Accepted Manuscript Public acceptability of population-level interventions to reduce alcohol consumption: a discrete choice experiment Rachel Pechey , Peter Burge , Emmanouil Mentzakis , Marc Suhrcke , Theresa M. Marteau PII:
S0277-9536(14)00304-9
DOI:
10.1016/j.socscimed.2014.05.010
Reference:
SSM 9461
To appear in:
Social Science & Medicine
Received Date: 23 July 2013 Revised Date:
7 November 2013
Accepted Date: 9 May 2014
Please cite this article as: Pechey, R., Burge, P., Mentzakis, E., Suhrcke, M., Marteau, T.M, Public acceptability of population-level interventions to reduce alcohol consumption: a discrete choice experiment, Social Science & Medicine (2014), doi: 10.1016/j.socscimed.2014.05.010. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT Manuscript Number: SSM-D-13-01794R1
Manuscript Title: Public acceptability of population-level interventions to reduce alcohol consumption: a discrete choice experiment
Authors:
RI PT
Rachel Pechey 1 Peter Burge 2 Emmanouil Mentzakis 3 Marc Suhrcke 1,4
Behaviour and Health Research Unit, Institute of Public Health, University of Cambridge,
Cambridge, UK
M AN U
1
SC
Theresa M Marteau 1
2
RAND Europe, Cambridge, UK
3
Economics Division, School of Social Sciences, University of Southampton, Southampton, UK
4
Health Economics Group, Norwich Medical School, University of East Anglia, Norwich, UK
Corresponding author: Theresa M Marteau, Behaviour and Health Research Unit, Institute of Public
TE D
Health, Cambridge CB2 0SR, UK. Email:
[email protected]; Tel: +44 1223 330562; Fax: +44 1223 762515.
Acknowledgments: The study was funded by the Department of Health Policy Research Programme
EP
(Policy Research Unit in Behaviour and Health [PR-UN-0409-10109]). The Department of Health had no role in the study design, data collection, analysis, or interpretation. The research was conducted independently of the funders, and the views expressed in this paper are those of the authors and not
AC C
necessarily those of the Department of Health in England. The final version of the report and ultimate decision to submit for publication was determined by the authors (TMM had final responsibility).
ACCEPTED MANUSCRIPT Abstract Public acceptability influences policy action, but the most acceptable policies are not always the most effective. This discrete choice experiment provides a novel investigation of the acceptability of different interventions to reduce alcohol consumption and the effect of
RI PT
information on expected effectiveness, using a UK general population sample of 1202 adults. Policy options included high, medium and low intensity versions of: minimum unit pricing (MUP) for alcohol; reducing numbers of alcohol retail outlets; and regulating alcohol
SC
advertising. Outcomes of interventions were predicted for: alcohol-related crimes; alcoholrelated hospital admissions; and heavy drinkers. Firstly, the models obtained were used to
M AN U
predict preferences if expected outcomes of interventions were not taken into account. In such models around half of participants or more were predicted to prefer the status quo over implementing outlet reductions or higher intensity minimum unit pricing. Secondly, preferences were predicted when information on expected outcomes was considered, with
TE D
most participants now choosing any given intervention over the status quo. Acceptability of MUP interventions increased by the greatest extent: from 43% to 63% preferring MUP of £1 to the status quo. Respondents’ own drinking behaviour also influenced preferences, with
EP
around 90% of non-drinkers being predicted to choose all interventions over the status quo,
AC C
and with more moderate than heavy drinkers favouring a given policy over the status quo. Importantly, the study findings suggest public acceptability of alcohol interventions depends on both the nature of the policy and its expected effectiveness. Policy-makers struggling to mobilise support for hitherto unpopular but promising policies should consider giving greater prominence to their expected outcomes.
Keywords: UK: Public acceptability; Alcohol; Health policy
1
ACCEPTED MANUSCRIPT Introduction Alcohol consumption and its consequences place a considerable burden on health services and policing, as well as imposing a myriad of less direct societal costs (Mohapatra et al., 2010; UK Government, 2007). Reviews of the evidence suggest three interventions that
RI PT
are likely to be particularly effective, which all happen to require government action:
increasing alcohol prices, reducing the availability of outlets selling alcohol, and restricting the marketing of alcohol (Jackson et al., 2010; NICE, 2010; WHO Regional Office for
SC
Europe, 2009). While policy decisions may be informed by evidence on the effectiveness of potential policies and the costs of implementing them, public acceptability of interventions
M AN U
likely plays an even greater role. In addition, the accurate assessment of public acceptability is important as it can affect whether an intervention will work in practice. A recent review of 200 studies of public acceptability of government interventions to change health-related behaviour, including a few studies on alcohol consumption, shows that
TE D
public support is strongest for policies that are least intrusive, with greater support for educational campaigns than for taxation or restrictions on sales (Diepeveen et al., 2013). Respondents’ own behaviour is also important, with those who do not engage in the targeted
EP
behaviour being more supportive of more intrusive intervention. For example, non- and ex-
AC C
drinkers are more supportive of interventions such as higher taxes, bans on sales of alcohol in corner stores and bans on TV advertising (Giesbrecht et al., 2007; Holmila et al., 2009; Wilkinson et al., 2009). Acceptability also seems to be patterned by respondent demographic characteristics, with women and older people finding more intrusive interventions more acceptable (Bongers et al., 1998; Giesbrecht et al., 2007; Greenfield et al., 2007; Holmila et al., 2009; Wilkinson et al., 2009), while the impact of socioeconomic status is less clear. The fact that existing studies are primarily based on standard opinion polls and surveys, along with the observed lack of public acceptability of increasing price to reduce 2
ACCEPTED MANUSCRIPT alcohol consumption has led to calls for research to better understand this (WHO Regional Office for Europe, 2009). A recent focus group study revealed scepticism about the effectiveness of minimum unit pricing with resultant limited support for this policy intervention (Lonsdale et al., 2012). Providing information about the effectiveness of policies
RI PT
can increase their acceptability, as shown in an experimental study concerning the use of financial incentives for smoking cessation and weight loss (Promberger et al., 2012). It is unknown, however, whether a similar effect will be seen for alcohol policies when the
SC
interventions may have a direct impact upon the public by, for example, making alcohol less accessible or less affordable.
M AN U
The aim of the current study is to examine public acceptability of three interventions to reduce alcohol consumption and to examine the extent to which acceptability is sensitive to information about the effectiveness of the interventions. We explore both the size of any effect as well as the domain in which it is reported (crime vs. health). The influence of
Methods
AC C
Participants
EP
also explored.
TE D
respondent characteristics, particularly drinking habits, gender and socio-economic status, is
1202 adults resident in England participated in the discrete choice experiment (DCE) during face-to-face interviews conducted in their homes in November and December 2013. Recruitment was carried out by a market research company at 30 sampling points across England, with quotas set on age, gender and occupational group (UK Registrar General’s classification: Higher Managerial and Professional (groups A&B); White Collar and Skilled Manual (groups C1&C2); and Semi-skilled and Unskilled Manual (groups D&E)). The sample was broadly representative of the English population (Table 1), including, 3
ACCEPTED MANUSCRIPT importantly, in terms of participants’ drinking habits (i.e. non-drinkers, moderate drinkers or heavy drinkers) (Health and Social Care Information Centre, 2012). Ethical approval for the study was obtained from the University of Cambridge Psychology Research Ethics
TABLE 1 ABOUT HERE
SC
Attributes and levels
RI PT
Committee (Ref: Pre.2012.61).
Participants were presented with a range of policy options alongside different policy
M AN U
outcomes with regard to three domains (alcohol-related crimes, alcohol-related hospital admissions, heavy drinkers). The key attributes were: (1) the type of intervention (minimum unit pricing, reduction in outlet density, regulation of advertising); (2) the intensity of the intervention (low, medium, high); and, (3) the magnitude of impact in each of the three
TE D
outcome domains (9 levels, representing the reductions in each domain in a community of 100,000 people over a year) (see supplementary materials for details [INSERT LINK TO
EP
ONLINE FILES]).
AC C
Policy options: Intervention type and intensity The three levels of intensity for each intervention type were selected such that they encompassed at least one option that was currently being discussed in the UK policy context, as well as looking at how increasing the intensity (and potential effectiveness) of these would impact on acceptability: setting a minimum unit price for alcohol of (a) 40p, (b) 70p or (c) £1; reducing the number of outlets that sell alcohol by (a) 10%, (b) 20% or (c) 40%; and regulating advertising of alcohol via (a) self-regulation, (b) a partial ban or (c) a complete ban. Advertising regulations were described as: Self-regulation: ‘Work with industry to make 4
ACCEPTED MANUSCRIPT sure they are following the current rules (For example, the government would look at ways to make sure that adverts for alcohol are not shown during programmes for children or teenagers)’; Partial ban: ‘Part ban on alcohol adverts (Adverts for alcohol would be banned
for alcohol would be banned)’.
Policy outcomes: Domain and magnitude of impact
RI PT
from TV and in cinemas)’; Complete ban: ‘Complete ban on adverts for alcohol (All adverts
SC
The impacts on each of the three domains were estimated using and extrapolating from the modelling of the Sheffield Alcohol Research Group (Purshouse et al., 2009; Purshouse et al.,
M AN U
2010) (see supplementary materials for details, Table S11 for estimates used [INSERT LINK TO ONLINE FILES]). As these estimates are based on several assumptions (see supplementary materials), in what follows we refer to those estimates as the best available outcome estimates associated with the interventions. Alcohol-related crimes and hospital
TE D
admissions were defined as in the Sheffield models (i.e. crimes: 20 types including theft, criminal damage, assault and causing death by dangerous driving; hospital admissions: both those wholly or partly attributable to alcohol, and due to both acute and chronic conditions).
EP
Heavy drinkers were defined as those drinking more than UK government guidelines (22
AC C
units or more per week for men, 15 units or more per week for women). The best available outcome estimates allow us to examine the acceptability of policies taking into account their predicted impact. However, in order to avoid confounding between intervention effects and outcome effects, the outcomes presented to participants in the study were not constrained to the most likely estimates for each policy, but were drawn independently from any of the nine estimates associated with that outcome.
5
ACCEPTED MANUSCRIPT Experimental design Three-way choices (choice sets) were presented to participants: two options describing types of intervention, alongside an option representing no change to the status quo (see Figure 1 for an example choice set). The ordering of the two active policy interventions
RI PT
in the scenarios was varied using a random draw. Each type of intervention was presented at one of the three levels of intensity, with the outcomes presented for all three domains for each option (each outcome at one of the nine levels of magnitude). Participants were asked to
SC
‘imagine the effects shown would be the real effects for the options’. For each choice,
participants decided which of the options (either of the two presented policy interventions or
M AN U
‘no change’) the government should choose to implement.
Such discrete choice experiments are based on the idea that the acceptability of a policy is characterised by certain selected attributes and that participants’ own acceptance of any policy depends on these attributes. Although there are some limitations inherent in such a
TE D
design (i.e. the preferences examined are hypothetical, and the attributes chosen cannot reflect the full spectrum of options), this does allow us to examine the influence of said attributes on participants’ (albeit hypothetical) preferences, and has been widely used in
EP
health-related studies (Ryan, 2004). See supplementary materials for more on the DCE
AC C
methodology used [INSERT LINK TO ONLINE FILES]. The DCE used 81 different choice sets (determined using an orthogonal main effects
fractional factorial design), divided into nine blocks of nine. Each participant completed one block. A blocking approach was used to specify the subsets of scenarios to present to each respondent in order to minimise the correlation between each individual attribute and the block selected; this ensured that each respondent was presented a range of scenarios with differences in both policy options and outcomes.
6
ACCEPTED MANUSCRIPT FIGURE 1 ABOUT HERE
Analysis Heteroskedastic conditional logit models (bootstrapped to correct for any
RI PT
specification error introduced through using multiple responses from each respondent) were used to model participants’ preferences between the different policy options. See
supplementary materials for model details [INSERT LINK TO ONLINE FILES].
SC
In the results section we present the predicted probabilities that different policy
M AN U
options would be chosen if the sample of respondents were to be offered the choice of that policy or no intervention. Two models are used: the first is estimated with common coefficients across all respondents to provide average values for the sample, the second takes into account statistically significant differences between subgroups. The models were then used to predict preferences in two situations: firstly, how participants would respond if they
TE D
had no information regarding the effectiveness of each policy, and secondly, how participants
Results
EP
would respond if they were informed of the best available outcome estimates for each policy.
AC C
Figure 2 shows the predicted relative utility of (ranked preferences for) each of the
policy options when taking into account the impact of withholding or providing the best available estimates of each of the outcomes. If outcome estimates are not taken into account, participants would prefer minimum unit pricing (MUP) of 40p and any regulation of advertising over the status quo (all p<0.001). Four policies are equally preferable to the status quo (MUP of 70p, and all three levels of outlet reduction), while MUP of £1 is less preferable than the status quo (p<0.001). 7
ACCEPTED MANUSCRIPT
FIGURE 2 ABOUT HERE
If outcome estimates are taken into account, the rankings of the different intervention
RI PT
types change (given that some interventions were expected to be more effective than others), with all policy options preferred over the status quo. The ranking of higher intensity MUP improves in particular.
SC
The percentage of the sample predicted to choose to implement a given policy option rather than choose to maintain the status quo is shown in Figure 3. Separate predictions were
M AN U
made for each intervention and for the various combinations of information on outcomes that may be available to individuals when making their choices. For example, if participants were choosing between a partial ban on adverts or making 'no change' to the status quo in the absence of any information on expected outcomes, approximately 70% of participants would
TE D
be predicted to favour the partial ban. Indeed, for the case where participants have no information on outcomes (i.e. effectiveness of intervention), 58% are predicted to choose MUP of 40p over the status quo, and between 66-77% would prefer advertising regulations to
EP
the status quo; the other policies would be chosen over the status quo by around 50% of the
AC C
sample, with the exception of MUP of £1, preferred by only 43% of respondents.
FIGURE 3 ABOUT HERE
By contrast, in cases where information on all three outcome estimates is available to participants, reductions in outlet density are predicted to be the least preferred options, with 53-57% choosing these over the status quo, compared to 59-63% for MUP interventions.
8
ACCEPTED MANUSCRIPT Advertising regulations remain the most acceptable, with 73-77% predicted to choose these over the status quo. While the type of intervention clearly impacts on its relative acceptability (with a difference of 34 percentage points predicted between the most and least acceptable policies in
RI PT
the case where no information is available regarding outcome estimates), presenting outcome estimates in all three domains would reduce this difference by 10 percentage points, and alter the relative ranking of policies. This is particularly evident for MUP interventions, the
SC
acceptability of which changes markedly in absolute and relative terms between these two cases: the proportion of the sample choosing MUP of £1 over the status quo increases by 20
M AN U
percentage points (from no outcome estimates known to all three known), and its ranking improves from last to fourth place, even higher than that of the less intensive MUP options.
Subgroup analysis
TE D
Taking into account statistically significant differences between subgroups, the most notable differences were by drinker status (see Figure 4), with non-drinkers placing a significantly higher value on all policies relative to the status quo, regardless of policy
EP
content (all p<0.1). In both cases where no outcome estimates are available and where all
AC C
three outcome estimates are available, 89%-95% of non-drinkers would choose a given policy over the status quo. By contrast, drinkers discriminated more between the different intervention types, with individual policies being preferred to the status quo by between 4276% of moderate drinkers and between 33-69% of heavy drinkers (in predictions for the case without available outcome estimates). Whilst non-drinkers are seen to be relatively insensitive to information on policy outcomes, the predictions show that in contrast drinkers are influenced by such information (in predictions including all three outcome estimates 5976% of moderate drinkers would choose a given policy over the status quo, compared to 459
ACCEPTED MANUSCRIPT 70% for heavy drinkers). In general, a lower proportion of heavier than moderate drinkers would prefer a given policy to the status quo, with an approximately 10 percentage point difference for most policies. Differences in model scale, however, show that non-drinkers had more unexplained error in their choices (i.e. were harder to predict) relative to moderate
RI PT
drinkers (parameter = 0.50, p<0.01), whereas hazardous and harmful drinkers were more certain in their choices compared to moderate drinkers (parameter = 1.35, p<0.01).
SC
FIGURE 4 ABOUT HERE
M AN U
Systematic analyses were conducted by demographic characteristics (see supplementary materials Table S14 [INSERT LINK TO ONLINE FILES]). Within the moderate drinkers group (comprising 64% of the study sample and 62% of the English population (Health and Social Care Information Centre, 2012)), males were more likely to
TE D
choose to maintain the status quo (p<0.05), whereas infrequent drinkers (drinking less than once a week) and those in the three highest occupational groups A, B & C1 were less likely to favour no change (both p<0.001). There were no other differences found by gender or
AC C
working status.
EP
occupational group, and no differences found by age, highest educational qualification or
In terms of outcomes, reductions in the numbers of alcohol-related crimes showed a
linear relationship with acceptability, while hospital admissions were linearly valued up to a reduction of 213 admissions, but with additional reductions beyond this point providing no extra value. In general, reductions in the number of heavy drinkers were not greatly valued by participants, with the exception to this pattern being amongst the female moderate drinkers (p<0.001).
10
ACCEPTED MANUSCRIPT Discussion The relative preferences for the three intervention types in the discrete choice models (minimum unit pricing, outlet reduction and regulating advertising) suggest that policies regarding regulation of marketing would be most acceptable to participants. Interestingly,
RI PT
however, whether or not information on the best available outcome estimates of these policy options is available influences the absolute and relative acceptability of the interventions, with acceptability increasing as more outcomes are taken into consideration. Indeed, in
SC
predictions where information on all three outcome estimates is available, a majority of participants are predicted to prefer any given policy to the status quo, and the gap in
M AN U
acceptability between regulating advertising and other policy options reduces compared to the case where no information on outcome estimates is available. In particular, MUP (which has greater expected effectiveness with regard to the investigated outcomes compared to the other intervention types) has similar levels of acceptability to outlet reductions in the
TE D
predictions where no information is available on outcome estimates, but is preferred to outlet reduction policies in the case where information is available on all outcome estimates. In absolute terms, for MUP of £1 (the least popular option when no information is available on
EP
outcome estimates), an additional fifth of the sample would prefer this policy over the status
AC C
quo should they have information on the outcome estimates, making this the preferred policy option after marketing regulations in this case. While there were significant differences in preferences by drinker status, with the models suggesting that heavier drinkers found all policies less acceptable, there were no marked differences by other respondent characteristics. The varying levels of acceptability by type of intervention reflect findings from surveys conducted in Canada and the US that showed more support for bans on TV advertising of alcohol than for changes in price and restrictions on sales (albeit in the forms 11
ACCEPTED MANUSCRIPT of increased taxes on alcoholic beverages and limits on the hours stores can sell alcohol) (Giesbrecht et al., 2007; Greenfield et al., 2007). Evidence for acceptability varying by effectiveness of intervention has similarly been shown in studies looking at acceptability of financial incentives (Promberger et al., 2012), and suggested in focus groups as a factor
RI PT
limiting public support with regard to alcohol policies (Lonsdale et al., 2012), but, to our knowledge, has not been empirically demonstrated in this context before, nor quantified in this way. The findings from the current study reflect one of very few experimental studies
SC
looking at acceptability, offering a more rigorous assessment of acceptability (by confronting respondents with trade-offs between different policy options) compared to the existing
hypothetical answers in opinion polls.
M AN U
acceptability literature which mainly relies on asking respondents to provide unrestricted,
The findings with regard to respondent characteristics also largely tie in with previous research (Diepeveen et al., 2013): primarily consisting of surveys suggesting that individuals’
TE D
behaviour is a key factor in predicting policy acceptability, with heavier drinkers being less supportive (Giesbrecht et al., 2007; Holmila et al., 2009; Wilkinson et al., 2009). Moreover, a small number of surveys have looked at attitudes to alcohol control policies by
EP
socioeconomic status (SES), but found weak, if any, patterns of support. Together with the
AC C
current study’s findings, this suggests that attitudes to alcohol control are largely not mediated by SES, but perhaps these mixed findings reflect the complex relationships between SES and drinking behaviour (Bloomfield et al., 2006; Erskine et al., 2010). The findings from this large sample, broadly representative of the English population,
provide novel experimental evidence of the relative acceptability of different policies across different demographic groups. A further advantage of the approach taken in this study was the assessment of topical interventions, combined with using estimates of likely outcomes (albeit based on several assumptions), to allow a timely and policy relevant appraisal of their 12
ACCEPTED MANUSCRIPT public acceptability. There are two main limitations to the current study, however: firstly, it focused in a somewhat stylized manner only on the impact of positive outcomes, whereas a policy discussion is likely to involve multiple and competing messages; secondly, participants were asked to ‘imagine the effects shown would be the real effects for the
RI PT
options’, while it is unlikely that the public place absolute trust in figures presented by parties with a stake in policy choices. Another possible concern is that participants were told that heavy drinkers comprised those ‘who drink more alcohol than the government advises’ but
SC
participants may have assumed different thresholds for heavy drinking if they were unaware of government recommendations. Further investigation of these influences would add nuance
M AN U
to the study findings in terms of the relative influence of likely policy outcomes on public acceptability.
The results suggest the most favoured policies in this study (restrictions on advertising) also had the lowest estimates of effectiveness (in terms of the investigated
TE D
outcomes). However, presentation of (beneficial) outcomes when conveying a policy may help boost its public acceptability; an implication that policymakers struggling in mobilising support for ex ante less acceptable but more effective interventions (e.g. MUP) could build
EP
on. Moreover, the patterning of support by drinking status provides an impetus for
AC C
government action, given that as alcohol consumption worsens, so too will the likely acceptability of any interventions to target relevant behaviours. Future research into the acceptability of policies to change behaviour should build on
these results to place them more in the context of real world policy deliberations, by taking into account how the message is framed (e.g., targeting particular groups of the population) and improving understanding of the role that varying degrees of uncertainty around the outcome estimates might play.
13
ACCEPTED MANUSCRIPT References Bloomfield, K., Grittner, U., Kramer, S., & Gmel, G. (2006). Social inequalities in alcohol consumption and alcohol-related problems in the study countries of the EU concerted
Alcoholism, 41, i26-i36.
RI PT
action 'Gender, Culture and Alcohol Problems: A multinational study'. Alcohol and
Bongers, I.M.B., Van de Goor, I.A.M., & Garretsen, H.F.L. (1998). Social climate on alcohol in Rotterdam, the Netherlands: Public opinion on drinking behaviour and alcohol
SC
control measures. Alcohol and Alcoholism, 33, 141-150.
Diepeveen, S., Ling, T., Suhrcke, M., Roland, M., & Marteau, T.M. (2013). Public
M AN U
acceptability of government intervention to change health-related behaviours: a systematic review and narrative synthesis. BMC Public Health, 13, 756. Erskine, S., Maheswaran, R., Pearson, T., & Gleeson, D. (2010). Socioeconomic deprivation, urban-rural location and alcohol-related mortality in England and Wales. BMC Public
TE D
Health, 10, 99.
Giesbrecht, N., Ialomiteanu, A., Anglin, L., & Adlaf, E. (2007). Alcohol marketing and retailing: Public opinion and recent policy developments in Canada. Journal of
EP
Substance Use, 12, 389-404.
AC C
Greenfield, T.K., Ye, Y., & Giesbrecht, N. (2007). Alcohol policy opinions in the United States over a 15-year period of dynamic per capita consumption changes: Implications
for today's public health practice. Contemp. Drug Probs., 34, 649.
Health and Social Care Information Centre. (2012). Statistics on Alcohol: England, 2012. Leeds: Health and Social Care Information Centre. Holmila, M., Mustonen, H., Österberg, E., & Raitasalo, K. (2009). Public opinion and community-based prevention of alcohol-related harms. Addiction Research & Theory, 17, 360-371. 14
ACCEPTED MANUSCRIPT Jackson, R., Johnson, M., Campbell, F., Messina, J., Guillaume, L., Meier, P.S., et al. (2010). Interventions on Control of Alcohol Price, Promotion and Availability for Prevention of Alcohol Use Disorders in Adults and Young People. London: The University of Sheffield, for NICE Centre for Public Health Excellence.
RI PT
Lonsdale, A., Hardcastle, S., & Hagger, M. (2012). A minimum price per unit of alcohol: A focus group study to investigate public opinion concerning UK government proposals to introduce new price controls to curb alcohol consumption. BMC Public Health, 12,
SC
1023.
Mohapatra, S., Patra, J., Popova, S., Duhig, A., & Rehm, J. (2010). Social cost of heavy
Public Health, 55, 149-157.
M AN U
drinking and alcohol dependence in high-income countries. International Journal of
NICE. (2010). Alcohol-use disorders: Preventing harmful drinking. NICE public health guidance. London, UK.
TE D
Promberger, M., Dolan, P., & Marteau, T.M. (2012). “Pay them if it works”: Discrete choice experiments on the acceptability of financial incentives to change health related behaviour. Social Science & Medicine, 75, 2509-2514.
EP
Purshouse, R.C., Brennan, A., Latimer, N., Meng, Y., Rafia, R., Jackson, R., et al. (2009).
AC C
Modelling to assess the effectiveness and cost-effectiveness of public health related strategies and intervention to reduce alcohol attributable harm in England using the Sheffield Alcohol Policy Model version 2.0. University of Sheffield. Report to the NICE Public Health Programme Development Group.
Purshouse, R.C., Meier, P.S., Brennan, A., Taylor, K.B., & Rafia, R. (2010). Estimated effect of alcohol pricing policies on health and health economic outcomes in England: an epidemiological model. The Lancet, 375, 1355-1364. Ryan, M. (2004). Discrete choice experiments in health care. Bmj, 328, 360-361. 15
ACCEPTED MANUSCRIPT UK Government. (2007). Safe. Sensible. Social. The next steps in the National Alcohol Strategy. London, UK: UK Government, . WHO Regional Office for Europe. (2009). Evidence for the effectiveness and costeffectiveness of interventions to reduce alcohol-related harm. Copenhagen: World
RI PT
Health Organisation.
Wilkinson, C., Room, R., & Livingston, M. (2009). Mapping Australian public opinion on
AC C
EP
TE D
M AN U
SC
alcohol policies in the new millennium. Drug and Alcohol Review, 28, 263-274.
16
ACCEPTED MANUSCRIPT Table 1. Participant characteristics (N=1,202)
n (%)
332 (27.6) 454 (37.8) 416 (34.6)
SC
383 (31.9) 527 (43.8) 292 (24.3)
RI PT
555 (46.2) 647 (53.8)
158 (13.1) 763 (63.5) 281 (23.4)
EP
TE D
M AN U
Variable Gender Male Female Age 18-34 35-54 55+ Occupational group A&B (Higher Managerial and Professional) C1&C2 (White Collar and Skilled Manual) D&E (Semi-skilled and Unskilled Manual) Drinking habits* Non-drinker Moderate drinker Heavy drinker Highest educational qualification No formal qualification GCSE or equivalent A-level or equivalent Degree or higher Other/still studying/don’t know Working status Working full-time Working part-time Not working Don’t know/refused
199 (16.6) 475 (39.5) 196 (16.3) 259 (21.5) 73 (6.1)
543 (45.2) 184 (15.3) 472 (39.3) 3 (0.2)
AC C
* Heavy drinkers: reported drinking more than UK government guidelines of 21 units a week for men and 14 for women over the past week; Moderate drinkers: reported drinking within guidelines over the past week; Non-drinkers: reported drinking no alcohol over the past 12 months
AC C
EP
TE D
Figure 1. Example choice set.
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
RI PT
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
outcomes (normalised to scale: 10=best, 0=worst).
SC
Figure 2. Utility of each policy option, modelled with and without information on
% choosing policy option over 'no change'
ACCEPTED MANUSCRIPT Complete ban of adverts Minimum Unit Price £1 Outlet reduction -40%
Partial ban of adverts Minimum Unit Price 70p Outlet reduction -20%
Self-regulation of adverts Minimum Unit Price 40p Outlet reduction -10%
100% 90% 80% 70%
RI PT
60% 50% 40% 30% 20%
0% No outcomes
Drinkers only
Admissions only
Crimes only
SC
10% Drinkers + Admissions
Drinkers + Crimes
Admissions + Drinkers + Crimes Admissions + Crimes
M AN U
Outcomes modelled (reduction in …)
AC C
EP
TE D
Figure 3. Proportion predicted to choose each policy option over status quo, by outcomes modelled.
ACCEPTED MANUSCRIPT Crimes only
Drinkers + Drinkers + Drinkers + Admission Admission Admission Crimes s + Crimes s s + Crimes 72% 73% 77% 56% 56% 59% 56% 53% 54%
AC C
EP
TE D
M AN U
SC
Policy Complete ban of adverts 66% 69% 67% 69% 70% Partial ban of adverts 71% 72% 71% 72% 73% Self-regulation of adverts 77% 77% 77% 77% 77% Minimum Unit Price 43% £1 49% 50% 51% 55% Minimum Unit Price 50% 70p 52% 53% 54% 55% Minimum Unit Price 58% 40p 59% 59% 58% 59% Outlet reduction -40% 49% 51% 50% 54% 52% Outlet reduction -20% 49% 50% 50% 52% 51% Outlet reduction -10% 52% 53% 52% 53% 53% To resize chart data range, drag lower right corner of range.
70% 73% 77% 58% 57% 59% 55% 52% 54%
73% 74% 77% 63% 59% 60% 57% 53% 54%
RI PT
Admission s only
No Drinkers outcomes only
Complete ban of adverts Self-regulation of adverts Minimum Unit Price 70p Outlet reduction -40%
Partial ban of adverts Minimum Unit Price £1 Minimum Unit Price 40p Outlet reduction -20%
100% 90% 80% 70%
RI PT
60% 50% 40% 30% 20% 10% 0%
Non-drinkers, no outcomes
Non-drinkers, all outcomes
Moderate drinkers, no outcomes
Moderate drinkers, all outcomes
SC
% choosing policy option over 'no change'
ACCEPTED MANUSCRIPT
Heavy drinkers, no outcomes
Heavy drinkers, all outcomes
M AN U
Drinker categorisation
AC C
EP
TE D
Figure 4. Proportion predicted to choose each policy option over status quo, by drinker categorisation.
AC C
EP
TE D
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT Policy Non-drinkers, Non-drinkers, Moderate no outcomes drinkers, Moderate all outcomes drinkers, Heavy no drinkers, outcomes Heavyall drinkers, outcomes no outcomesall outcomes Complete ban of adverts 89% 92% 66% 75% 58% 66% Partial ban of adverts 89% 91% 71% 75% 63% 67% Self-regulation of adverts 95% 95% 76% 76% 69% 70% Minimum Unit Price 89% £1 93% 42% 59% 33% 45% Minimum Unit Price 89% 70p 92% 49% 59% 40% 49% Minimum Unit Price 89% 40p 90% 59% 61% 50% 52% Outlet reduction -40% 89% 92% 48% 59% 39% 48% Outlet reduction -20% 89% 91% 49% 54% 40% 45% Outlet reduction -10% 89% 90% 51% 54% 42% 45%
ACCEPTED MANUSCRIPT Highlights Novel UK-representative study predicting the acceptability of alcohol interventions
•
Acceptability varied by policy: regulating advertising was the most popular option
•
Acceptability increased for all policies with information on expected effectiveness
•
Information on expected effectiveness changed the order of preference for policies
AC C
EP
TE D
M AN U
SC
RI PT
•
ACCEPTED MANUSCRIPT Supplementary materials: Public acceptability of population-level interventions to reduce alcohol consumption: a discrete choice experiment
Contents
RI PT
A. Producing estimates of outcomes ................................................................................................ 2 Table S1. Estimates of the effects of minimum unit pricing on heavy drinkers. ............................ 3 Table S2. Estimates of the effects of minimum unit pricing on alcohol-related hospital admissions and crimes. ................................................................................................................... 3
SC
Table S3. Studies considered in Purshouse et al (2009) to estimate the effect of a 10% reduction in alcohol outlets............................................................................................................................. 4 Table S4. The effect of MUP on overall consumption vs. hazardous and harmful consumption... 4
M AN U
Table S5. Estimates of the distribution of mean reduction in consumption across hazardous and harmful drinkers for outlet density reductions. ............................................................................. 5 Table S6. Estimates of absolute reductions in the numbers of hazardous and harmful drinkers for outlet density reductions. ......................................................................................................... 6 Table S7. Estimates of the effects of outlet density reductions on alcohol-related crimes and hospital admissions. ........................................................................................................................ 6
TE D
Table S8. Estimates of the distribution of mean reduction in consumption across hazardous and harmful drinkers for advertising regulations. ................................................................................. 7 Table S9. Estimates of absolute reductions in the numbers of hazardous and harmful drinkers for advertising regulations. ............................................................................................................. 7
EP
Table S10. Estimates of the effects of advertising regulations on alcohol-related hospital admissions and crimes. ................................................................................................................... 8 Table S11. Estimated reductions per 100,000 people per year in each domain used in the study 8 B: Design of the choice sets............................................................................................................... 9
AC C
Table S12. The attributes and levels used in the study. ................................................................. 9 Table S13. Use of a random draw to determine the order of presentation of the policy options. ...................................................................................................................................................... 10 C. Estimation of the discrete choice models .................................................................................. 11 Figure S1. Contribution to utility of intervention by alcohol-related hospital admissions........... 13 D. Results of analyses by demographic characteristics .................................................................. 15 Table S14. Revised choice model taking into account differences by sub-groups ....................... 15 References ........................................................................................................................................ 16
1
ACCEPTED MANUSCRIPT A. Producing estimates of outcomes
We have estimated the effects for England, using as our key references two publications produced by the Sheffield group:
RI PT
Purshouse et al (2010). Estimated effect of alcohol pricing policies on health and health economic outcomes in England: an epidemiological model. Lancet 2010; 375: 1355–64. Purshouse et al (2009). Modelling to assess the effectiveness and cost-effectiveness of
SC
public health related strategies and interventions to reduce alcohol attributable harm in England using the Sheffield Alcohol Policy Model version 2.0. Report to the NICE Public
M AN U
Health Programme Development Group
Those studies cover all three types of policies (minimum unit pricing, outlet density reduction, advertising regulation), but they do not cover all three levels of each of the three policies. In particular, there is no estimate of a £1 minimum unit price (MUP), and there is no
TE D
estimate in the literature for 20% or 40% reductions in outlet density, nor for self-regulation or partial bans on advertising. Plus in some cases, Sheffield uses other metrics than we require here: Purshouse et al (Purshouse et al. 2009; Purshouse et al. 2010) present the
EP
outcomes in terms of reduction in consumption rather than reduced prevalence of types of drinkers. Hence the need to add further assumptions. In what follows, we describe how we
AC C
derived the relevant values for each of the policies.
Estimating the effects of MUP: Purshouse et al (2010) provide estimates for the “full effect” (that would be reached at 10 years post implementation) resulting from a wide range of pricing policies on alcohol consumption (and other outcomes). Our estimates for heavy drinkers were derived as shown in Table S1, and estimates for hospital admissions and crimes were derived as shown in Table S2: 2
ACCEPTED MANUSCRIPT Table S1. Estimates of the effects of minimum unit pricing on heavy drinkers.
Baseline harm Percentage reduction in mean consumption per week (units) Absolute reductions **
40p 70p £1* 40p 70p £1*
Hazardous drinker 8,300,000 1.8% 18.5% 51.5% 149,400 489,700 1,535,500
Harmful drinker 2,900,000 4.5% 24.1% 55.3% 130,500 298,700 698,900
Total (hazardous+harmful) 11,200,000
279,900 788,400 2,234,400
RI PT
MUP
* Purshouse et al (Purshouse et al. 2009; Purshouse et al. 2010) did not model the effect of MUP=£1. We have estimated the associated outcomes according to a polynomial function, which resulted from plotting several
SC
MUPs against the estimated % reductions in Purshouse et al (Purshouse et al. 2010). The regression line gave a fit of R2=0.99 and thus allowed us to extrapolate beyond the MUPs in the article.
in number of hazardous/harmful drinkers.
M AN U
** Assuming that the percentage reduction in mean consumption translates into the same percentage reduction
Table S2. Estimates of the effects of minimum unit pricing on alcohol-related hospital
Hospital admissions
AC C
Crimes
1 0.7 0.5 0.4 0.3 0.2 1 0.7 0.6 0.5 0.4
EP
MUP
TE D
admissions and crimes.
Reduction per year 561,498 226,400 92,200 37,800 8,900 900 236,550 123,700 83,200 42,500 10,100
Reduction per year in a community of 100,000 1091 440 179 73 17 2 460 240 162 83 20
Note: the available estimates from Purshouse et al (2010) related to the MUPs from 0.2-0.7. We extrapolated the effect of the MUP=£1 from the polynomial function relating MUP to reductions in hospital admissions for the existing estimates and from the linear function between MUP and reduction in crimes. Both functions had a near perfect fit (R2>0.99).
3
ACCEPTED MANUSCRIPT Estimating the effects of outlet density reductions: The only available estimate from Purshouse et al (2009) related to a 10% reduction in outlets:
RI PT
we took the median estimate from the studies considered (Table S3):
Table S3. Studies considered in Purshouse et al (2009) to estimate the effect of a 10% reduction in alcohol outlets.
Effect on units of consumption (%)
SC
-2.3 -3.7 -0.3 -1 -1.9 -1.9
Effect on crime (in 1,000s)
M AN U
Blake & Nied (1997), model 3 Gruenewald et al (1993) Hoadley et al (1984) Schonlau et al (2008) Xie et al (2000) Median estimate
Effect on hospital admissions (in 1,000s) -24.6 -43.5 -3.2 -11.9 -22.6 -22.6
-60.9 -83.9 6.2 -22.8 -43.4 -43.4
Purshouse et al (Purshouse et al. 2009; Purshouse et al. 2010) present the effect of outlet density reduction on mean reduction in alcohol units consumed, but not on hazardous
TE D
and harmful drinking specifically, which was the focus of our study. In the absence of any further benchmarks, the assumption we adopted in transforming the relevant outcomes was that the distribution of the response (across different types of drinkers) to changes in outlet
EP
density is the same as that to MUP policies. Unfortunately, however, that distribution differs
AC C
between the pricing policies, e.g. between MUP=40p and MUP=70p (see Table S4).
Table S4. The effect of MUP on overall consumption vs. hazardous and harmful consumption (according to Purshouse et al (2009)).
MUP
40p 70p
Changes in overall mean consumption
2.40% 17.50%
Changes in hazardous consumption
1.80% 18.50%
Changes in harmful consumption
4.50% 24.10%
Changes in hazardous consumption as proportion of overall change 75% 105.7%
Changes in harmful consumption as proportion of overall change 187.5% 138%
4
ACCEPTED MANUSCRIPT Table S5 provides our estimates on how the mean reduction in consumption is distributed across hazardous and harmful drinking. For a fairly “low” intensity policy (assumed to be 10-20% outlet reduction), we assumed the same distribution (i.e. 75% of the mean goes to hazardous, 187.5% to harmful drinking) as in the fairly low intensity MUP
RI PT
interventions (assumed to be a 40p MUP). For higher intensity policies (assumed to be >20% outlet reduction) we assumed the same distribution as that for the high intensity MUP
intervention (MUP=70p). We then converted the percentages in Table S5 into absolute
SC
reductions (Table S6) (which we then converted into equivalent figures for a community of 100,000 - not shown here).
M AN U
We also had to make assumptions about the effects across all relevant outcomes resulting from outlet reductions higher than 10% (the only estimate available from Purshouse et al (2009)). We assumed a linear effect, to err on the cautious side. The alternative would have been a non-linear effect, as is the case for pricing policies, where the marginal effect
TE D
increases with the level of MUP.
Table S5. Estimates of the distribution of mean reduction in consumption across hazardous and
EP
harmful drinkers for outlet density reductions.
AC C
Outlet Mean alcohol Reduction in density consumption hazardous reduction reduction (%) drinkers (%) 10% 1.9 1.4 20% 3.8 2.9 30% 5.7 4.3 40% 7.6 8.0 50%
9.5
10.0
Reduction in harmful drinkers (%) 3.6 7.1 10.7 10.5 13.1
Assumptions
75 and 187.5% distribution 75 and 187.5% distribution 75 and 187.5% distribution 105.7 and 138% distribution 105.7 and 138% distribution
5
ACCEPTED MANUSCRIPT Table S6. Estimates of absolute reductions in the numbers of hazardous and harmful drinkers for outlet density reductions.
10% 20% 30% 40% 50% Baseline number of drinkers*
Reduction in hazardous drinkers 118,275 236,550 354,825 666,846 833,557 8,300,000
Total 221,588 443,175 664,763 970,368 1,212,960 11,200,000
SC
* Purshouse et al (2010)
Reduction in harmful drinkers 103,313 206,625 309,938 303,522 379,403 2,900,000
RI PT
Outlet density reduction
The effect on crimes and on hospital admissions “only” had to rely on the assumption of the
crimes/admissions (Table S7).
M AN U
linear effect, i.e. doubling the reduction resulted in double the reduced number of
Table S7. Estimates of the effects of outlet density reductions on alcohol-related crimes and
Absolute reduction in hospital admissions 22,600 45,200 67,800 90,400 113,000
AC C
10% 20% 30% 40% 50%
Absolute reduction in crimes 43,400 86,800 130,200 173,600 217,000
EP
Outlet density reduction
TE D
hospital admissions.
Estimating the effects of advertising regulation: The only directly related, available estimate in the Sheffield modelling results comes from a total ban effect. Purshouse et al (2009) emphasise that there are only two underlying studies, one of which finds a counter-intuitive alcohol consumption increasing effect and the other finding a rather notable consumption reducing effect. For our estimates we take the mean estimate of the two existing estimates Purshouse et al (2009) refer to (i.e. 12.5% reduction in mean consumption). Again we need to assume how the change in mean consumption is 6
ACCEPTED MANUSCRIPT distributed between harmful and hazardous drinking. Those assumptions are given below in Table S8 and again emulate the distribution that was estimated for the MUP interventions. -
For the partial ban effect, in the absence of any available estimates, we assumed a 50% lower effect than in the case of the total ban. For the self-regulation option, in the absence of any available estimates, we assume
RI PT
-
10% of the estimated total ban effect.
These percentages were then transformed into absolute reductions for the population of
SC
England (Table S9).
M AN U
Table S8. Estimates of the distribution of mean reduction in consumption across hazardous and harmful drinkers for advertising regulations. Advertising regulation
Change in mean consumption High Low Mean estimate estimate
Selfregulation Partial ban Complete ban
NA
NA
NA -29.90%
NA 4.90%
TE D
-1.3%
Change in hazardous drinkers (%) -0.94%
-6.3% -12.50%
-4.69% -13.2%
Change in harmful drinkers (%) -2.34%
Assumptions as to how to distribute the mean effect to hazardous and harmful consumption 75 and 187% distribution
-11.72% -17.1%
75 and 187% distribution 105.7 and 138% distribution
EP
Table S9. Estimates of absolute reductions in the numbers of hazardous and harmful drinkers for advertising regulations.
Hazardous drinkers
AC C
Advertising regulation Baseline harm* Self-regulation Partial ban Complete ban
8,300,000 -77,813 -389,063 -1,096,638
Harmful drinkers 2,900,000 -67,969 -339,844 -496,625
Total 11,200,000 -145,781 -728,906 -1,593,263
* Purshouse et al (2010)
To estimate the effect on admissions and crime we proceed similarly and obtain the following effects, in terms of absolute numbers for England (Table S10).
7
ACCEPTED MANUSCRIPT Table S10. Estimates of the effects of advertising regulations on alcohol-related hospital admissions and crimes.
Hospital admissions Crimes
High estimate
Low estimate
-279,900
60,800
-238,200
44,600
Mean -10,955 -54,775 -109,550 -9,680 -48,400 -96,800
RI PT
Advertising regulation Self-regulation Partial ban Complete ban Self-regulation Partial ban Complete ban
SC
Final Estimates: Table S11 shows the final estimates of reductions per 100,000 people per
M AN U
year in each domain used in the study.
Table S11. Estimated reductions per 100,000 people per year in each domain used in the study
(intervention and intensity on which each estimate is based in parentheses*)
19
AC C
2 20 3 84 4 94 5 161 6 188 7 240 8 296 9 460 Status quo 1000
Alcohol-related hospital Heavy drinkers admissions (RA: Self21 (RA: Self272 (RA: Selfregulation) regulation) regulation) (MUP: 40p) 44 (OR: 10%) 414 (OR: 10%) (OR: 10%) 73 (MUP: 40p) 544 (MUP: 40p) (RA: Partial ban) 87 (OR: 20%) 827 (OR: 20%) (OR: 20%) 106 (RA: Partial ban) 1360 (RA: Partial ban) (RA: Complete ban) 169 (OR: 40%) 1532 (MUP: 70p) (MUP: 70p) 213 (RA: Complete ban) 1654 (OR: 40%) (OR: 40%) 440 (MUP: 70p) 2721 (RA: Complete ban) (MUP: £1) 1091 (MUP: £1) 4342 (MUP: £1) 1600 22000
TE D
Alcohol-related crimes
EP
Outcome level 1
MUP: Minimum unit pricing; OR: Outlet reduction; RA: Regulating advertising * Estimates for each domain were drawn independently from any of the nine possible levels of each outcome, and were not constrained to the most likely estimates
8
ACCEPTED MANUSCRIPT B: Design of the choice sets
The first attribute in Table S12 was used to eliminate one of the three policy interventions from the choice set. In addition, a random draw was used to specify which order to present
any potential ordering bias (see Table S13).
Attributes
Number Levels of levels Minimum unit pricing; Outlet 3 reduction; Regulating advertising 3 40p; 70p; £1.00
EP
TE D
M AN U
Which intervention type does not appear in the choice set Intensity level of Minimum unit pricing Reduction in number of crimes for Minimum unit pricing Reduction in number of hospital admissions for Minimum unit pricing Reduction in number of heavy drinkers for Minimum unit pricing Intensity level of Outlet reduction Reduction in number of crimes for Outlet reduction Reduction in number of hospital admissions for Outlet reduction Reduction in number of heavy drinkers for Outlet reduction
SC
Table S12. The attributes and levels used in the study.
RI PT
the remaining two policy interventions (and their associated outcomes) in order to control for
Intensity level of Regulating advertising
AC C
Reduction in number of crimes for Regulating advertising Reduction in number of hospital admissions for Regulating advertising Reduction in number of heavy drinkers for Regulating advertising
9
19; 20; 84; 94; 161; 188; 240; 296; 460
3
21; 44; 73; 87; 106; 169; 213; 440; 1091 272; 414; 544; 827; 1360; 1532; 1654; 2721; 4342 10%; 20%; 40%
9
19; 20; 84; 94; 161; 188; 240; 296; 460
9
9
9 9 3 9 9 9
21; 44; 73; 87; 106; 169; 213; 440; 1091 272; 414; 544; 827; 1360; 1532; 1654; 2721; 4342 Self-regulation; partial ban; complete ban 19; 20; 84; 94; 161; 188; 240; 296; 460 21; 44; 73; 87; 106; 169; 213; 440; 1091 272; 414; 544; 827; 1360; 1532; 1654; 2721; 4342
9
ACCEPTED MANUSCRIPT Table S13. Use of a random draw to determine the order of presentation of the policy options.
0
No change
Outlet density reduction
0
No change
Regulating advertising
0
No change
Minimum unit pricing
1
No change
Outlet density reduction
1
No change
Regulating advertising
1
No change
Alternative B
Alternative C
Outlet density reduction Minimum unit pricing Minimum unit pricing Regulating advertising Regulating advertising Outlet density reduction
Regulating advertising Regulating advertising Outlet density reduction Outlet density reduction Minimum unit pricing Minimum unit pricing
AC C
EP
TE D
M AN U
Minimum unit pricing
Resulting design
RI PT
Random Alternative draw A
SC
Which intervention type does not appear in the choice set
10
ACCEPTED MANUSCRIPT C. Estimation of the discrete choice models
The discrete choice models developed are based on random utility theory and aim to replicate the stated choices made during the study. The models are based on the principle that
RI PT
each respondent acts to maximise their utility, i.e. chooses the policy alternative which they believe is the best of those on offer. Each choice alternative, i, in a discrete choice model is specified with a utility function for respondent, n, such that:
SC
U ni = Vni + ε ni
Vni is the systematic or measurable utility that the individual n receives from adopting choice
M AN U
i. In addition to these observed components, the utility functions contain error terms εni that account for the unobserved components of utility.
McFadden (1974) showed that working from the assumption that the individual is a utility maximiser when considering the alternatives, j, available to them, i.e.
TE D
Pni = Prob (V ni + ε ni > V nj + ε nj ∀ j ≠ i )
and that the error terms εni are assumed to be independent and identically distributed (iid)
EP
with a Gumbel distribution, that it is possible to derive a succinct closed form expression for the logit choice probability:
AC C
1
Pni
λ
Vni
= e ∑ eλ
1
Vnj
j
λ is the scale parameter and is inversely proportional to the standard deviation. Under the iid
assumption λ can be fixed (i.e. normalize standard deviation and variance). Relaxing this assumption and making scale a function of covariates we are able to characterize how the noise (i.e. unobserved component) in individuals' choices varies by individual characteristics.
11
ACCEPTED MANUSCRIPT This formulation is known as the heteroskedastic conditional logit (DeShazo and Fermo 2002). As there is an assumption of independence between observations1, the likelihood function is given by the product of the model probabilities that each individual chooses the
likelihood. Development of the specification of the utility functions:
RI PT
option that they are actually observed to select; and the models are fit through maximum
The models have been developed by exploring a range of different specifications of
SC
the utility equations used to represent the factors influencing the choice of each alternative.
M AN U
Firstly, the extent to which outcomes were valued linearly was tested. This was undertaken by defining the levels on each of the outcome attributes as a series of dummy terms within the model and plotting these estimates, along with their confidence intervals, to reveal the extent to which changes within the outcome level appeared to be valued linearly or
TE D
otherwise. For those cases where we believed that the valuations could be non-linear, piecewise linear specifications were tested to allow a point of inflexion. The preferred formulation was determined on the basis of tests of model fit, using the likelihood ratio test to
EP
compare alternative specifications to a linear specification.
AC C
The one outcome attribute for which a piecewise linear formulation was found to be preferred was the number of hospital admissions. For this we found that the parsimonious model providing the best model fit was one which included a point of inflexion at 213 per 100,000 alcohol-related hospital admissions; with the function showing gains in utility up until this point and then a plateau from thereon, suggesting no further value being placed on higher levels of outcome (see Figure S1).
1
An ex-post correction for the violation of this assumption was made through bootstrapping the models, sampling individuals rather than choice observations, to correct for any specification error introduced through the consideration of multiple responses from each respondent.
12
SC
RI PT
ACCEPTED MANUSCRIPT
M AN U
Figure S1. Contribution to utility of intervention by alcohol-related hospital admissions.
Secondly, we systematically compared model forecasts with observed choices to test whether respondent characteristics affected how participants valued certain policies or
TE D
outcomes. For cases where it appeared that the model did not accurately reproduce the observed choice behaviour, we then introduced additional covariates into the model to test whether the previous model had been underspecified and whether the new covariates were
EP
statistically significant. These tests were undertaken in a systematic manner to review the potential influence of each of the key demographic characteristics available within the
AC C
dataset. From these tests we identified that the respondent’s existing level of drinking played a major role in the valuation of the underlying attractiveness of each of the policies. We further observed statistical differences within the moderate drinkers group, with differences in propensity to choose the “no change” option by gender, social grade, and level of weekly alcohol consumption, and differences in the value placed on reducing the number of heavy drinkers according to gender.
13
ACCEPTED MANUSCRIPT Finally, correlations between alternatives were examined, to look for different substitution patterns. A nested logit formulation was tested which allowed different substitution rates between the two policy options and the “no change” option. However, the likelihood ratio test showed that this model was not found to provide a statistically better fit
AC C
EP
TE D
M AN U
SC
RI PT
to the data.
14
ACCEPTED MANUSCRIPT D. Results of analyses by demographic characteristics
Table S14. Revised choice model taking into account differences by sub-groups
Estimate
-0.4125 -0.6928 -0.3080 -0.4013 -0.4381 0.8043 0.5480 0.3151 -2.1023 0.2449 -0.5017 -0.4273
-0.5247 to -0.3003 -0.8330 to -0.5526 -0.4370 to -0.1791 -0.5386 to -0.2641 -0.5411 to -0.3350 0.6662 to 0.9424 0.4460 to 0.6500 0.1941 to 0.4361 -4.5165 to 0.3119 0.0292 to 0.4606 -0.7392 to -0.2643 -0.6827 to -0.1719
0.0006
0.3800 to 0.8032
0.0011
1.0322 to 2.5382
-0.0011
-1.0322 to -2.5382
0.0001
0.9398 to 3.0329
0.4973 1.0000 1.3468
0.1943 to 0.8003 n/a 1.0886 to 1.6050
M AN U
SC
Policies b Minimum Unit Price of 70p, drinkers Minimum Unit Price of £1, drinkers Availability -10%, drinkers Availability -20%, drinkers Availability -40%, drinkers Self-regulation on alcohol advertising, drinkers & non-drinkers Partial ban on alcohol advertising, drinkers Full ban on alcohol advertising, drinkers No change, Non-drinkers No change, Moderate drinkers – male No change, Moderate drinkers – social grade A, B, C1 No change, Moderate drinkers – drink less than once a week
TE D
Outcomes (Reduction in alcohol related …) Crimes (per crime per 100,000 people) Hospital admissions (per admission per 100,000 people), Up to 213 admissions c Hospital admissions (per admission per 100,000 people), Beyond 213 admissions c Heavy drinkers (per drinker per 100,000 people), Moderate drinkers - female d
EP
Model scale (error) Non-drinkers Moderate drinkers Hazardous and harmful drinkers
a
AC C
Summary statistics Observations Final Log Likelihood Rho²(0) Rho²(c)
95% CI a
RI PT
Coefficient
10818 -10605.5 0.108 0.072
Confidence intervals estimated from bootstrapped model Base is “no change” and MUP 40p for drinkers, and all policies other than self-regulation for non-drinkers, specified as such as differences between these were not found to be statistically significant c Admissions is added a piece-wise linear function, with no statistically significant value placed on improvements beyond 213 admissions per 100,000 people d All other categories were absorbed into the base as they were not-significantly different b
15
ACCEPTED MANUSCRIPT References Blake, D., and A. Nied. 1997. The demand for alcohol in the United Kingdom. Applied Economics 29: 1655-72. DeShazo, J.R., and G. Fermo. 2002. Designing Choice Sets for Stated Preference Methods: The
RI PT
Effects of Complexity on Choice Consistency. Journal of Environmental Economics and Management 44: 123-43.
Gruenewald, P.J., W.R. Ponicki, and H.D. Holder. 1993. The Relationship of Outlet Densities to Alcohol Consumption: A Time Series Cross-Sectional Analysis. Alcoholism: Clinical and
SC
Experimental Research 17: 38-47.
Hoadley, J.F., B.C. Fuchs, and H.D. Holder. 1984. The effect of alcohol beverage restrictions on consumption: a 25-year longitudinal analysis. American Journal of Drug & Alcohol Abuse 10:
M AN U
375-401.
McFadden, D. 1974. The measurement of urban travel demand. Journal of public economics 3: 30328.
Purshouse, R.C., A. Brennan, N. Latimer, Y. Meng, R. Rafia, R. Jackson, and P.S. Meier. 2009.Modelling to assess the effectiveness and cost-effectiveness of public health related
TE D
strategies and intervention to reduce alcohol attributable harm in England using the Sheffield Alcohol Policy Model version 2.0. University of Sheffield. Report to the NICE Public Health Programme Development Group.
Purshouse, R.C., P.S. Meier, A. Brennan, K.B. Taylor, and R. Rafia. 2010. Estimated effect of alcohol
EP
pricing policies on health and health economic outcomes in England: an epidemiological model. The Lancet 375: 1355-64.
AC C
Schonlau, M., R. Scribner, T.A. Farley, K.P. Theall, R.N. Bluthenthal, M. Scott, and D.A. Cohen. 2008. Alcohol outlet density and alcohol consumption in Los Angeles county and southern
Louisiana. Geospatial health 3: 91-101.
Xie, X., R.E. Mann, and R. Smart. 2000. The direct and indirect relationships between alcohol prevention measures and alcoholic liver cirrhosis mortality. Journal of Studies on Alcohol and Drugs 61: 499.
16