The study was conducted in three Climate Smart Villages (CSV) in Latin-American: Cauca, Colombia; Olopa, Guatemala; and Santa Rita, Honduras. Despite their distinct agro-climatic and geographical features, these villages exhibit similar farm socioeconomics, climate vulnerability, and adaptation strategies. The study used 2020 household survey data (Bonilla-Findji et al., 2020a,b,c). In brief, the CSV farming households are characterized by smallholder rainfed crop farming systems of staple crops, fruit trees and small livestock, with mostly a small-to-marginal farm size (Table 2). Furthermore, farming households from the three CSVs show similar adaptation strategies to climate change implemented through the climate smart village (CSV) approach framework of the Climate Change, Agriculture and Food Security (CCAFS) program (López et al., 2020). The analogous nature of the catalogues of CSA practices adopted by these three CSVs allows us to compare the impact of CSA adoption on capital-based livelihood vulnerability, which is the main focus of the paper, by combining regional data from three different locations.

The data used in the study corresponds to three published datasets containing household survey information from the “Integrated Monitoring Framework for Climate-Smart Agriculture” from the CCAFS research project (Bonilla-Findji et al., 2020a,b,c). The “Integrated Monitoring Framework for Climate-Smart Agriculture” was developed and used by CCAFS annually across the global network of Climate-Smart Villages to gather field-based evidence by tracking the progress on (i) adoption of CSA practices and technologies, as well as access to climate information services and (ii) their related impacts at household level (Bonilla-Findji et al., 2021).

The main objectives of the CSA monitoring framework are to (1) understand enabler and barriers for adoption of CSA technologies and practices; (2) understand gender-disaggregated perceived effects of CSA adoption, dis-adoption, control over resources and labour; and (3) understand CSA performance, synergies and trade-offs at the farm level. This framework introduces standardized descriptive indicators aimed at monitoring changes across various dimensions. These include five enabling dimensions that could influence adoption patterns, a core set of six indicators at the household level designed to evaluate the perceived impacts of CSA practices on aspects such as food Security, productivity, income, and climate vulnerability. Additionally, it incorporates four core indicators focusing on gender aspects, encompassing participation in decision-making, participation in implementation, access/control over resources, and work time. At the farm level, the framework includes seven core indicators to assess CSA performance, while also examining synergies and trade-offs among its three key pillars (Bonilla-Findji et al., 2021).

The final selection of monitoring household survey datasets from three Latin American countries were composed of 6 modules: (1) household demographics, (2) household agricultural activities, (3) exposure to climatic events, (4) access to weather and climatic information, (5) household food security and (6) adoption of CSA practices. From these modules within the CSV datasets, we selected the relevant information to construct the dataset for this study.¹ The CSA practices analyzed are listed in Table 3, it is important to note that according to the concept of CSA, practices and technologies should be implemented as portfolios considering context-specificities and acknowledging that sets of practices could actually contribute to climate smartness.

After selecting the final datasets, we proceeded with data processing by identifying the questions of the survey which addressed an aspect of a livelihood capital, from which a set of 23 unique mixed variables were constructed (Table 3). Each variable, whether numeric or categorical, represented one relevant aspect of a livelihood capital. All 23 variables were constructed based on previously used capital-based livelihood vulnerability indicators (Hahn et al., 2009; Pandey and Jha, 2012; Xu et al., 2020; Zhang et al., 2020; Zhang and Fang, 2020). We also consulted experts about factors that are known to influence capital-based livelihood vulnerability in Latin-American CSV farming households. The variable selection process resulted in a total of 23 mixed variables available for cluster analysis and farm type characterization. It is important to note that the survey dataset used covered relevant aspects of physical, financial, social and human capital extensively, whereas aspects of natural capital (such as soil or water quality) were less covered. For that reason, the natural capital variables selected represent qualitative proxies to estimate the state of natural resources in the farming household system.

Secondly, household records which contained no information on one or more of the selected survey questions were removed. Only households which had answered all the selected questions used to construct the 23 variables were included in the study. When there was more than one record per household, the answer from the head of household (man or woman) was prioritized and selected over the answers of other members of the household to ensure higher data reliability Despite the importance of assessing gender aspects in this study, the lack of systematic recording of two survey responses from both male and female head of household (especially in Cauca and Santa Rita CSVs) did not allow to include two responses per household and disaggregate the data by gender prior clustering. Finally, as a prerequisite for clustering, data quality checks were performed, which included removal of missing and out-of-range data, which unfortunately reduced the number of household records to use in the analysis, thus data did not allow enough reliability for performing gender analysis. Quantitative data were checked for outliers, which were removed to minimize bias and improve clustering. The average area destined for productive agricultural activities per household was restricted to a minimum of 200 m² and a maximum of 84,000 m² per household. Farming households which had a value outside that range were removed. No household had more than 10 combined number of tree species, crop species and livestock species.

Variable selection, data selection and data quality checks were carried out using Microsoft Excel (2020) software. The data cleaning process trimmed the initial datasets to a total of 267 households to be used in this study—71, 89 and 107 household records from Cauca, Olopa and Santa Rita, respectively. Unfortunately, due to the total number of female responses being rather low (30.3%), a separation of the total household sample by gender was not considered, as the sub-sample of women responses (81) was too low for the statistical analysis.

The statistical analysis of the data was divided into four steps: (1) multiple correspondence analysis (MCA); (2) computing the dissimilarity between observations using Gower’s metric; (3) agglomerative clustering using Ward’s minimum variance method and partitioning around medoids (PAM), and; (4) cluster validation using the silhouette width method. All the analyses were conducted in R version 3.6.3 (R Core Team, 2020), using the R packages “cluster,” “rtsne,” “ggplot,” “factoextra” and “factominer” for computing the clustering algorithms and respective visualizations. Table 4 shows the variables used in the cluster analysis.

First, to study the associations between the variables to be used in the cluster analysis, multiple correspondence analysis (MCA) was performed. MCA applies a dimensionality reduction algorithm to illustrate the principal components of the analysis in two dimensions. These two dimensions represent the combination of most influential variables in the analysis (Supplementary material). MCA allowed us to visualize the contributions of different variables to the combined first and second dimensions of the analysis, and therefore to identify the variables that most strongly influenced the variance of our data.

As a first step in the clustering Gower’s metric was used to compute the distance or dissimilarity between each pair of observations, since this distance metric is suitable for mixed data types and the datasets contained both categorical and numeric variables (Martin, 2016; Kassambara, 2017). As a prerequisite for clustering, Gower’s metric automatically normalized the data by rescaling it to a range between 0 and 1. Ward’s minimum variance method was used as the agglomeration or linkage method to compute the minimum distance between clusters. Ward’s method minimizes the total within-cluster variance, so that at each step of the clustering, the pair of clusters with minimum between-cluster distance are merged. Other authors also recommend using Gower’s metric and Ward’s linkage method for clustering mixed data (Chávez Esponda et al., 2010; Hendrickson, 2014; Martin, 2016). Once the distance information was obtained, agglomerative clustering was performed to obtain a tree-based representation of the objects (dendrogram). In agglomerative clustering, the cluster algorithm treats each single observation as a single cluster and pairs it with the next most similar cluster. This step is repeated until all clusters have been merged into one big cluster containing all objects, obtaining a dendrogram as a result (Kassambara, 2017).

After applying the clustering algorithm and computing the cluster distance information, the optimum number of clusters (k) that best represented the data was identified using the average silhouette method. This method allowed identifying the optimal number of clusters (k) which maximizes the average silhouette over a range of possible values for k (Kassambara, 2017). Once the number of optimum clusters (k) was selected, PAM partitioning (partitioning around the medoids) was performed, which partitions the data based on the most central point of each cluster (medoid). In k-medoids clustering, each cluster is represented by one of the data points in the cluster, the medoids, which correspond to the most centrally located point in each cluster. The medoids are therefore a representative example of the members of that cluster. As opposed to other types of clustering, the k-medoids statistical method is less sensitive to noise and outliers and it is compatible with the Gower distance metric (Martin, 2016). After PAM partitioning was performed, descriptive statistics of each cluster were obtained for interpretation and comparison of clusters. The above-mentioned steps for cluster analysis were performed first for all 267 CSV households. The same cluster analysis was then performed only for the 223 CSA farms which had adopted at least one CSA practice, in order to identify if implementation of at least one CSA practice has some effect on capital-based livelihood.

As a last step in the analysis, cluster validation was performed in order to evaluate the goodness of fit of the clustering algorithm results The silhouette width method was used as an internal cluster validation method, which measures how well an observation is clustered by estimating the average distance between clusters (Martin, 2016; Kassambara, 2017). The silhouette plot displays a measure of how close each point in one cluster is to points in the neighboring clusters. Observations with a large silhouette width (almost 1) can be considered well clustered, a small silhouette width means the observation lies within two clusters, and a negative silhouette width means the observation is probably placed in the wrong cluster (Kassambara, 2017).

The vulnerability analysis in this study was based on the sustainable livelihoods approach (SLA). SLA can be used to assess vulnerability at the local level by analyzing the status of the five livelihood capitals — financial, social, natural, human, and physical. These five capitals are the basis for assessing an individual’s or household’s ability to cope with risks. They also serve as a starting point identifying necessary interventions (Chambers and Conway, 1991). Zhang and Fang (2020) inspired our method to construct a Livelihood Vulnerability Index (LVI). They constructed an LVI integrating the 3 dimensions of IPCC’s definition of vulnerability—exposure, sensitivity, and adaptive capacity (IPCC, 2014), using the five livelihood capitals from the Department of International Development framework (DFID, 1999). To construct the LVI, each livelihood capital is represented by a number of indicators which belong to one of the three major vulnerability dimensions of livelihood capital: exposure, sensitivity and adaptive capacity. In this study, we first constructed an index of vulnerability for each livelihood capital, which represents the contribution of that specific capital to vulnerability, and then aggregated the scores of the five capital vulnerabilities into one LVI per group of households. This approach indicates how each capital contributes to livelihood vulnerability for each group of farming households.

The five capital vulnerability indices and final LVI per group were constructed using 25 sub-component indicators. We selected these indicators based on published studies that measure livelihood vulnerability (Eakin and Bojórquez-Tapia, 2008; Hahn et al., 2009; Pandey and Jha, 2012; Xu et al., 2020; Zhang et al., 2020; Zhang and Fang, 2020). They were also adapted from the variables used in the cluster analyses. To construct the LVI, a balanced weighted average approach was employed, which is based on the assumption that each sub-component equally contributes to the overall index (Hahn et al., 2009).

The whole process to construct the LVI includes four major steps. First, the sub-component indicators were normalized to a value between 0 and 1 by min–max normalization using Eq. 1:

where S_g is the original investigative indicator sub-component for households in the selected group g, and S_min and S_max are the minimum and maximum value for each sub-component, respectively.Second, after each sub-component was standardized, Eq. 2 was employed to calculate an index of each dimension of vulnerability (exposure, sensitivity and adaptive capacity) for each type of livelihood capital:

Where Ce, Cs and Ca correspond to the value of each vulnerability dimension (exposure, sensitivity or adaptive capacity) for each Livelihood Capital of each group g. Index sg represents the subcomponents, and n is the number of sub-components in each of the vulnerability dimensions of Livelihood Capital.

Third, after each dimension of vulnerability was calculated for each type of livelihood capital, Eq. 3 was employed to calculate the index of each Livelihood Capital’s contribution to vulnerability (CVI-IPCC):CapitalVulnerabilityIndex−IPCCg=Exposure−Adaptivecapacity×Sensitivity

Which can be expressed as:

where CVI−IPCCg is the value of the Livelihood Capital’s overall contribution to vulnerability for group g after incorporating each vulnerability dimension (Ce, Ca, Cs) of the IPCC’s vulnerability framework.

Fourth, the LVI at the group-level, which represents the value of the five livelihood capitals for a given group, was calculated using Eq. 4:

where w₁, w₂, w₃, w₄, and w₅ are the weights of the five livelihood capitals’ contribution to vulnerability—human capital vulnerability (HCV), natural capital vulnerability (NCV), physical capital vulnerability (PCV), financial capital vulnerability (FCV) and social capital vulnerability (SCV). In this study, the LVI ranged from −0.2 (low vulnerability) to 0.4 (high vulnerability). A detailed assessment process of the calculation of the LVI is illustrated in Supplementary material.

The LVI provides a comparable method for assessing vulnerability within groups of CSV farmers. To observe differences in vulnerability associated to CSA adoption, the LVI was constructed for the CSA farmers and the non-CSA farmers. This vulnerability analysis allowed us to observe the different contributions to vulnerability of each livelihood capital and overall LVI of each group of farming households.

The clustering exercise results shed light on three distinct clusters: one comprised of non-CSA farmers and the other two consisting of CSA farmers (refer to Supplementary material for detailed information on the clustering exercise). Within the non-CSA farmer cluster, individuals exhibited the lowest adoption rate of CSA practices, standing at 9.6% for all variables associated with each capital.

In the case of CSA farmers, we performed a two-cluster solution. This clustering approach revealed that the adoption of CSA practices played a pivotal role in delineating different groups within farming households. Subsequent rounds of clustering analysis underscored the primary divergence between the two groups of CSA farmers, which focused primarily on social capital. Cluster 1 showed superior performance in financial and human capital indicators compared to Cluster 2 (see Table 4). Conversely, the non-CSA farmer group displayed markedly lower values across almost all indicators when compared to the two groups of CSA farmers (see Table 5).

The multiple correspondence analysis (MCA) allowed to understand the variables with the largest influence in clustering the data, by underscoring the variables with the largest contributions to the combined first and second dimensions of the analysis (see Supplementary material). These included two variables from human capital (whether household increased food availability, PXW12, and food diversity, PXW13), Four variables from social capital related to acquisition of weather and climatic information and connection with social networks (CSS02, CSD02, CSS1 and CSD1), one variable from physical capital (PRAX_TOTAL: percentage of CSA practices adopted), and one variable from natural capital (SOIL, percentage of soil improving practices adopted). Based on this clustering result the following cluster descriptions were made:

The percentage of women responses was similar in the three clusters (35.4% in cluster 1, 28.4% in cluster 2, and 27.3% in the non-CSA cluster) and comparable to the percentage of women responses from the total household sample (30.3%). This result also indicated that, with the household data used in the analysis, gender as a variable for clustering did not show a significant influence in the partitioning of the data. It is possible that gender could have higher influence in the clustering than the one observed in this analysis, had the percentage of female respondents from the clustering sample been higher. Despite the relevance for this type of studies, the lack of systematic recording of one male and one female survey per household limited the possibility to analyse gender implications. The latter in addition to the significant amount of missing and incorrect data, which forced to discard many household records from the initial datasets.

Figures 1, 2 present the constructed capital vulnerability scores for each group of households, and the disaggregated indicator values per capital and per dimension of vulnerability. The final aggregated Livelihood Vulnerability Index (LVI) score for each group of households reveals that CSA Cluster 1 exhibits the lowest livelihood vulnerability (−0.048), followed by CSA Cluster 2 (0.063), with the non-CSA farmers group displaying the highest vulnerability (0.168).

In a comprehensive overview, CSA Cluster 1 demonstrates the lowest vulnerability across all types of capital, except for natural capital, where CSA Cluster 2 shows a slightly lower vulnerability (0,007 compared to −0.012, respectively). On the contrary, the non-CSA farmers group registers the highest vulnerability across all types of capital, despite scoring similarly to CSA Cluster 2 in social, financial, and physical capital (Figure 1). Regarding the impact of each livelihood capital on vulnerability, the three groups of farming households demonstrated comparable performance in terms of financial and physical capital. However, even within these two capital categories, CSA Cluster 1 outperformed the other two clusters, as illustrated in Figure 1.

The largest differences between the three groups of farming households were observed for social, human, and natural capital. For human and natural, the two CSA clusters showed similar and much lower vulnerability scores (human: 0.000 and 0.043; natural: 0.007 and 0.012, respectively) than the group of non-CSA farmers (human: 0.359; natural: 0.171). In terms of social capital, CSA cluster 1 showed by far the lowest vulnerability among groups (−0.152) whereas CSA cluster 2 and non-CSA farmers showed a similar and much higher vulnerability score in social capital (0.264 and 0.304, respectively) (Figures 1, 2).

The differences in human capital scores between CSA farmers and non-CSA farmers were mainly due to the different performance in the food availability and food diversity indicators, for which non-CSA farmers showed high vulnerability (Figures 1, 2). This result suggests that CSA interventions might be particularly effective in improving food diversification in the farming household for either on-farm consumption or market sales. Furthermore, a slightly higher percentage of CSA farmers achieved a higher education index. However, educational differences between farming households might have already been present prior to CSA interventions. A wider range of indicators for human capital could be included to characterize this type of capital. Then, we might observe larger differences between groups of farming households and where the differences come from. Other possible indicators to be included in the quantification of human capital could be the dependency ratio, the percentage of illiteracy in the household, the level of education, agricultural training of the members of the household, and the overall health of household members (Xu et al., 2020; Zhang et al., 2020; Zhang and Fang, 2020).

Non-CSA farmers did not adopt any soil-improving practices, which resulted in the main difference in natural capital scores among groups (Figure 2). However, is it possible that the group of non-CSA farmers has adopted or already implemented other soil conservation practices than the soil practices included in the CSA package, and therefore have a higher adaptive capacity in natural capital than what our indicator reflects. Another observed difference in natural capital between groups was the average area destined to productive agricultural activities per household, which was higher in CSA farmers compared to non-CSA farmers (Figure 2). This result suggests that agricultural land differences between CSA adopters and non-adopters might have already been present. However, since the indicator used corresponded to the area destined to productive agricultural activities and not the total farm size, it is also possible that CSA might have had a positive effect in supporting farmers in more efficiently utilizing their existing land for agricultural production. Nevertheless, this livelihood capital only included three indicators, one for each dimension of vulnerability, which greatly limits the ability to fully measure natural capital-related differences between groups. More indicators should be included to cover all aspects of this capital, including indicators about access to fresh water sources, quality of the owned land, exposure to climatic events, biodiversity degree, air quality, erosion protection, etc. (Xu et al., 2020; Zhang and Fang, 2020; Zhang et al., 2020).

The largest difference in physical capital between groups was the percentage of adopted irrigation practices, which was 0% for non-CSA farmers and 60 and 50% for the two groups of CSA farmers, respectively (Figure 2). This result suggests that CSA might have positively encouraged farmers to adopt improved water management practices. However, as for the adoption of soil-improving practices, it is possible that non-CSA farmers already have some irrigation practices or infrastructure in place that this indicator would not measure. The second largest difference in physical capital was the percentage of households which do not own the land dedicated to agricultural activities, which was higher in non-CSA farmers (50%) and relatively similar between the two groups of CSA farmers (33 and 29%). This result suggests that differences in land property rights might have already been present prior to CSA implementation. It is possible that households who are owners might have been at an advantage in their ability to take up CSA practices. The average number of household-owned goods was similar between the two groups of CSA farmers and higher than the non-CSA group (Figure 2). All three groups showed a similar percentage of households with cultivated land (above 90% in all cases), but a higher percentage of households in CSA cluster 1 had livestock compared to CSA cluster 2 and non-CSA farmers, suggesting the CSA focus on farm diversification improved resilience. The two CSA groups also showed a higher percentage of households with cultivated tree species. Although six indicators were used to measure physical capital, our set of indicators did not cover other relevant aspects for this type of capital, such as possession of agricultural machinery and other durable goods, the level of infrastructure, transport, access to markets, etc. To better quantify physical capital, further indicators which measure these aspects should also be included (DFID, 1999; Xu et al., 2020; Zhang and Fang, 2020; Zhang et al., 2020).

Adaptive capacity indicators described the largest differences in financial capital between groups. The two groups of CSA farmers were in general more able to borrow money or were able to generate extra sources of income compared to the group of non-CSA farmers. This result again highlights the potential positive effect of CSA implementation on food and income diversification. The percentage of households which were unable to save money in the past year was more similar among all groups but still higher for the CSA farmers than non-CSA farmers. A lower number of households did not receive income from agriculture in CSA cluster 1 (9.8%) compared to the other two groups (17.7 and 15.9%). Other possible indicators related to non-farm income sources are if the household experienced monetary loss or if the household is below the poverty line. The analysis could benefit from indicators on insurance, subsidies, total annual income or other types of financial support (DFID, 1999; Xu et al., 2020; Zhang and Fang, 2020; Zhang et al., 2020).

Social capital was the livelihood capital where the largest differences between groups of farming households were observed. Almost no household from CSA cluster 1 had difficulty accessing weather information whether daily or seasonal, whereas 70 and 50% of households in CSA cluster 2 and non-CSA, respectively, could not access this information (Figure 2). Furthermore, a high percentage of households in CSA cluster 1 acquired weather and climate information from social networks, which was not the case for the other two groups. This result suggests that CSA interventions might have been most successful in farming communities with already-established social networks. When these networks were not as strong, the positive CSA contribution in social capital might have been minimal. The involvement of households in social groups was much higher for CSA farmers (74 and 64%, respectively) than in non-CSA farmers (10%). Other possible indicators that could be included to better characterize social capital would be percentage of households with members working for the government, households receiving help from the government, number of community organizations joined, or diversity of agricultural product sales channels (DFID, 1999; Xu et al., 2020; Zhang and Fang, 2020; Zhang et al., 2020).