The current literature provides several tools designed to analyze farmers' preferences for maize crop such as participative research ( Bellon & Risopoulos, 2001 ; Ferro et al. , 2013 ), conjoint analysis ( Makokha et al. , 2007 ; Hirpa et al. , 2012 ), and the use of descriptive analyses ( Sibiya et al. , 2013 ).

In the early 1980s, the DCE was introduced as a technique for modeling consumer choices ( Louviere, 2001 ). The DCE relies on the Lancaster's theory of value ( Lancaster, 1966 ), which proposes that the utility of a product is decomposed into separable utilities for their characteristics. It is also based on the random utility theory ( Thurstone, 1927 ), which proposes that subjects choose among alternatives according to a utility function with two main components: a systematic (observable) component and a random error term (non-observable). The DCE has become the most sought after tool for analyzing individual behavior and choice. While four main choice modeling alternatives (choice experiments, contingent ranking, contingent rating, and paired comparisons) are available, only the choice experiment provides outcomes consistent with standard welfare economics ( Hanley et al. , 2001 ).

The DCE was first used in communication and transport studies ( Louviere, 1981 ); however, its use gradually spread to other areas such as market research ( Bastell & Louviere, 1991 ), environmental evaluation ( Hanley et al. , 1998 ), identifying attributes of products influencing consumers ( Lusk et al. , 2003 ), and agricultural multifunctionality ( Kallas & Gómez-Limón, 2007 ). Windle & Rolfe (2005) used this methodology to analyze alternatives for agricultural diversification in Australia. This method has also found use in organic agriculture ( Meas et al. , 2015 ), food traceability ( Wu et al. , 2015 ), maintenance programs for plants and animals ( Roessler et al. , 2008 ; Asrat et al. , 2010 ; Birol et al. , 2012 ), and provision of ecosystem services ( Villanueva et al. , 2017 ). However, empirical applications of DCE regarding farming innovations are few.

Furthermore, Birol & Villalba (2006) noted that a successful application of DCE in developing countries such as Mexico depends on a careful selection of election sets and an effective compilation of field data. In Mexico, choice experiments have been used in natural reserves ( Tudela et al. , 2009 ), trait selection of pig breeds ( Scarpa et al. , 2003 ), and assessment of transgenic corn crops in the states of Jalisco, Michoacán, and Oaxaca ( Birol & Villalba, 2006 ).

The choice modelling approach aims to identify the consumers' indirect utility function associated with the product attributes by examining the trade-offs they consider when making choices at a retail outlet. According to the random utility theory, the utility of an individual, n , choosing an alternative, j(Uⁿ_j,j = 1,...,J , is the sum of both components: Vⁿ_j , which is a function of the characteristics of the alternative (Xⁿ_j), individual characteristics ( S _n ), as well as another random component, εⁿ_j. Furthermore, the individual, n , will choose the alternative, j , if it provides a utility that is superior over any other alternative, i , available in the choice set.

To predict a subject's preference for attribute k , we need to define the "probability of choice" that an individual n chooses the alternative i rather than the alternative j (for any i and j within choice set T ). McFadden (1974) developed an econometric model that formalized respondents' decision-making process. This model is often referred to as the multinomial logit (MNL) model, which is considered the base model for DCE. In this model, the utility to person n from choosing alternative j on choice scenario t is given by:

where x_njt is a vector of observed attributes of alternative j , β is a vector of mean attribute utilities, and ε_njt is the "idiosyncratic" error term that follows independent and identically distributed (i.i.d.) type 1 extreme value distribution with scale parameter .

The probability (P_j│X_nt that an individual n will choose alternative j among other alternatives in an array of choice set T is formulated as follows:

where X _nt is the vector of attributes of all alternatives j=1, ...,J. In the case of estimating a MNL, the scale parameter σ_n is normalized to 1 for identification.

However, this model imposes homogeneity in preferences for the observed attributes. Thus, only average attributes' utilities can be estimated. Therefore, the MIXL (Mixed Logit Model) has been introduced to investigate the unobserved heterogeneity. However, it has been argued that much of the preference heterogeneity in MIXL can be are better captured by the scale term and thus known as "scale heterogeneity" ( Louviere & Mayer, 2007 ; Louviere et al. , 2008 ). According to Balogh et al. (2016) , the scale heterogeneity might be interpreted as the variation of randomness in the decision-making process over respondents, i.e ., the variance of the error term (and hence the degree of certainty) may differ across individual decision-makers. This is especially relevant for the stated preference data, where respondents interpret choice situations differently and pay varying levels of attention to the task presented ( Train & Weeks, 2005 ).

Among the various modeling approaches that include scale heterogeneity specification, Fiebig et al. (2010) proposed the GMNL model. According to this model, the utility of an individual, n , for selecting alternative, j , in a choice set, t , is given by

where γ is a mixing parameter between 0 and 1, whose value represents the level of independence or interaction between the scale term σ_n and the heterogeneity around the attributes' estimates ( n _n ). Fiebig et al. (2010) proposed that σ_n follows a log-normal distribution with mean equal to 1 and standard deviation τ. The GMNL estimates the τ term that captures scale heterogeneity across respondents. Further details about GMNL specification and estimation can be found in Fiebig et al. (2010) .

The usual procedure for calculating the WTP is estimating the distribution of utility coefficients and then deriving the distribution of WTP, which is the ratio of coefficients. However, Scarpa et al. (2008) described a method to estimate the distribution of WTP directly, which fits the data better, reduces the incidence of exceedingly large WTP values, and provides the analyst with greater control over the distribution of WTP. In the present study, we used the GMNL model in the WTP-space. In this case, the GMNL is reparametrized ( Greene & Hensher, 2010 ) by separating the variable price, ρ, and its coefficient, ßρ, n . By standardizing the price coefficient to 1, the WTP can be directly estimated. In this case, the mixing parameter (γ) turns to be a fixed parameter.

Finally, the DCE approach is similar to the PCE method used in this study ( Greene, 2012 ). The only difference is that the choice variable used was proportional data rather than individual choice data. That is, the choice variable consists of a set of sample proportions with values ranging from 0 to 1. This variable should sum to 1 over the alternatives in the choice set. Observed proportions may be equal to 1 or 0 for some individuals if they answered 100% for some alternatives within the choice set.

Regarding the analysis of preference heterogeneity, different techniques can be used. Within the choice experiment approach, the socioeconomic variables are typically interacted with the attributes. The LC model is one of the popular approaches for analyzing observed heterogeneity. Besides the relevance of socioeconomic variables in describing preferences, this model also provides a way to obtain information regarding the different segments of the market. To illustrate, the model begins by contrasting the "segmentability" of the population studied. The LC determines the probability of an individual belonging to a certain class and the probability of choosing one alternative conditional on the preferences within each class. Further details regarding this model are available in Greene & Hensher (2003) . In this study, we used the LC to analyze farmers' preferences. The "best" number of classes to be extracted was based on the comparison of the Bayesian information indicator (BIC), McFadden pseudo R ² , and plausibility of the results.

Data was collected from face to face survey with a sample of 200 farmers that was carried out in January and March of 2015; the sample was stratified by seed variety (creole and improved) and postal districts. Also, the interviews were made in a zone of potential corn production in the state of Chiapas: the towns of Villaflores, Chiapa de Corzo, Villacorzo, and La Concordia. In order to determine the sample size, information were used regarding the farmers who were registered in the Programa de Apoyos Directos al Campo (PROCAMPO), a program which is intended to promote and finance agriculture in the counties mentioned above. Notably, farmers enrolled in this program represent 98% of total corn farmers ( SIAP, 2016 ). The sample size was calculated as finite populations with 95% as significance level NS and an error of 6.87%. Table 1 represents the survey technical sheet.

Following Kallas et al. (2010) , the questionnaire was organized in two sections: the first included questions about the characteristics of the farmers (gender, education, age, experience), farm structure (location, farm size, soil type), farm management (input use and crop diversification), exogenous factors (output and input prices, market size, subsidies, information access, transition costs); the second part included the different choice sets to carry out the choice experiment. Analyses of the econometric models were performed with NLOGIT 5.0 software.

The application of the PCE can be summarized into the following steps: First, the characterization of the decision problem was predefined in terms of changes to the existing state, status quo, and base reference point. In this study, we placed values on the possible changes in the preferences of attributes when selecting the maize seeds and WTP for each seed type. The status quo in our case was, therefore, defined by the supply of improved and creole seeds.

Next, for the definition of attributes and their corresponding levels, we followed different studies and sources of information. To begin with, we analyzed the current farmer preferences when selecting crop seeds. The attributes that the farmers took into account when selecting a new variety are the corn ear shape ( Ferro et al. , 2008 ), number of grains per row ( Ferro et al. , 2008 , 2013 ), corn ear filling arrangement ( Ferro et al. , 2013 ), grain color ( Soleri & Cleveland, 2001 ; Benz et al. , 2007 ), ear size ( Ferro et al. , 2013 ; Sibiya et al. , 2013 ), ear height ( Ferro et al. , 2008 ), ear weight ( Ferro et al. , 2008 ; Sibiya et al. , 2013 ), resistance to disease ( Ferro et al. , 2008 ), ear diameter ( Ferro et al. , 2008 ), ear tightness ( Ferro et al. , 2008 ), stem thickness ( Ferro et al. , 2008 ), number of rows per cob ( Sibiya et al. , 2013 ; Ferro et al. , 2013 ), color of straw ( Ferro et al. , 2008 ), plant height ( Ferro et al. , 2008 , 2013 ), number of corn ears ( Ferro et al. , 2008 ; Sibiya et al. , 2013 ), cob diameter ( Herrera et al. , 2002 ), early maturity ( Sibiya et al. , 2013 ), yield ( Birol et al. , 2012 ; Ferro et al. , 2013 ; Sibiya et al. , 2013 ), grain size ( Bellon & Risopoulos, 2001 ; Sibiya et al. , 2013 ), flavor ( Sibiya et al. , 2013 ), tolerance to drought ( Bellon et al. , 2006 ; Sibiya et al. , 2013 ), tolerance to excessive rain ( Bellon et al. , 2006 ; Sibiya et al. , 2013 ), resistance to putrefaction of the corn ear ( Bellon et al. , 2006 ; Sibiya et al. , 2013 ), duration (cycle of growth) ( Bellon et al. , 2006 ; Sibiya et al. , 2013 ), plague resistance ( Bellon et al. , 2006 ; Sibiya et al. , 2013 ), resistance to storage plagues ( Bellon et al. , 2006 ; Sibiya et al. , 2013 ), and dough yield ( Bellon et al. , 2006 ). The product price is another important extrinsic attribute affecting the purchase decision ( Lockshin et al. , 2006 ).

The second step was to conduct a discussion group with researchers from the Instituto Nacional de Investigaciones Forestales, Agrícolas y Pecuarias (INIFAP, Mexico) ¹ to reduce the primary information obtained. Also, this group evaluated and verified the suitability of attributes. Subsequently, a pilot questionnaire was applied to test the validity of the attributes and to determine the level of the price vector.

Regarding the cost attributes and levels, the price vector was based on the average prices for a bag of 20 kg of seed, provided by the INIFAP and the Secretaría de Agricultura, Ganadería, Desarrollo Rural, Pesca y Alimentación ² . The price levels in the choice sets were selected to cover the central 90% of the observed values. However, the price was correlated with the type of seeds used. We used a labeled choice design to solve this problem, where each alternative choice was defined by the type of seed used. In Table 2, the main attributes and their levels are presented.

For the experimental design, forced and labeled choice sets representing the different varieties of maize seed were used. An efficient block design was used ( ChoiceMetrics, 2014 ), leading to 27 choice sets classified into three groups. Respondents were asked to set their preferences for the different alternatives. No evidence was found in the pilot or main survey preferring to reject all corn types in a choice set.

Before beginning the survey, the choice experiment was explained orally and in writing. Respondents were asked to set the percentage of preference for the different varieties of seed for maize cultivation this year; thus, the dependent variable in this study was a proportion of two mutually exclusive alternatives in each choice set. An example of a choice set is shown in Fig. S1 [suppl].

Table 3 represents a summary of the major sociodemographic characteristics of the respondents. The proportion of each stratum is similar across the population of farmers in the geographical area surveyed.

Results of the GMNL model in the WTP-space are shown in Table 4. The model showed a goodness-of-fit with an acceptable value of McFadden pseudo R ² that is equal to 0.169, similar to other studies that analyzed farmers' preferences through choice experiments ( Birol & Villalba, 2006 ; Kallas & Gómez-Limón, 2007 ; Ortega et al. , 2016 ). The log likelihood ratio was also highly significant at 99%. Results showed that the estimated coefficients of the majority of the levels of attributes are statistically significant. This result confirms that most of the attributes and levels considered in the model are significant and essential in predicting farmers' preferences.

The estimated parameters directly provide information about the WTP. These estimates should be multiplied by 100 as the price variable was divided by 100 during the estimation. Farmers showed a positive WTP for crop yield. The same trend occurred among the attributes of resistance to disease and ear length. Corn growers were willing to pay $15.80 for a 20 kg bag of improved corn seed to gain one centimeter in corn ear length. These results are in agreement with the findings of Kassie et al. (2017) . Maize varieties with medium- and large-sized cobs are preferred as the cob size has a lot to do with grain yield and marketability.

Furthermore, farmers were willing to pay $2.90 more to gain 1% of resistance to disease in the maize crop, and $39.89 more per bag to increase the crop yield by 1 ton. However, unexpectedly, the respondents did not give importance to the attribute "height of the corn,", despite the problems that this characteristic can cause. Nevertheless, in conformity with the finding of Hellin & Bellon (2007) , farmers gave more importance to corn stems for making fences and leaves as forage. Therefore, any corn type can be grown for forage, but the ones with a higher yield of biomass are the tall varieties. On the other hand, improved varieties have a little bud sport and usually produce less forage per unit of area ( Estrada et al. , 2015 ); thus, farmers are willing to accept a tall plant as long as it has a high yield and wind-resistant stalk.

In addition, the estimated coefficient of the alternative specific constant (ASC) of the improved corn was not significant. This result shows that the attributes and levels that were not included in describing corn may not be relevant.

Our results are consistent with other studies that considered high yields, resistance to diseases, and lower price as main drivers for selecting improved seeds ( Asrat et al. , 2010 ; Sibiya et al. , 2013 ; Waldman et al. , 2017 ).

According to Ajambo et al. (2010) , corn farmers in Uganda preferred drought-resilient varieties, with a short growth cycle and higher resistance to pests and diseases. Those farmers were willing to pay Ush 200–5,000/kg for a variety with such characteristics (1 US$ = 2,200 Ugandan USh). Comparatively, Kassie et al. (2014) pointed out that Zimbabwean farmers were willing to pay 1.75 times more to ensure tolerance to drought and a harvest of one more ton of crop. It was also found that producers were willing to pay 8.3 times the value to get a change in size from a small corn ear to a bigger one. Furthermore, the seed cost was also an important factor describing preferences. Our results are similar to those of other studies where the cost of seed is a main determining factor when choosing a variety ( Kyeyune & Turner, 2016 ). Kassie et al. (2017) indicated that the trait-based promotion and marketing of varieties constitutes a suitable strategy for the adoption of improved corn seeds.

Finally, regarding the scale factor, the estimate was high and significant, which confirmed a high level of unobserved heterogeneity and uncertainty in selecting the varieties. The results of our study show that the farmers demonstrate a high level of product uncertainty and randomness when choosing corn seed ( Fleming et al. , 2016 ).

A LC model was used to analyze farmers' observed heterogeneity. This model allowed us to classify corn growers into three types according to their preferences. The optimal number of segments, the BIC, the pseudo R ² , and probability of the result of each segment were computed ( Hu et al. , 2004 ). Therefore, the LC model with three classes was selected as the best fit. Out of 200 farmers surveyed, we found that 60.5% are innovators, 29.4% are transition farmers, and 10% are conservative (Table 5).

The first latent class was innovators who gave much importance to seed yield, resistance to diseases, and price. This segment is the most price-sensitive. The second latent class, transition farmers, considered yield as the most important attribute, followed by a lower preference for intensive seed type. Our results were similar to those of Ortega et al. (2016) who reported that Malawian farmers have a strong positive preference for maize grain yield. The third latent class, conservative farmers, considered improved seeds as unimportant; instead, they preferred creole seeds and gave importance to a large corn ear and seed price.

For classes 1 and 2, the ASCs were positive, but the ASC of class 3 were negative, which could have been due to the residual utility associated with the non-observed attributes. Farmers belonging to class 3 exhibited a negative utility for the improved seed. Farmers belonging to either class 1 or class 2 had a strong preference for improved seeds if these met the preferred attributes.

Finally, the price had negative effects; that is, the lower the price, the higher is the utility for farmers, and so a normal demand was consistent. It is important to stress how important fertilizers are when using improved seeds, because fertilizers are necessary to obtain better yields. The amount of fertilizer used is considerably higher when using improved seeds than when growing creole varieties ( Bernard et al. , 2010 ). In this respect, Gecho & Punjabi (2011) pointed out that the price of fertilizer lowers the probability of the adoption of improved corn. Furthermore, Salgado & Miranda (2010) stressed that the increase in corn productivity in Mexico in the coming years will be subject to the price of fertilizers.

Knowing the types of farmers who belong to each segment can help in the establishment of well-defined agricultural policies and local intervention strategies. To do so, we first described each segment using the sociodemographic characteristics. These characteristics included the farmer's age, number of generations in agriculture, number of generations in corn farming, year responsible for managing the exploitation, and year when corn farming began. Besides these sociodemographic variables, we also collected data relating to land management such as seed being used, corn sales, total surface, yield, total sales, distance from home to the exploitation field, and soil quality. In our study, soil quality was determined using a 11-point scale, where 0 suggested that the farmer considers the soil to be of bad quality, and 10 suggested that the farmer considers the soil to be of excellent quality.

Attitudes, opinions, and perceptions toward risk also play an important role in determining the adoption of seed varieties ( Howley et al. , 2015 ). Thus, in our profiling analysis, we also included the perception towards improved seed and risk attitude. Risk attitudes and opinions toward improved seeds were assessed via two principal component analyses (PCA) following the previous studies ( Asrat et al. , 2010 ; Birol et al. , 2012 ; Li et al. , 2012 ; Valdivia et al. , 2015 ). Table 6 shows the profiles of different segments.

According to our analyses, innovators were 56 years on average. They started cultivating corn in 1980 and showed acceptance of improved seeds. These farmers mainly cultivated improved seeds with a higher yield per hectare and achieved higher sales. These farmers own more land, consistent with Kalinda et al. (2014) finding that improved corn seed use is directly related to the size of land owned. However, these farmers are risk acceptors, as they had more resources to mitigate the effects of risks when adopting new technologies. Transaction costs per surface unit were lower than that for farmers owning small areas, consistent with the findings of Paredes & Martin (2007) study.

In contrast, transition farmers were aged 55 on average and were the fourth generation to grow corn. They gave less importance to soil quality, compared to other classes. They were, on average, risk takers and cultivated improved and creole seeds, depending on accessibility. They tried to use improved seeds on an experimental scale in their farms.

Finally, conservative farmers were 67 years on average and had more experience in crop management (they have been growing corn since 1971). This group of farmers used 89.9% of corn production for sales and the remainder for their own consumption. Most of them used creole seeds with lower yield, implemented smaller crops, and traveled a longer distance from their homes to their fields. These farmers were risk-averse, wherein their family members represented the main source of farming information. Thus, access to information can reduce uncertainty about the possible results of using new technology, as also noted by Honra et al. (2007) . For this reason, it is important that research, extension, and agricultural education work together to allow farmers to understand and appreciate the characteristics of new varieties ( Rivera & Romero, 2003 ).

Our results were similar to those of Villanueva et al. (2017) , who compared the characteristics of three groups of olive farmers in Andalusia, Spain (protesters, very high takers, and participants).