The study area is in the province of Padre Abad, one of the four provinces of Ucayali, the second largest region of Peru, situated in the Amazon (Figure 1). We focused on three of the seven districts within the province: Alexander Von Humbolt, Neshuya and Irazola. These three districts were selected due to the role of cocoa as a main economic activity and the rapid increase of its production over the last three decades across such territories as consequence of numerous ‘alternative development' programs that have had focused on the area to counter rising coca (Erythroxylum coca) plantations since the eighties (Zegarra, 2004). The climatic conditions in the province are typical of the humid tropics. Such ecological conditions jointly with the construction and successive improvements of the Federico Basadre highway connecting the regional capital Pucallpa, to the country capital Lima, have promoted internal immigration since the 1940s, especially during the 1981–1993 period, when the population in the province increased by 125% (INEI, 2009a). This rapid demographic transition precipitated considerable loss and degradation of forest in the province as a direct consequence of activities carried out by colonist farmers including livestock production, logging, planting of cash crops such as oil palm, cocoa, corn, pineapple and others (Porro et al., 2015).

To date, the zone is still an active forest frontier in the Peruvian Amazon where local claims for greater State support to agricultural activities and greater market access contrast with ecological concerns from some stakeholders due to the massive loss of forest across the province (MINAM, 2017). Forest loss has often been enabled by government-led interventions rolled out to achieve development goals. Despite these threats, natural ecosystems in the province still present high levels of biodiversity as the region overlaps the Andes-Amazon transition zone, identified as one of the most important biodiversity hotspots worldwide (Myers et al., 2000). This has led to establishment of legal protection over some areas to conserve biodiversity values present in these, as well as to promote sustainable land practices such as cocoa and oil palm agroforestry (Gobierno Regional de Ucayali, 2022). Despite such efforts, the study area features high levels of fragmentation across its landscape due to the loss of 90,437 ha – equivalent to ~44% of primary forests during the 2001–2023 period (MINAM, 2017)—which has severely affected the connectivity across existing habitats in the area, and harmed ultimately the region's biodiversity values. Mosaics comprised of scattered forest patches, cocoa, oil palm and pastures dominate the landscape across the study area (Clavo Peralta et al., 2022). It is estimated that around 2,000 cocoa farmers manage 9,205 ha of cocoa—corresponding to the 36% of Ucayali's cocoa harvesting area—are settled across the three districts comprising the study area. The cacao productivity reaches around 900 kg ha⁻¹ on average (Charry et al., 2023).¹ In addition, the most recent regional assessment of the cocoa sector indicates that around 5% of the cocoa production in Ucayali is produced under agroforestry systems, and that 95% of cocoa farmers in the region had obtained returns lower than PEN 12,000 (equivalent to US$ ~3 428) in 2019 (Gobierno Regional de Ucayali, 2019).

We conducted a discrete choice experiment (DCE) to model the preferences of cocoa farmers across the study area regarding attributes for a proposed biodiversity-oriented PES scheme. We apply two complementary approaches. The first approach is based on the assumption that the alternative chosen by each farmer represents the one generating the highest utility among those available and therefore has the highest probability of being selected (McFadden, 1980). The probability of choosing the alternative associated to the highest utility might be derived from an indirect utility function estimated through a Random Utility Model: U_ki = V_ki+ε_ki (Louviere et al., 2002), where U represents the kth farmer utility function for option i, which has a deterministic term, V; and an independent and identically distributed (iid) random term ε. This framework allows us to estimate their marginal willingness to accept (WTA) enrollment into the proposed PES scheme in relation to the status quo situation, this is:

Where WTA represents farmers' maximum willingness to accept the changes required to get enrolled into the PES scheme and move from the status quo situation. The term β_k reflects a proposed change in attributes of the PES scheme and α_k the corresponding amount that compensates for the proposed changes. In case the random term is proven to not be iid, we implement econometric specifications that do not require this condition (e.g. mixed logit) to calculate these parameters.

A second complementary approach derived from the first one helped us to explore potential variations in farmers' preferences due to heterogeneous characteristics. This was done by performing a latent class model, which provides insights regarding differences in preferences by identifying and characterizing classes within the sample (Boxall and Adamowicz, 2002). Different farmer classes are detected based on patterns of farmers' responses to choice scenarios after which farmers are allocated to classes using a probabilistic approach. This process implies that farmers with similar preferences are allocated to the same class. Preferences are assumed to be homogeneous within classes and may vary across classes (Doll et al., 2023). We extended our analyses by assessing if farmers' environmental attitudes and demographic characteristics might be a source of heterogeneity among classes. Considering N as the available option number in the set, and t as the choice situation, the probability of individual i selecting alternative j conditioned on the latent class c and scale class membership k might be represented as:

where β_cvX_itkt is the deterministic portion of the utility functions, with X capturing the attributes, including the compensation, and β_cv capturing class-specific marginal utilities for each class variable.

The survey instrument was designed using information collected through (1) semi-structured interviews with key regional actors involved in the forestry and cacao sectors, (2) two focus groups with eligible participants of the proposed initiative, and (3) secondary sources including publications and internal reports from organizations working in the study area. In the discrete choice experiment (DCE), we consulted about cocoa growers' preferences regarding three alternatives (i.e., Option A, B and C) per round. The following four attributes were consulted to participants of the DCE: (1) canopy cover (%) and number of tree strata required for areas enrolled in the PES, (2) monitoring modalities, (3) options for a connectivity bonus for boosting enrollment of neighboring farms in the initiative, and (4) the cash compensation amount for participating in the PES scheme to be paid annually after the fifth year of enrollment and conditional on successful verification of ecological commitments (Table 1). The selection of these attributes and corresponding levels is based on insights that we collected from the focus groups and semi structured interviews as well as from previous relevant studies [e.g. (Canessa et al. 2023), (Tadesse et al. 2021)]. In all choice cards presented to cocoa farmers, Option A represented the status quo reflecting the current situation regarding each of assessed attributes built using information from interviews conducted to local experts. Options B and C represented attribute combinations of a possible initiative aiming to enhance the capacity of existing cocoa plots to provide biodiversity related ecosystem services which would be compensated accordingly.

Based on consultations with local experts, we define one hectare (ha) to be the minimum size of a cocoa plot that could be enrolled in the proposed PES scheme. For the attribute related to canopy cover values required in PES enrolled cocoa plots, three levels (low, medium and high) were defined. The first level here refers to the status quo situation characterized by low canopy cover ( ≤ 30%), less than 25 shade trees ha⁻¹ and equal or less one stratum of trees in the cocoa plots. Meanwhile, the high level we proposed in the experiment is characterized by canopy cover above 50% or more than 100 individual trees (including both native and introduced species with the predominance of natives) within the cocoa plot, and at least three strata. The intermediate level for this attribute was a cocoa plot with 30–50% canopy cover, at least two strata of shade trees and between 25–50 shade trees ha⁻¹. These canopy cover values are based on existing scientific evidence that confirm the positive correlation between canopy levels and diverse biodiversity parameters in perennial crops (Bennett et al., 2022; Manson et al., 2024; Steffan-Dewenter et al., 2007; Wynter et al., 2025), and are also coherent with existing biodiversity standards currently being applied to cocoa agroforestry systems [e.g. Smithsonian (2023)]. To facilitate the understanding of this attribute during interviews, we use graphical material presented in (Orozco-Aguilar et al. 2021).

The three levels of the second attribute assessed in the DCE were related to farmers' participation in monitoring biodiversity for the proposed PES scheme. The first level reflects the status quo for this attribute in which no biodiversity monitoring occurs within cocoa plots. In the proposed second level, monitoring is exclusively performed by external actors commissioned by the implementer of the initiative (e.g. private investor or certification body) using remote sensing technologies as well as some field measurements (e.g. camera traps, acoustic recorders, etc.), meanwhile, in the third level, farmers are allowed to participate in biodiversity monitoring, which would imply some training as well as periodic reporting to complement external tracking.

The third attribute assessed through the DCE relates to mechanisms for boosting the enrollment of contiguous plots into the scheme following the assumption that numerous farms under agroforestry management nearby should provide greater habitat connectivity and therefore more robust biodiversity outcomes. The lack of connectivity can heavily compromise ecological processes within the landscape, such as seed and animal dispersals as well as local species abundance which, on turn, may lead to biodiversity loss (Grande et al., 2020; Hyseni et al., 2021). For this attribute, the first level (status quo) consisted of no mechanism for this purpose existing on the ground. Based on the existing literature (Le Gloux et al., 2025; Parkhurst et al., 2002), we proposed the following two levels for this attribute in addition to the status quo: a one-time cash bonus (PEN 300 or US$ ~86 per new participant, considering an US$/PEN exchange rate of 3.5) on top of the PES compensation that would be paid to individual farmers sponsoring the enrolment of nearby participants—those managing eligible cocoa plots up to 3 km away from the sponsor's enrolled plot; after the latter has complied with the rules of the PES scheme for five consecutive years to guarantee additional outcomes from implementing the proposed initiative. The third level for this attribute consists in that newer enrolment bonuses are added at the village level and then periodically delivered to corresponding organizations comprised of participants in the same village once newly enrolled farmers have stayed in the scheme for five consecutive years and complied with their environmental commitments. This strategy is analogous to an “agglomeration bonus” (Parkhurst et al., 2002) with the variant that, for this study, we did not establish any distribution rules in advance as such decisions should depend on each village-level organization. To facilitate the analysis of farmers' preferences between both approaches, we set the amount of this agglomeration bonus to be equal to the individual sponsorship bonus.

The fourth attribute assessed corresponds to the cash compensation amount that would be received by the participant per ha after five consecutive years of enrolment and verification of compliance with environmental commitments. The status quo for this attribute consists of receiving zero compensation [PEN 0]. In addition, we offered three compensation levels that reflect an increasing bid vector that would be introduced through the initiative [PEN 500, 1,800 and 3,000] (equivalent to US$ ~143, ~514 and ~857, respectively) as defined using information collected from interviews with local experts and focus groups who were consulted about their opinion regarding widely acceptable compensation in cash for conservation efforts and potential financial losses from changing agricultural practices. Finally, to facilitate and increase farmers' understanding of attributes and levels under consultation when conducting the DCE, local surveyors were trained by the first and third authors to use colloquial words for explaining technical terms as well as use diverse visual material previously designed for this purpose (figures, photographs, etc.). Table 1 summarizes the attributes and levels used in the DCE.

We followed guidelines from (Johnston et al. 2017) and designed the survey instrument to cover information related to individual environmental attitudes, preferences for the proposed intervention, and socioeconomic aspects. We also consider recommendations to reduce hypothetical bias. For instance, we include six modules in the survey: the first three sections include questions to inquire about survey participants' environmental attitudes including those related to biodiversity and agroforestry. The fourth survey module collected information about interviewees' land management practices as well as their economic activities and income sources. In the fifth module, the choice experiment was introduced by first providing information about assessed attributes and levels and then presenting the hypothetical scenario to be considered during selection of the most preferable options. Given the large number of possible combinations for the DCE (3 × 3 × 3 × 4 = 108) we apply an efficient D-design to optimize the number of choice cards to 20 representing combinations that are proven to be relevant for the proposed initiative (Walker et al., 2018). After asking and confirming whether the instructions about each attribute and corresponding levels have been well understood by survey participants, they received the instructions that they might select among three options each out of four times during the interview. Considering 2 alternatives to the status quo for each choice round and 4 rounds, each surveyed farmer was consulted for their preferences in relation to 8 alterative scenarios in total. Following each round, they were asked about their certainty about decisions made, and after all the rounds were conducted, some consequential questions (i.e. about the policy and survey) were also asked to mitigate hypothetical bias (Ready et al., 2010). We show an example of choice cards presented to participants in Supplementary material (Figure S1).

Between September and November of 2024, 300 surveys were completed by three local surveyors using the digital version of the survey installed into the ODK application for tablets. However, we report our results based on 223 farmers' responses who were certain of their answers and considered the proposed options or survey as realistic or consequential, respectively during all the voting questions. This subsample represents 75% of the collected sample and about 9% of the total number of cocoa producers settled within the three districts comprising our study area. To identify survey participants, we contacted natural resource managers working with local projects promoting cocoa agroforestry in the study area and consulted them about existing lists of cocoa farmers across the area from which we could randomly select 100 farmers per district. Selected farmers were contacted and visited by our surveyors. Although this was the most efficient way to contact farmers in the area, we acknowledge that this would also generate some bias as a disproportionate number of sampled farmers could have already been exposed to programs and information about biodiversity and cocoa agroforestry.

This study forms part of the Project Biomonitor4CAP (Project 101081964-2022) that obtained ethical clearance by the Independent Ethical Committee of the European Commission as requirement for its approval. The activities involving this research were conducted in accordance with the local legislation and institutional requirements. Research participants were enrolled on voluntary basis, and all of them provided their prior informed consent in written form. To ensure data confidentiality, all participants' responses were anonymized and presented in an aggregated way.

Out of the 300 original sample, 223 farmers were certain of their responses and form the final sample. Eighty one percent were men and were on average 47 years old (Table 2, Column 3). Due to our sampling procedure, most participants identified themselves as farmers (91%). Most farmers in the sample attained educational levels higher than primary school (56%) and manage less than 10 ha of land (58%). Small farmers in our sample reported an average annual on-farm income of 18 thousand PEN (US$ ~5.3 thousand), while annual on-farm incomes of larger farmers amount in average 68 thousand PEN (US$ ~19.5 thousand), more than three times the average income earned by the former. Also, we report incomes calculated by aggregating self-reported annual revenues from the most relevant on-farm economic activities across the study area besides cocoa production, mainly livestock and fish farming. Additional descriptive statistics about land uses present on managed lands of surveyed farmers are shown as Supplementary material (Table S1).

Table 2 allows us to compare the characteristics of our final sample with features of farmers settled across the study area who were included in the 2024 Peruvian National Agricultural Survey (2024 ANS) to explore the presence of representativeness biases. We observe that our sample feature a similar average age than farmers included in the 2024 ANS, but male farmers are slightly overrepresented in our sample. Although farmers in both samples are similar in terms of primary and secondary education attainment, farmers with superior studies and those without formal education are slightly overrepresented in our study sample. With respect to their land size distribution, our sample mostly consists of smallholders (< 10 ha) while the 2024 ANS included a larger share of farmers possessing larger holdings (>20 ha). This might be indicative that our results are better suited for design intervention targeting vulnerable smallholders. Disaggregated farmers' information per district is presented in Supplementary Materials (Table S2).

Our survey also captured important insights about sampled farmers' attitudes and beliefs toward biodiversity, farming and agroforestry. Slightly more than a half of farmers in the sample considers that they know what the term biodiversity refers to 56%, and 87% though that it is important. Sixty three percent of the sampled farmers agreed that a good farmer protects the forest, which is a slightly higher share than the percentage of farmers who expressed that at a good farmer is productive (59%) or does not pollute—for instance, by using chemical fertilizers (59%). In terms of their attitude toward agroforestry, 93% of sampled farmers agree that it both increases their total income, and it contributes to climate change mitigation. Seventy four percent of farmers believe that agroforestry improves ecosystem health, but most farmers in the sample think that this benefit could be in detrimental to their production as only the 48% of sampled farmers agree with the statement that agroforestry contributes to increased cacao yields and only 37% believe that agroforestry might help to control pests and diseases. Forty one percent of farmers in the final sample think that agroforestry improves soil fertility.

Table 3 firstly presents the results of the MCL and MXL models where the probability of uptake for an agroforestry-based PES scheme is described according to each of the consulted attributes. Both models were estimated using the 223 farmers subsample as they provide the better fit for our data (Table S3 in Supplementary Materials) as well as the most conservative results. Results shown in Table 3 indicate that the MCL model projects, as expected, that increases in the cash amount that participants would receive for engaging and complying with the PES scheme overall rise the probability that farmers will engage. Also, establishment of high canopy cover requirement is associated with reductions in farmers' utility and hence in the possibility to enroll in the proposed initiative. The agglomeration bonus paid collectively is less preferred than the individual sponsorship bonus among the farmers surveyed presumably due to individualistic productive behavior featured by most of farmers settled in the area, and incorporating a participatory monitoring mechanism in the PES scheme rather than one reliant on third-party verification increase the probability that farmers would enroll in the PES.

When testing for the Independence of Irrelevant Alternative (IIA) assumption to the MCL model, we found that it does not hold (χ² = 0.0363) for our sample. For this reason, we estimate the mixed logit model (MXL) as it relaxes IIA. Standard deviation coefficients estimated through the MXL model (Table 3, Column 3) allow us to confirm heterogeneous farmers' preferences as most of them present statistical significance. All the MXL parameters representing consulted attribute means (Column 2 in Table 3) are statistically significant and follow the same relationships with WTA found through the MCL model. The status quo is mostly rejected, and farmers are willing to accept compensation for participating in the initiative when high canopy cover is required in PES enrolled areas. We also confirmed that farmers prefer receiving an individual rather than collective bonus as mechanism to promote the connectivity of PES enrolled areas. However, the statistically significant standard deviation coefficients estimated through the MXL model suggest that not all farmers would obtain the same utility levels from being compensated or participating in monitoring tasks within the frame of the proposed PES scheme. We consider all the attributes consulted in surveys as random as such assumption provides the best explanations for our sample (Tables S4, S5 in Supplementary Material).

Although MXL captures heterogeneity in preferences through random parameters, it assumes homogeneous error variance across all decision-makers (Train, 2009). We thus extend this framework using the Generalized Multinomial Logit (GMNL) model whose results are reported in Column 4 of Table 3, which simultaneously accounts for both preference heterogeneity and scale heterogeneity (Fiebig et al., 2010). This econometric specification yields more robust and consistent estimates as evidenced by the superior Log Likelihood (LL) and the Akaike Information Criterium (AIC) values. It also shows that our participants exhibit different levels of choice consistency (τ = 1.246, p < 0.01). However, the non-significance of the mixing parameter (γ = −0.308, p > 0.10) indicates ambiguity into signaling the source of preferences variation (Keane and Wasi, 2013).

Regarding the amount required by participants to enroll in the proposed PES scheme, the models reported in this Section do not provide statistically significant PES compensation amounts. These were estimated using the delta method (Wooldridge, 2010) and are reported in Supplementary material (Figure 2). Given the ambiguity of the variation source in the GMNL model, we complement our analysis with a Latent Class (LC) Model to identify discrete behavioral segments in the population, as well as potential marginal willingness to accept (MWTA) for different groups. We report results from performing such a model in the next Section.

Since we found strong heterogeneity in farmers' preferences for levels in the amount required to compensate their enrollment in the PES scheme, we implement a LC model to investigate which factors might shape these differences. Based on theoretical suggestions (Boxall and Adamowicz, 2002) and prior empirical analyses [e.g. (Canessa et al. 2023)], we employ the presence of certain attitudinal, demographic and socioeconomic features defined posterior the data analyses as membership variables to select the optimal number of identified classes. We again use LL and AIC as main criteria (Table S6 in Supplementary Material) to estimate the optimal class number as three as shown in Figure 2.² The first identified class (Class I) represents the largest group (55.4% of the sample). They are more likely to enroll in the proposed scheme if the PES payment increases, and they are allowed to participate in monitoring tasks. Also, they are less likely to participate if high canopy level in enrolled plots is required or if the connectivity bonus is distributed collectively (at the village level) within the proposed PES. This class rejects the status quo scenario in larger magnitude than farmers in Class II as shown by corresponding coefficients and can be consider the “economically rational” class as they increase their enrollment when the payment per ha also increases. We estimated the average amount required to compensate utility losses of Class I farmers for adopting high canopy cover and collective connectivity bonus in PEN 737 and 807 (equivalent to US$ ~ 210 and ~ 230), respectively (Figure 2).

Meanwhile, farmers in Class II (34.3% of the sample) also reject the status quo but their marginal WTA for all the assessed attributes of the proposed PES scheme are statistically not different from zero. We presume that such results could signal the elasticity of this farmer group regarding levels in assessed attributes for the proposed PES scheme as consequence of altruistic values present to some extent in this class. On the other hand, Class III (10.3% of the sub sample) presents a positive and significant marginal WTA for high canopy values in their enrolled plots (352 PEN or US$ ~100 per ha). This group can be categorized as an individualistic class. Although they also present altruistic values to some extent, their responses indicates an average yet not significant negative marginal utility for the collective bonus, which could reflect some distrust on other village members. Based on the analysis of membership variables, the presence of positive attitudes toward forest conservation and being younger than 60 years old increases the likelihood of belonging to Class III compared to Class I. These results suggest that farmers younger than 60 years old with positive attitudes toward forest conservation are important features associated with this individualistic class. Other membership variables like possessing a small size farm, being a woman or educational attainment, do not correlate with probability of belonging to class II or III. These results remain robust even after using alternative specifications of the PES compensation amount, like estimations using net present values that incorporate an overall discount rate of 3.7% as suggested to evaluate profitability of national public projects with environmental contributions (MEF, 2011)—Table S8 in Supplementary Material.