Figure 1. Map of Brazil identifying the Caatinga biome from where the data of 1141 municipalities were used to understand the relationship between forest and food security.

We built a multidimensional food security index for two time-points (2006 and 2017) to understand how it changed over time and how it is related to forest cover change. We built one index using a Principal Component Analysis for each year to reduce a large number of variables into a few dimensions following a method proposed by Hummel et al(2016). We selected only the PCs with an eigenvalue greater than 1 (Cutter et al., 2003) and then we changed the cardinality of each PC depending on whether the variables that compose the PC contribute positively or negatively to food security (Supplementary Table 2). The final food security index for each year was calculated as the sum of the individual scores of each PC for each municipality based on Hummel et al(2016). We estimated the food security change in the municipalities by calculating the difference between the final index score between 2006 and 2017.

We gathered the data of native vegetation cover from the MapBiomas project (MapBiomas, 2020). We considered all categories of forest (forest, savanna, mangrove and forest plantations) and non-forested native vegetation (wetland, grassland, salt flats, rocky outcrops and other non-forest formations) to calculate the forest cover percentage for each municipality. We grouped all types of native vegetation into one class (i.e., native vegetation cover) because people in the Caatinga derive many uses related to food security from all types of native vegetation and not only from forests (Albuquerque et al., 2017). Then, we identified and spatialized the municipalities with synergies (win-win and lose-lose) and trade-offs (win-lose and lose-win) between forest cover and food security change, respectively. We considered all municipalities that gained food security from 2006 to 2017 as a ‘win’ group, irrespective of the scores’ size, as well as ‘lose’ if the food security score was below zero.

We first principal components analyses to reduce dimensionality of our Multidimensional Food Security Index (MFSI). We used z-transformation to standardise scores of the principal components and calculate the MSFI. Because we changed the cardinality of the dimensions to always increase MFSI, the index is the sum of both positive (increasing security) and negative values (decreasing security). Because the dimensions were formed by different variables across years we provided a table to help readers interpret what drives food insecurity (Table S1). The absolute changes of values of MFSI and forest cover were used to create maps that help to understand spatial distribution of both changes in forest cover and food security. To understand the relationship between forest cover and MFSI and its two most important principal components in 2006 and 2017 we built spatial regression models with a quadratic term of forest cover when testing the effect of forest cover on MFSI and its first principal component (PC-1) and without quadratic term for the second PC. In all models, spatial error was included to check for nonlinear relationships. We used the errorsalm function from the package ‘spdep’ in the R environment. Moran I test, using the function moran.test was then used to test whether after accounting for spatial error, there was still spatial autocorrelation of the residuals.

Results