The use of systems models to identify food waste drivers: Grainger et al. (2018a)

Grainger et al.’s paper investigated the drivers of household food waste using Bayesian networks to identify the impact of household characteristics and other variables on self-assessed food waste. Using EU-level Eurobarometer data from 2013, the study confirmed that the country, the age of the respondent, the status (student/non-student), and a belief that the family wastes too much are related to the level of self-assessed food waste. In addition, households from lower-income EU countries (e.g. Portugal, Greece, Bulgaria, Cyprus and Latvia), as well as students and young adults, tend to report higher levels of food waste.

However, the analysis found no evidence that food waste behaviours differ between people living in urban and rural areas, and little support of a difference between genders. These geographical and gender differences had been identified in previous literature as potential drivers of food waste (Wenlock and Buss, 1977; Sonesson et al., 2005; Barr, 2007; Koivupuro et al., 2012; Canali et al., 2014; Parizeau et al., 2015; Setti et al., 2016; Stancu et al., 2016). The additional insight provided by the application of Bayesian networks provides clarity to the researcher to understand which relationships have evidence within the currently available data. This insight can then be acted upon by the policy maker. In this case, the researchers suggested country-level policy measures targeting different age groups.

Model selection and averaging in the assessment of the drivers of household food waste to reduce the probability of false positives: Grainger et al. (2018b)

This paper used machine learning algorithms (Random Forests and “Boruta”) along with Generalised Linear Models to identify the key drivers of household food waste, while also reflecting the uncertainty inherent in the analysis of complex observational multidimensional data. The data investigated was household food waste data collected by WRAP (2012) which consisted of face-to-face in-home interview responses (categorical data) on socio- denrographic aspects of households and behavioural responses to food waste, along with data on the amount of waste collected from the kerbside for 1,770 households.

As the data set has over 50 variables, there would be over a quadrillion possible Generalised Linear Models to run. To simplify this, the “Boruta” and random machine learning algorithms were first used to refine and reduce the variable list. The “Boruta” algorithm adds randomness to the variable set by creating shuffled copies of all variables (these are called “shadow features”). It then runs a Random Forest classifier on the extended dataset, and assesses the mean decrease in accuracy to evaluate the importance of each variable (higher means are more important). At each iteration, “Boruta” assesses if each variable has a higher Z-score than the maximum Z-score of its shadow features. Variables with scores lower than shadow features are deemed highly unimportant, and removed from the set. The algorithm runs until all variables are confirmed or rejected (or it reaches a specified limit of runs — here, we used 500 trees maximum). The variables retained after applying the “Boruta” algorithm were then processed using a Generalised Linear Model to assess correlations between “avoidable household food waste” and the socio-demographic and behavioural variables.

The “Boruta” algorithm consistently identified household size, home ownership status, household composition, employment status and the presence of fussy eaters as significant drivers of food waste in all sets of variables. Household size was always the most important variable.

The final model contained household size, local authority, household composition, house type, home ownership status, employment status, the presence of fussy eaters, the presence of children aged between 3 and 11, age of the respondent, social grouping, checking cupboards for tinned food prior to shopping, and discard behaviours related to vegetables, cheese and food past its sell-by date. The variables with the largest positive effect (greater amounts of food waste) included the presence of fussy eaters, household size, and one particular local authority (individual local authority' identity was anonymised). Variables with the largest negative effect (reductions in food waste) included discard behaviours interacting with the presence of fussy eaters, employment status interacting with the presence of fussy eaters, four specific local authorities and home ownership status (owning a house outright).

As with Grainger et al. (2018a), the application of the machine learning algorithms has enabled new insight into the drivers of household food waste. Again, it is interesting to note that some of the drivers identified as important by previous literature, such as awareness of the food waste problem and shopping habits, here are found as not important.

Agent-based modelling

Agent-Based Models (ABMs) are computational systems that simulate the individual decision-making process of a large number of agents acting and interacting through a set of prescribed rules (Farmer and Foley, 2009). The output of an ABM are the emerging phenomena resulting from the interaction among agents’ choice on a large scale, both temporal and dimensional. The characteristics of ABMs lead to several advantages. On the one hand, they allow for a large degree of heterogeneity in agents’ characteristics and interaction rules; on the other hand, they' allow for the introduction of a well-defined institutional structure. Nevertheless, it is important to constrain the additional complexity to avoid generating models as difficult to understand as the reality studied.

The main tool to analyse ABMs are Monte-Carlo computer simulations, where a set of inputs is provided to the model, and the dynamics of the model are iterated many times with different sequences of random numbers. This allows the study of the statistical characteristics of the simulation output (means and standard deviations of the results, their distribution, and the occurrence of rare extreme events), separating random events from proper emerging properties of the simulated system. By' modifying the parameter sets, it is possible to check the robustness of the results and to assess the implications of a shift in one of the parameters. A well-developed model can be used as a virtual laboratory, as it allows the generation of alternative time-series under controlled “quasi-experimental” conditions. As such, ABMs can also be studied with regression techniques, exploring the correlations between different parameters and outcomes and the impact of different types of heterogeneity. Given that many relationships among variables are typically hard-wired, causation structures can also be studied. An alternative method of analysis frequently used to assess ABMs is the comparison of scenarios. Within this method, different initialisations and sets of rules are created to simulate specific known cases (such as two countries), or to study the expected impact of a policy intervention. Both the aggregate outcomes and the individual trajectories of the agents are then assessed comparatively. The analysis of the results frequently relies on graphs, such as plots and figures.

To design and develop an ABM, it is necessary to specify at least three elements: the entities (agents); their interaction rules; and the environment and institutions within which agents interact. The agents are the autonomous and discrete decision-making units whose behaviour is modelled. In socio-economic simulations, they are typically individuals, companies, or even nations. Their characteristics usually include: attributes (idiosyncratic or group-specific properties); rules of behaviour (assumptions made about their decision-making processes); memory (the possibility of recalling past actions and interactions and their results); and perception of the environment. The interaction rules are the constraints on how agents can interact. Depending on the type of model, they can be represented in game theoretical form (agents receive a payoff that depends on their actions and on those of other players), as economic exchanges (one or more individuals buy something that someone else sells in exchange for something else), or as exchanges of information. Exchanges typically happen on a defined interaction space. Finally, the environment and institutions define the external constraints that influence all agents (or groups of them), and their interactions.

Both ABMs described below (Grainger et al., 2018d), were developed in MatLab R2017a, while the Bayesian network of consumer food waste generation was developed in R. The integration of the two models was achieved through C++ in DOS, with externally controlled processes in both R and MatLab to allow the sharing of inputs and outputs.

An ABM of retail food waste

The retail ABM developed by Grainger et al. (2018d) aimed at simulating the interaction between the adoption of an innovation reducing food waste by retailers and resulting food waste levels. The challenge of this setting is represented by the fact that retailers earn a profit from the food wasted at home by consumers, thus profit-maximising retailers are not willing to innovate to reduce it. However, behavioural economics theory points out that additional concerns, such as reputation, can lead to non-trivial outcomes.

The ABM considers the market for a single food commodity, namely fresh fruit and vegetables, due to their high perishability. The introduction of a waste-reducing technology has an impact on the purchasing behaviour of consumers and on retailers’ marketing strategies. The market operates in imperfect conditions (e.g. asymmetric information and concentration).

Retail agents are modelled as belonging to three different groups: small shops, discounts and large-scale companies. Each agent can adopt only one of two different technologies: a baseline that generates a high amount of food waste (initially adopted by all retailers) or an innovative technology that reduces the amount of food waste generated either in store or by customers at home. Retailers decide whether to adopt the low-waste innovation based on a utility function which includes three main elements: (1) the profit earned, which depends on selling prices, innovation costs and the share of food wasted in store and by consumers after purchase; (2) environmental concerns, and reputational concerns linked to pro- environmental behaviours; (3) other retailers’ decisions.

To reduce complexity, consumers are modelled as homogeneous masses with shared attributes, or with attributes varying within a certain range, who at the onset of each simulation purchase from the same typology of retailers. Three groups of consumers are considered: (1) quality- oriented ones, who purchase from small shops, characterised by a low price elasticity; (2) unsophisticated ones, who purchase from large-scale companies, characterised by an average elasticity; (3) convenience-seeking consumers, who purchase from discounts, characterised by a high elasticity'. Consumers choose the retailer from which to purchase based on a set of parameters that do not vary inside groups, but may change between groups: elasticity to price; environmental concerns; their state of information about the existence of retailers which adopted the low-waste technology; and a satiation quantity', which is the same for all of the consumers and is technology-dependent (the quantity' of food necessary to achieve satiation is lower if the retailer a consumer purchases from has adopted the low-waste technology).

Within the model, time is divided into ticks during which decisions are assumed to be taken parallelly by all agents according to a set of steps. The intra-period steps of the retail model are the following:

  • 1. Each retailer (with a given probability) can decide to change the technology adopted, maximising its utility function.
  • 2. Given the previous decision, each retailer can change its selling price (small shops base the pricing decision on the behaviour of similar companies in their network, large and discount companies on the market share of adopting retailers).
  • 3. The consumers purchasing from a retailer that changes technology are assigned to the same retailer.
  • 4. A share of consumers becomes informed about the existence of the low-waste technology — according to the literature on innovation diffusion (Rogers, 2010), this share depends on information from external sources (e.g. advertising from retailers) and information circulating among consumers (e.g. word of mouth).
  • 5. A mass of consumers with similar characteristics decides to move to a different retailer based on the parameters listed previously, including their utility and information status.
  • 6. The market shares of each retailer are recalculated, and a new step can start.
< Prev   CONTENTS   Source   Next >