2.1 Research context
Peru is one of ten megadiverse countries and a centre of origin for crops important to the livelihoods of the poor, many of which are also of global importance. It possesses 184 species and hundreds of varieties of domesticated native plants, of which many species/varieties of these crops are considered “severely threatened” (FAO, 2015). There are over 5700 accessions of quinoa (Chenopodium quinoa Willd) conserved in seven genebanks, that have been characterised into 24 races (Tapia and Fries, 2007; Tapia et al., 2014), constituting thousands of varieties.
Many of these are at risk of disappearing (Kost, 2016) in large part as the national and international market is concentrated around only 15–20 mostly white varieties out of an approximate total of 3000 (Rabines/MINAGRI, pers. Com, 2014; Rojas et al., 2009). The resulting genetic erosion threatens Peru’s food and nutritional security, the sustainability of its high-altitude production systems and its ability to adapt to future climate change, along with emerging pests and diseases.
Furthermore, quinoa plays an important role in many Andean cultural traditions (Rojas et al., 2009) and its high profile in Peru in general makes it an ideal crop around which to explore its many non-market public good ecosystem service values and the general public’s willingness to support its in situ on farm conservation. Estimating the potential magnitude of such support and devising mechanisms to capture such values is critical, given that poverty rates in the arid Andean rural highlands can reach over 50% (INEI, 2015, p.63).
2.2 Choice experiment design
In a CE, respondents are presented with a series of choice tasks, known as choice sets, each containing a finite number of alternatives which describe an hypothetical environmental good or policy outcome in question (Hanley et al., 2001). CEs have been used extensively to evaluate farmer participation in schemes providing ecosystem services (e.g. biodiversity conservation: Sardaro et al., 2016; carbon sequestration: Aslam et al., 2017) or to gauge their preferences for crop traits improving livelihoods (e.g. Kassie et al., 2017; Maligalig et al., 2021); as well as to determine consumer/general public willingness to pay for ecosystem goods and services (Zander et al., 2013; and Martin-Collado et al., 2014, Blare et al., 2017; Müller et al., 2020; Drucker and Ramirez, 2020).
The alternatives presented in a CE vary with regard to the levels associated with each of the attributes and respondents are usually asked to choose their most preferred alternatives. By making this choice, respondents trade-off the attributes and the associated costs that come with the chosen alternative. A key component of the experiment is the definition of attributes used in the choice experimental design (Johnston et al., 2017). The attributes and levels for this study drew on the approach used by Zander et al. (2013) and Martin-Collado et al. (2014), and were adapted to the Peruvian crop genetic resource context in consultation with Peruvian genetic resources and agricultural experts. Each attribute represents a component of the TEV so that the sum of the separate attribute values may be used as a proxy for the TEV of the public good ecosystem service associated with the maintenance of quinoa diversity in farmers’ fields. The four attributes included Andean landscape conservation (includes ecological processes and aesthetics), insuring against the risk of agricultural production loss in the context of broader food security issues, quinoa diversity conservation and the maintenance of traditional knowledge and cultural practices – the latter including aspects of food culture (see Table 1).
As a monetary value, which is required for the calculation of welfare estimates, we selected a one-off donation (in Peruvian Soles) to a conservation programme for the crop in question. The use of one-off payment vehicles described as donations are common when evaluating environmental goods and services through respondents’ stated preferences (e.g. Veríssimo et al., 2009; Kragt and Bennett, 2011). Although one-off payments are criticised for not being incentive compatible (Johnston et al., 2017), we opted against the use of a non-voluntary tax contribution vehicle as many respondents may fall outside of the tax net. Nor did we select a repetitive payments vehicle as we did not want to make assumptions about how long payments are needed to successfully conserve crop varieties, which could potentially require support in perpetuity. The one-off payment vehicle also helps to simplify respondent understanding of the total cost of the CE alternatives.
[Table 1 here]
A generic design was used, and each choice set consisted of three alternatives from which respondents were asked to select their most preferred. One of the alternatives was always described as status-quo (SQ), while two others represented different scenarios under a quinoa crop diversity conservation programme. The SQ alternative did not involve a personal cost for respondents and can be interpreted as leaving things to business-as-usual and a consequent continuing erosion of quinoa diversity. The other two scenarios involved a one-off contribution towards a conservation programme and would result in benefits associated with an increase in such diversity (or at least avoiding any further decline). Given the number of attributes and their levels (Table 1), there would have been too many possible combinations (3^3*2^1*7^1 = 378) to use all of them in the survey and hence a choice experiment was designed which only included a fraction of these combinations. The use of qualitative levels for two of the attributes (Conservation of Andean Landscape - Improve, Stable, Decrease; and Risk of Agricultural Production Loss - High, Medium, Low), as have been used in other studies (Zander et al., 2013; Martin-Collado et al., 2014; Drucker and Ramirez, 2020) was necessary due to the challenges of articulating potential impacts in quantitative terms with regard to such multidimensional concepts as landscapes and food security. The design was pre-tested before the main survey started.
An important issue in experimental design is with regard to the identification of efficient designs capable of generating statistically significant attribute combinations associated with a given sample size (Rose and Bliemer, 2008). We generated a Bayesian efficient design (see Sándor and Wedel, 2001; Ferrini and Scarpa, 2007) of 24 choice sets which were blocked into three blocks using the software STATA. Each respondent was assigned one block of eight choice sets each (see Figure S1 in Supplementary Information for an example of a choice set). The design was based on prior parameter estimates that we assumed after expert consultation and literature review (Zander et al., 2013). Using prior parameter estimates leads to more reliable parameter estimates for a given sample size, even if the information on the parameters is scant and the priors mis-specified (Bliemer et al., 2009). While we did not know the exact values of the priors, we were quite certain about the expected signs.
2.3 Sampling and data collection
With a view to exploring how public willingness to support genetic resources conservation may vary between segments of rural and urban populations as they become more geographically distant from the genetic resource in question, population samples were selected around the important Andean quinoa producing regions of Puno and Cusco. These included the regional capital cities themselves, whose populations are respectively 135,300 and 437,500 (INEI, 2017), as well as the surrounding rural areas where incomes might be expected to be even more constrained compared to within the cities themselves. Surveys were also realised in the national capital, Lima (population 9.17m [INEI, 2017], which is distant from these quinoa producing areas but with higher average incomes.
With a view to limiting enumerators' risk due to the challenges of visiting households in Peru, a “second-best” convenience sampling method was used that involved enumerators randomly recruiting participants in central or communal areas, such as town squares, bus stations and markets. Although convenience sampling can result in the risk of selection bias (Moore, 2001) and unbalanced samples, given the experimental design and randomised treatment used here no major issues arising from demographic imbalances were anticipated. As can be seen in Table 1, there is in fact significant overlap between the sample and the actual demographics.
Sample size calculation used a cluster sampling approach (Walker and Adam, 2011) considering District population, an expected WTP contributor’s rate of 0.4 for Lima and 0.3 for the regional cities, a sample precision level between 0.1 (Lima and Puno) and 0.15 (Cusco) in a normal distribution z with p-value equals to 0.95. To take into consideration population heterogeneity, we considered a population heterogeneity of 0.15 in Metropolitan Lima and 0.1 in the regional capital cities, assuming a more heterogenous population in Lima. Finally, optimal cluster size was measured based on heterogeneity of population and cost of data collection. As a result, the minimum sample size was determined to be 471 (192 in Lima, 195 in Puno and 84 in Cusco). Four-hundred and ninety-one adult Peruvian resident respondents were interviewed between July and September 2017 in Cusco (91), Puno (200) and Lima (200). The interviews were administered in Spanish (and occasionally in local languages Quechua and Aymara) by three groups of trained enumerators in their respective locations. Subjects were not compensated for their participation, eliminating any selection bias related to financial incentives. Only adults were interviewed, and consent was established before each interview.
We used a three-part structured questionnaire. First of all, respondents were presented with questions related to their familiarity with and use of different varieties of quinoa. Secondly, they were presented with the CE tasks. In the third section, we asked for basic demographic information (gender, age, occupation, income, education, household composition, socio-economic status and wealth) and about the degree to which respondents make donations to good causes in general.
Information was provided with regard to agrobiodiversity in general. Prior to being presented with the choice sets, respondents were also reminded that achieving good conservation outcomes has a cost, that quinoa varieties are not the only crop that may require conservation funding, that there may be other good causes to support and that their household budgets need to cover other expenses too. This so-called cheap talk script helps minimise hypothetical bias that could lead respondents to overstate their willingness-to-pay (Ladenburg and Olsen, 2014). Having provided instructions on how to read the choice sets and make selections, eight choice sets were then individually presented.
2.5. Information framing
Framing is an effective way to increase awareness and potential WTP, as well as also increasing the scope sensitivity of welfare measures (Czajkowski and Hanley, 2009). This is because the value of environmental good or service not only depends on their physical characteristics, but also on the context within which they are located. In CE this refers to how the goods and services are described to respondents, in addition to their attributes. By providing different information to different sample treatment groups, respondents can be primed by the introduction of a stimuli before making their choices. This can trigger an emotional response, establish context, or change a subjects’ frame of reference (Weingarten et al., 2006).
Numerous case studies have shown framing to increase WTP in specific contexts for both direct and non-direct use products. Banerji et al. (2016), for example, found that nutritional information significantly increased WTP for vitamin-fortified millet in India. Bergstrom et al. (1990) found that framing increased WTP for American wetlands when respondents were reminded how different program attributes related to desirable consumption services. By contrast, Fox et al. (2002) found that Chinese consumers were willing to pay less for pork products when information about harmless irradiation was presented. These findings suggest that the effects of information framing can move WTP in both directions, depending on the person’s perception of the information included.
We used two different framing scenarios, one about the national identity (NI) significance of quinoa and one about food security (FS). The NI framing text contained a series of historical facts which detail quinoa’s native history to Peru and attempts by Spanish colonizers to eradicate the crop upon their arrival in the 16th century (Figure S2 in Supplementary Information). We hypothesised that this stimulus involving cultural nationalism will increase the appreciation of native crops, and hence respondents’ WTP. The FS framing utilised a series of questions regarding personal food security, under the hypothesis that heightened sensitivity to potential food shocks may increase the valuation of biodiversity and hence WTP for its conservation, given its role as an informal insurance mechanism.
The sample was split into three treatments with two groups of respondents being randomly presented with additional information either about the NI or the FS. A control group did not receive either of these additional information texts. All three treatment groups received basic information regarding what agrobiodiversity is, why it is important and current status/threats.
2.6 Data analysis
Choice experiments are based on random utility theory (Luce, 1959; McFadden, 1974) and the characteristics theory of value (Lancaster, 1966). One commonly applied method is the random parameter logit (RPL) model which was also used here to analyse the choice data. RPL models are able to account for panel-data, such as those obtained in this study with each respondent answering eight choice sets, allowing unobserved preference heterogeneity across individuals to be considered (see e.g. Hensher and Greene  for detailed model specifications). For attributes with three levels (see Table 1) the reference levels were the ones of the SQ alternative. Dummy variables for the other two levels were created and included in the models, so to model the preference for the change from the SQ.
Interaction terms were included between the SQ alternative and parameters that were assumed to have a significant impact on whether respondents preferred to donate to one of the two conservation program alternatives or the SQ. Such interaction terms related to location (Cusco, Lima, Puno), respondents’ age, gender, income and level of education, as well as the framing group they were in.
Results are presented for one baseline RPL model without interactions and one model in which we included the interaction terms between the SQ alternative and location. Both models were estimated using 2000 Halton draws.
Welfare estimates from the RPL model results were calculated using simulations. The simulated distributions were obtained by dividing draws from the distributions of the attribute coefficients by draws from the distributions of the coefficient of the monetary attribute. 10000 Halton draws were used in these calculations.
Stated preference methods such as choice experiments are often criticised for their insufficient sensitivity to scope regarding variations in the proposed scales of the environmental good or services to be improved (Czajkowski and Hanley, 2009). Here we followed the approach by Tavárez and Elbakidze (2019) and tested for scope effects to validate our results. We estimated a logistic regression model with the dependent variable indicating whether or not an alternative was chosen by a respondent (binary 0/1), the cost (one-off donation) and dichotomous variables indicating the number of improvements made through the conservation program relative to the status quo as the explanatory variables. With this we tested whether or not respondents were willing to pay more for more improvement as would have been expected under neoclassical economic behaviour.