We used an artefactual non-hypothetical experiment conducted in a realistic setting, specifically a mock/real brick-and-mortar supermarket, to increase the external validity of the study.
2.1. Study Design
First, we decided to use an artefactual experiment [
21] to ensure that the recruited participants were representative food purchasers and had experience with the concerned good [
22]. In addition, to ensure that the respondents had experience with the good, the target population consisted of participants who were responsible for the purchase of food products in their household and who consumed breakfast biscuits.
Second, we decided to conduct a choice experiment instead of other valuation methods because of its ability to value multiple attributes simultaneously, its consistency with the random utility theory, and the similarity of the choice task asked of the participants to their real purchase decisions [
23]. In addition, we designed a non-hypothetical experiment instead of using a hypothetical choice experiment to avoid hypothetical bias. Several papers have analyzed the hypothetical bias in choice experiments and compared results from both hypothetical and non-hypothetical versions ([
24,
25,
26,
27], among others). They have all provided strong evidence of hypothetical bias, suggesting the use of non-hypothetical experiments. In addition, Chang et al. (2009) [
24] found that non-hypothetical choices are a better approximation of true preferences than hypothetical ones, based on a comparison not only between hypothetical and non-hypothetical choice experiments but also with actual market shares. The interpretation of this finding is that the willingness to pay (WTP) values of non-hypothetical choice experiments (CEs) can be assumed to be the true values corresponding to actual payments in the marketplace [
24].
Third, as the proposed research has an empirical orientation to provide stakeholders (private and public) with information on consumers’ valuation of different nutritional claims, the external validity of the experiment is vital to allow for the generalization of the results. To increase the external validity of our experiment by increasing its ecological validity, the choice experiment was carried out in a close-to-real setting. In other words, we used a setting as similar to a real supermarket as possible. In particular, we used a mock/real brick-and-mortar supermarket. This mock/real supermarket is located in a logistic facility in the town where the experiment took place. This logistic space is available for companies to demonstrate how their product and service technology can help to create innovative solutions and improve productivity and competitiveness in the field of logistics. The logistic demonstration center is divided into several modules; Smart Store, Smart Point of Sale, Supply Chain Module, and Intelligent Transport Module. In this research, we used the Smart Point of Sale, which consists of a Smart Point of Sale Terminal and Smart Shelving for automatic inventory control to undertake the experiments in a close-to-real environment (
Figure A1 in the
Appendix A).
We designed a non-hypothetical choice experiment, introducing real economic incentives and real products. The participants received €10 because at the end of the experiment one choice set was randomly selected as binding and the respondents were required to purchase the food product chosen in the binding situation at the corresponding price.
We selected a box of half a kilo of breakfast biscuits with three different attributes: price, fiber, and fat content claims. The price levels were set, based on the market prices at the time of the experiment, at 0.5 €/box, 1.5 €/box, 2.5 €/box, and 3.5 €/box. The other two attributes had two options; the product carried the claim ‘high in fiber’ (FIBER) or the claim ‘reduced saturated fat’ (FAT) or did not carry a claim. The attributes selected and their levels are summarized in
Table 1.
The choice set design was generated following the Street and Burgess (2007) [
28] approach. For the main effects, three attributes were chosen with four, two, and two levels, respectively, as well as two options; we obtained eight pairs, and this design was 96.66% efficient compared with the optimal design. Thus, each respondent was asked to make four choices because we randomly split the choice sets into two blocks. Each choice set included three alternatives; two designed alternatives consisting of different products and a non-buy option.
2.3. Implementation Procedure
The experiment was conducted by the research team as follows. On arrival, the consumers received information on the nature of the experiment and signed informed consent for participation. An ID number was assigned to each respondent to guarantee anonymity. The monitor provided a general overview of the working session and informed the participants that, at the end of the experiment, they would receive €10 to purchase a box of biscuits (the one that they chose in the binding choice set) at the corresponding price. The monitor insisted that it was in their best interest to choose only the product that they were really interested in purchasing because this product could be the selected type of biscuit in the binding choice set. In addition, the participants received all this information in clear written instructions together with the information on the biscuits and the attributes presented in the different choice tasks. Then, the respondents were asked to choose four times between two boxes of biscuits or the non-buy option in front of the supermarket shelves with the real biscuit boxes (
Figure A1 in the
Appendix A). Afterwards, they went to the cashier and another monitor asked them to select randomly one card out of four cards numbered from 1 to 4 (choice sets) to determine the binding choice set (see
Figure A1 in the
Appendix A). Then, the respondent received €10 to purchase the box of breakfast biscuits selected in this binding choice set at the corresponding price.
The respondents were also required to complete a brief questionnaire with the following structure: (i) food and breakfast product purchase and consumption; (ii) objective nutritional knowledge and use of nutritional information; (iii) interest in healthy eating; (iv) weight, height, and health status; and (v) socio-demographic and economic characteristics (gender, family size and composition, age, educational level, and income range).
2.4. Measures
The measurement of consumers’ choice were made by asking the respondents to choose four times between two boxes of biscuits or the non-buy option in front of the supermarket shelves, as mentioned above. In the questionnaire the participants were asked first whether they were responsible for food purchases. In addition, consumers were required to rate on a seven-point scale the importance that they attached to different attributes when shopping for food products. Finally, the respondents were asked about their frequency of consumption of cereals and breakfast biscuits; the options included never/once a month or less, 2–3 times a month, 1–2 times a week, 3–4 times a week, 5–6 times a week, once a day, and more than once a day.
The objective nutritional knowledge was measured following Grunert et al.’s (2010) [
34] scale based on the knowledge of dietary recommendations. The participants were asked about their knowledge of health expert recommendations (should eat more, about the same, less, or try to avoid) regarding a series of nutrients or substances.
The use of nutritional information was assessed using four items on a seven-point Likert scale (e.g., ‘I usually pay attention to nutrition information when I see it in an ad or elsewhere’) based on Moorman (1998) [
35]. The Cronbach’s alpha for both four-item measures was 0.78, indicating good internal consistency reliability. Interest in healthy eating was measured on a seven-point Likert scale using the Roininen et al. (1999) [
36] scale (e.g., ‘The healthiness of food has little impact on my food choices’) The Cronbach’s alpha for both eight-item measures was 0.84, indicating very good internal consistency reliability.
Finally, the participants were required to report, apart from their socio-demographic and economic characteristics, their weight, height, and health problems. With this information, each participants’ Body Mass Index (BMI) was calculated, and the participants were classified into different groups following Aranceta-Bartrino et al. (2016) [
37]. Details on how each item is measured can be found in the results section.
2.5. Data Analysis: Discrete Choice Modeling
The data gathered in the choice experiment were used to estimate a utility function derived from the Lancastrian consumer theory of utility maximization [
38]. Lancaster (1966) [
38] proposed that the total utility associated with the provision of a good can be decomposed into separate utilities for their attributes. However, this utility is known to the individual but not to the researcher. The researcher observes some attributes of the alternatives, but some components of the individual utility are unobservable and treated as stochastic (random utility theory by McFadden, 1974) [
39]. Thus, the utility is taken as a random variable for which the utility from the
nth individual facing a choice among
j alternatives within choice set
J on each of
t choice occasions is represented as follows:
where β is a vector of parameters associated with the vector of explanatory variables
Xnjt, and ε
njt is an independent identically distributed (i.i.d.) error term over time, people, and alternatives. Traditionally, it was assumed that consumers were homogeneous in terms of taste, and conditional logit models were fitted [
39]. However, numerous empirical papers using choice experiments have found that consumers’ preferences for food products are heterogeneous. In this case the specification of the model should allow the parameters to vary in the population. Two alternatives have gained popularity when addressing this issue of heterogeneity, the random parameter logit model (RPL) and the latent class logit model (LC), both of which are versions of a mixed logit model [
40]. In the RPL, each individual has a unique set of preferences and estimates of the utility function. Then, heterogeneity is included by adding a vector of parameters that incorporates individual preference deviations with respect to the mean preference values; β in (1) is not constant but varies across individuals, β
n. However, if the preferences are assumed not to be ‘unique’ for each individual but rather distinct for a determined number of individual classes, the LC suits the modeling of choices better. In this model, consumers are assumed to belong to different segments or classes, each of them characterized by different class-specific utility parameters. In other words, within each segment, consumers’ preferences are homogeneous, but preferences vary between segments, reflecting a ‘lumpy’ spread of preferences and allowing a more in-depth understanding of heterogeneity [
40]. The latter modeling approach has gained popularity and has recently been used in several studies on consumers’ valuation of food products [
41,
42,
43,
44,
45,
46].
In the LC model, the utility of individual ‘
n’ choosing alternative
j on the
tth choice occasion is:
where β
s is the parameter vector of class
s associated with the vector of the explanatory variable, and
Xnjt and ε
njt|s are error terms that follow a Type I (or Gumbel) distribution.
Thus, the probability that an individual will select alternative
j, conditional on being in segment
s, can be expressed as follows:
where
Pns is the allocation of individual
n to the
s class (probability of class
s), and
Pnjt|s is the choice probability that individual
n, conditional on belonging to class
s (
s = 1, …,
S), chooses alternative
j from a particular set
J, comprising
j alternatives, on a particular choice occasion
t [
47].
The parameters for the attributes are estimated by maximizing the likelihood function in the state of incomplete prior information on class membership or choice probabilities [
46]. Then, the number of segments is endogenously determined jointly with the utility coefficients. The latent class model was estimated using NLOGIT 5.0 (Econometric Software Inc., Plainview, NY, USA).
In our empirical specification, the utility function includes the product attributes as explanatory variables, as well as an alternative-specific constant (α) representing the non-buy option. The utility function is specified as follows:
The constant α represents the alternative-specific constant coded as a dummy variable that takes a value of 1 for the non-buying option and a value of 0 otherwise. It is expected that the constant α will receive a negative and significant value, indicating that consumers obtain a lower level of utility when they select the non-buying option than they do when selecting the other two alternatives (A and B). The price was defined by the price levels in the design. The other two variables (FIBER and FAT) were defined as dummies.
One of the key issues in latent class modeling is the selection of the number of segments to be considered. As Swait (1994) [
48] stated, the optimal number of latent segments must be selected by looking at different multiple statistical criteria but also by assessing whether additional segments provide any further economic information, with the overall aim of attaining segment parsimony. To determine the best number of classes, we calculated four information criteria; the Akaike Information Criterion (AIC), the modified Akaike Information Criterion (AIC3), the Bayesian Information Criterion (BIC), and the ρ
2, called the Akaike Likelihood Ratio Index [
49]. The preferred model should be the one with the lowest AIC, AIC3, and BIC and the highest
.
Using the estimated parameters, the marginal WTP was calculated as the negative ratio of the partial derivative of the utility function with respect to the attribute of interest, divided by the derivative of the utility function with respect to the variable price.
The marginal WTP was calculated for each of the obtained segments.