects,
there was significant variation in drinking participation and drinking volume by birth cohort. In particular, male cohorts
born between the 1965 and 1974 and female cohorts born between 1955 and 1974 reported higher rates of drinking
participation (P < 0.05), while the most recent cohorts (born in the 1990s) had lower rates of participation (P < 0.01).
Among drinkers, the most recently born cohort also had sharply lower average consumption volumes than older cohorts
for both men and women (P < 0.01). Conclusion Recent birth cohorts (born between 1995 and 1999) in Australia
report significantly lower rates of both drinking participation and drinking volume than previous cohorts, controlling
for their age distribution and overall changes in population drinking. These findings suggest that the recent decline in
alcohol consumption in Australia has been driven by declines in drinking among these recently born cohorts. These trends
are consistent with international shifts in youth drinking.

Keywords Age period cohort, alcohol, Australia, surveys, trends, youth drinking.

Correspondence to: Michael Livingston, National Drug and Alcohol Research Centre, 22 King Street, Randwick, NSW, 2031, Australia.
E-mail: m.livingston@unsw.edu.au
Submitted 20 September 2015; initial review completed 21 December 2015; final version accepted 17 March 2016

INTRODUCTION describing overall trends. However, they cannot be disag-

Alcohol is related to an array of health and social problems
in Australia, with associated costs estimated most recently
at $15 billion per annum [1]. Alcohol consumption trends
are dynamic, with estimated Australian per-capita alcohol
consumption showing a steady increase in the years
following the end of World War 2, peaking in the early
1970s, before declining sharply in the 1980s and stabiliz-
ing through the 1990s and 2000s (Fig. 1). Consumption
has declined steadily during the past decade, with per-
capita consumption in 2013–14 reaching its lowest level
since the early 1960s. Per-capita consumption estimates,
derived from tax and importation records, are useful for
gregated to explore trends in subgroups of the population.
One way to examine these population subgroups is to
analyse data from population surveys.
Australian survey data have been used to explore the
disparity between recent trends in consumption and corre-
sponding trends in alcohol-related harm rates [2,3] and
trends in drinking by young people [4], but analyses of
long-term drinking trends are few. The Australian Institute
of Health and Welfare (AIHW) examined drinking trends
from a range of Australian survey data, finding an increase
in risky drinking between 1995 and 2007, with little vari-
ation by age group [5]. A recent analysis of data from the
Australian Bureau of Statistics (ABS) National Health

© 2016 Society for the Study of Addiction Addiction, 111, 1590–1598


Figure 1 Estimated per-capita alcohol
consumption, Australia, 1960–61 to 2013–14.
Source: Australian Bureau of Statistics [30]
Survey found an increase in consumption between 2001
and 2011 across all age groups [6], although this study
has been criticized for the alcohol measures used [7] and
the estimate of a trend based on just two data points [8].
To our knowledge, there has been no attempt to explore
whether birth cohort (or generational) effects can provide
insight into population trends in alcohol consumption in
Australia, although other researchers have examined
drug-related outcomes such as heroin overdose [9].
Age–period–cohort (APC) models can be used to
unpack trends in overall alcohol consumption into three
component factors: an overall change in population behav-
iour at a particular point in time (a ‘period’ effect), changes
in the age structure of the population where the behaviour
varies across the life-course (an ‘age’ effect) or changes in a
behaviour among particular generations that continue
through their life-course (a ‘cohort’ effect).
For example, Kerr and colleagues analysed alcohol con-
sumption in the United States between 1979 and 2005,
finding that drinking peaked among cohorts born between
1976 and 1985 [10]. Similarly, researchers in Germany
found large declines in drinking from the 1940s through
to 1970s birth cohorts, followed by increases in more
recently born cohorts [11]. These findings are supported
by a comprehensive review of studies conducted up to
2009 which found that, particularly in the United States,
more recent birth cohorts were heavier drinkers [12]. In
contrast, more recent analyses of Swedish data show sharp
reductions in drinking among most recently born cohorts
(born in the 1990s), following high-consuming cohorts
born in the 1980s [13]. These Swedish findings match
results of a study of UK drinking between 1984 and
2009 [14]. Similarly, an analysis of teenage drinking in
the United States over 32 years using annual survey data
found sharply declining rates of drinking in birth cohorts
born in the 1990s [15]. The evidence that drinking among
young people has fallen in recently born cohorts is
© 2016 Society for the Study of Addiction
Trends in Australian alcohol consumption 1591
consistent with school survey data from Europe and the
United States, which show declining teen drinking in 20
of 28 countries [16]. Overall, these findings reflect broadly
similar patterns, with increasing drinking up until the
1980s birth cohorts followed by (where it has been studied)
declining drinking in cohorts born more recently.
In this study we decompose 18 years of Australian
survey data on alcohol consumption in terms of age, period
and cohort components to test whether recent changes in
consumption are occurring at the period level across the
whole population or are being driven by recently born
cohorts. This information is not only important in relation
to the ongoing debates about trends in alcohol consump-
tion and harm in Australia [4,6,8], but will also provide
useful comparative results with the growing body of inter-
national research described above.
METHODS
Data
We obtained data from seven waves of the Australian
National Drug Strategy Household Survey (NDSHS). These
surveys have been conducted approximately every 3 years
since 1985, with various changes in the questionnaire
and survey methodology during the study period. To
reduce the influence of these changes we limit our analyses
to the seven waves of the survey from 1995 to 2013, com-
prising respondents aged 14–79 years (excluding n = 5175
aged 80 or over). Surveys from 2004 onwards collected
data from 12- and 13-year-olds, but these respondents
(n = 2132) were excluded due to their exclusion from sam-
pling in the earlier waves. We have excluded respondents
aged 80 or over due to the lack of disaggregated age data
for these respondents in some survey waves. A summary
of these survey waves is provided in Table 1.
The changes in methodology during the study period
raise some concerns about comparability, but differences
Addiction, 111, 1590–1598
1592 Michael Livingston et al.

Table 1 Summary of survey data used in this study. number of drinks in various common containers. While
the wording of these questions has been consistent dur-
Year Method Sample size Response rate
ing the study period, the response categories provided
1995 Face-to-face 3783 57% have varied slightly. The response options for drinking
1998 Face-to-face 10 238 52% frequency and usual quantity for each of the survey
2001 A mix of face-to-face, 26 211 50% waves are laid out in Tables 2 and 3, showing that there
drop-and-collecta were two distinct sets of response options for frequency
and CATI
(1995–98; 2001–13) and three for quantity (1995–
2004 A mix of drop- 27 789 46%
and-collect and CATI
98; 2001–07; 2010–13).
2007 A mix of drop- 22 089 49% Exploratory analyses suggested that the changes to the
and-collect and CATI response options had little overarching impact on estimates
2010 Drop-and-collect 25 187 51% of drinking volume in each of the surveys, but dummy
2013 Drop-and-collect 22 364 49% variables for each set of frequency and quantity response
options were included in the final models to ensure these
a
Drop-and-collect involves face to face recruitment of a respondent and changes did not bias the outcomes. Even with these
subsequent self-completion of a printed questionnaire, which is collected
1–2 weeks after recruitment. CATI = computer-assisted telephone dummies, if changes to the response items (or to the survey
interviews. methods) affected how respondents estimated their
consumption, these changes would be evident only in the
‘period’ effects.
in responses by study method have found little evidence of
An estimate of annual total drinking volume was
bias in consumption estimates [17]. We excluded responses
derived for each respondent by multiplying the mid-point
with incomplete or inconsistent responses to alcohol con-
of their responses to each of these questions. Thus, for
sumption questions (n = 6944) or no response to the ques-
example, a respondent who reported drinking 2–3 days
tion on education level (n = 6147). Rates of missing data
per week and drinking three to four drinks when they
were consistent across survey waves (between 5 and
drank would have an estimated 2.5 × 52 drinking occa-
10%). The final models for current drinking include
sions (130) of 3.5 drinks each, leaving a total estimated
124 440 (69 193 females and 55 257 males) respondents,
volume of 455 standard drinks. Responses to the top
while the models for drinking volume (among current
category of volume were coded as 21 drinks when the top
drinkers only) include 106 461 respondents.
category was 20+ and 15 drinks when the top category
was 13+ (15 was used rather than 14, based on the distri-
bution of drinks in the years where the survey had more
Measures
detailed options). Volume was converted from Australian
We present models here for two measures of drinking— standard drinks (10 g of pure alcohol) to litres of pure
any drinking during the past 12 months (current drinker) alcohol. Final analyses were conducted on the natural
and volume consumed in the past 12 months. The first logarithm of estimated annual volume to address the
measure is based on two questions that have been asked highly skewed distribution of drinking volume.
consistently across the seven survey waves under analysis:
1 Have you ever had a full serve of alcohol?
2 Have you had an alcoholic drink of any kind in the last
Table 2 Usual frequency categories used in National Drug Strategy
12 months?
Household Survey (NDSHS) surveys, by survey wave.
Respondents who responded ‘no’ to either of these
questions were classed as abstainers, while respondents Survey wave
who answered ‘yes’ to both were treated as current Frequency
options 1995 1998 2001 2004 2007 2010 2013
drinkers.
The measure of total volume of alcohol consumed in Every day X X X X X X X
the past 12 months was calculated only for respondents 5–6 days X X X X X
who were current drinkers. Again, the questions used to 4–6 days X X
derive this measure were consistent over the seven waves 3–4 days X X X X X
under analysis: 2–3 days X X
1–2 days X X X X X
1 How often do you have an alcoholic drink of any kind?
1 day X X
2 On a day that you have an alcoholic drink, how many
2–3 per month X X X X X X X
standard drinks do you usually have? 1 day per month X X X X X X X
Respondents were provided with a clear explanation less X X X X X X X
of the definition of Australian standard drinks and the

© 2016 Society for the Study of Addiction Addiction, 111, 1590–1598

 Trends in Australian alcohol consumption 1593

Table 3 Usual quantity categories used in National Drug Strategy statistical models to be estimated, it does not resolve the
Household Survey (NDSHS) surveys, by survey wave. underlying issue of the dependence of the three measures
Survey wave [18,19]. Recently, Yang & Land have argued that data on
Quantity repeated cross-sections is modelled more accurately using
options 1995 1998 2001 2004 2007 2010 2013 a hierarchical structure, with individuals (and their age)
nested within periods and birth cohorts [20,21]. Thus,
20+ X X
they argue that APC models based on appropriately speci-
16–19 X X
fied CCREMs produce robust and reliable estimates of the
13–15 X X
13+ X X X X X effects of age, period and cohort on a particular outcome.
11–12 X X X X X We followed Yang and Land’s approach, modelling alcohol
9–12 X X consumption as a function of fixed effects of age (and
9–10 X X age-squared) at level 1, with random effects of cohort
7–10 X X X and period at level 2.
7–8 X X X X
Concerns about the reliability of CCREM models of APC
5–6 X X X X X X X
effects have been raised [22]. In particular, they have
3–4 X X X X X X X
1–2 X X X X X conducted a simulation that demonstrates that CCREMs
2 drinks X X are not robust when period or cohort effects are driven by
1 drink X X continuous trends [22]. While this cannot be assessed
Less X X definitively outside an APC-modelling framework, descrip-
tive analyses of our data suggest that this is unlikely. For
period, the overall mean per-capita consumption shows a
In our models we included age as a continuous slight increase over the period, but any trend is small
variable, along with age-squared to allow for a non-linear (especially when 1995 is excluded, as it is in some of our
relationship between age and consumption [13]. Respon- sensitivity analyses) and countered by recent declines.
dents over the age of 79 were removed from our analyses Cohort effects are also unlikely to be driven by a linear
due to small numbers. We grouped the sample into 5-year trend. For example, drinking participation rates between
birth cohorts, spanning 17 cohorts from 1915–19 to the ages of 14 and 17 for the most recent five cohorts are
1995–99. Because of small numbers in the earliest birth 69, 68, 67, 57 and 37%, suggesting a steep, non-linear
cohorts, we combined respondents born between 1915 shift in recent cohorts.
and 1924 into a single cohort. Period effects were modelled We ran linear CCREMs for the log-volume model
as random effects in our cross-classified random-effects and logit CCREMs for the model of current drinking.
models (CCREM). The significance of the period and cohort effects were
Analyses controlled for the highest level of education assessed by conducting likelihood-ratio tests comparing
attained by respondents, grouped into five categories: still the full APC models with models excluding survey year
at school (4.9%), did not finish high school (24.5%), (period) or cohort [23]. Initial descriptive models found
completed high school (12.6%), non-university post-school significant interactions between sex and cohort, and
qualification (32.2%) and a degree or higher (21.3%). sex and survey year, so all CCREMs were run separately
for men and women. All analyses were conducted in
SAS [24].
Analysis

Simple descriptive analyses were conducted to summarize

the data. Means (for drinking volume) and percentages
(for current drinking prevalence) (both with 95% confi-
dence intervals) are presented. The modelling of age, period
and cohort effects in a series of repeated cross-sectional
surveys is complicated by the linear dependence of the
three variables under analysis (period = age + cohort).
These dependencies mean that the same data-generating
process could be modelled accurately by a number of differ-
ent combinations of age, period and cohort effects [9,18].
It is relatively straightforward to develop statistical
models that will converge (e.g. by grouping age into 5-year Figure 2 per capita alcohol consumption (litres of pure alcohol) by
categories, period into 3-year categories and cohort into year, National Drug Strategy Household Survey (NDSHS) survey and
7-year categories). However, while this allows the Australian Bureau of Statistics (ABS) recorded consumption

© 2016 Society for the Study of Addiction Addiction, 111, 1590–1598

1594 Michael Livingston et al.

RESULTS
Figure 2 compares the estimates of per-capita consump-
tion derived from the seven survey waves with official
data on alcohol available for consumption produced by
the ABS.
As is typical, survey estimates are much lower than
official measures of consumption [25,26]. In general, the
survey data track the trends in official consumption well,
with coverage varying between 45 and 47% from 1998
to 2013. The survey data for 1995 stand out, accounting
for just 37% of the official estimated consumption. There
are no obvious reasons for the 1995 survey to underesti-
mate consumption more than the subsequent surveys, so
it was included in the main models presented here. A
further round of sensitivity analyses were conducted using
only data from 1998 onwards to ensure that any problems
with the 1995 data do not overly affect our results.
Descriptive statistics of the study sample disaggregated
by gender are provided in Table 4 to provide context for
the subsequent modelling.
The results of the logistic CCREMs for drinking preva-
lence are presented in Table 5. The significant fixed effects
for age (presented here as odds ratios) in both models indi-
cate linear and non-linear associations between age and
prevalence for both men and women, while the significant
random effects for cohort and period imply significant
variation in prevalence across period and cohort. Random
effects are not estimated directly in the CCREMs, so only
parameters for their standard deviations are included.
The effects of age, period and cohort on drinking participa-
tion are understood more clearly via Fig. 3, which presents

Table 4 Description of study sample by period, age and cohort.

People Men Women

Drinking Mean Drinking Mean Drinking Mean

Period n prevalence volume (SD)a n prevalence volume (SD)a n prevalence volume (SD)a

1995 3783 80.1% 4.67 (7.3) 1670 84.1% 6.79 (9.0) 2020 76.6% 2.65 (4.3)
1998 10 238 84.8% 5.50 (8.3) 4227 87.3% 8.06 (10.3) 5658 82.7% 3.30 (5.1)
2001 26 211 84.1% 5.43 (8.0) 11 056 86.9% 7.62 (9.7) 14 603 81.8% 3.56 (5.5)
2004 27 798 84.8% 5.65 (8.2) 11 622 88.1% 7.93 (10.0) 15 719 82.3% 3.75 (5.6)
2007 22 089 83.4% 5.83 (8.4) 9196 86.8% 8.05 (10.2) 12 427 80.8% 3.95 (6.0)
2010 25 187 82.5% 6.26 (9.8) 10 842 85.6% 8.81 (12.1) 13 908 80.0% 4.02 (6.5)
2013 22 364 81.4% 5.99 (9.3) 9445 83.6% 8.27 (11.0) 12 483 79.5% 4.07 (7.1)
Age group (years)
14–17 4134 59.4% 2.89 (5.6) 3291 57.5% 3.38 (6.6) 3665 61.2% 2.49 (5.3)
18–24 11 095 89.2% 6.02 (8.9) 5466 89.3% 8.29 (10.8) 6972 89.1% 4.23 (6.5)
25–29 9334 88.8% 5.30 (8.4) 4312 91.6% 7.66 (10.2) 6204 86.8% 3.57 (6.1)
30–39 22 998 87.9% 5.13 (8.1) 10 834 90.1% 7.46 (10.0) 15 341 86.3% 3.42 (5.8)
40–49 19 996 87.7% 6.07 (9.1) 10 130 89.8% 8.27 (10.9) 12 662 86.0% 4.23 (6.8)
50–59 18 086 84.6% 6.53 (9.6) 9586 88.4% 9.32 (11.8) 11 794 81.5% 4.07 (6.1)
60–69 14 516 79.6% 6.31 (8.9) 8595 85.3% 8.81 (10.8) 9652 74.4% 3.76 (5.3)
70–79 8711 71.7% 5.75 (7.6) 5844 80.6% 7.69 (8.9) 6305 63.5% 3.48 (4.6)
Birth cohort
1916–24 1169 66.4% 5.18 (6.9) 551 77.3% 7.01 (9.3) 618 56.6% 2.95 (3.6)
1925–29 2980 71.3% 5.73 (7.5) 1468 80.8% 7.63 (8.6) 1512 62.1% 3.32 (4.4)
1930–34 5047 71.7% 5.87 (7.7) 2365 80.5% 7.86 (9.1) 2682 63.8% 3.67 (4.9)
1935–39 6989 75.4% 6.02 (8.1) 3443 82.8% 8.14 (9.6) 3546 68.3% 3.51 (4.8)
1940–44 8770 79.9% 6.20 (9.5) 4090 86.2% 8.64 (10.2) 4680 74.4% 3.74 (5.2)
1945–49 11 026 83.0% 6.54 (9.4) 5088 87.7% 9.41 (11.6) 5938 79.0% 3.80 (5.6)
1950–54 11 195 84.8% 6.14 (9.0) 5025 88.4% 8.77 (11.1) 6170 81.9% 3.83 (5.7)
1955–59 11 452 86.5% 6.30 (9.3) 5081 89.4% 8.73 (11.3) 6371 84.2% 4.25 (6.5)
1960–64 12 911 88.1% 5.82 (8.7) 5684 89.9% 8.03 (10.6) 7227 86.6% 4.01 (6.4)
1965–69 12 684 88.8% 5.67 (8.8) 5433 90.9% 8.00 (10.8) 7251 87.3% 3.86 (6.3)
1970–74 13 001 88.7% 5.22 (8.4) 5319 90.9% 7.56 (10.2) 7682 87.1% 3.52 (6.2)
1975–79 11 172 87.7% 5.40 (8.3) 4656 89.5% 7.86 (10.4) 6516 86.3% 3.58 (5.7)
1980–84 9923 85.3% 5.48 (8.5) 4267 85.8% 7.64 (10.5) 5656 85.0% 3.84 (6.0)
1985–89 7037 80.9% 5.23 (8.5) 3096 80.1% 6.98 (10.1) 3941 81.5% 3.89 (6.8)
1990–94 3871 72.2% 4.83 (9.1) 1805 71.4% 6.04 (10.3) 2066 72.8% 3.79 (7.8)
1995–99 1426 43.6% 2.50 (5.8) 687 41.2% 2.96 (6.1) 739 45.7% 2.12 (5.7)

a
Mean volume expressed in litres of pure alcohol per year for drinkers only. SD = standard deviation.

© 2016 Society for the Study of Addiction Addiction, 111, 1590–1598

 Trends in Australian alcohol consumption 1595

Table 5 Cross-classified random effects logit models of drinking participation, 1995–2013.

Men Women

Estimate (odds ratio) 95% CI Estimate (odds ratio) 95% CI

Fixed effects
Age 1.130** 1.114–1.147 1.071** 1.059–1.085
2
Age 0.999** 0.999–0.999 0.999** 0.999–0.999
Intercept 0.465** 0.308–0.701 1.052** 0.704–1.573
Random effects Estimate 95% CI Estimate 95% CI
Period (SD) 0.104 0.054–0.198 0.082 0.041–0.167
2 2
LR χ = 31.75** LR χ = 22.77**
Cohort (SD) 0.340 0.235–0.490 0.400 0.275–0.582
2 2
LR χ = 362.28** LR χ = 416.26**

SD = standard deviation; CI = confidence interval; LR = likelihood ration. *P < 0.05; **P < 0.01.

Figure 3 Predicted age, cohort and period effects on drinking participation, men (top row) and women (bottom row)

the predicted probability of drinking participation across
age, period and cohort groups with 95% confidence inter-
vals. The significance of the random effects were assessed
using random-effects contrasts, which compare each level
of the random variables (year and cohort) to their mean.
The effects presented in Fig. 3 demonstrate marked
variation in drinking participation at the cohort level.
While the period effects overall are statistically significant,
the variation across years is relatively minor. For men,
2004 was significantly above average (88.1% drinkers,
P = 0.002) and 2013 was significantly below average
(84.6%, P = 0.010), while all other years were not signifi-
cantly different from the average. For women, 2004 was
also above average (82.4%, P = 0.008), while only 1995
was significantly below average (77.6%, P = 0.007). In
terms of cohorts, participation rates peaked for the two
cohorts born between 1965 and 1974 (both significantly
higher than average participation, P = 0.010 and
P = 0.014) for men, before dropping sharply in the
most recent cohort, born between 1995 and 1999
(P < 0.001). For women, the earliest cohorts (between
1916 and 1934) had significantly lower drinking partici-
pation rates, while the four cohorts born between 1955
and 1974 had significantly higher participation rates. As
with men, the most recent cohort had significantly lower
participation (P < 0.001). The age effects were similar
for men and women, with an increase in participation
until the late 20s and early 30s and then declining
participation for respondents aged more than 50,
particularly for women.
The results of the CCREMs modelling the drinking
volume are presented in Table 6 and the predicted values
from the model in Fig. 4 to illustrate the age, period and
cohort effects. Note that the predicted values in Fig. 4 are
smaller than the raw means presented in Table 4 due to
the use of logged data in the regression models.

© 2016 Society for the Study of Addiction Addiction, 111, 1590–1598

1596 Michael Livingston et al.

Table 6 Cross-classified random-effects models of drinking volume (l pure alcohol/year), 1995–2013.

Men Women

Estimate 95% CI Estimate 95% CI

Fixed effects
Age 0.058** 0.049 to 0.068 0.014** 0.005–0.022
Age 2 0.001** 0.001 to –0.001 0.001** 0.001 to –0.001
Intercept 0.301* 1.177 to –0.355 0.102 0.319 to 0.115
Random effects
Period (SD) 0.066 0.033 to 0.129 0.115 0.064 to 0.206
2 2
LR χ = 21.06** LR χ = 84.66**
Cohort (SD) 0.188 0.126 to 0.282 0.124 0.076 to 0.203
2 2
LR χ = 160.23** LR χ = 59.21**

CI = confidence interval; SD = standard deviation; LR = likelihood ratio. *P < 0.05; **P < 0.01.

Figure 4 Predicted age, cohort and period effects on volume of alcohol consumed (litres of pure alcohol per year), men (top row) and women
(bottom row)

The effect of age is significant for both men and women,
with consumption peaking in middle age and being lower
for teenagers and older respondents. The pattern of cohort
effects is similar for men and women, with significantly
lower average volumes in the most recently born cohorts
(P < 0.001 for both men and women). This suggests that
even members of these cohorts who choose to drink are
drinking less than drinkers in previous birth cohorts.
Women drinkers born between 1955 and 1964 reported
significantly higher average consumption levels, while
male drinking peaked for the cohorts born between 1975
and 1989. Year effects were significant for both men and
women, with 1995 lower than average in both cases
(P < 0.001 for women, P = 0.006 for men).
We conducted a series of sensitivity analyses. First,
models were re-run with data from 1995 excluded;
cohort and age results were unchanged. Secondly,
models were run with a categorical age-group variable
based on 5-year categories. These models produced
broadly similar results, with a sharper increase in partic-
ipation and volume over the early age groups. Thirdly,
simple logistic and linear regression models (i.e. not
multi-level models) were run with survey weights incor-
porated and the age, period and cohort patterns were
unchanged. Finally, we ran models excluding the
youngest cohort, as they appear in only two waves of
the surveys. These analyses found significantly lower
drinking participation for the cohort born between
1990 and 1994 for both men and women, but not lower
volume among drinkers. The results of these supplemen-
tary analyses provide broad support for the robustness of
the models presented here.

© 2016 Society for the Study of Addiction Addiction, 111, 1590–1598


DISCUSSION
Our findings accord with previous Swedish and UK work
[13,14]. Recent declines in per-capita consumption appear
to be driven by two major changes: (1) the ageing of
heavier drinking cohorts into lighter drinking stages of
the life-course and (2) sharp reductions in drinking among
recently born cohorts. This reduction has seen falls in both
the probability of participating in drinking and the amount
consumed by drinkers, suggesting that a significant
generational shift is under way. This trend spans many
high-income countries [16], but little research has been
undertaken to understand the factors driving it [27]. It is
particularly noteworthy that we found both declining
participation in drinking among these cohorts, but also
sharp declines in average consumption among those who
do drink—other research examining these trends interna-
tionally has focused largely on abstaining rates [4,16].
Understanding this decline in drinking by young people
(i.e. recently born cohorts), which has occurred in a period
of increasing alcohol availability, is critical to inform policy.
The potential for these changes to result in long-term
public health gains is high, given the body of research
showing that drinking patterns while young are strong
predictors of drinking problems later in life [28]. It is also
worth noting the changes in participation by birth cohort
for women, with a steady increase in drinking from the
earliest cohorts through to those born in the 1950s and
1960s. This closing of the gender gap in participation
reflects a broad change in the role of women in Australian
society, and particularly the social norms and expectations
related to women’s drinking [2].
The age patterns we found show that consumption
peaks among middle-aged Australians (between the ages
of 40 and 60 years). While these data are based on rela-
tively simple survey questions and do not capture fully
the effects of drinking patterns, these findings still point
to an overlooked population group. Most public education
and information campaign focus on teenagers and young
adults, but we have shown that older Australians are
drinking at levels that put themselves at risk of harm [29].
Our study has a number of important limitations. First,
the data are based on self-reports of alcohol consumption,
which have an array of weaknesses [26]. Further, these
data come from surveys with relatively low response rates
and with changes in both methodology and question struc-
ture over the study period. Any impact of changes to the
survey should be represented only in the period effects of
the above models, but caution is needed when interpreting
the age and cohort effects. Response rates are stable, but
the typically low response rates of young people may have
implications for our age and cohort effects (particularly for
the most recent cohorts). Furthermore, the most recently
born cohort in our models appear in only two waves of
© 2016 Society for the Study of Addiction
Trends in Australian alcohol consumption 1597
the survey, raising some concerns that the sharp declines
in participation and drinking may not be stable cohort
effects. We conducted sensitivity analyses excluding this
cohort and found significant declines in participation for
the next-youngest cohort, but not in drinking volume. This
provides some reassurance that a real generational shift in
drinking participation is occurring, but more data are
required to confirm fully the volume changes found.
Finally, the limitations of all models that try to incorporate
the interdependent variables of age, period and cohort
should be acknowledged. As noted by others [22], even
using the CCREM approach implemented here does not
resolve fully the identification problem inherent in APC
modelling. To alleviate these concerns, we conducted
sensitivity analyses, based on a variety of different model
specifications, all of which produced the same broad results
as those presented here.
Declaration of interests
None.
Acknowledgements
The Australian Institute of Health and Welfare manage
the data collection and dissemination of the National
Drug Strategy Household Survey and we are grateful to
them for facilitating access to the data via the Australian
Data Archive. This research was part-supported under
Australian Research Council’s Discovery Projects
funding scheme (project number DP150101024). M.L.
is supported by an NHMRC Early Career Fellowship
(APP1053029), P.D. by an NHMRC Senior Research
Fellowship (APP1004140). The Centre for Alcohol
Policy Research is funded by the Foundation for Alcohol
Research and Education, an independent, charitable
organization working to prevent the harmful use of
alcohol in Australia http://www.fare.org.au.
References
1. Collins D., Lapsley H. The Avoidable Costs of Alcohol Abuse in
Australia and the Potential Benefits of Effective Policies to
Reduce the Social Costs of Alcohol. Canberra: Australian
Department of Health and Ageing; 2008.
2. Livingston M. Recent trends in risky alcohol consumption and
related harm among young people in Victoria, Australia.
Aust NZ J Public Health 2008; 32: 266–71.
3. Livingston M., Matthews S., Barratt M. J., Lloyd B., Room R.
Diverging trends in alcohol consumption and alcohol-related
harm in Victoria. Aust NZ J Public Health 2010; 34: 368–73.
4. Livingston M. Trends in non-drinking among Australian
adolescents. Addiction 2014; 109: 922–9.
5. Australian Institute of Health and Welfare (AIHW) Risk
Factor Trends: Age Patterns in Key Health Risk Factors Over
Time. Canberra: AIHW; 2012.
6. Yusuf F., Leeder S. R. Making sense of alcohol consumption
data in Australia. Med J Aust 2015; 203: 128–30.
Addiction, 111, 1590–1598
1598 Michael Livingston et al.
7. Donath S. Estimated alcohol consumption in the 1995
National Health Survey: some methodological issues.
Aust NZ J Public Health 1999; 23: 131–4.
8. Livingston M., Dietze P. Comment on Yusuf and Leeder.
Med J Aust 2015; Available at: https://www.mja.com.au/
journal/2015/203/3/making-sense-alcohol-consumption-data-
australia#comments
9. Hall W. D., Degenhardt L. J., Lynskey M. T. Opioid overdose
mortality in Australia, 1964–1997: birth-cohort trends.
Med J Aust 1999; 171: 34–7.
10. Kerr W. C., Greenfield T. K., Bond J., Ye Y., Rehm J.
Age–period–cohort modelling of alcohol volume and
heavy drinking days in the US National Alcohol Surveys:
divergence in younger and older adult trends. Addiction
2009; 104: 27–37.
11. Pabst A., Kraus L., Piontek D., Mueller S. Age, period, and
cohort effects on time trends in alcohol consumption in the
German adult population. Sucht 2010; 56: 349–59.
12. Keyes K. M., Li G., Hasin D. S. Birth cohort effects and gender
differences in alcohol epidemiology: a review and synthesis.
Alcohol Clin Exp Res 2011; 35: 2101–12.
13. Kraus L., Tinghög M. E., Lindell A., Pabst A., Piontek D.,
Room R. Age, period and cohort effects on time trends in
alcohol consumption in the Swedish adult population
1979–2011. Alcohol Alcohol 2015; 50: 319–27.
14. Meng Y., Holmes J., Hill-McManus D., Brennan A., Meier P. S.
Trend analysis and modelling of gender-specific age, period
and birth cohort effects on alcohol abstention and consump-
tion level for drinkers in Great Britain using the General
Lifestyle Survey 1984–2009. Addiction 2014; 109: 206–15.
15. Keyes K. M., Schulenberg J. E., O’Malley P. M., Johnston L. D.,
Bachman J. G., Li G., et al. Birth cohort effects on adolescent
alcohol use: the influence of social norms from 1976 to
2007. Arch Gen Psychiatry 2012; 69: 1304–13.
16. de Looze M., Raaijmakers Q., ter Bogt T., Bendtsen P.,
Farhat T., Ferreira M., et al. Decreases in adolescent weekly
alcohol use in Europe and North America: evidence from 28
countries from 2002 to 2010. Eur J Public Health 2015; 25:
69–72.
© 2016 Society for the Study of Addiction
17. Roy Morgan Research. National Drug Strategy Household
Survey 2004—Technical Report. Melbourne: Roy Morgan
Research; 2005.
18. Smith H. L. Advances in age–period–cohort analysis. Sociol
Meth Res 2008; 36: 287–96.
19. Kupper L. L., Janis J. M., Karmous A., Greenberg B. G.
Statistical age-period-cohort analysis: a review and critique.
J Chronic Dis 1985; 38: 811–30.
20. Yang Y., Land K. C. Age–period–cohort analysis of repeated
cross-section surveys: fixed or random effects? Sociol Meth
Res 2008; 36: 297–26.
21. Yang Y., Land K. C. A mixed models approach to the
age-period-cohort analysis of repeated cross-section surveys,
with an application to data on trends in verbal test scores.
Sociol Methodol 2006; 36: 75–97.
22. Bell A., Jones K. Another ‘futile quest’? A simulation study of
Yang and Land’s Hierarchical Age-Period-Cohort model.
Demogr Res 2014; 30: 333–60.
23. Skrondal A., Rabe-Hesketh S. Generalized Latent Variable
Modeling: Multilevel, Longitudinal, and Structural Equation
Models. Boca Raton, FL: CRC Press; 2004.
24. SAS Institute The SAS System for Windows, v9.4. Cary, NC:
SAS Institute; 2014.
25. Livingston M., Callinan S. Under-reporting in alcohol surveys:
whose drinking is under-estimated? J Stud Alcohol Drugs
2015; 76: 158–64.
26. Gmel G., Rehm J. Measuring alcohol consumption. Contemp
Drug Prob 2004; 31: 467–540.
27. Pennay A., Livingston M., MacLean S. Young people are
drinking less: it’s time to find out why. Drug Alcohol Rev
2015; 34: 115–18.
28. McCambridge J., McAlaney J., Rowe R. Adult consequences of
late adolescent alcohol consumption: a systematic review of
cohort studies. PLOS Med 2011; 8: e1000413.
29. Hunter B., Lubman D. I., Barratt M. Alcohol and drug misuse
in the elderly. Aust NZ J Psychiatry 2011; 45: 343.
30. Australian Bureau of Statistics Apparent Consumption of
Alcohol, Australia, 2013–14. Canberra: Australian Bureau
of Statistics; 2015.
Addiction, 111, 1590–1598