Food system sustainability metric

The protocol used to build the metric of food system sustainability has been developed and presented in detail in [18]. What follows below is a brief technical overview of how the metric and the country’s associated sustainability scores were built. The first step was a thorough literature review that was used to identify the relevant peer-reviewed articles, documents, and reports published in English between 2000 and 2018 that discuss indicators and metrics of food system sustainability. The search was conducted using multidisciplinary databases including Google scholar, JSTOR, and Scopus, and searching for the keywords “food system(s)” AND (“sustainability” OR “sustainable”) in the title or abstract of the documents. After removing of duplicates, 83 documents were retained (see references in S1 Table for the complete list). The review of those 83 document indicated that four dimensions of sustainability are generally acknowledged in the literature related to food systems: environment, economic, social and food security & nutrition. The food system indicators mentioned in the 83 documents were then systematically recorded. One hundred and ninety-two different indicators were identified and documented, along with the dimensions, components and, when available, domains of the metric to which those indicators were linked.

The list of 192 indicators was refined using nine inclusion/exclusion criteria specifically considered to address methodological and conceptual issues revealed by the literature review–see [18] for details. Exclusion criteria included: evidence of cross-correlation between indicators, use of latent variables, composite, and non-comparability nature of the indicators. Inclusion criteria included: conceptual relevance, global scale, global validity, time period (2000–2018), and clear methodology. The detailed definitions of those inclusion/exclusion criteria and the issues they addressed are provided in S2 Table. Applying the inclusion/exclusion criteria to the pool of 192 indicators resulted in a short-list of 27 indicators (see diagram in S1 Fig for a PRISMA flow diagram of the selection process).

The next step was to compute countries’ sustainability scores. Indicators were first normalized using Box-Cox transformation to avoid issues associated with potential heteroskedastic datasets, and then normalized using a standard min-max transformation with a [0, 1] range. Based on the literature on composite indicators and multi-criteria analysis (see e.g. [19–21]), a formula was developed for the computation of the sustainability score. Two criteria were considered for this: (i) the potential degree of compensation that exists between the four dimensions of the metric; and (ii) the cross-correlations between indicators within each dimensions of the score. A minimal level of compensation was assumed between the four dimensions of the metric (which means that no particular dimension of the sustainability metric was expected to be fully substitutable by any other). Based on this reasoning and the results of the cross-correlation matrixes estimated for each of those dimensions (see diagram in S2 Fig), the sustainability score formula was: (1) where SuScore is the country sustainability score, Ind₁ … Ind_n, Ind_m, Ind_k, and Ind_l represent the groups of indicators associated with the four dimensions of the metric (Env, Soc, Econ, and Food&Nutr). ‘Geometric’ and ‘Arithmetic’ indicate the types of mean used within and between the different elements of the metric to aggregate the indicators.

One critical issue—albeit rarely discussed in papers dealing with the construction of global metrics—is that it is not possible to maximize both dimensions of the metric at the same time: a trade-off exists between the number of countries that can be included in the metric and the number of indicators used to build that metric; the larger the numbers of indicators, the lower the number of countries for which all indicators are available, and vice versa (see diagram in S3 Fig for an illustration of that trade-off). A three-criterion decision approach is then used to identify what combination (countries–indicators) is optimal, given the current availability of data. The three criteria were: (i) the number of countries that are dropped out of the metric when the number of indicators is increased by one; (ii) the standard deviation of the aggregated mean of the countries’ scores when the number of indicators included in the metric is increased or decreased by one; and (iii) the degree to which countries’ ranks shift when a new indicator is included in the metric–see [18] for details. Using a simple minimal function approach, [18] was able to determine that the combination of 20 indicators– 97 countries is the optimal combination at present given the current datasets available.

Food system drivers

The second step in the analysis is the identification of food system drivers. For this, peer-reviewed articles and gray literature discussing food systems were scanned, with the objective to identify which amongst those documents also treated the question of food system drivers. The keywords used for this search were “food system(s)” AND “driver(s)” in the title, abstract, or main body of the document (we acknowledge here that the term “driver(s)” is not (yet) systematically used in the food system literature and that its use as keyword may have led to the missing of a few potentially relevant articles). We used the same reference databases used by [18] for the sustainability indicators above (Google Scholar, JSTOR, and Scopus) and the search was limited to the same time period (2000–2018) and to English (language) literature. Thirty-three documents were identified after elimination of duplicates.

An inductive framework [22] was then used to determine whether a core set of drivers could be identified from these 33 documents. For this, we started from the definition of food system drivers proposed by [5] (drivers of food systems are “endogenous or exogenous processes that deliberately or unintentionally affect or influence a food system over a long-enough period of time so that their impacts result in altering durably the activities [or actors’ behaviors], and subsequently the outcomes, of that system” p.152 –modified by us). We then applied the inductive framework: statements made in the 33 documents on what processes influence food system changes were systematically recorded. Convergences and similarities in those statements were then established and a series of propositions on what driver(s) were generating these patterns, were produced. This process allowed us to isolate a subset of archetypal (or key) drivers.

For each of these key drivers, potential links with indicators (of the magnitude of change associated with drivers) were identified from publicly available datasets. In addition to their conceptual relevance, two additional criteria were used to determine the suitability of the indicators as drivers of food systems: (i) their global coverage (i.e. covering low-, middle- and high-income countries); and (ii) their availability for a period of time long enough for potential changes related to the cumulative and/or lagged impact of these drivers to be observable and estimated. This last criterion stems from the premise that what constitute the driving elements for most of the transitions or transformations that are observed in food systems are not just the results of the current (or past) levels of the process considered but the change in those levels over time. For instance, it is not just the temperature or the intensity of precipitation that is thought to lead to adaptations (or transformations) amongst the actors of food systems, but the change in temperature or intensity of precipitation relative to past conditions [17]. Likewise it is not the level of consumer income per se that is said to drive the rise in demand for animal-based protein as currently observed in middle income countries, but the relative increase in that consumer income [23,24]. An 8 to 15-year window (depending on the availability of the data) was therefore used to compute those changes in levels.

Correlation between drivers and level of sustainability

The correlations between individual drivers and sustainability scores were established by computing Spearman correlation coefficients (ρ) between the two country-level series after outliers were removed (using the R package OutliersO3). Spearman correlation was chosen because it does not require continuous-level data (as it uses ranks) and does not assume normal distribution of the variables; as such it is more robust than the Pearson correlation test [25].

The statistical significance of the correlation coefficients (ρ) were computed at α = 0.05. The analysis allowed us to test the null hypothesis H_o (namely that there is no monotonic relation between drivers and sustainability scores). In this context, a p ≥ 0.05 would signify that Ho could not be rejected, i.e. that no significant correlation exists. In parallel, the strength of the correlation coefficients ρ was assessed using a 5-scale system as proposed by [26] where a case of 0.0 ≤ IρI < 0.2 indicates a very weak correlation; 0.2 ≤ IρI < 0.4 = weak; 0.4 ≤ IρI < 0.6 = moderate; 0.6 ≤ IρI < 0.8 = strong; 0.8 ≤ IρI ≤ 1.0 = very strong correlation. Note that the two tests are independent in the sense that the level of statistical significance of the Spearman correlation does not bear any implication about the strength of the relationship (in that a relationship can be strongly significant but weak–or vice versa).

Principal component analysis

While a one-to-one correlation analysis between the food sustainability score and drivers provides a solid representation of the dynamics between the two factors taken individually (although the directionality of the link may not be obvious), it does not control for potential confounding factors and/or multicollinearity between drivers. Multivariable analyses would be necessary to address those issues. However, because the set of indicators used to proxy the drivers varies from one country to the next, running a multi-variable correlation analysis with the 20 drivers as explanatory variables would be possible only with 37 countries–which would significantly reduce the robustness and the representativeness of the analysis. Instead we performed a Principal Component Analysis (PCA). The motivation was to go beyond the one-to-one basis of the Spearman correlation and to consider all drivers together in order to identify potential interactions between those drivers. We used the ‘PCA’ command from FactoMineR library in R^® to perform the analysis, and the model is computed using the correlation matrix of the variables (drivers).

Assessing the sustainability of food systems

The first component of our analysis was therefore the assessment of the sustainability of food systems. For this, we replicated the approach developed and presented in their recent paper by [18]. Four dimensions of sustainability were included in the analysis to capture food system’s sustainability: environmental, economic, social, and food security & nutrition dimensions [2,27–29]. All four dimensions are complex and compound in nature, and two of them (the environmental and food security and nutrition dimensions) are usually further decomposed into components. For environment, the most frequently discussed components are: air, water, soils and land, (bio)diversity, and energy [6,30–32], and for food and nutrition, those components are: food security, food safety, food waste and losses, and nutrition [33–35].

Following the protocol designed by [18], the four dimensions and their corresponding components were combined to represent the first two levels of the food system sustainability metric. Specific domains were then added to each dimension. For food security, the four domains conventionally referred to in the literature are: availability, accessibility, utilization, and stability [36,37]. For water, and soils and land, two domains are usually considered: ‘quality’ and ‘use’. Those were therefore incorporated in the metric, while for biodiversity, wildlife and agrobiodiversity were included. For the social dimension, the literature indicates that issues of equity, gender and inclusion are commonly considered [38,39]. Less consensus exists regarding the representation of economic sustainability. Financial performance (creation of value added), employment rate, and economic inequality were selected for this study [31,35]. For the nutrition component, diet and the triple burden of malnutrition (undernutrition, overweight and obesity, and nutrient deficiency) were included [37,40,41]. In total, this results in 27 domains covering the four dimensions of food system sustainability (see Table 1). For each of these 27 domains, available indicators were compiled from publicly available databases (e.g. World Bank, IMF, UN-agencies)–see fourth column in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The four dimensions, 27 domains and associated indicators used to build the metric of food system sustainability.

The 20 indicators indicated with an * are those which were eventually retained by [18] to calculate the 97 countries’ sustainability scores.

https://doi.org/10.1371/journal.pone.0231071.t001

Cursory inspection of the existing databases reveals that the same indicators are rarely available across the full range of low, middle, and high-income countries. In other instances, indicators are available, but only for a sub-set of countries that lack coverage with the rest of the indicators. As a result, given the current availability of data at the global level, it is not possible to build a full metric of food system sustainability including all countries in the world. A compromise needs to be made between the number of countries and the number of indicators to be included in the metric. Based on the decision criteria presented in the methodology section, [18] show that the optimal combination that balances number of countries and dimensionality of representation is 20 indicators and 97 countries.

Using formula (1) we then computed individual country’s sustainability scores for the 97 countries for which the 20 indicators were available. The resulting map is displayed in Fig 1. It is similar to the one presented in [18]. The proportions of low, middle, and high-income countries amongst those 97 countries are respectively 18% (L), 51% (M), and 31% (H), which is remarkably close to the proportions observed for the 218 countries and other regions listed in the World Bank 2017 list, respectively: 14% (L); 49% (M); 37% (H). The subsample of 97 countries included in this analysis is therefore a good representation (in terms of proportions of low, middle and high-income countries) of the 2017 situation in the world.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Food system sustainability score calculated for 97 countries and 20 indicators covering the four dimensions of food system: Environment, economic, social, and food security & nutrition.

The list of indicators used to build the map is provided in Table 1. Country individual scores are provided in S4 Table. Reprinted from [18] under a CC BY license, with permission from Scientific Data, original copyright 2019.

https://doi.org/10.1371/journal.pone.0231071.g001

Drivers of food system sustainability

The second step in the analysis consisted in integrating the major drivers of food systems changes. Drawing on the review of food system drivers conducted by [5] and the methods presented earlier, 12 key drivers were identified from existing available datasets. Those 12 drivers are listed in the second column of Table 2, organized according to their contribution to the three major elements of food system: production-supply, trade-distribution, and demand-consumption.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Food system drivers and their indicators.

https://doi.org/10.1371/journal.pone.0231071.t002

On the consumption/demand side, three distinct but related drivers are widely acknowledged in the literature: population growth, increases in (consumers’) income, and urbanization [22,30,42–43]. A fourth key driver that is expected to have increasing influence on the consumer/demand component in the future is the growing attention paid by the different actors of the food systems to diet and health-related issues [44]. Although only marginal influence is seen to date (see e.g. [45–47] for some examples), attention paid to diet and health-related issues is expected to become one of the key drivers on the consumption/demand side in the future.

On the production/supply side, several forms of technological innovation (agriculture mechanization, irrigation, plant breeding, etc.) along with intensification and homogenization of the agricultural sector have been identified as two major drivers of food systems changes [48–50]. Conjointly, improved access to local infrastructure (e.g. power grid, roads) and to urban centers [51–53] as well as to local supermarkets [54,55] have led to significant changes, especially for the smaller and medium-size producers in low and middle income countries (LMICs) [56–58]. In parallel, general degradation of soil quality and other agro-ecosystem conditions is widely recognized as a key driver of decrease in agricultural production [13,14,30]. The literature also identifies increases in temperature combined with higher intensity and severity of extreme weather events (which all fall under the generic category of “climate change”) as an important driver of food systems [3,59–63].

At the intersection of supply/production and distribution/trade, policies and other related national and international processes established to facilitate or mitigate trade expansion and exportation of agricultural products (e.g. subsidies in OECD countries) are another significant driver of food system changes [64–71]. On this distribution-trade side, the two other key drivers often mentioned in the literature are: internationalization of private investments (leading to, amongst others, the “supermarketization” of food systems [72]); and the growing concern from local and national policy makers/governments for food safety [73–75].

Once the drivers were identified, we looked for indicators or proxies for each of them in publicly available datasets. Twenty indicators were identified (meaning that for some drivers, more than one indicator/proxy could be identified). Those are listed in the third column of Table 2, along with the databases from which they were extracted, the time period and the number of countries for which they are available, and the selected periods and methods used to compute the changes in those indicators.

Correlating levels of sustainability and drivers

Once the indicators of each key driver were identified, individual Spearman rank correlations (ρ) were estimated to evaluate the strength of the relationships between the drivers and the 97 countries’ sustainability scores. The results of the Spearman tests are presented in Table 3. They indicate that amongst the 12 drivers, three of them show moderate (0.4 ≤ IρI < 0.6) to strong (0.6 ≤ IρI < 0.8) levels of correlation with the sustainability scores, and seven show weak correlation (0.2 ≤ IρI < 0.4). No very strong correlation (IρI of 0.8 or greater) were observed. The statistical significance of the Spearman coefficients (determined for α = 0.05) indicates that the existence of those correlations is confirmed in nine out of the 20 correlations tested (Table 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Spearman correlation tests between food system drivers and country food system sustainability scores.

Results ranked from the strongest to the weakest coefficients ρ (in absolute values). Drivers above the dotted line are those displayed in Fig 2.

https://doi.org/10.1371/journal.pone.0231071.t003

The strongest correlation was observed between food system sustainability and change over time in merchandise and services trade per capita (Fig 2A). This indicator was introduced as a proxy for the internationalization of private investments in food systems. The correlation was found to be positive and strong (ρ = 0.70, p < 0.0001), suggesting that countries with large increase in merchandise and services trade over the last decade (2004–2015) are also amongst the countries with high sustainability scores. The color-code used to discern between low-, middle- and high-income countries reveals that middle-income countries are amongst those that contribute the most to this strong correlation. Around 3,000 USD of trade per capita the relation starts plateauing and beyond this point (mainly for high-income countries), an increase in trade is not associated with any further increase in food systems’ sustainability.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Relationship between country food system sustainability scores and key drivers.

Only drivers displaying a Spearman correlation coefficient (ρ) greater than 0.3 are shown: (a) change in merchandise and services trade; (b) change in population growth; (c) changes in agricultural area; (d) change in women employment in services; and (e) change in urban population. Trends (blue lines) estimated using linear or polynomials of degree 2 functions with a jacknife resampling method, are for illustration purpose only.

https://doi.org/10.1371/journal.pone.0231071.g002

The second strongest correlation was found with change over time in population growth (Fig 2B). In this case, a strong and negative association with countries’ food system sustainability scores is observed (ρ = -0.66, p < 0.0001), suggesting that the countries which have experienced the fastest growth in population in the last decade are also amongst those with the lowest food sustainability scores. The figure indicates that those countries are for their majority low-income countries. On the other end of the spectrum, high-income countries, which are characterized by lower levels of population change, are also amongst those displaying higher sustainability scores.

The third correlation is moderate in strength and, again, negative (ρ = -0.47, p < 0.0001). It concerns the changes over time in agricultural area, used as a proxy for the intensification of agriculture sector (Fig 2C). Countries on the right hand side of the 0 value along the horizontal axis (i.e. with positive values) are countries which have continued to expand their agricultural “frontiers” in the last decade. Those are also amongst the countries with the lowest sustainability scores. The figure indicates that they are mainly middle- and low-income countries. In contrast, countries with negative values (essentially high-income countries) are also characterized by higher sustainability scores.

Two more drivers display Spearman coefficients that are greater than 0.30 and at the same time statistically significant. Those are the change over time in women employment in services (ρ = -0.34, p = 0.013) and the change over time in urban population (ρ = -0.33, p = 0.001)–both introduced in the analysis as indicators of urbanization and its associated changes in life style (Fig 2D and 2E). For those two correlations, the relationship is negative, and the distinction low / high income, although still observable, is less marked.

The interpretation of all these different results will be discussed in greater detail in the following section.

Beyond the five drivers mentioned above, four other indicators appear to have statistically significant correlations with the food system sustainability scores. These are: changes in fertilizer use (negative); (ii) change in cereal yield (positive); (iii) change in foreign direct investment (positive), and (iv) change in GDP per capita (negative). Their Spearman coefficients are below 0.30 however, indicating that these correlations are relatively weak–see Table 3 for details.

Principal component analysis

To strengthen our understanding of the relationship between drivers and indicators, a Principal Component Analysis (PCA) was then implemented, with the main objective to verify that the results obtained through the univariate framework of the Spearman correlation were robust and not distorted by the existence of collinearity between drivers. The results of the PCA (performed using the drivers’ correlation matrix) are summarized numerically in Table 4 (eigenvalues and % variance for the 10 first principal components).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Results of the principal component analysis⁽¹⁾.

https://doi.org/10.1371/journal.pone.0231071.t004

Table 4 reveals that the first two principal components (PCs) captures only a relatively modest proportion of the overall variance (24%) and that the % of variance along the 10 first PCs is characterized by low and very similar values decreasing only very slowly across the PCs. This suggests that the variables (drivers) are “scattered” in the variable’s space and essentially uncorrelated. This is confirmed by the cross-correlation matrix computed for the 20 drivers (displayed in diagramme in S4 Fig) which shows very little cross-correlations between the drivers. Overall, this indicates that the risk of multi-collinearity in our analysis is minimal and therefore that the Spearman correlation results are robust.