Systematic review on effects of bioenergy from edible versus inedible feedstocks on food security

Achieving food security is a critical challenge of the Anthropocene that may conflict with environmental and societal goals such as increased energy access. The “fuel versus food” debate coupled with climate mitigation efforts has given rise to next-generation biofuels. Findings of this systematic review indicate just over half of the studies (56% of 224 publications) reported a negative impact of bioenergy production on food security. However, no relationship was found between bioenergy feedstocks that are edible versus inedible and food security (P value = 0.15). A strong relationship was found between bioenergy and type of food security parameter (P value < 0.001), sociodemographic index of study location (P value = 0.001), spatial scale (P value < 0.001), and temporal scale (P value = 0.017). Programs and policies focused on bioenergy and climate mitigation should monitor multiple food security parameters at various scales over the long term toward achieving diverse sustainability goals.

3 of 46 effect", "Positive effect" and "Both negative and positive" represented by purple, blue, green and yellowgreen respectively.
The first alluvial diagram (Supplementary Figure 1) includes all 224 journal articles. We can trace the flows backwards (right to left) to see how the bioenergy overall effect is related to the Bioenergy feedstock, SDI group, and food factors. We can see that the majority of the articles that indicated a negative overall effect are in the food price, food production and multiple food factor levels, and are spread across all SDI groups, although more heavily concentrated in the higher SDI, multiple or global SDI groups and roughly equal in all bioenergy feedstocks. The articles that found no effect indicated food production and there were various SDI groups as well as bioenergy feedstocks. Positive overall effects were not found in papers that indicated food price was a factor and were more often in lower SDI groups.
Both negative and positive overall effects were more likely found in papers with multiple food factors and varying SDI groups and bioenergy feedstocks.

of 46
As there were only four papers with bioenergy feedstock not specified, these papers were removed from the dataset and a new alluvial diagram generated. The patterns visible did not drastically change with the removal. However, the papers that discussed edible bioenergy with a low SDI group were more easily visible with the removal of those four papers. To summarize the data, tables were created that included counts of papers by food factor, bioenergy feedstock by overall effect on food security.

Supplementary Table 2: Counts of journal articles for each bioenergy feedstock classification by
bioenergy overall effect when food availability as a food factor was mentioned in the article.

Edible Inedible Both Edible and Inedible
None -found no effect 3 3 3 Negative effect 4 2 2 Positive effect 2 4 1 Both negative and positive 1 1 0 6 of 46 Supplementary Table 3: Counts of journal articles for each bioenergy feedstock classification by bioenergy overall effect when food price as a food factor was mentioned in the article.

Edible Inedible Both Edible and Inedible
None -found no effect   Since there are several categories with small counts (less than 5) and expected cell counts were lower than required for the parametric test, primarily in the bioenergy feedstock "Not Specified" category and the "Both negative and positive" overall effect category here, a non-parametric (permutation-based) Chi-Squared test was used to test for this relationship. More details on the permutation test used here can be found in (Greenwood (2020)). A randomization seed was fixed to ensure reproducibility of results.  The mosaic plot above shows that none of the standardized residuals are greater than 2 in absolute value, indicating that none of the combinations of bioenergy feedstock and overall effect are much greater than or less than expected if there was no relationship between the bioenergy feedstock and the overall effect on food security. This is as to be expected as the statistical test also indicated that we are not able to conclude that there is a feedstock effect. This plot also helps to visualize the combinations with small cell counts compared to the more common combinations of "Negative effect" and "Edible".

of 46
Alternately, studies were removed from the model with bioenergy feedstock "Not Specified".

of 46
Supplementary  The mosaic plot above shows that none of the standardized residuals are greater than 2 in absolute value, indicating that none of the combinations of bioenergy feedstock and overall effect are much greater than or less than expected if there was no relationship between the bioenergy feedstock and the overall effect on food security. This is as to be expected as the statistical test also indicated that we can't conclude that there is a bioenergy feedstock effect. This plot also helps to visualize the combinations with small cell counts compared to the more common combinations of "Negative effect" and "Edible".
Next, we analyzed overall effect by food factors (Food Price, Food Production, and Food Availability, or Multiple Factors).

of 46
Supplementary Again there are a few combinations with a small number of expected journal articles based on marginal totals, primarily in the "Both positive and negative" overall effect category and in the "Price" food factors 15 of 46 category. Again, a non-parametric Chi-Squared test of independence will be used to test for the relationship between food factors and the overall effect on the food security.

Supplementary Figure 8: Histogram of permuted test statistics for the Non-Parametric Chi-Squared test of a relationship between food factors and overall effect on food security. The vertical red line marks the observed Chi-Squared test statistic.
Both the parametric and non-parametric Chi-Squared tests indicate that there is strong evidence against there being no relationship between food factors and the overall effect on food security mentioned in the journal articles ( 2 = 44.8289, permutation-based p-value < 0.001).

of 46
Supplementary Figure  The mosaic plot above indicates that there are three combinations with a standardized residual that is larger than would be expected if there was no relationship between the food factors and the overall effect on food security. The combinations include food price and negative effect, food availability and positive effect, and multiple factors and both positive and negative effects. This seems to indicate that if a paper indicated both positive and negative effects, the paper also likely considered multiple food factors. Also there were more papers than would be expected that indicated the overall effect was negative that also considered food price as a factor. Lastly there were more papers than would be expected that found an overall positive effect that considered food availability as a factor. The rest of the combinations were about what you would expect if there was no relationship between food factors and overall effect on food security.

of 46
Next, we analyzed overall effect by socio-demographic index (Low, Low-Middle, Middle, High-Middle, High, Multiple, Global SDI) group.
Supplementary  There are a few combinations with a small number of expected journal articles based on marginal totals, primarily in the "Both positive and negative" and "Positive effect" overall effect categories as well as in the "Nonefound now effect" category. Every SDI group has at least one cell with a small expected number of journal articles. Again, a non-parametric Chi-Squared test of independence will be used to test for the relationship between SDI groups and the overall effect on the food security. From the mosaic plot we can see combinations contributing the most to the Chi-Squared statistic. The standardized residual for the low SDI group and positive overall effect is much larger than would be expected if there was no relationship between SDI group and overall effect on food security. The residual for low-middle SDI and Negative effect is smaller than would be expected, meaning there were fewer papers reporting a negative effect for low-Middle SDI countries than would be expected if there was no relationship between SDI group and overall effect on food security.

of 46
Then, we analyzed overall effect by spatial scale (Household, Community, Regional, National, Multinational, Global, and Multiple scales).  There are many combinations with a small number of expected journal articles based on marginal totals, primarily in the "Both positive and negative" and "Positive effect" overall effect categories. Again, a nonparametric Chi-Squared test of independence will be used to test for the relationship between spatial scale and the overall effect on the food security.

Supplementary Figure 12: Histogram of permuted test statistics for the Non-Parametric Chi-Squared test of a relationship between spatial scale and overall effect on food security. The vertical red line marks the observed Chi-Squared test statistic.
Both the parametric and non-parametric Chi-Squared tests indicate that there is strong evidence against there being no relationship between spatial scale and the overall effect on food security mentioned in the journal articles ( 2 = 47.7596, permutation-based p-value < 0.001).

of 46
Supplementary Figure 13: Mosaic plot of standardized residuals for the spatial scale by overall effect.

The cell that is shaded dark blue indicate there are is one standardized residual that is greater than 4 which is much larger than would be expected if there was no relationship between spatial scale and overall effect. One cell is shaded light red, meaning its standardized residual is between -4 and -2 and is smaller than would be expected if there was no relationship between spatial scale and overall effect. Cells with solid borders indicate positive residuals whereas cells with dashed borders indicate negative residuals.
From the mosaic plot we can see combinations contributing the most to the Chi-Squared statistic. The standardized residual for the household scale and positive overall effect is much larger than would be expected if there was no relationship between spatial scale and overall effect on food security. The residual for household and negative effect is smaller than would be expected, meaning there were fewer papers reporting a negative effect for household countries than would be expected if there was no relationship between spatial scale and overall effect on food security.

of 46
Next, we analyzed overall effect by temporal scale (Short-term, Long-term, Both short,-and long-term, and Term not specified). There are a few combinations with a small number of expected journal articles based on marginal totals, primarily in the "Both positive and negative" and "Positive effect" overall effect categories. Again, a nonparametric Chi-Squared test of independence will be used to test for the relationship between temporal scale and the overall effect on the food security.

Supplementary Figure 14: Histogram of permuted test statistics for the Non-Parametric Chi-Squared test of a relationship between temporal scale and overall effect on food security. The vertical red line marks the observed Chi-Squared test statistic.
Both the parametric and non-parametric Chi-Squared tests indicate that there is moderate evidence against there being no relationship between temporal scale and the overall effect on food security mentioned in the journal articles ( 2 = 19.1639, permutation-based p-value = 0.017).

of 46
Supplementary Figure 15: Mosaic plot of standardized residuals for the temporal scale by overall effect.

The cell that is shaded light blue indicate there are is one standardized residual that is between 2 and 4 which is larger than would be expected if there was no relationship between temporal scale and overall effect. Cells with solid borders indicate positive residuals whereas cells with dashed borders indicate negative residuals.
From the mosaic plot we can see combinations contributing the most to the Chi-Squared statistic. The standardized residual for the short term and positive overall effect is much larger than would be expected if there was no relationship between temporal scale and overall effect on food security.
Next, we analyzed overall effect by dataset type (Observed, Predicted, and Both Observed and Predicted).

of 46
Supplementary

of 46
There are three combinations with a small number of expected journal articles based on marginal totals, all in the "Both" category for the dataset type. Again, a non-parametric Chi-Squared test of independence will be used to test for the relationship between spatial scale and the overall effect on the food security.

Supplementary Figure 16: Histogram of permuted test statistics for the Non-Parametric Chi-Squared test of a relationship between dataset and overall effect on food security. The vertical red line marks the observed Chi-Squared test statistic.
Both the parametric and non-parametric Chi-Squared tests indicate that there is moderate evidence against there being no relationship between dataset and the overall effect on food security mentioned in the journal articles ( 2 = 13.4662, permutation-based p-value = 0.036).

of 46
Supplementary Figure  In this situation, there are no cells that provide a clear contribution to the test statistic that found moderate evidence of a relationship between data set type and bioenergy overall effect. It is interesting that the "observed" data sets had fewer negative effects than expected and the "predicted" ones had more than if no relationship were present, but the sizes of the differences from expected are not that large.
Statistical learning methods provide a way to explore papers for common characteristics from the multivariate information available using dimension reduction and clustering techniques to visualize and identify common patterns among the papers. These explorations seek to identify common patterns on bioenergy feedstock, food security parameter, temporal scale, spatial scale, and SDI group (see alluvial diagram in Supplementary Figure 21 for details on levels) and then explore potential connections of those identified clusters with the overall effect identified in the papers. This analysis excludes four papers that did not clearly specific Edible or Inedible bioenergy feedstock, so used n=220 observations. Gower's dissimilarity (Gower (1971)) is used to compare the multivariate categorical aspects of the papers using a weighted combination of the differences on each variable. In the calculation of Gower's dissimilarity, papers in the same category on a variable are measured as a difference of 0 and different categories (any combination) treated at 1, and these are averaged across all the variables to create the overall dissimilarity measure. Three techniques are used to explore connections among the papers. First, papers are clustered using hierarchical cluster analysis with Ward's method (Murtagh and Legendre (2014)), then the clusters are visualized in a two-dimensional map using multi-dimensional scaling with minimum spanning trees (Oksanen et al. (2019)) (minimum spanning trees show connections among closest neighboring papers), and finally alluvial diagrams are used to explore connections of the cluster identifiers and the features used to create them to more fully characterize the clusters and explore links between them and the type of overall effect found in the papers. Note that the type of overall effect was not used in the dimension reduction or cluster analysis so that the clusters could be compared to that characteristic.
Cluster analysis provides a tool for identifying groups of manuscripts with similar characteristics within a cluster and different ones among the clusters (see Everitt et al. (2011)). Hierarchical cluster analysis depends on the choice of dissimilarity (Gower's), the agglomeration method which we chose as Ward's linkage method (Murtagh and Legendre (2014)) as it created more distinct clusters with these data than some other linkage methods we explored, and the choice of the number of clusters. Based on the heights of mergers in the dendrogram, there appear to be four distinct clusters. Given the results for the four clusters, we can proceed to try to understand the characteristics of those clusters, by using the twodimensional map presented in Supplementary Figure 18  Classical multi-dimensional scaling is a dimension reduction technique for visualizing similarities/differences among observations in a map based on the relative proximities of the papers -here that is based on bioenergyFeedstock, foodFactors, temporalScale, spatialScale, and SDI information that is also used in the cluster analysis. The Euclidean distance in the plots provides information on the relative close-ness/far-ness of the papers which is created based on the Gower's dissimilarity matrix.
There is no clear interpretation of the axis directions as equivalent maps can be rotated, flipped, and re- and both types of bioenergy, that mainly contain production or multiple food factors with long term time scales and regional and national spatial scale, a mix of SDI scales with more of the higher levels and a mix of different types of effects. Cluster 4 (green) is mainly Edible bioenergy with large proportions in price, production, and multiple factors for food factors, short or long term temporal scales with mostly national or multinational scales and mix of SDI levels. Cluster 4 had a high proportion of negative effect and then no effect outcomes.
Supplementary Figure 21: Alluvial plot of the bioenergy feedstock, the food factors and the overall effect of the bioenergy on food security from a review of 220 journal articles, with clusters.
This following section describes a machine learning approach to uncover the effects of bioenergy on food security from the published studies described above. The use of machine learning is robust to the relatively small sample size of the present study and can account for the very large number of interactions between paper-level variables that need to be accounted for to treat the full scope of the relationships.
Natural language processing techniques (e.g. TFIDF, non-negative matrix factorization to project the text into a metric space) on text-heavy variables were explored but were adjudged to be unsuitable due to limitations in the vocabulary within these variables. Likewise, a maximum likelihood estimation approach was rejected for two reasons. First, the dataset is relatively small and each sub-dataset is smaller yet, and relying on the asymptotic properties of frequentist hypothesis testing with these data seemed questionable.
Second, the explanatory variable effects are likely dependent on interactions between them. This implies a potentially large number of terms to include in the model: 63 interactive terms to capture all interactions in the original six variables and more than 1 million for an exhaustive treatment of the binary indicators.
This would also have complicated post-analysis interpretation.
The methods used to conduct a study likely depends on the dependent variable studied. To account for this, three models were tested for the three major categories of dependent variables found within the papers in the sample: food price, food production, and food availability. By separating data into subsets that focused on only a single type of dependent variable (price, production, availability) the problem of modeling the dependent variable became a straightforward multi-categorial problem. Six explanatory variables were included in each of the models: SDI, bioenergy type, bioenergy generation, spatial scale, temporal scale, and data type. Data were transformed into multiple binary indicators, one for each category in the domain of the variable.
The relationships between overall effect and the paper-level variables was extracted using a machine learning approach. Random forest models are decision-tree ensembles in which each tree is constructed via bootstrap aggregation and sampling from the explanatory variable space (Breiman (2001)). They are suitable for capturing complex interactions between explanatory variables. The impact of each binary indictor created from the original six variables was assessed via partial dependence plots (Friedman (2001)).
The study-level variables on the found effects from the sample of papers reveals different effects across the three food security categories. This suggests that global conclusions about the effect of bioenergy on food security are not warranted and that effects depend on the specific aspect of food security examined.
The following summary focuses on effects that are robust in the sense they are obtained irrespective of the other variables they interact with.
In modeling the effect of study-level variables on price, studies that focused on edible biofuels or that took a global view were more likely to find a negative effect of bioenergy production (Supplementary Figure 22). Those that included middle SDI countries were more likely to lead to ambiguous effects with nearly all combinations of study-level variables resulting in clear increases in the probability of either both positive and negative effects. Finally, price effects derived from bioenergy are difficult to detect by focusing on low SDI countries, or by studying the household or regional levelinclusion of either of these was associated with finding no effect.
The papers that included low SDI countries or that analyzed the household level were more likely to find positive effects of bioenergy on production (Supplementary Figure 23). In terms of food availability, those papers that included household level analysis were more likely to find positive effects of bioenergy on food security while those that included low-middle SDI countries, that focused their work at the community level, or that studied first generation biofuels were more likely to find negative effects

of 46
The R package bookdown (Xie 2020a) was used to create this report document using the R language (R Core Team 2020). In addition, the following were packages used for the analysis and/or formatting of this • circlize (Gu 2020) • RColorBrewer (Neuwirth 2014) • viridis (Garnier 2018)