2.1. Choice Experiment
Choice experiment (CE) has been widely used to research consumer behaviors, preferences, and willingness to pay (WTP) for different goods. The method has been applied to various topics, such as studying the effect of additional product information [
16] and predicting market performance for new products [
7]. CE presents participants with options that have various levels of attributes and ask participants to choose among the alternatives. Attribute levels vary across the alternatives and are designed in the way that ensures there is always a trade-off between alternatives. With different combinations of attributes in each choice set, CEs replicate a consumer’s rational decision-making process and are useful in estimating consumers’ utility function.
One possible drawback for hypothetical CE is that participants do not pay real money when making decisions, and the choices they make are therefore not real. However, the validity of CEs in consumer studies has been proved. Carlsson and Martinsson [
17] suggested that CE responses are statistically indistinguishable across hypothetical and non-hypothetical (real purchasing) treatments. Lusk and Schroeder [
18] demonstrated that the biases associated with estimated marginal WTP using hypothetical CE are reduced when CE questions are framed in a way that is similar to actual purchasing settings (non-hypothetical settings). They also found that although total WTP (relative to opt-out) was upwardly biased in hypothetical choice experiments, the marginal WTP (the difference in WTP for two products) was consistent. Yue and Tong [
6] found the difference between hypothetical and non-hypothetical choice experiment estimation of WTP is 7.5–9.0%, which outperformed most contingent valuation studies examined by List and Gallet [
19].
We conducted two versions of CEs in this study. The first CE focuses on consumers’ WTP for general sustainable farming attributes and farmers’ engagement, and the second CE focuses on consumers’ preference for different sustainable agriculture program features. In both versions of CEs, participants were asked to choose between two options of canned sweet corn (15.5oz/3.5 servings) with varying prices. We chose canned sweet corn due to its popularity and its reduced perceived heterogeneity from respondents. Also, although this study focused on a particular product, the majority of the sustainable program features are general.
In the first version of the CE, respondents were asked to choose between canned sweet corn with different production/processing methods, certification methods, food miles, and prices. We used three types of production/processing methods (conventional, organic, and sustainable), four types of certifications, four levels of farm miles, and three levels of prices (The average price of canned sweet corn, according to the US Department of Agriculture (USDA), is about
$0.48 per cup in 2016, and is about
$0.52 after adjusting to 2019 USD. Thus, the price for a 15.5 oz canned sweet corn is about
$1. According to USDA, organic corn prices are generally two to three times higher than conventional corn prices. The prices were chosen to cover the price range of conventional and organic sweet corns.). Respondents were provided with the definition of sustainable agriculture “sustainable agriculture is an integrated system of plant and animal production practices having a site-specific application that will last over the long term (1) enhance environmental quality and the natural resource base upon which the agricultural economy depends, (2) make the most efficient use of nonrenewable resources and on-farm resources and integrate, where appropriate, natural biological cycles and controls, (3) sustain the economic viability of farm operations, and (4) enhance the quality of life for farmers and society as a whole.”
Table 1 summarizes the levels of the attributes used in the first CE. We include a column that specifies the coding for each variable. Price and food miles are treated as continuous variables, while production methods and certifications are treated as dummy variables, with conventional and no certification as the base level.
In the second version of CEs, respondents were asked to choose between canned sweet corn with five farm program attributes. These attributes covered five major aspects of sustainability programs, i.e., farmers’ engagement, the role of science, marketing promotions, sustainability measurements, and communication on sustainability practices. These five aspects of sustainable agriculture programs are selected based on the sustainable agriculture programs listed as partners of the USDA National Institute of Food and Agriculture (Available at
https://nifa.usda.gov/program/sustainable-agriculture-program). Promoting farmers’ engagement is listed as a program feature of the Regional Sustainable Development Partnerships in Minnesota. Communicating and funding sustainable-agriculture-related science is mentioned in the National Institute of Food and Agriculture Annual Report of 2016. Marketing promotion is a common focus of the program marketing guide, such as the farmers’ market guide of North Central Region Sustainable Agriculture Research & Education (SARE) (An example of a farmers market guide can be found at
https://www.northcentralsare.org/Resources-and-Learning/Books/The-New-Farmers-Market.). Lastly, several sustainable agriculture programs have developed very specific sustainability measurements and guidelines for sustainability practices (An example of a sustainability guideline can be found at
https://silt.org/sustainability-guidelines-and-requirements/). We expect that consumers would demand concrete measures of sustainability and good communication of sustainable practices.
To capture consumer preferences, we designed a choice experiment with various sustainable practices to evaluate if some program practices would induce a WTP premium or not. The attribute levels are presented in
Table 2. Similar to in
Table 1, we also provided the coding for each variable, where the price is treated as a continuous variable, and program features are treated as dummy variables with a certain feature as the base level. The details of the attribute levels are described as follows.
First, regarding farmers’ engagement, the lowest level of farmers’ engagement is that managers dominate the certification process, i.e., managers tell farmers what is required to participate. We included two higher levels of farmers’ engagement, i.e., farmers participate in learning what is required to meet consumer demands and farmers advise program managers on program requirements and activities, and the latter one has the highest level of farmer engagement.
Second, in terms of the role of science, we used the case where the program plays a passive role, i.e., farmers must seek out scientific information related to sustainable agriculture on their own, as the base level. We then include two more levels for the role of science, i.e., the program provides scientific information to farmers, or the program funds science. Both levels were program practices introduced in the National Institute of Food and Agriculture 2016 Annual Report (2016).
Third, regarding marketing promotion, we used the typical practice of sustainable agriculture programs, i.e., helping farmers produce crops more sustainably (without a marketing focus), as the base level. We are interested in what consumers think about a sustainable program that does marketing promotion. Thus, we included two alternative program practices, i.e., the program helps farmers to create new market opportunities, or the program helps farmers and processors reach more consumers.
Fourth, we expected having measurements of sustainable practices to be an essential attribute, as studies demonstrated that the concrete information on sustainable agriculture has a significant impact on consumer preferences [
9,
10]. Thus, we designed variations in sustainable practice measurement. We set the base level to be that farmers declare that they are sustainable with no on-farm measures required. We then included two more levels, i.e., farmers in the program must demonstrate the use of sustainable practices, or measures of on-farm practices and consumer buying decisions are used to measure sustainability. The last level not only considers farm practice but also marketing performance as sustainability measurement.
Lastly, in addition to communicating program practices to consumers, we expected that consumers may care about how programs communicate with food processors and grocery chains because such communication promotes sustainable program transparency. We thus used the current farm practice, i.e., sustainability scorecards, as the base level, and the included program provides popular materials, or program provides facts on farmer use of sustainable practices, as alternative practices.
To mimic actual shopping, we included an opt-out option for both of the CEs. The two versions of experiments were generated using optimal D-efficiency fractional factorial design with two blocks. Each participant randomly answered one of the two blocks of questions. Each block consisted of 6 or 8 questions for the first and the second version of the CEs, respectively. We presented an example of the first and the second version of CE in
Table 3 and
Table 4, respectively. Besides the choice scenarios, we also asked questions about participants’ socio-demographic backgrounds and economic status. The CEs were conducted online, and participants were randomly selected across the United States recruited by Qualtrics™, a professional survey company. A pretest of the survey was conducted to validate the experiment design. Only consumers who were primary grocery shoppers were allowed to take the surveys.
2.2. Econometrics Model
A mixed logit model was employed to analyze the CE data. A mixed logit model allows consumers’ taste parameter to vary by some distribution, and it does not need the independence of irrelevant alternatives (IIA) assumption [
20,
21].
We assumed a linear utility function as Equation (1):
In Equation (1), individual
selects alternative
j with the preferred attributes and price combination among a set of
alternatives (
. The individual needs to make choices for
choice scenarios. Additionally,
is a vector of observed attributes, including a price level, and
is the random coefficient vector following some density function
where
is a vector of the parameters that define the distribution. In this study, we assume the density function
is multivariate normal, and thus
includes the mean vector and a variance matrix of a multivariate normal distribution. To reduce the number of parameters, we also assumed the variance matrix as being diagonal.
is the random term assumed to be independently and identically distributed following type I extreme value distribution.
denotes individual
choosing alternative
in choice scenario
. The probability of individual
choosing alternative
in choice scenario
given
is
The likelihood function could then be defined as:
where
is the density function of a normal distribution with parameter
. To simulate the integral in Equation (3), we applied 1000 Halton draws of
from the distribution
. The parameters were estimated by using Maximum Likelihood. The mixed logit models were estimated using the R package “mlogit” [
22].
The WTP for attribute
k was then defined as
where
is the marginal utility of attribute
k and
is the marginal disutility from price. Since both
and
were estimated, the standard error of WTP was then estimated using the Delta Method.
Furthermore, the main focus of this study was to understand consumer preferences for sustainable agriculture programs. We employed a latent class logit model (see Boxall and Adamowicz [
23] and Greene and Hensher [
24]) to identify market segments based on consumers’ preference heterogeneities and demographics. The latent class analysis assumes that consumers can be segmented into a few classes, and the preferences are heterogeneous across different classes, while members of each class have homogeneous preferences. The purpose of the latent class logit model is to identify market segments in terms of consumer preferences for sustainable agriculture programs.
Suppose individual
belongs to class
with the probability
In Equation (5),
denotes vectors of parameters with
set to 0, and
denotes the demographics of individual
. Since individuals within the same group are assumed to have a homogeneous taste parameter, the distribution of taste parameter
for individuals in class
can be specified as
. We can then write individual
’s contribution to the likelihood function as
The parameters were estimated by using maximum likelihood. The latent class logit model was estimated with R package “gmnl” [
24].