Scottish Government

2 METHODS

This section briefly outlines the conceptual framework used as the basis of the Sheffield Alcohol Policy Model. Comprehensive details of the original version of the model (1-1), developed for the UK Department of Health in 2008, can be found in Brennan et al (2008). However the model was recently revised and extended (as version 2) in a fresh set of intervention analyses for the National Institute for Health and Clinical Excellence ( NICE) in 2009, and this later version is used as the basis for the Scottish adaptation (Purshouse et al, 2009). An overview of the differences between version 1-1 and version 2 are provided here. Details of all Scotland-specific adaptations are also described. The section concludes with an itemisation of the set of policies analysed using the Scottish adaptation, in terms of both baseline analyses and sensitivity analyses.

2.1 CONCEPTUAL FRAMEWORK

A conceptual framework for modelling interventions aimed at reducing levels of alcohol misuse is shown in Figure 2.1. At its most fundamental, the conceptual framework has two components:

1. The impact of an intervention on patterns of alcohol consumption at a population level

2. The impact of changes to such patterns of alcohol consumption on societal outcomes.

This is a suitable framework for representing the impact of policies which aim to reduce harmful outcomes through reductions in alcohol consumption (such as the pricing policies considered here). It is less appropriate for policies which may reduce harm without necessarily reducing consumption, such as staggering closing times for on-licensed premises.

Figure 2.1: High-level conceptual framework

In this study, the first component of the conceptual model is extended further, as shown in Figure 2.1, to consider how interventions affecting alcohol pricing and price-based promotions lead to a change in price, and how the change in price leads to a change in consumption. Other causal pathways (such as the psychology of 'getting a deal') are not explicitly represented.

The spectrum of societal outcomes to be considered by the model depends on the adopted perspective. The original study for DH considered a range of health, crime and workplace outcomes (both to individuals and to institutions in the public and private sector), based on the Cabinet Office (2003) assessment of the costs of alcohol misuse in England, together with a set of other outcomes (consumer spending, industry revenue, government revenue) that are not part of a traditional economic analysis. Other impacts, such as transitional costs to industry, lost welfare to the drinker, and outcomes for the family and friends of dependent drinkers were considered out of scope. This perspective is retained in the Scottish analysis.

2.2 SHEFFIELD ALCOHOL POLICY MODEL STRUCTURAL ASSUMPTIONS

The conceptual model described above is implemented using two distinct modelling methodologies:

• An epidemiological model of the relationship between consumption and health, crime and workplace harmful outcomes (known as the 'consumption-to-harm' model)
• An econometric model of the relationship between price and consumption (known as the 'price-to-consumption' model).

The two models are described in more detail below. Note that some of the text and schematics in this section have been extracted from Brennan et al (2008) and Purshouse et al (2009).

2.2.1 Modelling the relationship between consumption and harm

The model relates changes in the prevalence of alcohol consumption to changes in the prevalence of experiencing harmful outcomes. Risk functions relating consumption (however described) to level of risk are the fundamental components of the model.

2.2.1.1 Alcohol-attributable fractions and potential impact fractions

The methodology is similar to that used in Gunning-Scheper's (1989) Prevent model, being based on the notion of the alcohol-attributable fraction ( AAF) and its more general form, the potential impact fraction ( PIF).

The AAF of a disease can be defined as the difference between the overall average risk (or incidence rate) of the disease in the entire population (drinkers and never-drinkers) and the average risk in those without the exposure factor under investigation (never-drinkers), expressed as a fraction of the overall average risk. For example, the AAF for breast cancer is simply the risk of breast cancer in the total female population minus the risk of breast cancer in women who have never drunk alcohol, divided by the breast cancer risk for the total female population. Thus, AAFs are used as a measure of the proportion of the disease that is attributable to alcohol. While this approach has traditionally been used for chronic health-related outcomes, such an approach can in principle be applied to other harms (not just in the health sector).

The AAF can be calculated using the following formula:

Equation 2.1: Alcohol-attributable fraction

where RR_i is the relative risk of exposure to alcohol at consumption state i, p _i is the proportion of the population exposed to alcohol at consumption state i, and n is the number of consumption states.

If the reference category is abstention from alcohol then the AAF describes the proportion of outcomes that would not have occurred if everyone in the population had abstained from drinking. Thus the numerator is essentially the excess expected cases due to alcohol exposure and the denominator is the total expected cases. In situations where certain levels of alcohol consumption reduce the risk of an outcome (eg. coronary heart disease) the AAF can be negative and would describe the additional cases that would have occurred if everyone was an abstainer.

Note that there are methodological difficulties with AAF studies. One problem is in defining the non-exposed group - in one sense 'never drinkers' are the only correct non-exposed group, but they are rare and usually quite different from the general population in various respects. However, current non-drinkers include those who were heavy drinkers in the past (and these remain a high-risk group, especially if they have given up due to alcohol-related health problems). Several recent studies show that findings of avoided coronary heart disease risk may be based on systematic errors in the way abstainers were defined in the underlying studies. For example, Fillmore et al (2007) reanalysed data from previous studies and concluded that if ex-drinkers had been excluded from the abstainer group, then no protective effects of moderate consumption would have been observed.

The potential impact fraction ( PIF) is a generalisation of the AAF based on arbitrary changes to the prevalence of alcohol consumption (rather than assuming all drinkers become abstainers). Note that a lag may exist between the exposure to alcohol and the resulting change in risk. The PIF can be calculated using the following formula:

Equation 2.2: Potential impact fraction

where is the modified prevalence for consumption state i and state 0 corresponds to abstention.

In the model, alcohol consumption in a population sub-group is described non-parametrically by the associated observations from population surveys. For any harmful outcome, risk levels are associated with consumption level for each of the observations (note that these are not person-level risk functions). The associated prevalence for the observation is simply defined by its sample weight from the survey. Therefore, the PIF is implemented in the model as:

Equation 2.3: Potential impact fraction (as implemented in the model)

where w _i is the weight for observation i, is the modified risk for the new consumption level and N is the number of samples.

2.2.1.2 Derivation of risk functions

The impact of a change in consumption on harm was examined using four categories of risk functions:

1. Relative risk functions already available in the published literature

2. Relative risk functions fitted to risk estimates for broad categories of exposure (common for chronic health harms)

3. Relative risk function derived from AAFs for partially attributable harms

4. Absolute risk functions for wholly attributable harms.

Risk functions fitted to risk estimates for broad categories of exposure

While it may be possible to use risk estimates from broad categories of exposure assuming essentially flat relative risks across each consumption category, this does not allow the examination of the effects of relatively small shifts in patterns of consumption. Continuous risk functions were therefore fitted when risk estimates were available using polynomial curves.

One limitation of the approach is that risk estimates are available for only a few exposure groups which may underestimate or overestimate the risk beyond the last data point. This was notably the case in chronic health harms. Thus, an upper threshold was applied for conditions where the predicted estimates were unlikely to match the anticipated behaviour. Essentially, this results in a flat risk after this upper threshold. This assumption was made in the absence of consensus in the literature (Booth et al, 2008).

Deriving a relative risk function from the AAF

For some types of harms, such as crime and acute health harms, evidence is available for AAFs but not risk functions. Such evidence can be used to derive a relative risk function assuming the relationship described in Equation 2.1 since the AAF is a positive function of the prevalence of drinking and the relative risk function.

Two assumptions are necessary to compute a relative function from an AAF: (i) assumptions about the form of the curve (or risk function); (ii) assumptions about the threshold below which the relative risk is unity (ie. harm is not associated with alcohol). A linear function was selected for the analysis due to the lack of data in the literature.

The consequences of alcohol consumption tend to be distinguished in terns of those due to average drinking levels (chronic harms) and those due to levels of intoxication (acute harms). Different thresholds were thus used according to the link between harms and drinking pattern:

• The risk was assumed to start from 3 units per day for males and 2 units per day for females for harms related to mean consumption. These thresholds were derived from UK Department of Health guidelines for maximum intake (in weekly terms, 21 units for men and 14 units for women).
• The risk was assumed to start at 4 units for males and 3 units for females for harms related to peak consumption (measured as units drunk on the heaviest drinking day during the past week). These thresholds deliberately do not correspond to the intoxication definition (more than 8 and 6 units for men and women respectively) because this would imply that the risk for those drinking at the threshold would be the same as the risk of abstainers, which contradicts published evidence on acute harms. The use of 4 units for men and 3 units for women (the recommended UK Department of Health daily limits) appears a sensible choice, since it is also unlikely that the risk starts increasing from zero units of alcohol.

The resulting relative risk function is therefore a function of consumption (for which a slope is defined) and threshold as follows:

Equation 2.4: Relative risk linear function

where c = consumption level, T = threshold and ß = slope parameter.

Estimating absolute risk functions for wholly attributable harms

While it was possible to estimate relative risk functions for most harms, it was impossible to derive such functions for wholly attributable harms (with an AAF of 100%) due to the absence of a reference group.

An alternative approach was thus adopted: absolute risk functions were calculated based on the number of events, the drinking prevalence, and the total population. As for relative risk functions, assumptions were necessary about the curve form and the starting threshold. The same assumptions as for relative risks were used for consistency.

2.2.1.3 Mortality model structure

A simplified version of the model structure for mortality is presented in Figure 2.2. The model is developed to represent the population of England in a life table. Separate life tables have been implemented for males and females.

Figure 2.2: Simplified mortality model structure

The life table is implemented as a linked set of simple Markov models with individuals of age a transitioning between two states - alive and dead - at model time step t. Those of age a still alive after the transition then form the initial population for age a+1 at time t+1 and the sequence repeats.

The transition probabilities from the alive to dead state are broken down by condition and are individually modified via potential impact fractions over time t, where the PIF essentially varies with consumption (mean for chronic conditions and maximum daily for acute conditions) over time:

Equation 2.5: Potential impact fraction, as implemented in the model, showing time variation

where PIF_t is the potential impact fraction relating to consumption at time t, i = survey sample number, N = number of samples in sub-group, r _i,t is the risk relating to the consumption of survey sample i at time t, r _i,0 is the risk at baseline, and w _i is the weight of sample i.

Note that the PIF can be decomposed to enable different population groups at baseline - for example, moderate, hazardous and harmful drinkers - to be followed separately over the course of the model.

The model computes mortality results for two separate scenarios (a baseline - implemented as 'no change to consumption' in the analysis herein - and an intervention). The effect of the intervention is then calculated as the difference between the life tables of two scenarios: enabling the change in the total expected deaths attributable to alcohol due to the policy to be estimated.

Outcomes from the mortality modelling are expressed in terms of life years saved.

2.2.1.4 Morbidity model structure

A simplified schematic of the morbidity model is shown in Figure 2.3. The model focuses on the expected disease prevalence for population cohorts and as such is quite simple. Note that if an incidence-based approach were used instead then much more detailed modelling of survival time, cure rates, death rates and possibly disease progression for each disease for each population sub-group would be needed.

Figure 2.3: Simplified structure of morbidity model

The morbidity model works by partitioning the alive population at time t, rather than using a transition approach between states as previously described for the mortality model. Alive individuals are partitioned between each alcohol-related condition to be included (and an extra condition representing overall population health, not attributable to alcohol).

As in the mortality model, the PIF is calculated based on the consumption distribution at time 0 and t and risk functions. The PIF is then used to modify the partition rate (ie. the distribution across the alcohol-related conditions for alive individuals) to produce person-specific sickness volumes. These volumes then form the basis for estimating both health service costs and health related quality of life.

Quality adjusted life years ( QALYs) are examined using the difference in health-related quality of life (utility) in individuals with alcohol health harms and the quality of life measured in the general population (or 'normal health'). Utility scores usually range between 1 (perfect health) and 0 (a state equivalent to death), though it is possible for some extreme conditions to be valued as worse than death. The utility scores are an expression of societal preference for health states with several different methods available to estimate them. Note that because a life table approach has been adopted, the method to estimate QALY change for morbidity also encompasses the mortality valuation.

2.2.1.5 Time lag effects for chronic harms

For acute conditions it seems reasonable to assume that any change in consumption is immediately followed by a change in the risk of harm. However for chronic conditions this relationship may not be instantaneous: a 'time lag' may exist between change in consumption and change in risk.

Only one study (Norstrom & Skog, 2001) was identified that provided evidence on population-level time lags. The authors suggest an overall lag of 4 or 5 years (for combined chronic and acute conditions). More evidence was found concerning the time lag between onset of high levels of consumption and onset of disease in individuals, although the exact onset of such consumption is recognised as difficult to establish. The lag to full effect varies (by condition) between 5 and 15 years for most conditions; for certain cancers the lags were reported to be between 15 and 20 years. Given the lack of consensus, a mean lag of 10 years is assumed for all chronic conditions in the model with linear progression to 'full effect' on risk.

2.2.1.6 Crime model structure

The crime model considers how changes in consumption impact on changes in the volume of offences per annum, for a defined set of offence types. As for the health model, the main mechanism is the PIF, which is calculated based on the consumption distribution at time 0 and time t and an estimated risk function. The PIF is then applied directly to the baseline number of offences to give a new volume of crime for time t. The model uses the consumption distribution for the intake in the heaviest drinking day in the past week (peak consumption) since crime is assumed to be a consequence of acute drinking rather than average drinking (and so there is no time delay between change in exposure to alcohol and subsequent change in risk of committing a crime).

Figure 2.4: Simplified structure of crime model

Outcomes are presented in terms of number of offences and associated cost of crime and QALY impact to the victim. The outcomes from 'do nothing' and the policy scenario are then compared to estimate the incremental effect of the implementation of the policy.

2.2.1.7 Workplace model structure

The model focuses on two types of workplace harm: absenteeism from work and unemployment. The Cabinet Office (2003) study on the cost of alcohol-related harm also considered lost outputs due to early death; however these are excluded from the model to avoid double-counting the social value of life years lost already estimated in the health and crime models.

The absenteeism model is linked to the unemployment component in a dynamic approach (such that a change in consumption is associated with a change in the working population and thus the absenteeism in this population) as shown in Figure 2.5. Based on baseline consumption, consumption at time t and risk functions derived above, a PIF is calculated and applied to the absence rate. Absenteeism is assumed to be related to acute drinking and so maximum daily intake is applied as the consumption measure and it is assumed that there is no time delay between change in exposure to alcohol and subsequent change in risk of absenteeism. A similar approach is adopted for unemployment, although the latter is assumed to be associated with average drinking.

Figure 2.5: Simplified structure of workplace model

The number of days absent from work is then calculated based on the absence rate, the mean number of days worked and the number of working individuals in each age group/gender sub-group. Days absent from work are then valued using daily gross income.

Outcomes for two scenarios - do nothing and policy implementation - are computed separately. The difference is then taken to estimate the incremental effect of the policy.

2.2.2 Modelling the relationship between price and consumption

The pricing model uses a simulation framework based on classical econometrics. The fundamental concept is that (i) a current consumption dataset is held for the population; (ii) a policy gives rise to a mean change in price; (iii) a change in consumption is estimated from the price change using the price elasticity of demand; (iv) the consumption change is used to update the current consumption dataset. Due to data limitations, the change in levels of peak consumption has to be estimated indirectly via a change in mean consumption.

2.2.2.1 Drinking preferences for population sub-groups

The population sub-groups - defined by gender, age group and baseline consumption status - form the building blocks of the price-to-consumption model. For each sub-group, a 16 element beverage preference vector is defined. The vector describes how mean consumption is split, on average, between different categories of beverage. Beverage categories are defined by three dimensions: beverage type (ie. beer/cider, wine, spirit and RTD), retail type (ie. off-trade or on-trade) and price point (ie. higher and lower, about a threshold defined as the 25th percentile of the cumulative price distribution). Hence beverage categories range from lower-priced off-trade beer/cider through to higher-priced on-trade RTD.

For each beverage category, a detailed price distribution is defined in terms of £ per unit. Since pricing policies may affect price distributions in quite complex ways, a non-parametric representation is preferred.

Impact of a minimum price on the price distribution

For each price observation that is below the defined minimum price threshold, the price is inflated to the threshold.

Impact of a discount ban on the price distribution

For each price observation that is at a discounted price, the price is inflated to the corresponding list price. Since individual price observations are not defined as promoted or otherwise (rather this is based on separate evidence), some detailed manipulation of the distribution is required as shown below:

• For every off-trade price observation (with price P, purchase volume V and sample weight W) for beverage Y:
• Find the corresponding promotional price range R
• Look-up the proportion of sales of beverage Y in range R that are promoted (0<= d<=1, where d=0 indicates zero sales on promotion in this price range and d=1 indicates all sales are on promotion in this price range)
• If d>0, split price observation into two separate observations: { P, d* V, d* W} and { P, (1- d)* V, (1- d)* W}
• For the first observation, look-up the conditional distribution of list prices associated with promotions at this sales price [ c _R,…, c _n] where n is the total number of price ranges, where 0<= c _i<=1 with associated multipliers to list price [ m _R,…, m _n]. Split the observation into further separate observations if c _i>0
• For each new observation, i, adjust the price P to the minimum permitted price P= P* m _i
• Replace the original observation with the new set of observations in the price distribution.

2.2.2.2 Econometric model

An econometric model has been developed to examine the relationship between the purchasing of units of the 16 beverage categories, and of other non-durable goods, (on the left hand side) and their prices, the income of the individual and covariates around gender, ethnicity, age, education, region, household composition, household size and employment status (on the right hand side). The econometric model is described in more detail in Brennan et al. (2008). The resulting system of equations is analysed using iterative three-stage least-squares regression to estimate coefficients for all relevant terms. Elasticities of demand can be computed for the various products from these coefficients. In particular, a 16x16 matrix of price elasticities is obtained.

The elasticities provide information on the responsiveness of the population to price changes. They inform the scale of expected reduction in purchasing of a category of alcohol if its price changes. They also inform the knock-on effects on purchasing of other products, via the so-called 'cross elasticities' for price, enabling an assessment of the potential scale of switching to increased purchasing of a second category of alcohol (eg. lower-priced off-trade wine) if the price of the first category of alcohol (eg. lower-priced on-trade beer/cider) increases.

Elasticities can also be estimated for income, enabling an assessment of the potential change in purchasing of alcohol with changes to income.

Note that insufficient data is available to estimate elasticity matrices that are specific to Scotland. Therefore the existing elasticities for England are reused in this analysis. The matrix for moderate drinkers is shown in Table 2.1. The matrix for the aggregation of hazardous and harmful drinkers is shown in Table 2.2. The change in consumption for each beverage category for each sub-group can then be calculated using the matrices together with the changes in mean price for each beverage category faced by the sub-group.

Table 2.1: Price elasticity of demand for 16 beverage categories (moderate drinkers)

Table 2.2: Price elasticity of demand for 16 beverage categories (hazardous and harmful drinkers)

2.2.2.3 Regression model linking mean consumption to peak consumption

The Expenditure & Food Survey provides evidence on purchasing of alcohol by individuals in both the on-trade and off-trade, but does not contain a measure of binge drinking. Whilst it would seem sensible to assume that on-trade purchasing is directly associated with consumption, it is clearly not reasonable to assume that off-trade purchases are consumed on the same day and by the individual purchasing the alcohol. EFS data can therefore provide only a very incomplete picture of binge drinking, which is essentially an estimate of the extent of 'on-trade bingeing' ignoring any off-trade consumption. This has significant limitations as it is recognised that significant proportions of binge drinking occurs at home or involves a combination of both on-trade and off-trade consumption (Hughes et al, 2008). Attempts to produce on-trade binge elasticities failed due to insufficient observations in the data. Therefore it has not been possible to construct estimates of the price elasticity of bingeing behaviour (in terms of either frequency or magnitude of bingeing).

For a population survey containing data on both mean consumption and peak daily consumption, it is possible to map the scale of bingeing from the mean intake using standard statistical regression model techniques, using age and gender as covariates. Separate linear models are constructed for each drinker type due to the anticipated differences in behaviour of moderate, hazardous and harmful drinkers. The models predict the peak daily intake from the average daily intake of alcohol. The ratio of predicted peak intakes for mean consumption levels before and after an intervention are then used to adjust the actual baseline peak consumption level for each sample in the model.

2.3 Scottish adaptation

This section describes in detail the adaptations of the existing English policy model to enable estimates to be made for the population of Scotland.

2.3.1 Quantification of alcohol consumption

Population surveys provide the main approach to assessing alcohol consumption in the population of Scotland, and serve as detailed non-parametric distributions of alcohol consumption patterns in the model.

2.3.1.1 Scottish Health Survey

The Scottish Health Survey ( SHeS) is a cross-sectional household survey of around 11,500 individuals living in households in Scotland. Respondents are asked how often over the last year they have drunk each of a number of different types of drink, and how much they have "usually" drunk on any one day. The method used for calculating average weekly consumption is to multiply the number of units of each type drunk on a usual drinking day by the frequency with which it was drunk. Respondents are also asked about the number of units consumed on the heaviest drinking day in the past week. The SHeS raw data on volumes of alcohol consumption is analysed and transformed into units of alcohol consumed.

The main questions on alcohol consumption allow estimation for each individual of:

• Number of weekly units consumed (split by beer/cider, wine, spirit and RTD) - used as a proxy for average consumption
• Units consumed on the 'heaviest drinking day' during the past week - a measure of peak consumption which provides a proxy for heavy episodic drinking (also known as binge drinking)
• Detailed population distribution by characteristics such as age, sex and income.

Data has been obtained and analysed for 2003, which is the most recent year available from the UK Data Archive (Joint Health Surveys Unit et al, 2006). To take account of changes in the strength of some alcohol products, the Office for National Statistics (Goddard, 2007) undertook a review of the existing methodology for converting volumes into units in the General Household Survey and the Heath Survey for England (the SHeS uses the same methodology as the latter). These updated conversion factors have been used in analysing the SHeS 2003 data. Market research data obtained by the Scottish Government (2008b) suggests that the volume of ethanol purchasing in Scotland have changed little between 2005 and 2007, so use of the SHeS data as a baseline should be a reasonably robust assumption.

In 2003, 8,611 individuals had data for both the mean weekly consumption and the maximum consumption one day over the past week, excluding outliers (individuals with a mean weekly intake over 210 units were removed after inspection of the data). Drinkers aged 16 years old and over in Scotland had an average weekly intake of 20.6 units for males and 10.3 units for females. The numbers of units drunk on the heaviest drinking day are 7.0 and 4.1 respectively. Figure 2.6 and Figure 2.7 present the distributions of weekly and peak alcohol consumption for males and females in Scotland. The 2003 age and gender-specific distribution of alcohol consumption for adults (18+ years) in Scotland is presented in Appendix 1.

Figure 2.6: Distribution of the mean weekly intake among individuals aged 16 years old and over ( SHeS 2003)

Figure 2.7: Distribution of the maximum unit drunk one day the last week among individuals aged 16 years old and over ( SHeS 2003)

2.3.1.2 Scottish Schools Adolescent Lifestyle and Substance Use Survey

Information on childhood drinking is available from the Scottish Schools Adolescent Lifestyle and Substance Use Survey ( SALSUS), a cross-sectional school survey. Data has been extracted from the UK Data Archive for SALSUS 2006 ( BMRB, 2008). The survey covers secondary school pupils in years S2 and S4, with ages ranging from 12 to 15 years. Since the model includes 11 to 15 years old as the youngest population sub-group, the SALSUS data provides a suitable representation of this group in Scotland. The 2006 survey includes data from 23,180 pupils in Scotland. Analogous to the analysis of adults' alcohol consumption based on SHeS, updated conversion factors were also applied for alcohol consumption in children.

In 2006, the alcohol consumption questions related to:

• the frequency of drinking (from never to almost every day)
• past-week quantity consumed broken down by beverage type.

SALSUS does not cover information on peak drinking. Therefore the daily maximum consumption was estimated based on the weekly consumption of each child. The peak consumption model (see Section 2.3.2.4) was applied, assuming that the relationship between weekly and peak consumption of 11 to 15 year olds is the same as for 16 to 17 year olds.

2.3.2 Modelling the relationship between price and consumption

The existing econometric framework is reused, but with Scottish data where available. The transaction level diary data in the Expenditure & Food Survey contains detailed purchasing information for the Scottish population, although the sample size is considerably smaller than the equivalent English data (Office for National Statistics and Department for Environment Food and Rural Affairs, 2008). Market research data on price distributions and price-based promotion distributions was not available for the purposes of this study, and therefore indirect approaches were necessary in constructing Scottish price and price-based promotion distributions for the model.

2.3.2.1 Price distribution

The baseline information for the Scottish price distributions for each population sub-group in the model, broken down into beer/cider, wine, spirit and RTD in both the on-trade and off-trade, comes from the EFS. However, from the previous analysis for DH and NICE it is known that the EFS distribution differs - in some cases substantially - from gold standard price distribution data available from AC Nielsen (2008) and CGA Strategy (2009). In particular, the prevalence of very cheap alcohol is lower in the market research data than in the self-reported survey. Therefore it is prudent to adjust the raw Scottish EFS data prior to use for policy analyses.

For each beverage type in both the off-trade and on-trade, the scaling factor between the cumulative raw English distribution and the cumulative raw Scottish distribution is calculated at deciles of the distribution. The price distribution used in the existing English model (raw EFS, adjusted for both Nielsen and CGA data) is then modified using the series of scaling factors to produce a derived Scottish distribution. This latter distribution is then decomposed into the population sub-groups and used directly in the Scottish model. The before and after price distributions in both England and Scotland are shown in Figure 2.8 and Figure 2.9 for the off-trade and on-trade respectively.

Figure 2.8: Comparison of Scotland and England price distributions for off-trade beverages

Figure 2.9: Comparison of Scotland and England price distributions for on-trade beverages

Most beverage types have similar price distributions between Scotland and England based on the comparison of EFS raw price data. The prices are higher for off-trade wine and lower for off-trade RTD in Scotland compared to England. Scotland has significantly lower prices for on-trade spirit. By examining the EFS data, it can be seen that England has a higher proportion of 'spirits with mixer' (which are more expensive in terms of unit price than spirit alone) sold under the on-trade spirit category (65% for England versus 45% for Scotland) which may explain the difference.

2.3.2.2 Price-based promotion distribution

In the England model, the extent of off-trade discounts are represented as four matrices (beer/cider, wine, spirit and RTD) derived from Nielsen data (see Table 2.3 for the discount matrix for off-trade beer/cider; matrices for other beverage categories are provided in Appendix 2). The price ranges shown have inclusive lower bounds and exclusive upper bounds. Given the actual sales price, the matrix provides the distribution of the original list prices for the fraction of products on promotion. For example, of all beer/cider sold at between 25p per unit and up to (but not including) 30p per unit (fourth row of Table 2.3), 63.5% was sold at a discount. Of this promoted quantity, 30.8% had a list price in the same price bracket, whereas 34.6% had a list price in the 30p to 35p bracket, 19.9% in the 35 to 40p bracket and so on.

Table 2.3: Extent of off-trade beer/cider discounts based on Nielsen data for England & Wales (derived from data © Nielsen 2008)

As no off-trade discount information for Scotland was available to the study, we assumed that the Scottish market is characterised by the same pattern of off-trade discounting as the English market, in terms of the cumulative price distribution. This assumption is operationalised by adjusting the original set of 10 price ranges for England so that the corresponding cumulative price distribution ranges match for England and Scotland. Table 2.4 shows the original England price ranges and adjusted price ranges for Scotland. For each range shown, the lower bounds are inclusive of the price shown and the upper bounds are exclusive.

Table 2.4: Original English and mapped Scottish price ranges, used to estimate a price-based promotion distribution for Scotland

2.3.2.3 Preferences for on/off trade alcohol

The preferences for on/off trade alcohol purchasing (ie. the proportions of total consumption of each beverage that are purchased in the off-trade or on-trade) for each sub-group population are an important model input. To retain the Scottish specific preferences for on/off trade alcohol from the EFS raw data, the weights of each sample that describes the derived Scottish price distribution were adjusted so that they correctly reflected Scottish preferences.

Table 2.5: Comparison of preference for off-trade alcohol between Scotland and England

2.3.2.4 Relationship between change in mean consumption and change in peak consumption

As in the England model, a standard statistical regression model was built to map the scale of peak consumption from the mean daily alcohol consumption. Regression models are built separately for moderate, hazardous and harmful drinkers and the coefficients are presented in Appendix 3. For illustration, the three models are plotted for males aged 25 to 34 years in Figure 2.10.

Figure 2.10: Illustrative example in males aged 25 to 34 years old

2.3.3 Modelling the relationship between consumption and harm

The Scottish model uses the existing model structure (based on the potential impact fraction) and broad scope of harms, but uses a distinct set of alcohol-related health conditions and crimes, together with mortality, disease prevalence, crime, absence and unemployment rates for the Scottish population.

2.3.3.1 Health conditions in the model

The model includes 50 alcohol attributable health conditions based on those specified in the ISD-Scotland report (Grant et al, 2009) ², as shown in Table 2.6. The health conditions are classified as wholly attributable to alcohol (ie. with 100% AAFs) or partially attributable (ie. with <100% AAFs). The conditions can also be classified as chronic (due to prolonged intake of alcohol) or acute (due to acute alcohol intake).

Remarks:

V$: V12-V14 (.3 -.9), V19.4-V19.6, V19.9, V20-V28 (.3 -.9), V29-V79 (.4 -.9), V80.3-V80.5, V81.1, V82.1, V82.9, V83.0-V86 (.0 -.3), V87.0-V87.9, V89.2, V89.3, V89.9
V$$: V02-V04 (.1, .9), V06.1, V09.2, V09.3

Table 2.6: Health conditions included in the model

Compared to the England model, the health conditions of diabetes mellitus and methanol poisoning are excluded. New conditions for the Scottish model include Wernicke's encephalopathy, portal hypertension, excessive blood level of alcohol, accidental poisoning by and exposure to noxious substances, intentional self poisoning by, and exposure to alcohol, poisoning by and exposure to alcohol, undetermined intent, evidence of alcohol involvement determined by blood alcohol level and evidence of alcohol involvement determined by level intoxication.

2.3.3.2 Mortality model parameters

The mortality rates are derived from GROS 2007 data. For partially attributable chronic conditions, the relative risk functions for both mortality and morbidity are based on the same body of literature as the England model (see Table 2.6). For wholly attributable conditions (acute and chronic), absolute risk functions are estimated using the same method as for the England model (see Section 2.3.2.3 of the England report), considering the Scottish mortality and morbidity rates and the Scottish specific maximum daily (for acute conditions) or mean (for chronic conditions) drinking prevalence. For partially attributable acute conditions, relative risk functions of both mortality and morbidity are estimated applying Scottish AAFs (based on those reported by Grant et al (2009)) and Scottish specific peak drinking prevalence. The AAFs and risk functions are given in Appendix 4 and 5.

2.3.3.3 Morbidity model parameters

The morbidity rates are derived from Scottish 2007 hospitalisation data (see Appendix 4). An individual may have more than one alcohol-attributable discharge in one year, and more than one alcohol-attributable diagnosis within a discharge. The ISD-Scotland method to avoid double counting has been applied (Grant et al, 2009):

• For each individual, identify all alcohol-attributable diagnosis codes from their discharge records
• For each individual, identify the earliest hospital discharge in the year
• In the event of there being two or more alcohol-attributable diagnoses within the same discharge, select the condition with the highest position within discharge record.

The model requires inputs on costs, utilities and hospital admission multipliers for each health condition. Since Scottish-specific data is not available, the England model inputs were used. For the health conditions that exist in both the Scotland and England models, the costs, utilities and multipliers are assumed to be the same. The few Scottish health conditions that did not appear previously in the England model have been matched to similar health conditions already included in the model, following consultation with clinical experts (see Table 2.7). It is assumed that the matched conditions share the same costs, utilities and multipliers. For the new health conditions of "evidence of alcohol involvement determined by blood alcohol level" and "evidence of alcohol involvement determined by level intoxication", it is assumed that these are supplementary codes that have been used due to a lack of a primary diagnosis (eg. the person is drunk, but otherwise in normal health). Therefore, for these two conditions, the utilities are assumed to be the same as in the general population (ie. no loss of utility); the costs are assumed to be the same as A&E admission cost as per "ethanol poisoning"; and the multiplier is also assumed to be the same as ethanol poisoning. The utilities, costs and multipliers of Scottish health conditions are given in Appendix 6 and 7.

Table 2.7: Matching Scottish new health conditions with existing conditions

2.3.3.4 Crime model parameters

The definition of crime categories in Scotland is different from England. Therefore, a different set of crime categories were used in the Scottish model (see Table 2.8). Apart from the police recorded crime volumes, other crime model inputs (including multipliers, costs and AAFs) are based on English data. Matching between Scottish crime categories and existing modelled English crime categories is necessary to apply the England model inputs.

Table 2.8: Matching Scottish crime categories with existing English crime categories

The police recorded crime volumes were based on the latest data collected by the Scottish Government (Scottish Government, 2008c). The multipliers used to uplift the recorded crime volumes to actual crime volumes were based on the British Crime Survey ( BCS) 2003, following Dubourg et al (2005), and are shown in Appendix 8. The Scottish Crime and Victimisation Survey ( SCVS) was not used to derive the multipliers due to the small sample size and large confidence intervals. Figure 2.11 illustrates the confidence intervals (re-scaled to have a mean of unity) of some crime multipliers estimated based on BCS 2007/8 (Kershaw et al, 2008) and SCVS 2006 (Brown and Bolling, 2007).

As for England, the police reported crime volumes do not provide a breakdown of offences by age and gender. Therefore, the method used in the England model was adopted to split the crimes into different population sub-groups (see Section 2.6.2 of Brennan et al (2008) for further details). Appendix 9 presents the breakdown of total estimated offences by age and gender in Scotland. These raw volumes are shown graphically in Figure 2.12 and Figure 2.13. Note that a large contribution to the total volume of offences for each crime is made by males and people aged under 25.

Figure 2.11: Comparison of confidence intervals of estimated crime multipliers between BCS 2007/8 and SCVS 2006.

Figure 2.12: Estimated total crime volumes for Scotland for higher volume crime categories included in the model (greater than 100,000 offences per annum)

Figure 2.13: Estimated total crime volumes for Scotland for lower volume crime categories included in the model (fewer than 100,000 offences per annum)

The AAFs of each crime category due to alcohol consumption were estimated using the youth offending data from the 2006 Offending Crime and Justice Survey ( OCJS) - a survey of people aged from 10 to 25 living in private households in England and Wales (Home Office et al, 2008). The baseline AAFs were based on the same assumption as adopted by the England model: ie. drinking is mentioned as one of the reasons for committing the crime. Sensitivity analyses were performed using AAFs based on (1) drinking was mentioned as the only reason for committing the crime (providing a lower bound for AAFs) and (2) alcohol was consumed before committing the crime, regardless of whether or not it was mentioned as a reason for the crime (providing an upper bound for AAFs). Table 2.9 gives the three sets of AAFs, using the different assumptions, by gender and age group.

Table 2.9: Crime AAFs used in the Scottish model (derived from OCJS 2006)

The relative risk functions were estimated based on the AAFs and the Scottish peak consumption prevalence using a similar method as for acute partially attributable health conditions (see Figure 2.14 to Figure 2.17 and Appendix 10). Note that although some of the relative risks appear substantial (particularly for females), they may be associated with low absolute volumes of crime (as shown previously in Figure 2.12 and Figure 2.13).

Figure 2.14: Relative risk functions in males aged less than 16

Figure 2.15: Relative risk functions in males aged 16 to 25

Figure 2.16: Relative risk functions in females aged less than 16

Figure 2.17: Relative risk functions in females aged 16 to 25

In the Scotland model, the same source (Dubourg et al 2005) was used to extract the unit crime costs as in the England model. Costs also include the physical and emotional impacts on direct victims which are based on work by Dolan et al (2005) to obtain estimates of the quality of life impact of different crimes. For non-property crimes (eg. violence), the same assumption was used as in the England model, which values the quality-adjusted life year ( QALY) loss due to crime using £81,000 per QALY (as previously discussed with Home Office experts and based on Carthy et al (1999)). For property crimes (eg. theft and criminal damage), the Scotland model represents the physical and emotional impacts on direct victims as direct financial costs. The costs and utilities of each crime category are given in Appendix 11.

2.3.3.5 Workplace model parameters

Inputs to populate the workplace model were mainly extracted from the Scottish sample contained within the 2008 Labour Force Survey (Office for National Statistics and Northern Ireland Statistics and Research Agency, 2009): for absence rate, number of days worked, annual gross income and working rate (see Table 2.10). The participation rate was calculated using a similar definition as in MacDonald and Shields (2004), considering both the economically active and inactive population aged 16 years and over. Non-workers were derived from the following variables in the LFS: ILO unemployed and inactive. Such a definition thus takes into consideration people looking after their home families and people who are permanently sick.

Table 2.10: Workplace model inputs

Using Scotland-specific alcohol consumption prevalence (mean consumption prevalence for unemployment and peak consumption prevalence for absenteeism) from SHeS 2003, the Scotland model adopts the same method to estimate relative risk functions for unemployment and absenteeism as in the England model (see Section 2.7.1.2 and 2.7.2.2 of the England report). The relative risk functions for unemployment and absenteeism are shown in Figure 2.18 to Figure 2.21 and Appendix 12. As in the England model, the workplace model excludes people age 65 and over.

Figure 2.18: Risk functions for unemployment in males

Figure 2.19: Risk functions for unemployment in females

Figure 2.20: Risk functions for absenteeism in males

Figure 2.21: Risk functions for absenteeism in females

2.4 Policies appraised

The Scottish adaptation of the model has been commissioned to consider the impact of minimum pricing policies in isolation, a total ban on price-based promotions (ie. short-term discounting from list price) and minimum pricing policies working in tandem with a discount ban. Ten separate thresholds for a minimum price are explored (25p to 70p in steps of 5p), aiming to cover a range of levels of outcome, in terms of consumption, harm and financial impacts. 21 polices are appraised in total.

2.5 Sensitivity analysis

The analysis of pricing policies includes a set of sensitivity analyses that attempt to account for the uncertainty in the representation of both current alcohol purchasing and consumption in Scotland and how changes to price might influence consumer behaviour. Key uncertainties around the relationship between alcohol consumption and the population-level risk of coronary heart disease, and between alcohol consumption and population-level risk of crime are also explored. Descriptions of the different sensitivity analyses are provided here; for results see Section 3.3.

Sensitivity analyses included:

• Probabilistic sensitivity analysis - considers the impact of uncertainty in the parameter estimates from the econometric model, from which elasticities are derived
• Differential responsiveness of heavy drinkers - considers the implications of a what-if? scenario in which hazardous and harmful drinkers are comprehensively less responsive to price changes than moderate drinkers
• Preferences for off-trade consumption - considers the implications of using alternative evidence (to the Scottish data in the EFS) for the proportion of alcohol consumption that is purchased in the off-trade
• Protective effects of alcohol for coronary heart disease - considers the impact of using alternative risk functions for CHD, which offer increased protective benefit for some levels of alcohol consumption
• Attribution of alcohol to crime - uses different definitions of attribution to construct alternative AAF estimates and hence alternative risk functions for various types of alcohol-related crime.

2.5.1 Probabilistic sensitivity analysis

The impact of alcohol pricing policies on society is quite extensive (even within individual sectors, such as healthcare where over 50 separate conditions are considered in Scotland to be related to consumption) and as a result the model contains a large number of model parameters which must be estimated. All of these parameters are subject to uncertainty as to their true value. In this analysis, probability distributions are fitted to the core econometric elements of the overall model since the price elasticity of demand is the key active ingredient for estimating pricing policy impacts. Fitting probability distributions to all model parameters is not feasible within the scope of the current study, and is arguably not a priority since alcohol policy modelling is also subject to considerable structural uncertainty (ie. the errors that are introduced when real-world processes are represented in a mathematical model).

The three-stage least-squares regression of the system of equations used to estimate price elasticities produces a series of variance-covariance matrices. In these circumstances, assuming conditions of multivariate normality, Cholesky decomposition can be used to sample alternative parameter estimates (from which own-price and cross-price elasticities can directly be derived). The model is then re-run with the new parameter estimates to generate fresh outcomes. The process is repeated a large number of times (100 here, but ideally more) to produce a distribution of outcomes. From this, the likelihood of exceeding a particular threshold for an outcome can be estimated.

Due to time constraints, the model runs have been restricted to just consider the impact on consumption (rather than going on to consider the subsequent impact on harms) for three policy options: a 40p minimum price, an off-trade discount ban, and the combination of these two policies. Estimates of the 95% confidence interval around consumption reductions have been obtained.

2.5.2 Differential responsiveness of heavy drinkers

The differential impact of pricing policies on the consumption of moderate versus heavier (hazardous or harmful) drinkers estimated by the original Sheffield model has come under external scrutiny. In an analysis of the model methodology and results, funded by the brewer SAB Miller, the Centre for Economics and Business Research (Centre for Economics and Business Research, 2009), suggested that the implied overall elasticities for a 10% across-the-board price increase (0.35, 0.47 and 0.45 for moderate, hazardous and harmful drinker respectively - based on results from the original study) were inconsistent with other findings from the literature. This is because the results suggest that moderate drinkers are less responsive to price than heavier drinkers.

Caution is required when comparing elasticities in the literature, since the demand metric can vary between studies. This is particularly the case for the meta-analysis of elasticities for heavy drinkers conducted by Wagenaar et al (2008), where several of the elasticities in the individual studies related to the frequency or magnitude of heavy episodic drinking (or bingeing). Comparing these findings against elasticities based on mean levels of consumption may lead to invalid conclusions since the bases of demand are different. However studies do exist which suggest that price responsiveness may reduce with increasing levels of mean consumption. Manning et al (1995) identified a non-linear relationship between consumption and price elasticity, with moderate (but not light) drinkers exhibiting the greatest elasticity. However the data used to generate the estimates relates to a survey of the US population in 1983 and its relevance to England or Scotland in 2009 is open to question.

Most of the estimates available in the literature consider a limited decomposition of beverage types. These may arguably be unable to represent the heterogeneity in consumer response (for example, the most popular beverage in a country is often found to be the least price elastic) and certainly offer limited support to the requirement to understand substitution between beverage types, beverage quality, and the on-trade and off-trade. The 256-element elasticity matrix used in the model was specifically designed to facilitate such an analysis. A what-if? sensitivity analysis is considered here in which the combined hazardous and harmful drinker matrix is attenuated across all elements by comparison to the moderate drinker matrix. The Chisholm et al (2004) assumption that heavy drinkers are one third less responsive than moderate drinkers is used. The revised hazardous-harmful matrix is shown in Appendix 13.

2.5.3 Preferences for off-trade consumption

The split of consumption between off-trade and on-trade for each sub-group in the model is based on purchasing data from the EFS. For the Scottish subset of data in the survey, overall almost 73% of alcohol (measured in terms of units of ethanol) is consumed in the off-trade. Some variation exists between beverages: 47% of beer/cider is consumed in the off-trade but the corresponding figure for wine (including fortified wine) is 93%. This evidence is based on self-reported data, aggregated over the period 2001/02 to 2005/06. Alternative data for Scotland on the split - from AC Nielsen - has been made available to the research team (Scottish Government, 2008b). A comparison with the EFS data is shown in Table 2.11. Overall, the market research data (for 2007) indicates that the off-trade represents a smaller proportion of consumption overall. At beverage category level, the picture is more mixed: there is a reasonably good match for beer/cider between the two data sources, but wine and spirit off-trade preferences are lower for Nielsen, whilst RTD preferences are higher.

The reason for the discrepancy is not fully understood. The Nielsen data is based on a combination of census and survey and so could be argued to represent a gold standard. One hypothesis could be that the modelling assumption that two weeks' purchasing in the EFS is equivalent to two weeks' consumption is not always appropriate; a second hypothesis could be that off-trade purchasing is recorded more accurately (eg. via till receipts) than on-trade consumption (eg. which is subject to memory recall) and it is recognised that recall methods tend to underestimate actual consumption levels.

Table 2.11: Comparison of preferences for off-trade alcohol between EFS and Nielsen data sources

The impact of using the alternative Nielsen evidence has been tested by proportionately adjusting all sub-group off-trade preferences to reflect the alternative overall preferences shown in Table 2.11.

2.5.4 Protective effects of alcohol for coronary heart disease

There is some debate in the literature over the nature of the relationship between alcohol consumption and risk of coronary heart disease - in particular over the degree of protective benefit that might be afforded by some degree of consumption. In the basecase model, evidence from the meta-analysis by Corrao et al (2000) is used, whereby the risk function is adjusted for both gender and geographic area (Mediterranean or non-Mediterranean) ³. The non-Mediterranean version was used in the original model for England; however it could be argued that the set of countries considered 'non-Mediterranean' in the meta-analysis are not particularly representative of England or Scotland and therefore no adjustment should be made for region. Therefore a sensitivity analysis has been run using an adjustment for gender alone. The equations for alternative risk functions are shown in Table 2.12. Corresponding plots are shown in Figure 2.22.

Table 2.12: Alternative relative risk functions for coronary heart disease used in the model (derived from Corrao et al, 2000)

Figure 2.22: Alternative CHD risk functions for (a) males; (b) females

2.5.5 Attribution of alcohol to crime

The basecase model makes use of evidence from the Offending, Crime and Justice Survey to estimate the fraction of cases of various types of crime that can be considered attributable to alcohol. The survey invites respondents to state (i) why they committed a crime and (ii) whether or not they were intoxicated at the time. Respondents can select multiple reasons under part (i), which include 'don't know' and 'other'. Following criminologist expert opinion, in the basecase model attribution is assumed if the respondent selects 'drunk' as one of (possibly several) responses. As sensitivity analyses, both worst-case and best-case scenarios are also considered:

• Worst case - attribution is based on respondent selecting 'drunk' only
• Best case - attribution is based on respondent saying they were drunk at the time.