Local level estimates of food, drink and tobacco expenditure for Great Britain

We present expenditure estimates for 106 product categories across Great Britain for the years 2008–2016. Estimates are at the Local Authority District level (n = 380) and the categories cover all food, drink and tobacco commodities. Reliable, local level expenditure estimates are crucial for understanding broader market trends, assessing economic stability and for projections. This is especially important for commodities such as alcohol, tobacco and unhealthy foods due to their role in the prevalence of non-communicable diseases. There has been relatively little research into local area spatial patterns of expenditure, with existing estimates often of insufficient resolution for informing planning decisions. We use spatial microsimulation to create an archive of expenditure datasets. This was achieved by linking socio-demographic foundations with detailed datasets on individual expenditure. Whilst initially developed to aid investigations into sociodemographic trends in the meat industry, the data have reuse potential in a number of disciplines, including public health, economics, retail geography and environmental management. The framework could be applied to other regions with appropriate data.


Background & Summary
Over the past 50 years, the UK has experienced major shifts in dietary patterns due to changes in agricultural practice, trade policies and food industry marketing 1 . Further changes may be on the horizon in the context of a UK exit from the European Union 2 . Against this backdrop of continuous national level change, there is substantial local level variability in food consumption and expenditure patterns, which has been attributed to spatial variations in factors including socio-economic status [3][4][5] and demographics 6 . These changes are reflected by individual expenditure patterns as surveyed annually by the Living Costs and Food Survey [7][8][9][10][11][12][13][14][15] , with corresponding results published to the regional level across the UK 16 (i.e. 12 geographical zones). To help understand the local level variability of expenditure, and to form a baseline for future projections, we present an open access archive of expenditure datasets for Great Britain for the years 2008 to 2016. Each annual dataset consists of an expenditure estimate for 106 food, drink and tobacco categories for every Local Authority District (LAD) (n = 380).
Robust estimates of local level spatial patterns of food and drink expenditure are crucial for understanding broader trends, for assessing market stability and for future projections. It has long been argued that the most powerful theoretical models for explaining human behaviour operate at the individual person level 17 , with emergent higher-level properties giving the best opportunity to understand the entire system at all levels. Reliable and detailed information on the spatial distribution of food, drink and tobacco expenditure is also key for research in the fields of public health, environmental impact and retail geography. This importance is highlighted by the prevalence of non-communicable diseases such as cardiovascular disease, cancer and diabetes which currently account for 70% of all deaths worldwide and 90% in the UK 18 . As these diseases share key modifiable behavioural risk factors such as tobacco use, unhealthy diet and the harmful use of alcohol, it is clear that understanding expenditure patterns associated with key commodities is of great value for public health research. There is also an increased awareness of the environmental impact of food production, with livestock production responsible for 14.5% of all anthropogenic greenhouse gas emissions in 2004 19 , whilst 71.2% of deforestation in South America between 1990 and 2005 was for conversion to pasture 20 .
There has been relatively little research into local area spatial patterns of food and drink expenditure in the UK. Whilst expenditure data is routinely collected by companies and organisations, these are seldom open-access, often at a coarse spatial resolution and only provide a snapshot of specific products or socio-economic groups. In the UK, comprehensive estimates of food and drink expenditure are published annually by the Department for Environment, Food & Rural Affairs (DEFRA) 16 , representative of the population. However, these data are only available at the regional level (12 geographical zones) and as such are not at a sufficient resolution for informing planning decisions related to public health infrastructure, retail or the environment at the local level. For example, concerns over access to healthy foods 21 cannot be assessed using regional level data. Furthermore, there is little information on associated individual level socio-demographics, which have been shown to be strongly linked to expenditure 3 .
This study aims to bridge the identified data gap between the published regional level estimates of expenditure 16 and known drivers of local level variation 3-6 by producing local area level datasets of expenditure using the technique of spatial microsimulation. Increasing computational efficiency and falling costs combined with the improved availability of survey microdata have increased the ability to produce such datasets 22 . As such, using the code developed by Lovelace and Dumont 23 , an archive of expenditure datasets has been created. This process used the most recent census and survey microdata available to the authors at the time of writing, alongside a range of geospatial datasets.
We believe spatial microsimulation techniques of the type described in this paper hold great potential benefits for a range of disciplines including economics, retail geography and public health. Whilst this study focusses on Great Britain, the framework here could be applied to any location with the appropriate data sources.

Methods
Spatial microsimulation. This study uses spatial microsimulation to generate expenditure estimates under the framework shown in Fig. 1. Spatial microsimulation involves 'the creation, analysis and modelling of individual level data allocated to geographic zones' 23 , and has been used in the fields of health care demand 24 , educational attainment 25 , commuting patterns 26 , and population projections 27 amongst others. For a comprehensive overview of the microsimulation process the reader is directed to Birkin and Clarke 28 and a guide to implementation can be found in Lomax and Smith 29 .
As with most spatial microsimulation models, the input data for this study consists of microdata -a non-geographical individual level dataset -and constraint tables, which provide aggregate counts for each geographical zone (LAD). The framework is split into two separate microsimulation models as shown in Fig. 1; a comprehensive 'adult' model for those aged 16 and over and a 'child' model for those aged 15 and under. Once each of the models have completed, the results are merged to generate a full synthetic population and consequently LAD level estimates of expenditure.
Specifically, this study employs Iterative Proportional Fitting (IPF), implemented within the R programming environment (https://www.r-project.org). IPF works by adjusting a large array of weights -rows corresponding to individuals and columns corresponding to the geographic zones (e.g. LADs) -iteratively, to maximise the fit between simulated and known (e.g. census/survey) data. The mathematics of IPF are covered by Fienberg 30 a guide to implementation is provided in Lomax and Norman 31 whilst the code used here for implementing IPF in R was developed by Lovelace and Dumont 23 .

Microdata -Living cost and Food Survey (LcF). Microdata are taken from the Living Cost and Food
Survey (LCF), the most comprehensive survey on household spending in the UK, covering approximately 12,000 respondents from 6,000 households each year. The LCF is carried out by the Office for National Statistics (ONS) and has been running in its current format since 2008. The LCF is designed to be representative of people living in households in the UK, using a multi-stage stratified random sample with clustering approach. The survey is weighted to compensate for non-response and also to ensure the sample distribution matches the population distribution in terms of region, age group and sex. The LCF runs continuously throughout the year to avoid seasonal variation 32 .
The LCF comprises an expenditure diary detailing purchases over a two-week period and an interview covering socio-demographic characteristics, income and regular items of household expenditure. Respondents are required to record all expenditure over the two-week period (regardless of outlet), thus providing a comprehensive account of household expenditure. Commodities recorded in the LCF diary (and consequently in this study) are grouped by category, based on The Classification of Individual Consumption by Purpose (COICOP) coding framework. COICOP groups products into homogenous categories for which food, drink and tobacco constitute 106 separate groups. Categories may define a specific product (e.g. 1.1.6.2.1 = Bananas -fresh) or a homogenous group of products (e.g. 1.1.1.4.1 = Cakes and puddings). The framework structure also allows easy aggregation to higher levels (e.g. 01.1.2.5 = Dried, salted or smoked meat and edible meat offal; 01.1.2 = Meat; and 01.1 = Food). The full list of 106 food, drink and tobacco expenditure codes used in this study can be found in Supplementary File 1. The 2016-2017 LCF survey reported some commodities (specifically those consumed away from home) only to an aggregate level, resulting in fewer categories for our 2016 dataset (n = 80). These aggregated categories are included in Supplementary File 1. Whilst grouping products in this manner may mean that analysis related to specific products is restricted, the 106 categories provide sufficient detail for most applications. Whilst various other coding frameworks are available, COICOP was specifically developed by the United Nations Statistics Division to analyse individual expenditures, and was therefore adopted by the ONS for use in the LCF. Datasets presented in this study can be directly compared with others which use the COICOP framework, whilst the detailed descriptions provided in Supplementary File 1 allows cross-referencing with alternative frameworks if required.
The LCF is geocoded at a coarse level, detailing which of the 12 government regions each individual resides in (Scotland, Wales, Northern Ireland, South East, London, North West, East of England, West Midlands, South West, Yorkshire and the Humber, East Midlands, North East). As discussed previously, this is insufficient for informing www.nature.com/scientificdata www.nature.com/scientificdata/ planning decisions related to public health infrastructure, retail or the environment at the local level. Whilst it would be technically possible to constrain the microsimulation model using these data (i.e. only individuals sampled in the South West region of England from the LCF would be able to be assigned to LADs in the South West region), this would result in a much-reduced sampling pool insufficient for spatial microsimulation. As such, no initial geographical constraints are used in the microsimulation model although the regional information is used to account for relative regional price levels and for model validation purposes, as discussed in due course.
It should be noted that whilst the LCF survey includes individuals from Northern Ireland, insufficient constraint variables were available for microsimulation within Northern Ireland (see below). As such, whilst microsimulation outputs presented in this paper are restricted to Great Britain, individuals from Northern Ireland (from the LCF) are included within the sampling pool and may be allocated to any LAD in Great Britain if their socio-demographic characteristics are appropriate. www.nature.com/scientificdata www.nature.com/scientificdata/ Formatting the survey microdata. The LCF contains a wealth of information, much of which is not required for the purposes of this study and can thus be discarded. As the microsimulation process requires common variable classes for the microdata and corresponding constraint dataset, re-formatting is required to generate the appropriate classes. Table 1 lists the LCF variables used in this study and their categorisation. Table 2 provides an example extract of the formatted socio-demographic microdata, and Table 3 shows an example of the diary information.
From 2015 onwards the LCF reporting window moved from a calendar year (January to December) to a financial year (April to March). To maintain consistency of our datasets, we use a calendar year throughout (i.e. our 2015 dataset represents LCF data from January 2015 to December 2015). As the LCF details when each survey was completed during the year, we achieve this by removing and appending records from each year as appropriate. The 2015-2016 LCF survey also includes additional records from January to March 2015, making it possible to construct a seamless data series. constraint variables. As with other microsimulation applications, the model presented here is underpinned by the assumption that the target variable (expenditure) is associated with the geographical constraint variables. Constraint variables were chosen following the guidelines of Lovelace and Dumont 23 , based upon relevance to the target variable (expenditure) and data availability. Table 4 details the constraint variables selected, the source datasets and their temporal coverage. The microsimulation is split into two separate sub-routines: a comprehensive 'adult' microsimulation model for those aged 16 and over and a simpler 'child' microsimulation model for those aged under 16. This is because many of the variables are not available and/or not applicable for those under the age of 16 (e.g. unemployment). As noted previously, many of the constraints listed in Table 4 are not available for Northern Ireland and as a result the microsimulation presented here was restricted to Great Britain.
constructing the baseline population. Microsimulation requires the baseline population of each constraint (i.e. the total number of people in each zone) to correspond to the population from which the microdata has been sampled. For the LCF, this is all people living in households aged 16 and over (for the adult model) or aged 15 and under (for the child model). The IPF algorithm also requires the baseline population to be identical across all constraint variables. To meet these requirements, our baseline population for each year is taken from the Office for National Statistics mid-year population estimates, with residents living in communal establishments removed to result in only residents living in households. All other constraints are scaled to this baseline population, as described by Lovelace and Dumont 23 . Table 5 shows an extract of the final 2008 age-sex constraint table (household residents only) for three local authorities.
With counts of communal establishment residents only available for the year 2011, we assume that this population is unchanged throughout the study years (2008 to 2016). This is a reasonable assumption, as communal establishment populations are usually fairly stable in terms of their size and demographic structure. For example, an elderly care home will contain a similar group of individuals from year to year. This stability is recognised by the ONS, who treats communal establishment populations as a different and more stable group to the household population when producing the mid-year estimates 33 . Furthermore, any deviation from the 2011 counts will have a negligible impact on the model output as communal establishment residents account for a small proportion of the overall population -just 1.7% in 2011 34 . Formatting the ethnicity constraint. Annual estimates of the number of people aged 16 and over per ethnic group for each LAD are taken from the Annual Population Survey (APS) ( Table 4). These data are categorised to correspond to the LCF microdata classes (Table 1) and scaled to the baseline population. An extract of the final 2008 dataset is shown in Table 6. As the APS sample already excludes most communal residents, we assume that the proportions of each ethnic group is consistent between the baseline population and the APS sample.
Formatting the unemployment constraint. Annual estimates of the number of unemployed people aged 16 and over in each LAD are taken from ONS Model Based Estimates of Unemployment (Table 4). These data are scaled to the baseline population, with an extract of the final 2008 dataset shown in Table 6. The unemployment estimates are derived from the Labour Force Survey, which excludes most communal establishment residents 35 . As such, we assume that the proportion of those unemployed is consistent between the baseline population and the model based estimates. www.nature.com/scientificdata www.nature.com/scientificdata/ Formatting the student status constraint. Annual estimates of the total number of students aged 16 and over are taken from the Annual Population Survey ( Table 4). As the LCF microdata does not sample students who reside in halls of residence, these students are removed from the constraint estimate. This is achieved using 2011 Census estimates of the numbers of people aged 16 and over who live in student accommodation ( Table 4). As these records are only available for 2011, we assume that the number of students in halls of residence remains constant throughout 2008-16, a reasonable assumption due to the transient nature of the population. These resulting data are scaled to the baseline population, with an extract of the final 2008 dataset shown in Table 6.
Formatting the income constraint. Annual estimates of gross weekly pay (pre-tax) for each LAD is taken from the Annual Survey of Hours and Earnings (ASHE) ( Table 4). This provides data on the pay levels and distribution of UK employees aged 16 and over. The ASHE is based on a sample of employee jobs taken from Her Majesty's Revenue and Customs Pay As You Earn (PAYE) records and as such does not include those who are self-employed. The initial data is provided in terms of percentiles with data available for P10, P20, P30, P40, P50, P60, P70 and P80 (Table 7). Each percentile indicates the value below which a given percentage of the observations fall; for example a P20 value of £219.80 indicates that 20% of the sample has an income of less than £219.80.
To make the ASHE data compatible with the microsimulation model, it is first necessary to estimate the total number of people covered by the sample. This is achieved using employment status estimates from the Annual Population Survey (Table 4). This provides estimates of the number of employees (i.e. those covered by PAYE records) per LAD (Table 7).
Once the number of persons in each category has been estimated, a constraint table is generated containing the income brackets for each LAD (in £s) and the number of employees within each category ( Table 8). As with other constraints the values are scaled to match the baseline population. The same income brackets are used to categorise the LCF microdata for each individual LAD as shown in Fig. 1.

Formatting the household characteristics constraint. Data on the household characteristics of each
LAD is taken from the 2011 Census (Table 4). The dataset covers all individuals aged 16 and over living in a household, providing a description of household type (age and number of people and dependent children living in the household). The categories are grouped to correspond with those in the LCF microdata ( Table 1). As information is available only for 2011, we assume that the proportion of each household type remains constant, being scaled to the baseline population each year. Table 9 shows an extract of the final household characteristics constraint table. child microsimulation model. For people aged under 16 years of age, a simpler microsimulation model is employed as many of the constraint variables are not applicable or not available (e.g. unemployment). A simpler model is also deemed appropriate as children contribute a negligible amount of total expenditure, accounting for just 0.78% in 2016-17 according to the Living Cost and Food Survey 15 . As with the adult model, ONS mid-year population estimates are used in conjunction with 2011 Census estimates of communal residents to create a baseline population. The child model uses an age sex constraint with age categories of 0-9 years (male), 10-15 (male), 0-9 years (female) and 10-15 (female), as shown in Fig. 1. Missing variables. Whilst constraints of age-sex and household type are available for all 380 local authorities across Great Britain, other constraints (students, unemployment, ethnicity and income) are unavailable for a minority of LADs due to small sample sizes or missing data. For example the Isles of Scilly have a total population of just 2,292 people (2014 estimate), meaning that some constraints would be disclosive if published. Whilst this is not deemed an issue in terms of model robustness, the model needs to be able to cope with missing data. This is achieved by dynamically adjusting the final constraint table for each LAD depending on which variables (and categories within) are available. Whilst the student and unemployment variables are binary (either available or not available), the variables of ethnicity and income may be partially complete (e.g. there may be an estimate of the number of individuals of black ethnicity but no estimate for those of mixed ethnicity). In these cases, the constraint (and microdata) is re-categorised to utilise the available data. For example, if an estimate for those of mixed ethnicity is unavailable for a particular LAD, new categories of 'black' , 'white' or 'other' will be created.
In most circumstances a complete suite of constraints are available allowing for a full microsimulation model. As all LADs have complete age-sex and household type variables the microsimulation model will run on these as a minimum. Table 10 shows the number of LADs with each constraint available for each year.
Accounting for relative regional consumer price levels. It is well known that the price of goods and services varies throughout the UK 36 . In 2016 food and non-alcoholic beverages in London cost 2.2% more than the UK average whilst in Scotland they cost 0.2% below average 36 . This is a potential problem for the  www.nature.com/scientificdata www.nature.com/scientificdata/ microsimulation model as the process allows an individual from the LCF microdata to be assigned to any LAD in Great Britain, according to the constraint variables. For example, if an individual from Scotland (from the LCF microdata) is assigned to a London LAD, their expenditure will likely be under-estimated.
To account for this, ONS Relative Regional Consumer Price Levels (RRCPLs) data 36,37 are used to adjust expenditure values depending on their source region (from the LCF microdata) and their destination region (as assigned by the microsimulation model). Pre-microsimulation expenditure values are scaled to a 'UK average' price before being adjusted back to regional levels according to the region in which the microsimulation model assigns them to. The ONS provides an aggregate RRCPL value for each of the 12 regions (for all products) and provides more detailed category level values for London, Scotland, Northern Ireland and Wales. As such we use the detailed category level values where available and the aggregate value for all categories where not, as shown in GiS expenditure datasets. Once the model is complete, GIS expenditure datasets may be created by joining the expenditure tables with spatial boundaries. Figure 2 shows examples of selected datasets for the year 2012. Cumulative categories are generated by summing the appropriate individual COICOP categories (e.g. all food and drink: Fig. 2a, alcoholic drinks: Fig. 2b, tobacco and cigarettes: Fig. 2c) whilst individual COICOP categories can also be mapped (e.g. bacon and ham purchased for household supplies: Fig. 2d). For visualisation and integration with other datasets, the GIS vector shapefiles may be converted to a spatial grid of data cells in a similar manner to other spatial datasets (e.g. James et al. 38 ).

Data Records
The local level expenditure datasets described in this article are publicly and freely available through Figshare 39 .  Table 4. Constraint variables and source datasets for the adult and child microsimulation models. *The age and sex constraint forms the baseline population to which all other constraints are scaled to.

Technical Validation
The validation of microsimulation models has received much attention in the literature due to the dangers of using incorrect model data to inform policy 40,41 . Validation of microsimulation models presents a substantial challenge since detailed spatial microdata are seldom available -in fact it can be argued that if such data were available the microsimulation process would be redundant. There are a variety of methods available for validation, broadly categorised as internal validation (ensuring the model makes sense in reality given the input data) and external validation (ensuring the model coincides with external reality).

internal validation. Internal validation is the most common form of microsimulation model evaluation and
is the process of comparing the model's output against data that are internal to the model itself 23 . We carried out internal validation in a similar manner to Lovelace, et al. 26 , by calculating for each LAD the correlation between the aggregate counts from the constraint variables and those generated in our spatial microsimulation. In our models, the results were affirmative; the lowest correlation for a single zone for all years was 0.9876 and in many cases was perfect (at least with an approximation to 4 decimals). The high correlation coefficients throughout give us confidence that the microsimulation process has worked correctly and common issues such as empty cells 42 and incorrectly specified constraint variables are not present.
However, internal validation needs to be viewed in context as IPF microsimulation always converges towards the optimal solution for known constraint variables: it is the unknown cross-tabulations and target variables that we are trying to simulate with spatial microsimulation, so external validation should also be used 42       www.nature.com/scientificdata www.nature.com/scientificdata/ External validation. In addition to internal validation, we use two methods for external model validation: a) by comparing the simulation results at the aggregate level with estimates from a dataset external to the model, and b) by aggregating-up the small area estimates provided by the simulation to compare the results with expenditure data that is provided at higher geographies.
Previous studies have shown that there are relationships between socio-economic status/deprivation and expenditure on certain commodities. Total food expenditure and consumption of fruit and vegetables has been shown to be greater in more affluent households 3,43 whilst smoking prevalence is often greater in more deprived areas 44 and among those of lower socio-economic status 45 . As such, we explored the relationship between our LAD level estimates of product expenditure and deprivation as measured by an external dataset. We used the Index of Multiple Deprivation (IMD) (https://www.gov.uk/government/statistics/english-indices-of-deprivation-2015), the official measure of relative deprivation for local authorities in England. IMD is based on seven different domains of deprivation: income deprivation; employment deprivation; education, skills and training deprivation; health deprivation and disability; crime; barriers to housing; and services and living environment deprivation. Whilst some of the domains are similar in nature to the constraints used in the microsimulation model (e.g. income deprivation), there are no common datasets between the IMD and the microsimulation, with metrics calculated in different ways. Furthermore, many of the IMD domains are completely absent from the microsimulation (e.g. crime, health deprivation and disability) and vice-versa, resulting in a minimal risk of circularity when exploring relationships. As the majority of datasets used in the IMD were collected in 2012 and 2013 we use the    www.nature.com/scientificdata www.nature.com/scientificdata/ 2013 microsimulation results to test for a correlation against the IMD. As the IMD is a ranked dataset, we use the Spearman's test of rank correlation, with the results shown in Table 12.
All correlations were significant with a strong positive correlation between IMD and estimates of 'all food and drink' and 'fruit and vegetable' expenditure, suggesting that expenditure on these categories is less in more deprived areas, as found by previous research 3,43 . Conversely, there is a negative correlation between tobacco and cigarette expenditure and IMD, suggesting expenditure is greater in more deprived areas, in agreement with previous studies 44 . These results are encouraging, suggesting the model is accurately capturing variation in expenditure for different product categories at small area level. www.nature.com/scientificdata www.nature.com/scientificdata/ As noted previously, the LCF is geocoded to a limited extent, with information provided on which of the 12 regions each individual resides. This information is used by the Department for Environment, Food and Rural Affairs (DEFRA) to estimate the expenditure of products at the regional level, as published in the annual Family Food report series 16 . As this regional geographic information is not used in the microsimulation model other than for adjusting for relative regional price levels, this presents a useful form of validation by comparing the aggregated microsimulation results at the regional level to the corresponding values estimated directly from the LCF. Whilst both methods estimate the same parameter (expenditure by region), they are generated in completely different ways. The LCF averaging approach (as used by DEFRA 16 ) takes the average weighted expenditure of surveyed individuals in each region, whilst the microsimulation approach generates a synthetic population with every individual assigned an expenditure profile which is then aggregated to the regional level. Figure 3 shows the results of the microsimulation model (aggregated to the regional level) alongside the corresponding values from the LCF, including error bars (±1.96 standard errors). Results are shown for a grouped category (all products) and an individual COICOP category (bacon and ham for household supplies). Corresponding 2012 maps for these categories are shown in Fig. 2. Figure 3 demonstrates that the aggregated microsimulation results shows good correspondence with the LCF regional averages for both grouped categories (all products) and single COICOP categories (bacon and ham). Reassuringly, the majority of microsimulation estimates fall within the 95% confidence limit of the original LCF averages and the general trends in the LCF time series are reflected in the aggregated microsimulation results. This suggests that the simulated microdata correspond well to the 'real world' survey data. Whilst some regions such as the South East show a very close alignment between the aggregated microsimulation results and the LCF regional averages, others are not as closely aligned (e.g. North East). There are a number of possible reasons for these differences observed, including localised factors not accounted for by the microsimulation constraint variables or insufficient detail of input information. For example, the relative regional consumer price levels used to account for spatial differences in product pricing are only available for 2010 37 and 2016 36 and only at the aggregate level for regions within England (Table 11). It should also be noted that the confidence limits referred to here for the LCF only cover sampling errors and not non-sampling errors (systematic errors and random errors) 32 , and as such this uncertainty in the LCF should also be considered when comparing the estimates.
The internal and external validations presented here suggest that the microsimulation estimates of expenditure are capturing real differences. As such, we believe spatial microsimulation techniques of the type described in this paper hold great potential benefits for a range of disciplines including economics, retail geography and public health. Whilst this study focusses on Great Britain, the framework here could be applied to any location with the appropriate data sources.

Usage Notes
The expenditure estimates produced in this study are based on data from the LCF, Census and other official sources. Therefore, the outputs provided are subject to any biases or errors contained in the source datasets. Household surveys such as the LCF have the potential to suffer from issues such as non-response from specific groups (e.g. low income households 46 ) and measurement error 32 , especially in relation to products consumed away from home 47 and alcohol 48 . Whilst the ONS has a robust sampling, weighting and quality control framework for the LCF 32 , the end user should be aware of the potential biases and errors, especially when considering specific commodities or socio-demographic categories.
code Availability