Measuring sustainability efficiency at farm level: a data envelopment analysis approach

Abstract

1. Introduction

The Committee on Twenty-First Century Systems Agriculture (NRC, 2010, p. 4) characterises sustainable agriculture as one that satisfies human food, feed, fibre and biofuel needs; enhancing the quality of the environment and resource base; ensuring the economic viability of the agricultural sector and improving the quality of life of farmers, farm workers and society. Agricultural policies in developed countries have promoted adoption of such sustainable practices on agricultural holdings. The European Union’s Common Agricultural Policy (CAP) has been no exception. Since its inception, it has undergone different reforms that reflect changing political priorities over time. Initially, the CAP essentially aimed at guaranteeing food security by stimulating agricultural production and protecting farmers’ quality of life. A succession of changes has reformulated the CAP into a policy that embraces food safety, animal welfare, land management, rural development, environmental development and pollution control. In short, the CAP leans towards promoting a more sustainable agricultural sector.

Sound implementation of sustainability practices requires appropriate tools to measure farms’ success in achieving policy goals. Since the pioneering work of Farrell (1957), the production economics literature has developed efficiency indices that can be extended to assess this success. While the literature on efficiency measurement initially focused on the desired output’s production technology, as the relevance of environmental sustainability increased, firm-performance studies were extended to include environmental concerns and the literature is now rich (Färe et al., 2005; Coelli, Lauwers and Van Huylenbroeck, 2007; O’Donnell, 2007; Murty, Russell and Levkoff, 2012). However, it is only recently that efficiency measures have been extended to quantify the third pillar of sustainability, namely the social dimension of firm performance (Chambers and Serra, 2018), and the meagre literature available has not yet been applied to farm sustainability measurement.

The objective of this article is to derive farm-level efficiency measures that include economic, environmental and social considerations building on the method proposed by Chambers and Serra (2018). No previous study has ventured into quantifying farm efficiency as a provider of social outputs, which represents our first contribution to the literature. This requires extending the Chambers and Serra (2018) model to allow for the stochastic nature of agricultural production, which represents our second contribution. Assessing the environmental and social dimensions of performance and accounting for stochastic conditions, requires data that are not usually available, especially at farm-level. We elicited this information through a survey in our application for Catalan arable crop farms.

The remainder of this article is organised as follows. In the next section, a brief review of the most relevant literature is provided. We then present the methodological framework employed, followed by a description of the data. The fourth section focuses on results and the final section presents a conclusion.

2. Literature review

The ‘triple bottom line’ captures the essence of firm sustainability performance by going beyond economic measures to include environmental and social considerations (Elkington, 1998). However, over the past three decades, the research community has produced numerous studies that focus almost exclusively on the economic and environmental dimensions of firm performance, while relatively little research has been conducted on firms’ social impacts (Sarkis, 1998; Lundin et al., 2004; Hervani, Helms and Sarkis, 2005; Ferri and Pedrini, 2018).

With increased relevance of corporate social responsibility (CSR), firms have been increasingly taking responsibility for their impacts not only on the natural environment, but also on society, and have become better corporate citizens by adopting CSR strategies (Carroll, 2000). Growing social demands for CSR practices influence all economic and business activities, especially in the food sector. Hartmann (2011) suggested the food sector is one of the most sensitive businesses in terms of CSR given its effects on the environment, animal welfare and society. By being implemented by the food companies, CSR and sustainable practices are now having an impact all along the food supply chain, from farm-to-fork (Weiss, 2012).

The main methodological challenge to include the social dimension of sustainability in performance measurement is to identify key indicators that reflect this dimension and to be able to measure them. For agriculture, Lebacq, Baret and Stilmant (2013) suggest that social sustainability involves a combination of indicators that include education, working conditions, wellbeing, health, quality of rural areas, acceptable agricultural practices and product quality. Van Calker et al. (2007) provide a definition of farm-level social sustainability based on working conditions and societal sustainability issues. Although there is no commonly accepted definition of social sustainability (Allen et al., 1991), there is a consensus that there are two types of social sustainability: private social sustainability that focuses on the farm community (e.g. farmers’ well-being and working conditions) and public social sustainability that relates to society as a whole, in the form of contributions to public goods (e.g. quality of rural areas, local employment, product responsibility) (Diazabakana et al., 2014).

Because of the many subjective aspects that affect social sustainability, measuring social performance is not trivial (Edum-Fotwe and Price, 2009). Several researchers suggest that measurements of the different aspects of social sustainability can be enhanced by combining both quantitative and qualitative data (Van Calker et al. 2007; Trainor and Graue, 2013). Although they are somewhat subjective, indicators for private social sustainability are relatively simple to elicit (e.g. number of days off or farmers’ subjective opinion on their quality of life). In contrast, assessing public social sustainability is more challenging, due to difficulties in identifying public goods and quantifying how agriculture contributes to their production (Ait Sidhoum, 2018). Due to data limitations, our empirical application for Catalan crop farms focuses on private social sustainability. More precisely, two outputs associated to the social dimension of sustainability are considered, namely the level of work satisfaction as perceived by farmers and the number of work injuries. We also include farmer’s working conditions as an input.

In terms of methods, the literature on farm sustainability proposes various indicators of farm performance that can account for the economic (e.g. farm output to cost ratio), environmental (e.g. share of pasture area on the farm) and social (e.g. farmer’s number of working hours) pillars and combines these indicators in various ways, such as composite indicators or radar charts (Barnes and Thomson, 2014; Latruffe et al., 2016). One shortcoming is that the three pillars of sustainability are usually assessed differently and combined in arbitrary ways. By contrast, the productive efficiency literature, which initially aimed at evaluating economic efficiency, extended its methods to assess both the economic and environmental performance of firms in a more consistent way (Zaim and Taskin, 2000; Agrell and Bogetoft, 2005; Cuesta, Lovell and Zofío, 2009; Dakpo, Jeanneaux and Latruffe, 2017). However, very few studies within this literature have addressed the social dimension. To the best of our knowledge, there is only one study that quantified the social dimension of firm performance through efficiency and this is not in agriculture (Chambers and Serra, 2018). This study used the non-parametric method Data Envelopment Analysis (DEA) to estimate firm efficiency that accounts for CSR activities. Our analysis departs from this approach by allowing for production risk.

Given the stochastic nature of agricultural production, it is crucial to account for production uncertainty in farms’ efficiency analyses (O’Donnell, Chambers and Quiggin, 2010). Most empirical studies on efficiency ignore this and use the level of realised output to measure farm performance. This, however, implies that poor outcomes arising from the stochastic nature of production may be confounded with an inefficient use of the technology. Our model allows for the stochastic conditions in which production takes place. We follow the proposal by Chambers and Quiggin (1998, 2000) and model risk using the state-contingent approach. This approach, which has its foundations in pioneering works by Debreu (1959) and Arrow (1965), differentiates outputs according to the state of nature in which these are realised. The production technology output is then defined in terms of the distribution of (planned) ex ante outputs, instead of the realised ex post output. As this requires specific data, only few empirical studies of production and efficiency use the state-contingent approach. Through our farm-level survey, we elicited ex ante production data to empirically represent the state-contingent technology of our sample farms. Our article thus takes the Chambers and Serra (2018) proposal one step further by modelling agricultural production under risk using the state-contingent approach.

3. Methodological framework

3.1. Production technology

The extension of production efficiency measures to account for the environmental impact of economic activities has not been without debate. Murty, Russell and Levkoff (2012) and Coelli, Lauwers and Van Huylenbroeck (2007) propose environmental efficiency measures based on the materials balance concept. However, these models have limitations, as underlined in Dakpo, Jeanneaux and Latruffe (2016) and Førsund (2018). In this context, both Serra, Chambers and Oude Lansink (2014) and Dakpo (2016) extend Murty, Russell and Levkoff's (2012) work, which models a polluting technology with two independent sub-technologies: one sub-technology that produces the desired output and one polluting sub-technology. Serra, Chambers and Oude Lansink (2014) and Dakpo (2016) each introduce an approach that overcomes the assumption of independent sub-technologies made in Murty’s approach. Dakpo (2016) proposes dependence constraints linking the various sub-technologies with the help of intensity variables. Serra, Chambers and Oude Lansink (2014) rely on the state-contingent approach and the materials balance principle to model the stochastic nitrogen runoff as a function of stochastic yields. Here, we use Serra, Chambers and Oude Lansink's (2014) approach, as it is consistent with the fact that farms face stochastic production conditions. We extend the latter approach to integrate (private) social outputs in the production technology. In this way, our model also extends on Chambers and Serra (2018) by allowing for stochastic production conditions. Our approach is based on the assumption that efficient farms maximise desirable agricultural outputs (agricultural output in our case study) and positive social outcomes (farmers’ satisfaction in our case study), and minimise two unintended environmental outputs and an unintended social output (nitrogen and chemical input pollution and work injuries, respectively, in our case study).

Our representation of T assumes constant returns to scale (CRS)² and meets materials balance conditions requirements (Coelli, Lauwers and Van Huylenbroeck, 2007; Serra, Chambers and Oude Lansink, 2014). In this regard, the applications of organic and chemical fertilisers (r_k) equal the quantity absorbed in the production of intended outputs plus the runoff by-products. Fertiliser runoff is random since the quantity of fertiliser absorbed by plants depends on plant growth and can be represented by $rk=q˜k+z˜k,$ where $q˜k$ is the quantity of fertiliser input $rk$ absorbed by agricultural production and $z˜k$ represents the runoff. Only the quantity of fertiliser that remains on the crop (⁠$q˜k$⁠) has an impact on the quantity of crop produced (Serra, Chambers and Oude Lansink, 2014). We further assume that any application of PHI results in pollution, whether it stays on or within the plant or runs off.

Each of the five sub-technologies that integrate the global production set is associated to a specific output derived from a set of inputs that interact with each other, as described in the following paragraphs. The first sub-technology $(TY)$ evaluates economic (productive) performance by capturing how productive inputs $(xn)$⁠, PHI applications $(cd)$⁠, the absorbed quantity of nitrogen $(q˜k)$ and working conditions $(wa)$ contribute to increase the desired agricultural output $(y˜h)$⁠. We assume free disposability, which implies that, given an input set, output can be reduced by any desired amount, free of charge.

The second sub-technology (⁠$TZ)$ evaluates environmental performance and measures the nitrogen that has not been absorbed by the crops, which is costly to dispose. Nitrogen runoff (⁠$z˜k$⁠) is assumed to increase with fertiliser application (⁠$rk$⁠). Following Serra, Chambers and Oude Lansink (2014), we allow fertiliser pollution to depend on productive inputs (x_n) such as labour and capital. To model x_n we assume fertiliser is used by the sample farms to reduce production risk. In this regard, small farms are assumed to be more risk-averse than larger farms and thus to use fertiliser more liberally. We further assume that an improvement in working conditions (⁠$wa$⁠) improves labour performance, as well as farmers’ judgements regarding the need to apply fertilisers, which can contribute to moderate consumption of polluting inputs and nitrogen runoff.

The third sub-technology $(TP)$ also measures environmental performance and reflects the assumption that any application of PHI $(cd)$ results in pollution $(p)$⁠, which cannot be abated without a decline in farm activity levels (weak disposability assumption). We also allow conventional inputs $(xn),$ such as area of land sprayed, to increase the total environmental impact of PHI. In this regard, larger farms tend to be more extensive in their production methods, which leaves more room for weed growth and requires a more intensive herbicide application. An exception is the amount of seeds, as a larger quantity of seeds implies a higher crop density, thus less space for weeds, which should reduce the need for herbicides.³ Better working conditions $(wa)$ are also assumed to improve farmers’ judgements in applying polluting inputs and reduce environmental impacts.

Following Chambers and Quiggin (2000), uncertainty is represented through the state space Ω, which contains a number of states (⁠$ω=1,…,Ω$⁠) randomly chosen by nature. We assume that the farmer chooses a technically feasible combination of inputs and random outputs before ‘nature’ makes a choice. For example, the vector of non-stochastic input quantities, $x∈ℝ+N,$ is used to produce the state-contingent desirable output vector, $y˜∈ℝ+Ω$⁠, where $y˜={ye:e∈Ω}$⁠, with $ye$ the ex post value if nature chooses state e. Based on the assumptions discussed in this subsection, our DEA model specification is presented in the next subsection.

3.2. DEA models

3.3. Computation of overall efficiency

The overall sustainability efficiency index in equation (13) is a simple average of the five sub-technologies’ scores. A special feature of this index proposed by Murty, Russell and Levkoff (2012) is that a farm is considered as efficient if and only if a score of 1 is obtained for each sub-technology, i.e. $ξ1ω=ξ2ω=ξ3=ξ4=ξ5=1.$

4. Data

4.1. Database

Our analysis is based on cross-sectional farm-level data collected in 2015 from a sample of 173 agricultural holdings specialised in the production of cereal, oilseed and protein (COP) crops and located in the region of Catalonia in Spain. The Spanish COP crop production reached nearly 4 billion Euros in 2015 and represented more than 13 per cent of the total crop production in the country (Spanish Ministry of Agriculture, 2016). The production is generated by more than 130,000 holdings, namely 13 per cent of all Spanish agricultural holdings (INE, 2013), on the largest proportion of the utilised agricultural area (UAA) in Spain: the UAA in Spain totalled 23.3 million hectares in 2013 (INE, 2013) of which 32 per cent were being devoted to COP crops.

As noted by Chambers and Quiggin (2000), the key challenge to construct empirical representations of state-contingent technologies is the lack of information on the ex ante distribution of the random variables. We follow Chambers, Serra and Stefanou (2015) and use survey-elicited ex ante outputs to empirically represent the stochastic technology. For this purpose, we conducted a survey before the beginning of the agricultural season (October 2015) to collect point estimates of anticipated yields for three alternative states of nature: bad, normal and ideal growing conditions, that is to say $y=(y1,y2,y3)$ (see Chambers, Serra and Stefanou (2015) and Serra, Chambers and Oude Lansink (2014) for further details). We also collected detailed information from each farm on planned (ex ante) input use, including arable land (⁠$x1$ in hectares), capital in terms of replacement value (⁠$x2$ in Euros), paid and unpaid labour (⁠$x3$ in hours), energy expenses $(x4$ in Euros) and crop-specific inputs, namely applications of PHI (⁠$cd$ in litres), seed expenses (⁠$x5$ in Euros) and quantities of nitrogen (⁠$rk$ in kilograms). In order to estimate the sub-technologies representing farm social outputs, we also collected information on farmers’ degree of work satisfaction⁴ (⁠$s$-Likert scale from 1 to 4) and information on the accidents and work injuries (⁠$i$⁠) occurring in the farm.

Table 1 provides summary statistics for the variables considered in this study. It shows that COP crop output value per farm fluctuates from less than 30,000 to more than 63,000 Euros, depending on the state of nature, with 46,000 Euros being the average in normal conditions. On average, our sample farms cultivate 72 hectares of arable land, have a capital replacement value of 145,000 Euros, devote slightly less than 800 labour hours per year to the farm activities and spend around 4,400 and 3,900 Euros on energy and seeds, respectively. Derivation of the pollution and social variables was a little more involved, which we discuss in the following two sub-sections.

4.2. Pollution proxies

On average, sample farms apply 80 litres of PHI, which corresponds to a rate that is slightly higher than 1 litre per hectare. Schreinemachers and Tipraqsa (2012) report an average value of around 1.7 kilograms of active ingredients per hectare in Spain in the period 1999–2001, indicating that our sample farms are below the national average. To quantify PHI pollution (p) we use the environmental impact quotient (EIQ) developed at Cornell University that provides an estimation of the environmental and health impacts derived from PHI applications (Kovach, Petzoldt and Degni, 1992; Eshenaur et al., 2015).⁵ More specifically, to estimate PHI pollution generated by a farm, we multiply the amount, in litres, of each active ingredient applied on the farm, by its corresponding EIQ and add the total. The resulting quantity is taken as the estimate of p (the output of the PHI pollution sub-technology). Since EIQ is a unit-free index, p is expressed in litres.

In order to estimate nitrogen pollution (⁠$zk$⁠), we follow Serra, Chambers and Oude Lansink (2014) and use the quantities of chemical and organic fertilisers applied and convert them into nitrogen quantities $(rk)$⁠. While for chemical fertilisers the quantity of nitrogen can be easily found in the commercial product label, we use Mercadé, Delgado and Gil (2012)’s coefficients to approximate the quantity of nitrogen contained in organic fertilisers and the Spanish Ministry of Agriculture Fisheries (2010) coefficients to quantify the nitrogen content in seeds. We then calculate nitrogen balance. Implementing the nitrogen balance constraint requires estimation of crop nitrogen removal $(q˜k)$⁠, which depends on yields. As our desirable output is collected under three alternative states of nature, we compute three possible nitrogen removal quantities per farm (⁠$q1,q2,q3$⁠) (see Serra, Chambers and Oude Lansink (2014) for further details). By computing the difference between nitrogen applied $(rk)$ and nitrogen removed $(q˜k)$⁠, three possible nitrogen balances (one for each state of nature) are obtained (⁠$z1,z2,z3$⁠). The nitrogen balance per farm fluctuates from 5,900 kg (in bad crop growing conditions) to 3,500 kg in good crop growing conditions (see Table 1), which is compatible with higher amounts of nitrogen being absorbed by crops under good crop growing conditions.

4.3. Social proxies

Regarding the extent of farm’s worker injuries (i), since farms in our sample are mainly family-based farms employing a very small number of workers (mainly members of the manager’s family), very few injuries were reported by farmers yielding an average of 0.35 injury per farm. As noted by Sueyoshi and Sekitani (2009), DEA models need to treat zeros in the data carefully. In order to avoid zero values in our data, we build a new injuries variable as follows: we give an initial score of 100 to each farm; then, for each minor injury reported we remove 5 points, while we remove 20 points for a serious injury.⁶ For example, a farm with one minor injury and one serious injury will have a score of $i=100−(5+20)=75$⁠. The resulting average score above 97 (Table 1) confirms the low level of injuries in our sample farms. Since the higher the score, the fewer the injuries, we need to flip the inequality sign in the last equation in model (12).

Farmers’ work satisfaction (s) was obtained by asking farmers to value their overall degree of satisfaction with their work on a Likert scale, from 1 to 4,⁷ with 1 being the lowest and 4 the highest degree of satisfaction. Table 1 shows that the average work satisfaction degree is 3.4, indicating a relatively high satisfaction level.

To derive a quantitative measure of working conditions (⁠$wa$⁠), we also used a four-point Likert scale. Farmers were asked to value 16 items reflecting different dimensions of working conditions: workload, difficulty of the work, creativity, skills development, freedom in decision making, flexibility of schedules, work motivation etc. To reduce the number of netputs and improve the discriminatory ability of DEA, we perform a principal component analysis (PCA) on these 16 working condition items. Hence, PCA is used here as a descriptive technique, allowing for the variables to be of any particular type (Jolliffe, 2011: 339). PCA results reveal that the underlying structure of the working conditions can be largely explained by three independent components: skill discretion, decision autonomy and physical and psychological demand. The factor ‘skill discretion’ summarises the information related to creativity at work, skills and versatility in the workplace. ‘Decision autonomy’ represents the capacity and freedom of farmers to make decisions without external constraints. Finally, ‘physical and psychological demand’ represents requirements related to workload. Only the items with a significant loading in each of the three factors are retained for the analysis (namely 12 items, see Table 2). Each component is quantified by summing the score points of the corresponding items to obtain three continuous variables (Rubin et al., 2007; Spurrier et al., 2008). The average scores of skill discretion and decision autonomy are 12.71 and 13.77, respectively while physical and psychological demand shows a lower average of 9.72.

5. Results

Efficiency scores are derived using the General Algebraic Modelling System (GAMS) software and are shown in Table 3. Figure 1 summarises efficiency scores through histograms and nonparametric kernel density functions for each sub-technology. The results show heterogeneity in farm performance across the different sub-technologies. The overall efficiency averages 0.741 (last row in Table 3) and encompasses technical (or economic), environmental and social efficiency scores equation (13). Environmental efficiency, measuring the farm performance in minimising pollution caused by both PHI and nitrogen, is the lowest among the three components; on average, 0.526. This confirms an inefficient use of environmentally detrimental inputs (Reinhard, Knox Lovell and Thijssen, 2000). In contrast, the desired output technical (or economic) efficiency is 0.879 on average. These results are consistent with previous studies that showed that a relatively high degree of technical efficiency goes together with a relatively low environmental performance (Dakpo, Jeanneaux and Latruffe, 2017). Reinhard and Thijssen (2000) found nitrogen efficiency levels around 0.56 for a sample of Dutch dairy farms. Asmild and Hougaard (2006), who analysed a sample of Danish pig farms, found that half of the farms were environmentally inefficient with average efficiency scores in the range between 0.34 and 0.56. Still in the livestock sector, Dakpo, Jeanneaux and Latruffe (2017) obtained an average greenhouse-gas-emission efficiency of 0.51 and good output efficiency of 0.74 for sheep meat breeding farms in grassland areas in France. Guesmi and Serra (2015), who assessed pesticides efficiency of a sample of Catalan arable crop farms, found relatively low efficiency levels (from 0.57 to 0.42).