Change in alcohol outlet density and alcohol-related harm to population health (CHALICE): a comprehensive record-linked database study in Wales

Summary of cross-sectional studies

A 2009 systematic review of 44 studies of alcohol outlet density found that the most frequently investigated alcohol-related harm was violent crime, especially assault. 22 The majority of these studies were from the USA. High outlet densities were associated with higher rates of assault, self-reported injuries, motor vehicle accidents and high rates of pedestrian collisions. Higher outlet density was also shown to be associated with increased domestic violence and child abuse.

Of two additional American studies not included in the review,22 one found a cross-sectional association between higher outlet density and an increased rate of alcohol-related hospital admissions during 1 year (1996) in San Diego County, CA, USA,42 and the other, set in California and Louisiana, found an association with increased self-reported liver disease, sexually transmitted infections (STIs) and violence. 43 Very few studies have investigated alcohol consumption. In this review, two out of three studies found an association between outlet density and high levels of consumption. 22

Longitudinal studies of outlet density and alcohol-related harm

The review article of longitudinal studies, which included papers published up to 2006, discussed the results of studies investigating change in outlet density in three groups of study designs: (1) interrupted time series; (2) natural experiments from alcohol bans; and (3) changes in licensing arrangements. 41 None was set in the UK. Many of these studies were old and few accounted for spatial autocorrelation in their analyses (i.e. the lack of independence of proximate small areas). This review did not include any methodological assessment of the measurement of outlet density. The overall conclusion was that an increase in outlet density was associated with increased consumption and interpersonal violence, but the evidence was weaker for an association with motor vehicle crashes.

We found a further 18 papers23–40 that investigated associations between change in outlet density and a limited range of outcomes, namely alcohol consumption,24–27 violence,23,28–33,36,37 STIs,34,35 suicide38 and other causes of death. 39,40 Of these, seven23,24–26,33,39,40 were published after this study [change in alcohol outlet density and alcohol-related harm to population health (CHALICE)] was funded.

Outlet density, neighbourhood deprivation and population migration

Two US papers investigated the association between outlet density and neighbourhood deprivation. 44,45 Higher outlet densities were found in more deprived areas but paradoxically one study suggested that levels of consumption were highest in less deprived areas, suggesting that the mismatch between supply and demand could result in alcohol-related harm being disproportionately higher in people living in deprived neighbourhoods in proximity to alcohol outlets. 44

A study set in New Zealand also found an association between higher outlet density and areas of deprivation. 46 A cross-sectional study set in Glasgow found that the spatial association between outlet density and deprivation did not vary systematically, highlighting the importance of local context in the design of future studies. 47 Two studies, set in Australia32 and the USA,23 have found evidence that outlet density is more important in areas with high pre-existing outlet numbers and in areas of high neighbourhood deprivation.

Although it is clear from published research that selective population migration may account for a spurious widening of health inequalities over time,48,49 there are no studies that explicitly investigate this possibility for alcohol-related harm outcomes.

Measurement of outlet density

In the longitudinal studies, the primary measure of outlet density was estimated as the number of outlets per resident population in 331,33,39 of the 10 studies that estimated a density. 23,30,31,33–37,39,40 One study modelled the number of outlets with the resident population as an offset. 32 Five studies estimated outlet density as number of outlets per road-mile30,34–37 and one study estimated density per square mile. 23 Two studies compared densities estimated per road-mile with densities estimated per population and per square mile and reported no difference between the methods. 35,36 Another variant was using the distance from each residence of study participants to the nearest outlet,24–27,40 or the number of outlets per administrative area. 28,29,38

Although most cross-sectional papers used resident population as the denominator, one study set in California noted that there is no standard for measurement of outlet density and estimated outlet density using several different measures for the denominator, including resident population, geographical area (e.g. square mileage) and other more spatial measures of exposure to outlets including Euclidean distance from home to nearest outlet and number of outlets per 0.5-mile ‘buffer zone’, that is, the geographical area defined by 10–15 minutes’ walking time. 44 In this analysis there was little difference in the results for these measures. A New Zealand study used 10 minutes’ travelling time by car to define a neighbourhood area with the density measure as the number of outlets within each neighbourhood. 46 The only published UK research, from Glasgow, estimated outlet density using two different measures: outlets per 1000 residents and the mean network distance from the centroid of each small area of interest to the nearest outlet. 47

Welsh Index of Multiple Deprivation 2008

We used the Welsh Index of Multiple Deprivation (WIMD) 2008 calculated at LSOA level as the measure of neighbourhood deprivation in this study. 53 The WIMD is available for 2005, 2008 and 2011. Because each version includes a different set of variables and there are therefore differences in the index construction, the three scores are not directly comparable. We therefore chose the WIMD 2008 version as it was the version that was closest in date to the start of the study time periods. The WIMD 2008 contains eight weighted domains of deprivation (Table 2). The overall index is constructed by ranking the LSOAs in each domain from 1 to 1896, exponentiating the [transformed (0,1)] ranks, which are then summed and weighted as in Table 2 to give the final LSOA WIMD score, with high scores indicating high levels of deprivation.

The main criticism of the WIMD is that it is largely driven by data availability and subjective assessment of its quality and utility rather than any by a priori considerations (although income and employment account for nearly one-half of the final score). The inclusion of a health domain, which might be considered as outcome data and subject to circularity, in fact makes little difference in analyses of the WIMD and health outcomes. 54 The WIMD is very highly correlated with other deprivation indices such as the Townsend score55 and has the advantage of being updated regularly in the intercensal period.

Because the final WIMD scores have little intuitive meaning and their distribution is highly skewed (Figure 1), we chose to create five groups, or quintiles, of deprivation with an equal count of LSOAs in each group (Table 3). This was also necessary as an anonymised categorisation in quintiles is permissible in the Secure Anonymised Information Linkage (SAIL) Databank (see Chapter 4).

FIGURE 1.

Distribution of WIMD 2008 scores. SD, standard deviation.

Figure 2 shows a map of Wales and the LSOA WIMD 2008 quintiles.

FIGURE 2.

Map of Wales showing the LSOA WIMD 2008 quintiles. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

Rural–urban settlement type classification

We also used the 2004 rural–urban classification published by the ONS56 as the measure of settlement type for 2001 census LSOAs.

For LSOAs there are six settlement types, categorised into three groups:

• urban (population > 10,000)

• town and fringe

• village, hamlet and isolated dwellings.

Figure 3 shows a map of LSOAs in Wales with the three settlement types.

FIGURE 3.

Settlement type classification. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

Ordnance Survey MasterMap Integrated Transport Network™ layer

The ITN layer is a data set representing the road network and road routing information for Great Britain. 66 It was launched in November 2001 as part of OSMM, the OS flagship digital mapping product. The road network element of the ITN data set contains all classifications of roads, from motorways down to local streets. The road routing information contains information such as height, width and weight restrictions, traffic calming information, turn restrictions and one-way roads. Although precise details on how the software is derived are scarce because of the sensitive commercial nature, the OS website suggests that a combination of ground surveys and photogrammetry has been used. 66 Overall OSMM accuracy is quoted as being 1.0 m or better with a 99% confidence level for urban areas at a scale of 1 : 1250 and 2.5 m or better with a 99% confidence level for rural areas at a scale of 1 : 2500. 66 The ITN data in this project have been used to model the shortest routes between residences and alcohol outlet locations for two different scenarios: walking and driving.

Integrated Transport Network™ driving network

The ITN data, as supplied by the OS, contain information on road types and route restrictions, which enables the modelling of a route between two locations in a Geographic Information System (GIS). Unfortunately, average road speeds and rural–urban designation were not provided as part of the data set. To model road speeds for the network a broad estimate of road speeds was developed using a nominal road speed applied to a road section based on road classification, for example motorway [57 miles per hour (mph)], A road (57 mph or 27 mph) and B road (27 mph).

The final step was to assign a travel time to each section of road based on the estimated road speed using the road length. This was achieved by using the following conversion equation:

Refining the road network to better represent drive times in urban and rural areas

During data validation, the default drive time estimations were found to be too generous. The default network model assumes that there are no time penalties when performing turns on the network or when crossing junctions where there may be traffic lights. As a result, the distances travelled by car in 10 minutes far exceeded what was reasonable in the given time. To address this issue the underlying network model was adjusted to better reflect reality, using local knowledge of driving in rural and urban areas. Time delays at different points in the road network (e.g. road junctions, right turns and left turns) were introduced to the model to better represent a typical car journey through the network. Figure 5 (Cardiff, urban area) and Figure 6 (Pembrokeshire, rural area) illustrate how time penalties for certain road manoeuvres (e.g. turning right onto a road) and crossing junctions reduce the service area for a fixed point. As a result there was a 90% reduction in the modelled alcohol outlet availability for both urban and rural test locations.

FIGURE 5.

Drive time of 10 minutes for urban areas, showing increasing time penalty adjustments and decreasing range. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

FIGURE 6.

Drive time of 10 minutes for rural areas, showing increasing time penalty adjustments and decreasing range. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

Integrated Transport Network™ walking network

The ITN walking model primarily used the same data as the driving model except for the addition of urban paths. As the name suggests, this models urban-based footpaths in major conurbations, thus allowing the calculation of non-vehicular routes for pedestrians. The addition of these data improved the modelling for urban areas for walking, as illustrated in Figure 7, by reducing the travel time and distance from a household to an outlet through the use of footpaths. A universal speed of 5 km per hour (kph) was assigned to the walking model, as this is widely accepted as the average walking speed for an adult.

FIGURE 7.

Integrated Transport Network^TM urban paths. © Crown copyright and database rights 2013 Ordnance Survey 100050829.

AddressBase Premium

AddressBase Premium62 is the most comprehensive address data set for Great Britain and is based on three data products: the local government’s National Land and Property Gazetteer (NLPG), OSMM Address Layer 2 and the Royal Mail Postcode Address File^® (PAF). Collectively this is known as the National Address Gazetteer, from which the ABP products are extracted. ABP provides a complete view of an address or property, including its full lifecycle from build to retirement. It contains all current addresses, alternative addresses, provisional properties and historical property information when available. The data were received in comma-separated value (CSV) text file format and were processed into a database for (1) the geocoding of outlet data and (2) identifying the precise location of residential properties for the density modelling process.

This level of detail was essential so that both outlet availability and residential housing availability and locations could be modelled during the study period. By selecting residential property types, using the supplied ABP classification lookup table (which includes type and start and end dates for each address) and the start and end dates for the quarterly time periods defined for the project, we were able to model almost all of the residential housing availability for each of the quarters during the study period. There were a couple of limitations found within the data with regard to the start dates for properties when the local authority had batch processed a number of residential properties, giving them the same start date. As many of these properties as possible were investigated and when the date was found to be inaccurate it was adjusted to a more appropriate date. For example, a large number of residences in a unitary authority had a start date in late 2007 (possibly at the inception of the NLPG and ABP). On investigation it became clear from alternative data sources [e.g. Google Maps and Google Street View (Google Inc., Mountain View, CA, USA)] that the housing stock predated this date by decades and, therefore, these records were reassigned a start date prior to the study period so that these residences were included in the density calculations.

Points of Interest database

Points of Interest (POI) is a database covering all of Great Britain and containing over 4 million different geographical features. 67 The POI database includes a number of data sets. An indicative summary is supplied by the OS (Table 5).

Each record is georeferenced and classified according to the following main classifications:

• accommodation, eating and drinking

• commercial services

• attractions

• sport and entertainment

• education and health

• public infrastructure

• manufacturing and production

• retail

• transport.

Although the POI database contains information on the location of alcohol outlets we could not use it systematically because no information is provided on opening and closing dates. In addition, we found that there was under-reporting of outlets in urban areas and over-reporting of outlets in rural areas.

Georeferencing outlets

Each alcohol outlet had to be geographically located so that (1) a simple count of outlets per LSOA could be derived and (2) the distance between each residence and each outlet could be computed using a GIS for the walking and driving outlet density calculations. We had to assign an accurate (x,y) co-ordinate to each outlet by geocoding the outlet address. Geocoding is the process of matching text-based address data to known geographical co-ordinates, typically via a GIS. A number of studies have assessed geocoding techniques, positional accuracy68–70 and the impacts of match rates and positional accuracy on resulting analyses. 71–73 In particular it has been shown that, whenever possible, high-resolution address data (building level) should be employed as the reference data set, particularly when fine-scale analysis is being performed. 73 To estimate the alcohol availability for 10-minute walking and driving buffers around a residence, it was important to use the best-quality address data available for the study area.

The text-based address data received from the local authorities were parsed and broken down into address components (organisation name, building name, street number, thoroughfare, post town and postcode) as defined by the Royal Mail. 74 Each category was indexed before the reference data were searched and the best match was recorded using software written for this particular project. The ABP data set was used as the reference data set. Each successfully fully matched record was allocated the corresponding UPRN from the ABP data, as well as the corresponding (x,y) co-ordinates. The ABP data set contains a point for each residence, which is located within the footprint of the residence. The buildings are surveyed with a spatial accuracy of ±1–2 m in urban areas. For records that were not fully matched, a probabilistic matching algorithm was implemented to identify addresses based on partial address matches. For example, premises were identified from street numbers and names when a premise name could not be matched. The final automatic geocoding step filtered for matches based on the postcode provided in the licence data. In Great Britain there are, on average, 18 properties per postcode. Filtering identified the premise or estimated the location.

The results of the geocoding process varied across the 21 local authorities that supplied historical licence data. One local authority was unable to provide us with historical data – the current licence register was provided with the required geographical references. Table 6 provides an overview of the geocoding process and the percentage of outlets that we were able to geocode using the geocoding software. The average automatic ABP match rate for the local authorities in Wales was 50.5%.

The main reasons for a failure to match an address record were an absent or inaccurate premise name or street name or an absent or inaccurate postcode. For example, a premise was recorded with the following address: ‘Market Square, NPXX 4XX’; however, the correct address was in fact ‘Market Street, NPXX 4XX’. Small errors in the main address elements make it difficult for a machine to disambiguate possible matches. This is particularly true when the error in the address could partially match to an alternative address. This example illustrates this because both Market Square and Market Street exist in the locality, making identifying the correct address impossible for an automated system.

The geocoding match rate was markedly higher for records obtained from more urbanised local authorities than for records obtained from those local authorities in a more rural setting. The match rate generally improved as local authorities became more urbanised (Figure 8).

FIGURE 8.

Comparison of geocoding match rates between local authorities.

Manual georeferencing

We developed a manual procedure using Google Maps for those outlets not successfully automatically geocoded. There is a precedent for using web-based mapping technologies to fill gaps and perform neighbourhood audits. 75–77 Google Maps and Google Street View were examined for the outlet location using the information contained in the licence record, for example outlet name, street and locality. The outlet was identified as a point on Google Maps and the latitude and longitude of the outlet location were extracted using the information for the point contained in the URL. When the record was unclear because of differences in urban and rural geocoding, the appropriate street location was identified and an approximate outlet location was assigned by assessing the buildings along the street.

Location difference errors, introduced as part of the manual match process, were calculated from the two sets of co-ordinates using a Euclidean distance difference measure. The final data set contained the match type (geocoded, manually matched and approximated) recorded against each outlet. This enabled purely spatial research focused on urban areas to omit the approximated locations.

Verification of manually matched locations

As a large proportion of the licence files had to be geocoded manually it was decided to validate the methodology using known addresses that had been successfully matched as part of the geocoding process. A sample of 1604 known outlet locations – those that had been successfully matched to an ABP record – was matched manually. This resulted in a data set with two sets of co-ordinates: the known outlet location and the manually matched location. These two sets of co-ordinates allowed us to perform some basic analyses on any errors introduced as part of the manual match process. Table 7 provides a summary of the verification results and shows that the manual matching process was a reasonable method to fill in the gaps in the licensed premise locations.

TABLE 7 - Verification results

Distance	% manually matched locations within distance
100 m	79
500 m	91
800 m	93

We also compared the assignment of the correct LSOA to an outlet using the manually matched method and the geocoded method. Figure 9 shows where the manual matching process resulted in a different LSOA being assigned to the outlet. The maps show that there is a spread across urban and rural areas, with no particular error pattern being attributable to either an urban or a rural setting. The error rate for incorrect LSOA assignment was 6%.

FIGURE 9.

Comparison of geocoded with manually matched LSOA assignment. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

Methodology

The methodological steps used to create a network outlet distance measure for each residence were as follows:

1. We created the network data, residence data and outlet data GIS feature layers using the ABP data set (Figure 10a–c).

2. We loaded the residence and outlet data onto their computed network locations [the nearest network feature to the (x,y) co-ordinate of the point feature] (Figure 10d).

3. We calculated a residence–outlet matrix for each residence within the study. This calculates the distance between two network locations within a given threshold (Figure 10e).

4. Some residences and outlets were located a short distance from the official road network, for example along a private road or a long driveway. We added the distance that a residence or outlet was located from the road network to the residence–outlet matrix network distance. We then removed from the calculation any outlets that had moved beyond the 10-minute network threshold.

5. We calculated metrics required to compute the density scores for each residence and took the LSOA mean score.

FIGURE 10.

Steps in the outlet density calculation. (a) Network; (b) residential properties; (c) outlets; (d) residences and outlets; (e) residence–outlet matrix; and (f) outlets in residence–outlet matrix. © Crown copyright and database rights 2013 Ordnance Survey 100050829. Local Government Information House Limited copyright and database rights 2013 100050829.

Formulae for the outlet density computation

The method can be written as follows:

1. For each residence, j, calculate the distance (d_ij) to each outlet (O_i).

2. Define the size of the buffer in terms of a distance (β) (depending on speeds and whether walking or driving).

3. Define the access to O_i from the residence, j, as:(2)11+γdij 2k

4. if d_ij < β and 0 otherwise.

5. Define the accessibility of outlets from RALF, j, as:(3)aj=∑dij<β11+γdij 2k.

The parameters in Equations 2 and 3 can be specified as follows.

Butterworth filter (γ and κ)

The Butterworth filter has its origins in signal processing but has been used before as a filter in gravity models in accessibility research. 79 The shape of the Butterworth filter lends itself to accessibility modelling as it allows closer locations to have similar weightings up to a threshold distance before a distance decay is encountered. It was hypothesised that an individual would be prepared to travel to any outlet within about 200 m of his or her origin before distance became a motivational factor. The Butterworth filter models this initial distance more closely than other more commonly used filters (Figure 11).

FIGURE 11.

Plots of the reciprocal function, exponential function and Butterworth function with β = 2, α = 0.5 and κ = 3 respectively (for illustration).

To derive the weighting scores the distance decay was modelled at two given distances and then γ and κ were determined from Equations 4 and 5:

⁽⁴⁾ k = 0.5 × log ⁡ ( 1 − ε 2 ε 2 × ϕ 2 1 − ϕ 2 ) log ⁡ ( β α )

⁽⁵⁾ γ = 1 − ε 2 ε 2 × 1 β 2 k ,

where the accessibility was set to be ε at the edge of the buffer, at a distance β, and ϕ at some other distance, α. The values were set as 0.1 at β = 833 m and 0.6 at a distance α to obtain the values κ = 1.81 and γ = 2.577 × 10^–9 for walking calculations. For the 10 minutes of driving time, with a buffer at 10 miles, the accessibility could be set to be 0.1 when β = 10 miles and to be 0.8 when α = 1 mile. The values of κ and γ are 1.123 and 0.5625, respectively.

Calculation of outlet density

To compute a LSOA score for each quarter in the study period the mean of the density scores for every residence in a LSOA was taken. An (x,y) co-ordinate within the LSOA was then calculated based on the average density location within the LSOA to produce a density-weighted centroid to which the density scores could be attached. For the other non-density metrics shown in Table 8 (e.g. nearest outlet distance) it was decided to use a median score to aggregate to the LSOA. The median value was chosen as the method of aggregation to reduce the impact of the outliers on the variables being calculated and to give a more representative score.

Optimisation of computing power

The calculation of the density scores and other variables required a substantial amount of computing power. To efficiently manage the process and produce the results in a reasonable amount of time it was necessary to employ some advanced computing techniques. Parallel computing is the process of breaking a computational problem into chunks for processing, either on one of the processor’s cores or through the distribution of data and code across a network of computers.

It was found to be necessary to use each core of a quad core machine to achieve the required efficiency gains. The data were easily split into manageable chunks, first by geography in the form of local authorities and second by time in the form of years and quarters. Splitting the residential data resulted in two major benefits. First, the smaller chunks of data made the processing of the density scores much more manageable given the system resources and stability and, second, the chunks acted as a failsafe so that if the system crashed then it would require only the last chunk to be reprocessed rather than the whole data set.

Figure 12 illustrates the conceptual parallel processing framework that was employed. Each local authority was processed by splitting the data into yearly chunks, with each quarter for that year being processed in parallel. The parallel processing framework and refinement of the algorithms resulted in massive efficiency savings and cut the processing time down from approximately several days to 8 hours for the walking calculations and from several weeks to 3 days for the driving calculations. The difference in processing time between driving and walking scenarios can be accounted for by examining the size of the data that each outlet density matrix produced for the different scenarios. Table 8 illustrates the differences for four local authorities and how the increase in the number of calculations required in urban areas is substantially more than in rural areas.

FIGURE 12.

Parallel processing framework. Q, quarter.

When the data shown in Table 8 are aggregated for the whole study period it shows the amount of computing power required to complete the density calculations. For Cardiff alone the number of residence–outlet pairs was approximately 600 million for the whole study period. Twenty variables were calculated for each residence–outlet pair resulting in approximately 12 billion calculations for driving density and 1.3 billion calculations for walking density. For rural areas this figure dropped to a more modest 950 million and 236 million calculations respectively. The final output was a CSV text file.

Interpretation of the density score

The density scores model the impact of localised geography on potential access to alcohol outlets by taking into account the ‘attractiveness’ (the likelihood that someone would visit) of an outlet at household level. This attractiveness is modelled using a geographical gravity model, which uses the road and footpath network distances between a residence and an outlet and weights the distance using the Butterworth filter. However, variations in local geography mean that it is not straightforward to interpret the density score. Tables 9 and 10 are reference tables showing the required change in the number of outlets at a distance interval that gives a unit change in density score. The tables were created using 1000 iterations for each distance interval (e.g. 0–150 m) and averaging the score to simulate the variety of possible distances between a household and an outlet and aggregating to LSOAs. These tables are indicative only but can be used as reference tables to interpret a density score change.

Number of outlets

The total number of outlets open in Wales on 1 January 2006 was 9610, rising to 11,732 by 2011 quarter 4 (Figure 13).

FIGURE 13.

Quarterly trend in count of outlets open on the first day of the quarter, 2006–11. Q, quarter.

Figure 14 shows the highly skewed nature of the LSOA distribution of counts in 2006 quarter 1.

FIGURE 14.

Distribution of LSOA numbers of outlets open on the first day of 2006 quarter 1. SD, standard deviation.

The descriptive statistics for the number of outlets per LSOA open on the first day of a quarter are shown in Table 11 for quarter 1 of each year and Table 12 for quarter 4 of each year. Both show a similar trend of an increasing maximum number of outlets and mean number of outlets per LSOA over the time period.

Table 13 shows the descriptive statistics for the weighted number of outlets per LSOA open during quarter 4.

In 2006 quarter 4 there were 201 LSOAs with no outlets, falling to 164 by 2011 quarter 4. Around 90% of LSOAs had ≤ 10 outlets at the start of 2006 quarter 4, falling to 85% at the end of the study period. This also suggests a small increase overall in the number of outlets (Table 14).

Simple outlet density: static population and changing population

Simple outlet density was calculated as the weighted number of outlets time open ‘at risk’ (i.e. days open during the quarter or year) divided by the LSOA population aged ≥ 16 years. For ‘static population’, the denominator was the 2006 quarter 1 population, whereas for ‘changing population’ the population estimate for each successive quarter was used as the denominator. Tables 15 and 16 show the annual descriptive statistics for the static population changing population respectively. The overall trend is for greater alcohol availability over time with increasing LSOA variability.

Figures 15 and 16 show the distribution of values for the 2011 static population outlet densities and the 2011 changing population outlet densities, respectively, confirming the negative skew described in Tables 15 and 16.

FIGURE 15.

Histogram of 2011 static population outlet densities.

FIGURE 16.

Histogram of 2011 changing population outlet densities.

Figure 17 shows a map of the simple outlet density for the static population for 2011 quarter 4.

FIGURE 17.

Map of simple outlet density: static population, 2011 quarter 4. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

Figure 18 shows a map of the simple outlet density for the changing population for 2011 quarter 4.

FIGURE 18.

Map of simple outlet density: changing population, 2011 quarter 4. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

The correlation matrix in Table 17 shows a high degree of correlation between the annual static population and the changing population outlet densities.

Figure 19 shows a scatterplot of the static and changing population outlet densities for 2011. The scatterplot suggests that two of the 1896 LSOAs have an outlying difference between the static population and the changing population outlet densities. These two LSOAs are W01000743 (Castle 2, Swansea, difference = 24.2 per 1000 population) and W01001701 (Butetown 3, Cardiff, difference = 17.6 per 1000 population). The difference in values is explained by large population increases, from 1627 to 2604 in Castle 2 and from 1683 to 2917 in Butetown 3.

FIGURE 19.

Scatterplot of static and changing population outlet densities, 2011.

Overall, the differences between the static population and the changing population were small (Table 18).

Nineteen LSOAs showed a difference of more than ±2 per 1000 population between 2001 and 2006 (Table 19) and this accounts for the differences between the static population and the changing population.

Overall, there was no justifiable reason not to use the changing population in the main analyses. Comparison with the static outlet density values will provide some measure of the impact of population migration but this will effectively be limited to the effect of very few LSOAs and is unlikely to be important.

Walking outlet density

Table 20 shows the descriptive quarterly data for variation in the LSOA measure of walking outlet density.

Figure 20 shows the distribution of 2011 quarter 4 walking outlet density values.

FIGURE 20.

Histogram of 2011 quarter 4 walking outlet density values. Q, quarter; SD, standard deviation.

Figure 21 shows the LSOA spatial variation in walking outlet density.

FIGURE 21.

Map of Wales showing the LSOA spatial variation in walking outlet density. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

Driving outlet density

Table 21 shows the descriptive quarterly data for variation in the LSOA measure of driving outlet density.

Figure 22 shows the distribution of 2011 quarter 4 driving outlet density values.

FIGURE 22.

Histogram of 2011 quarter 4 driving outlet density values. Q, quarter.

Figure 23 shows the LSOA spatial variation in driving outlet density.

FIGURE 23.

Map of Wales showing LSOA spatial variation in driving outlet density. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

Table 22 shows the correlations between the measures of outlet density for the start and end density values. Clearly the three measures are measuring different aspects of alcohol availability.

Figure 24 shows the linear association between the walking and the driving outlet density measures.

FIGURE 24.

Linear association between the walking and driving outlet density measures. Q, quarter.

The essential difference between simple and walking outlet density is that urban areas tend to have more dense networks and hence higher values of walking outlet density. Figure 25 shows the difference between the mean simple and the mean walking density for each local authority and the differences are smaller, or negative, in the urban authorities. The pattern is the same for driving outlet density (not shown).

FIGURE 25.

Differences between mean simple and walking density for local authorities.

Alcohol availability process: current and historical availability

If the alcohol availability data are thought of as a vector A(t) = [a(t),. . ., a(t – 7)] (say, the LSOA outlet density over the current and previous seven quarters), we are interested in determining what function of this vector best explains alcohol consumption C(t + 1) at the next measurement occasion. It might be something simple such as ‘consumption depends only on most recent past availability’:

that is, the value for the previous quarter.

Alternatively, ‘consumption depends on the most recent and earlier availability’, perhaps through a linear combination of all elements of A(t – 1):

It can also be hypothesised that ‘consumption depends on the maximum historical availability’, which is very non-linear (and non-differentiable in places) and could be written as:

Equivalently, ‘consumption depends on the minimum historical availability’.

Finally, it can be hypothesised that ‘consumption depends on the maximum or minimum availability within the previous four quarters’.

All of these cases are functions of the history of the availability process. It will then be possible to compare the various possible models above using the above measures by assessing the best fit of the availability process.

Change in outlet density

Change in outlet density can be estimated simply by subtracting an earlier quarterly value from a later quarterly value. As the degree of change will be some function of the absolute starting values, it is sensible to calculate change relative to the baseline value. However, as inevitably there will be densities that do not change between quarters, this will provide a surfeit of zeros.

Relative change was, therefore, estimated as the absolute annual change between two successive years’ quarter 4 values divided by the mean of the five quarterly values. This resulted in a stable set of values for change. Two separate variables for change in outlet density were then derived: one variable for a zero or positive change and a second variable for a zero or negative change.

Because this method will constrain the variance to be small where the baseline and final outlet density values are large, the variance was stabilised by dividing the absolute change by the square root of the mean of the five quarterly values.

Volatility

The next question was how to derive a variable for the volatility (or equivalently the inverse of stability) of the five quarterly values. The general shapes might be a regular increase or decrease, shown as series 2 (green) and series 3 (blue) in Figure 26, or no change, shown as series 1 (black).

FIGURE 26.

Hypothetical change in quarterly outlet densities. Q, quarter.

A more complex pattern of change between quarters may exist, summarised in Figure 27, in which densities display a ‘V’ or inverse ‘V’ pattern but the change is zero in both cases.

FIGURE 27.

Hypothetical zero change in quarterly outlet densities. Q, quarter.

The pattern may also be more complex, such as a ‘M’ shape (Figure 28). The previous quarter values are 1 for black and 0 for green. Here, the relative change for the series 1 black densities is 1/2 = +0.5 and for the series 2 green densities = –1/2 = –0.5, but these values ignore the variability in quarterly measurements, given that they are equally volatile. The same would hold for an inverse ‘M’ shape.

FIGURE 28.

Hypothetical complex change in quarterly outlet densities. Q, quarter.

A ‘volatility’ variable was therefore estimated, equal to the sum of the absolute differences between the five quarters/the mean of the five quarters. So, for the black densities, volatility = (4 + 2 + 1 + 2)/2 = 4.5, and for the green densities, volatility = (2 + 1 + 2 + 4)/2 = 4.5. The variance of the volatility measure was stabilised by dividing by the square root of the mean of the five quarterly values.

This combination of three measures attempts to capture as much as possible a complex, real-world process based on administrative data.

Change in alcohol outlet density

The difference in annual outlet density was computed between 2011 quarter 4 and 2010 quarter 4, between 2010 quarter 4 and 2009 quarter 4, and so on up to between 2007 quarter 4 and 2006 quarter 4, for each of the simple, walking and driving outlet densities.

Change in walking outlet density

Table 29 shows the descriptive statistics for the last period of change (2010 quarter 4 to 2011 quarter 4).

TABLE 29 - Descriptive statistics for walking outlet density alcohol availability for 2010 quarter 4 to 2011 quarter 4

SD, standard deviation.

Summary	Previous quarter	Absolute change	Relative change	Total change	Volatility
Mean	2.93	0.047	0.021	0.274	0.149
SD	4.15	0.370	0.172	0.652	0.311
Minimum	0	–1.634	–0.747	0	0
25th centile	0.77	–0.003	–0.003	0.002	0.002
Median	1.55	0	0	0.029	0.027
75th centile	3.15	0.04	0.033	0.279	0.181
Maximum	56.2	7.61	2.68	13.6	3.8

Figures 29–31 map the walking outlet density positive and negative change and volatility respectively.

FIGURE 29.

Walking outlet density, positive change, 2011 quarter 4. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

FIGURE 30.

Walking outlet density, negative change, 2011 quarter 4. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

FIGURE 31.

Walking outlet density, volatility, 2011 quarter 4. © Crown copyright and database rights 2013 Ordnance Survey 100019153.

Changes to the Welsh Health Survey alcohol questions

Between the 2007 and 2008 surveys, the questions on alcohol consumption and the method of conversion of drinks to units changed. This was to achieve consistency with the new ONS methodology first introduced in 2006. 87 These changes and their impact on the survey estimates are discussed in this section.

In response to the question included in the 2005/6 and 2007 surveys (Table 35),88 the participant may have under- or overestimated their consumption of alcohol due to the lack of inclusion of wine glass size, advice to not include cans and bottles and the onus being on the participants to calculate their own total units consumed. This may have resulted in an underestimate of consumption from calculation error and from reporting consumption at a desirable level, that is, social desirability bias. 89–91 Additionally, wine consumption was defined only as a small glass. This is no longer the standard measure provided in pubs and restaurants and may also have resulted in an underestimate of consumption. Alternatively, respondents may have overestimated their consumption from counting all bottles as two units.

The new alcohol question, first used in the 2008 survey83 and shown in Table 36, removes the responsibility to complete the calculations from the respondent. Respondents are now required only to enter the number of each size of each drink type consumed on their maximal drinking day and the calculations are completed by the survey coders. This allows the respondent to navigate the question quickly and provides space for the respondent to document/itemise the drinks consumed, reducing the calculation burden on respondents, which may lead to a more accurate number of units being recorded. However, a respondent may still under- or overestimate the numbers of drinks consumed.

Under-representation of Lower Layer Super Output Areas

The 2005/6 and 2007 surveys included respondents from fewer than 61% of the LSOAs, in contrast to the surveys from 2008 onwards (Table 40).

It is clear that, because of the effects of the question change and the under-representation of LSOAs before 2008, the analysis should use survey data from 2008 to 2012.

Modelling the number of units consumed

Here we used a multilevel Poisson model, in the family of generalised linear mixed models (GLMMs), fitting the count of units consumed by each individual as a function of the alcohol availability process, adjusting for individual- and LSOA-level covariates. This model predicted consumption in the year following the alcohol availability process measured in the preceding year.

The multilevel structure was year at level 1 with individuals nested within households, LSOAs and unitary authorities. We first estimated a null model of the unexplained random variation at each level. We then fitted the alcohol availability process in four separate models as the (1) previous quarter for current availability; (2) the historical maximum; (3) the historical minimum; and (4) the highest of the four previous quarters. We compared the model fits using the Akaike information criterion (AIC) and selected the model with the best fit. To this best-fitting model we added terms for positive and negative change in outlet density and volatility.

We then added individual-level covariates to adjust for age group in 10-year bands, sex, and the head of household NS-SEC3 (3-part National Statistics Socio-Economic Classification), the three-class variable defined by the NE-SEC as a measure of socioeconomic position. 93 We also fitted a term to allow for different linear trends in consumption by year. We fitted terms for LSOA deprivation quintiles and the three-category rural–urban settlement type. In our previous analysis of this data set94 we found significant interactions between age group, sex and deprivation quintile and we also assessed this in these models. We then explored whether or not the effect of a change in outlet density on consumption varied with deprivation by modelling the interaction between change in outlet density and deprivation quintiles. We also explored a stratification approach in which the data set was divided into deprivation quintiles 1 and 2 (low deprivation) and deprivation quintiles 4 and 5 (high deprivation).

We accounted for missing NS-SEC data using multiple imputation with chained equations,95 based on all data set variables in the imputation specification and five imputation models. In total, we assessed four models: two for simple outlet density with static and changing populations and two for the measures of walking and driving outlet density. All models were fitted in R software (version 3.0.2, The R Foundation for Statistical Computing, Vienna, Austria).

Geographically Weighted Regression spatial analysis of binge drinking

We then proceeded to a Geographically Weighted Regression (GWR) model with binge drinking as the outcome, also adjusting for the same individual-level covariates associated with alcohol consumption. We used a pragmatic approach to bandwidth selection as the size of the data set was too large for an adaptive process to run, deciding on 150 nearest LSOA neighbours. We adjusted for multiple comparisons using the false discovery rate of Benjamini and Hochberg96 (a conservative approach to multiple significance testing). We presented the model outputs in thematic maps for the intercept model-predicted binge drinking coefficients followed by the measures of alcohol availability that showed geographical variability.

Alcohol consumption outcome measures

A total of 74,069 (96.1%) respondents responded to the questions on alcohol consumption. The mean number of units consumed, including the never drinks/none categories, was 5.2 [standard deviation (SD) 7.7, range 0–306]. The median number of units consumed was 2.5 (interquartile range 0–8). The distribution was negatively skewed as expected (Figure 32).

FIGURE 32.

Distribution of units consumed on the heaviest drinking day in the last week. SD, standard deviation.

Clearly it is not possible to consume 306 units in 1 day and so a ceiling had to be applied to define a plausible level of consumption. Previous research on occasion-based drinking patterns has considered a maximum threshold of 450 g of ethanol per occasion, which is approximately equivalent to 56 units (at 8 g per unit). This also equates approximately to a maximum blood alcohol content of 0.5%, estimated using the Widmark formula,97 which is usually incompatible with life. 98

We therefore excluded the 141 (0.19%) respondents reporting a consumption of > 56 units, leaving 73,928 respondents in the analysis. Of these 73,928 respondents, 9800 (13.3%) were ‘never drinkers’. The full categorisation according to the Department of Health definition92 is shown in Table 44.

Of the 64,128 respondents who were not ‘never drinkers’, the mean number of units consumed was 5.8 (SD 6.9), ranging from 0 to 56. The median number of units consumed was 4 (interquartile range 0–9). Table 45 shows the descriptive statistics by survey year.

Around 32% of ‘ever drinkers’ did not report drinking in the previous week and a further 65% consumed between 1 and 20 units. Table 46 shows the number of respondents reporting consumption of each successive possible number of units from 0 to 20 units.

Table 47 shows the descriptive statistics for the categories of consumption by survey year.

Response by demographic and socioeconomic variables

Overall, 39,561 (53.5%) respondents were female and 34,367 were male. The proportions of respondents by sex varied little by year. The overall response by age group is shown in Table 48. There was little variation by year and so the data are presented aggregated across the six surveys.

The breakdown of respondents into the categories of socioeconomic position is shown in Table 49.

The breakdown of respondents into deprivation quintiles is shown in Table 50.

The breakdown of respondents into the categories of LSOA settlement type is shown in Table 51.

Table 52 shows the descriptive statistics for the mean number of units consumed on the heaviest drinking day and the categories of consumption for the demographic and socioeconomic individual variables. Table 53 shows the descriptive statistics by LSOA deprivation quintile and settlement type and Table 54 shows the descriptive statistics by local authority.

Distribution of consumption by Lower Layer Super Output Area

Table 55 shows the descriptive statistics for the mean number of units per LSOA for 2008 to 2012, for all respondents consuming ≤ 56 units.

In total, 1768 LSOAs had survey respondents for each year. Of the 128 missing LSOAs, one LSOA had three missing years and four LSOAs had two missing years. The other 123 LSOAs were missing data for 1 of the 5 years.

Simple outlet density: static population

Table 56 shows the parameter estimates and standard errors (SEs) with p-values for the final model of consumption and simple outlet density, static population. The previous quarter outlet density was significantly and positively associated with consumption in the following year. Change and volatility were not significant in the model. The patterns of association with the other model covariates were as expected.

We found that the interaction of age group*sex*deprivation quintile was not significant. Table 57 shows the random-effects variance (as the SD for direct comparison with the fixed effects) in the null and final models. Clearly the majority of the unexplained variation in alcohol consumption is attributable to the household and its magnitude was large in comparison with the fixed-effect model estimates. There was little difference between the null model and the final model. The individual-level variance in a Poisson model is constrained to be the mean and is therefore not estimated as a separate parameter.

Simple outlet density: changing population

Table 58 shows the results for the final model of consumption and simple outlet density, changing population. As in the static population model, the previous quarter outlet density was significantly and positively associated with consumption in the following year. Change and volatility were not significant.

Table 59 shows the random-effects variance in the null and final models. As with the static population model, the majority of the unexplained variation in alcohol consumption was attributable to the household and its magnitude was large in comparison with the fixed-effect model estimates.

There were no substantive differences in the estimates for the measures of outlet density with and without correcting for population migration.

Walking outlet density

Table 60 shows the results of the final model of consumption and walking outlet density. The previous quarter outlet density was significantly and positively associated with consumption in the following year. The positive and negative change parameters were not significant in the model, although the direction of association of these main effects was plausible. Volatility was not significantly associated with consumption in the following year.

Table 61 shows the random-effects variance in the null and final models. As with the simple outlet density models, the majority of the unexplained variation in alcohol consumption was attributable to the household and its magnitude was large in comparison with the fixed-effect model estimates.

We also fitted the walking density model excluding the 109 LSOAs with an English border to investigate the possibility of bias from boundary effects, as an English alcohol outlet might lie within a 10-minute buffer zone but would not be counted. The estimates were very similar (data not shown).

Table 62 shows the alcohol availability estimates for the final model of consumption and walking outlet density for deprivation quintiles 1 and 2 (least deprived LSOAs).

Table 63 shows the alcohol availability estimates for the final model of consumption and walking outlet density for deprivation quintiles 4 and 5 (most deprived LSOAs). There was little difference between the estimates in the deprivation-stratified models.

Driving outlet density

Table 64 shows the results of the final model of consumption and driving outlet density. The previous quarter outlet density was significantly and positively associated with consumption in the following year. Change and volatility were not significant.

Table 65 shows the random-effects variance in the null and final models. As with the previous models, the majority of the unexplained variation in alcohol consumption was attributable to the household and its magnitude was large in comparison with the fixed-effect model estimates.

Walking outlet density: non-linear dependence

We further explored the shape of the relationship between the previous quarter walking outlet density and consumption by fitting a cubic spline with three degrees of freedom. 99 Three degrees of freedom allowed sufficient flexibility in the shape of the relationship without overfitting the models. The shape of the relationship is shown in Figure 33, with five plots, one for each imputation model. The slope is steeper at outlet density values between 0 and 1, which includes 712 (37.5%) LSOAs, and is much shallower at values > 1. We therefore refitted the walking density model with a separate term for above and below a natural cut-point of outlet density = 1 and the estimates are shown in Table 66.

FIGURE 33.

Shape of the relationship between the previous quarter walking outlet density and consumption. Pred, predicted consumption; prevqrt, previous quarter.

The rate ratio for the values of outlet density between 0 and 1 was larger than that for higher values and suggests that in these less dense LSOAs a unit change in outlet density was associated with a 6% increase in consumption. A unit change in outlet density at values > 1 was associated with an additional 0.7% change in consumption. The combined terms for previous quarter were globally significant (p < 0.001).

Binge drinking: walking outlet density

In the initial global model we fitted the probability of individual binge drinking as a function of alcohol availability, adjusting for age group, sex, the interaction between age group*sex, WIMD 2008 deprivation quintile and settlement type (Table 67).