Farm-level socio-economic data was collected through comprehensive surveys by FEDEGAN from 2011 to 2019 (N = 3,644). The survey comprised 10 main modules including farm characteristics (e.g., size, type of production), management practices (e.g., use of technology, rotation), productivity (e.g., liters of milk or kgs of meat), environmental information (e.g., presence of springs or forest patches), social information about the farmer/manager (e.g., demographics of the farmer, region), financial information about the farmer/manager (e.g., income, assets) (Supplementary Figure S1). We selected the socio-economic explanatory variables for this study from these variables based on factors influencing adoption of SPS in Latin America identified through a systematic literature review (Supplementary Figures S1, S2; Supplementary Table S2; Chamorro-Vargas, 2024; Chamorro-Vargas et al., 2025, submitted manuscript). To reduce dimensionality of the data, some variables were combined based on the themes identified in the systematic literature review. For example, we summed binary survey data (yes/no responses) within themes (e.g., number of sources of information for the project, number of best management practices) (see Supplementary Table S3 for full list). We excluded variables that had many missing values for the surveyed farms (Supplementary Table S4). We obtained additional variables of interest from secondary sources, such as the gross domestic product (GDP) of the departments where the farms were located (national census data, DANE, 2023; Supplementary Table S5). We calculated some variables using the distance matrix algorithm in QGIS (QGIS Development Team, 2022), including distance of farms to the closest demonstration farm and distance to the nearest SCR neighbor. We included department size to validate if the barriers to the adoption of SPS changed with the inclusion of this variable.

Informed by the systematic literature review of factors influencing adoption of SPS in Latin America (Chamorro-Vargas, 2024; Chamorro-Vargas et al., 2025, submitted manuscript), we gathered environmental information on climate change (e.g., expected changes in temperature and precipitation in the upcoming years, environmental vulnerability), forest cover, soil conditions (e.g., soil erosion) and water variables (e.g., water demand, humidity) from open-access data provided by governmental institutions of Colombia such as the environmental information system (SIAC), Geographic Institute Agustin Codazzi (IGAC Datos Abiertos), and Institute of Hydrology, Meteorology and Environmental Studies (Geoportal IDEAM). Data were downloaded from the relevant websites (Supplementary Table S5), set to the same projection (EPSG: 3116 MAGNA-SIRGAS/Colombia Bogotá zone) and cropped to the extent of the SCR area (Figure 1). We extracted the target variables of interest (Supplementary Table S5) associated with each farm coordinate from the vector shapefiles using the function st_join in the R sf package (Pebesma and Bivand, 2023). We explored visualizations of the distributions of each variable of interest and only included variables in the models that showed some variation among sites. Finally, farm level environmental information on presence of forest or water springs collected in the surveys by FEDEGAN was extracted for the environmental data frame to complement the environmental analysis with finer scale data. The final data frame containing farm environmental information included 31 variables for a total of 2,945 farm centroids.

During the SCR project land use information was collected quantifying the area of each SPS land use type on the farm (i.e., of iSPS, dispersed trees, fodder banks, live fences, and fodder hedges). Although forest protection and natural regeneration were practices targeted under the SCR project, they were not considered in our analysis of adoption of SPS. For the majority of the farms, this data collection was done at the beginning of the project. Some farms were also re-surveyed in a second round of data collection at the end of the project; in total 3,188 farms had land use data from both the beginning and end of the project.

We calculated the SPS adoption variable as the total area on which an SPS practice was implemented for each farm. Here we aimed to understand the process of adoption of all types of SPS together rather than the adoption of each specific feature of SPS (Jara-Rojas et al., 2020), thus, adoption was measured as the sum of hectares of all types of SPS present in the farms at the end of the SCR project (SPS_end). For live fences and fodder hedges measured as linear features in meters, data were transformed to reflect the area these SPS features occupy in the farm. To calculate area, live fences were assumed to be 3.5 m wide and fodder hedges 1 m wide (Heiber Pantevez, personal observations). iSPS, dispersed trees and fodder banks were measured in hectares. We chose SPS_end as the adoption variable instead of the difference of SPS hectares from the beginning of the project to the end (SPS_change) because the goal of the study was to understand the total adoption or maintenance of SPS on farms. Further, we deemed this measure more interpretable. Nevertheless, we explored the SPS_change and found that most farms increased their SPS coverage through the project (N = 2,843), while a smaller number did not change it (N = 270) or decreased (N = 75) (Supplementary Table S6).

We merged the adoption dataset (N = 3,188 farms) with the socio-economic survey (N = 3,644 farms) and environmental (N = 2,945 farms) data with the merge function in R using the farm ID, which resulted in a sample of 2,986 unique farms with adoption, socio-economic, and environmental data. We identified and removed outliers for the SPS adoption variable using a custom outlier removal function implemented in R. This function uses an interquartile range criterion where data points outside the set bounds were considered outliers. The upper and lower bounds for outliers were defined with the equations [Q1 − multiplier * IQR and (Q3 + multiplier * IQR)], and a multiplier of 6 was used as default, where Q1 = quartile 1, Q3 = quartile 3 and IQR = interquartile range (Q1–Q3). This process removed 60 farm sites leading to a final data set of 2,926 farms for the analysis.

We used linear mixed effects models fitted by maximum likelihood using lmer function in lmer4 R package version 4.4.2 (Bates et al., 2015) to study the effect size of socio-economic (Model 1) and environmental (Model 2) variables on adoption of SPS (fixed effects Table 4). Both models included region (Figure 3; Supplementary Table S1) and call year (i.e., the year when the farm joined the project, Tables 2, 3) as random effects, since values within each group were expected to be more similar than values between groups due to similarities within regions and within calls. Covariates were also included to account for their effects on the model, such as the amount of SPS coverage in the farm in the baseline (initial) measure of the project (SPS_baseline). The response variable for the models (hectares of SPS adopted at the end of the project) were log-transformed, due to their right-skewed distribution (Speekenbrink, 2023). To avoid multicollinearity, correlation matrixes, correlation tests, and principal component analyses were done on the original selection of variables of each model using the R package factoextra (Kassambara and Mundt, 2020). For correlated variables, we selected the subset that permitted a complete set of variables based on the systematic literature review of factors influencing the adoption of SPS in Latin America, prioritizing inclusion of variables with highest response rate (Supplementary Table S4; Chamorro-Vargas, 2024; Chamorro-Vargas et al., 2025, submitted manuscript). Our final models included 18 and 11 variables respectively, which falls far below the maximum number of variables that could be estimated in theory, based on the rule of thumb of >10 observations per estimated parameter (Harrison et al., 2018). The assumptions of linear regressions were validated visually using the distribution of residuals (Harrison et al., 2018; Speekenbrink, 2023). We also conducted several additional analyses on data subsets, including analysis of (1) each region separately (Table 2), and (2) each of the specific SPS practices (Dispersed trees, Live fences, iSPS) adopted during the project, to determine if the socio-economic and environmental factors affecting adoption changed.

Linear Mixed Effects Model equation including dependent variable, fixed effects, and random effects. To make the variables comparable all the numeric variables in the dataset were scaled using the scale function in R.

Where log(Yij) represents the log-transformed SPS adoption in hectares for the ith farm in the jth region and tth call year. β₁, …, β_K are fixed-effect coefficients for the individual-level socioeconomic or environmental predictors (see Table 4 for variables used for each model). β 0 is the intercept, which is made up of a mean component γ 00 , a random intercept for each region u 0 j and a random intercept for each call year of the survey v 0 t . ∈ ijt is the residual error term for each observation.

To assess the importance of the factors influencing adoption we used a model selection algorithm with the dredge function in R package MuMIn (Bartoń, 2023) for Model 1 and Model 2. This function uses the Akaike Information Criterion (AIC) metric which balances model fit and complexity. The AIC aids in identifying the variables to include in the most parsimonious model, which are expected to be the most important ones to explain the outcome variable (Bartoń, 2023). Then, we utilized the glmnet function (Friedman et al., 2010) to run the Least Absolute Shrinkage and Selection Operator (LASSO) implementation technique, a regularization method widely employed for variable selection and regularization in high-dimensional data settings (Tibshirani, 1996). LASSO penalizes unimportant variables by shrinking their coefficients to zero and keeps the coefficients of the important variables (Tibshirani, 1996). Additionally, using, the bartMachine R package (Kapelner and Bleich, 2016), we conducted Bayesian Additive Regression Trees (BART), a Bayesian ensemble approach for modeling the unknown relationship between a vector of observed responses y (SPS adoption in this case) and a set of predictor variables, without assuming any parametric functional form for the relationship. For BART, the var_selection_by_permute function was used to identify the variables that split the branches of the decision trees more often as they were expected to be important predictors of the adoption of SPS (Bleich et al., 2014). By using three different methods to assess variable importance, and exploring similarities and differences, we increase the robustness of our analysis since each technique has different strengths and limitations.

For the socio-economic model (Model 1), PES had the largest significant positive effect on the number of hectares adopted by farmers during the project, whereas access to credit and farm area did not significantly affect adoption (Figure 4). Farm characteristics including farm machinery and GDP of the region where the farm was located had a positive significant effect on adoption. Various farm management aspects were also influential on the adoption of SPS. Farms that used more inputs like fertilizers and pesticides adopted fewer hectares of SPS than farms that used less inputs (Figure 4). In contrast, farms that already used best management practices (BMPs) such as water and forest protection, paddock rotation and division, and crop diversification, had more adoption. In contrast, vegetation diversity (i.e., the number of tree and shrub species in pasture per hectare on the farm) did not significantly affect adoption of SPS. Most of the information transfer variables had a null effect on adoption, except for distance to the closest SCR farm, which had a significant positive effect on adoption, meaning that farms adopted more hectares of SPS when they were further from other SCR farms. Farmers’ characteristics such as gender, age and education did not have a significant effect on adoption. As expected, the covariate hectares of SPS_baseline had a large positive effect on the amount of SPS at the end of the project. We found similar results when including department size in our model, except that while department size is a significant predictor (driven largely by a single, large department with significant adoption), local GDP becomes insignificant (Supplementary Figure S3).

We found differences compared to the full dataset when separately analyzing adoption in each region (Supplementary Figure S4) and for each SPS practice (Supplementary Figure S5). For instance, economic incentives such as PES were only positively significant in three of the studied regions: Magdalena, Cesar and Orinoco (Supplementary Figure S4) and for two of the SPS adoption practices: dispersed trees and live fences (Supplementary Figure S5). Farm characteristics, such as farm area, had a positive significant effect on adoption of SPS in most of the regions (Supplementary Figure S4) and for most practices (Supplementary Figure S5). While the specific variables for management, information transfer and farmers’ characteristics varied slightly, the regional and practice level models resembled the general model trend of good management practices being positive for adoption, and that most information sources and farmers characteristics were not significant.

The most important classes of variables from Model 1 were economic incentives and farm characteristics (Table 5). PES was ranked as important in all three variable selection procedures (i.e., AIC, LASSO, BART), followed by farm machinery and GDP, which appeared in two of the methods (Table 5). Farm area was another important variable in the farm characteristics category, although it only appeared in the BART result (Table 5). Management variables are important in at least one of the variable selection methods were number of BMPs and crop diversity (Table 5). Information transfer and famer characteristics variables were not ranked as important by any of the selected methods, confirming their lack of significance. As expected, the covariate SPS_baseline had the largest fixed effect, coefficient, and variable inclusion proportion. The variable selection results are in accordance with the results of the model (Figure 4).

For the environmental model (Model 2), farm characteristics such as soil conditions, forest presence, water conditions and vulnerability to climate change had a significant effect on adoption of SPS (Figure 5). For example, soil erosion was found to have a positive significant effect on adoption in Model 2. The presence of forests and springs on the farm also had a positive influence on adoption. In contrast, humidity, water demand and environmental vulnerability to climate change negatively affected the number of SPS hectares adopted by farmers during the SCR project (Figure 5). Hydric precipitation anomaly, precipitation changes by 2040, temperature change by 2070 and hydric vulnerability did not significantly affect adoption, although the trend of their effect was negative (Figure 5). Similarly to Model 1, we found that the results of this model varied across regions (Supplementary Figure S6) and practices (Supplementary Figure S7). For instance, the vulnerability variables that had a significant negative effect on adoption in the full dataset changed across regions and constrained adoption (Supplementary Figure S6). Hydric vulnerability had a negative significant effect in Cesar and Orinoco, whereas precipitation anomalies affected Boyacá & Santander, and environmental vulnerability to climate change affected Magdalena and the Coffee region (Supplementary Figure S6). The presence of forest and arid conditions had a negative significant effect on the adoption of dispersed trees, while precipitation change by 2040 had a positive significant effect on the adoption of iSPS (Supplementary Figure S7).

The most important variables from Model 2 were water conditions and vulnerability to climate change (Table 6). Water demand was the most important variable being ranked in all the methods used, followed by humidity and high precipitation anomaly, which appeared in two of the methods (Table 6). Water spring presence was another important variable, although it only appeared in the AIC methodology (Table 6). As expected, the covariate SPS_baseline had the largest fixed effect, coefficient, and variable inclusion proportion. Although the model found soil erosion and presence of forest on farm as significant predictors of adoption (Figure 5), these variables were not selected with the methodologies used to rank variable importance. Further, precipitation anomalies experienced by the farms were not significant in the model (Figure 5) but were ranked by variable selection methods (Table 6).

Studies on adoption of agriculture technology around the world show how economic incentives are an effective way of overcoming initial reluctance from farmers to adopt new practices (Liu et al., 2018; Begho et al., 2022; Priya and Singh, 2024). Furthermore, studies on scaling out SPS in Colombia have found that lack of capital for investment is one of the main barriers for adoption in the country (Calle et al., 2013; Lerner et al., 2017; Jara-Rojas et al., 2020). We found that PES was the strongest positive socioeconomic predictor for adoption of SPS, and that it was consistently ranked as an important predictor of SPS adoption. Our study thus concords with other works showing that economic incentives are an important mechanism for promoting adoption of SPS, by helping farmers overcome initial economic barriers, such as the costs associated with planting (Cerrud Santos, 2004; Casasola Coto et al., 2007; Pagiola et al., 2007; Calle et al., 2013; Lerner et al., 2017; Calle, 2020; Jara-Rojas et al., 2020). Furthermore, PES can be seen by farmers as a support for their stewardship values, encouraging them to continue their ongoing care for their land and forest (Chapman et al., 2020). In contrast, we did not find that access to credit influenced SPS adoption. A possible reason for the lack of importance of access to credit was that other economic variables such as PES, farm machinery and farm area overshadowed any effect access to credit had on adoption. Further, the variable access to credit used in this study was based on whether a farmer applied for a credit but did not indicate whether they obtained credit. There were multiple steps for credit approval, and many SCR farmers were subsequently denied credit. Thus, the variable that we had access to from the survey did not fully capture actual access to credit. Our work suggests that future SPS promotion projects can use economic incentives such as PES to increase the probability of adoption of SPS, potentially including both initial uptakes, as well as increase in amount (hectares) of adoption. Access to credit should be further investigated for its potential to increase adoption.

Previous studies on agriculture adoption have also found that lack of capital for investment affects farmers differently (Holguín et al., 2003; Cerrud Santos, 2004; López et al., 2007; Zabala et al., 2022). For instance, studies have found that wealthier farmers who own larger farms and more machinery tend to have more capital for investing in SPS and can afford to devote a larger area of their lands to SPS (Frey et al., 2012; Cárdenas Gutiérrez, 2014; Bussoni et al., 2015; Gosling et al., 2020; Babi, 2021; Apan-Salcedo et al., 2022). In accordance with this expectation, we found that farms with more machinery adopted more SPS, and that farm area had a significant effect on adoption for most regions and SPS practices. Machinery can help farmers with the management of SPS, making tasks such as preparing soil for planting, pruning trees, mowing grass and watering plants more efficient and less labor-intensive. Furthermore, we observed that farms located in departments with higher GDP had higher adoption of SPS, which was expected according to the literature on adoption of agriculture innovation which finds that wealthier regions have higher capacity for investment (Feder et al., 1985; Mwangi and Kariuki, 2015). Both farm machinery and GDP were ranked as important variables in at least two of the variable selection methods used, demonstrating their importance for the adoption of SPS. These results highlight that the implementation cost of SPS is likely an important barrier to adoption for farms with less wealth and capital for investment (Calle et al., 2013; Lerner et al., 2017; Jara-Rojas et al., 2020). Future projects and research should create initiatives and interventions that consider the heterogeneity of farmer populations in each region and develop appropriate programs to reach socially and economically vulnerable groups for which adopting SPS is more challenging.

Given the constraints climate change imposes on agricultural production, environmental factors, such as poor soil and water conditions and less predictable climates leading to higher vulnerability to climate change, were expected to have a negative effect on adoption. For instance, some authors have found that the fear of investing in trees that might not survive intense climatic events keeps farmers from adopting silvopasture (Calle et al., 2009; Calle, 2020; Lee et al., 2020; Apan-Salcedo et al., 2022). We found that most environmental factors did constrain adoption of SPS, except for presence of forest and water springs on farms and high levels of soil erosion, each of which promoted adoption. The presence of forest and water springs on the farm can be related to the environmental values of farmers that are in close contact with nature which could motivate them to adopt more hectares of SPS. For instance, the protection of water sources, biodiversity and forests is frequently mentioned as a motivation to adopt SPS by farmers in literature of Latin America (Calle, 2020; Rizo-Chavarría et al., 2022; Timoteo et al., 2023; Chamorro-Vargas, 2024; Chamorro-Vargas et al., 2025, submitted manuscript), and may also help farmers adapt to climate extremes of drought or flood. Additionally, the expectation of soil recovery is another strong motivator for adoption of SPS in Latin America (Oliva et al., 2018; Olival et al., 2022; Rizo-Chavarría et al., 2022). Thus, farmers experiencing high levels of soil erosion could be motivated to adopt SPS.

Farmer characteristics, management practices and preferences are some of the most studied factors in the literature on adoption of SPS in Latin America (Chamorro-Vargas, 2024; Chamorro-Vargas et al., 2025, submitted manuscript). In general, no conclusive trends exist regarding the influence of age, gender or education on the adoption of sustainable practices (López et al., 2007; Garbach et al., 2012; Cancino et al., 2016; Torres, 2016; Rasch et al., 2021). However, farmers’ preferences and management practices were found to be important to the adoption of SPS (Torres, 2016; Sibelet et al., 2017; Stefano et al., 2020). The results of this analysis are in accordance with previous studies that found a null effect of age, gender and education on adoption of SPS (Patiño et al., 2012; Apan-Salcedo et al., 2022). Furthermore, we found that management variables had significant effects on adoption, where farmers who use more sustainable management practices (less external inputs, more crop diversity, and more use of BMPs) adopted more SPS than farmers that use less environmentally friendly practices, with crop selection and use of BMPs ranked as important variables. Management decisions taken by farmers can be influenced by farmers’ preferences, stewardship and environmental values which lead them to have strong personal motivations to improve their farms and pass them to the next generations demonstrating long-term motivations (Calle, 2008, 2020; Smith et al., 2022). Furthermore, SPS_baseline had a large positive significant effect on the amount of SPS adopted, which means that farms with SPS have great potential for further adoption. Promotion projects can target farmers who already have strong environmental values and that already have adopted SPS regardless of their age, education, or sex to achieve a greater area of adoption of SPS.

Communication among farmers, sources of information and short distances between farms were expected to have a positive influence on adoption because communication facilitates information transfer about implementation of new systems and can increase the likelihood of neighbors to adopt (Anfinnsen et al., 2009; Frey et al., 2012; Apan-Salcedo et al., 2022). However, in this study, most of the variables related to information transfer did not have a significant effect on adoption, except for distance to closest SCR farm, which had the opposite effect than expected, where farms that were more separated from each other adopted more SPS area. A possible reason for this result is that larger farms are expected to be farther apart from each other, since the distance between farms was measured from a central coordinate of each farm. While economic factors had greater explanatory power on SPS adoption than social factors in our study, it is nonetheless well documented in the literature that social networks are important for creating synergies and communication, impacting collaboration, promoting social acceptance, and acting as a main driver of adoption (Liu et al., 2018; Tapasco et al., 2019; Blesh et al., 2023).

Furthermore, studies on factors influencing adoption of agroforestry practices consistently found technical assistance (TA) to be a crucial factor promoting adoption (Cerrud Santos, 2004; Nascimento et al., 2014; Lerner et al., 2017; Tapasco et al., 2019; Babi, 2021; Smith et al., 2022). TA helps farmers overcome one of the main barriers for adoption, which is lack of knowledge (Calle et al., 2013). In this study we found a null effect of additional TA obtained by farmers on adoption, which may be because TA was provided by the SCR project to all of the farmers. While the TA provided by the project most likely had a significant positive effect on adoption, its influence could not be assessed because it was provided to all farms in the sample. In spite of our finding, based on other literature, future studies and projects can use TA and extension programs to promote SPS adoption, given that lack of knowledge and awareness of SPS is one of the main current barriers to adoption (Cancino et al., 2016; Castillo Ruíz, 2019; Tapasco et al., 2019; Calle, 2020; Tarbox et al., 2020), and that SPS implementation is a complex process requiring substantial learning and experimentation by each practitioner (personal observations). Furthermore, future studies could explore in detail how the amount and quality of TA and farmer participation in TA workshops influences SPS adoption.