3.2. Stimuli: The Discrete Choice Experiment
Choice experiments (CEs) have become widely used in the field of food choice research. This approach is useful for understanding the demand for a new product with new attributes and also for examining behavioural issues [
27]. When carefully designed, a CE conducts a temporal evaluation of attributes over a range of choices in order to reveal any significant relationships between choice and available attributes.
As well as being useful for the analysis of product demand, a CE provides an attractive, yet not widely explored, feature in allowing for the estimation of the choice probability for a product alternative, conditional upon the depth and content of the consideration set available to the individual when being asked to make the choice. In the context of food choice, it has been reported that CE estimates are sensitive to the dimensionality of the experimental design. Gao and Schroeder [
25] and Caputo, Scarpa, and Nayga [
28] found that the stability of preferences for cue attributes was affected by the number of attributes. They also reported that this effect existed for credence-type attributes but not for so-called independent attributes (i.e., aspects related to the physical nature of the product, information on which can be directly observed by the consumer). These results corroborate findings by Hensher [
29] and suggest that the information processing strategy of individuals relates to the functional relationship between attributes available to individuals in the choice situation. Therefore, a partial profile design was created for the CE in this study. The partial profile design, which was first described by Green [
30], allows for more realism in the decision-making process, as product comparisons in real-life situations are less likely to always include identical labelling attributes. To replicate this presence/absence of certain information, the attribute levels within our CE were set as binary. In this way, at each choice (the temporal structure), a set of unique labels (i.e., cues) was made available. This aspect of the partial profile design makes the CE more reflective of the cue-based decision-making model by Hamlin [
9].
The partial profile design presents choice tasks that vary only in the levels of a subset of all attributes, which distinguishes it from a full-profile design (across all choice tasks and concepts, all attributes are present, although the levels of each attribute vary according to the experimental design). A full-profile design is not representative of more realistic in-store choice situations, as beef products typically differ in the extent of the labels presented on the package. The full-profile design has, however, been widely used in applied research, especially when the number of attributes is not too large or when there are only a few levels per attribute, or both (as recommended by Green and Srinivasan [
31]).
The use of a partial profile design is not without problems, however, as previous findings suggest that the importance of the price attribute may be reduced, leading to inflated willingness-to-pay estimates. Based on research by Hensher [
29] showing that the likelihood of misspecified estimates increases with a narrower attribute range, it was therefore deemed relevant to include a rather wide array of price levels.
An unlabelled choice task approach was taken in the discrete choice experiment (DCE). The heading of each alternative (within choice tasks) was generic (i.e., beef alternative 1, beef alternative 2, etc.), and the only way to discriminate between the alternatives was through the attributes. An example of a choice task used for the DCE is given in
Figure 1. The respondents were asked to consider at most six attributes (middle and right-hand concept in
Figure 1) and at least four attributes (left-hand concept). In this example, information (labels) referring to reference code, farm animal welfare, organic production, and health impact were not presented. The respondents were initially instructed that the DCE was related to beef products such as minute steak, pepper beef, roast beef, sirloin steak, and tenderloin. The respondents were told to assume that all other mandatory information regarding the choice was always present and that the alternatives presented in each choice task only differed in the attributes presented (
Table 3). The food labelling rules are set at the European Union level for all member states, and the general labelling requirements are currently set out in Regulation (EU) 1169/2011. This regulation outlines the mandatory information that must be included on all food product labels, including the product name, ingredient list, use-by date, and any specific instructions or conditions of use.
In each choice task, there were three alternatives. Two alternatives always included the specific COO denomination so as to allow for trade-offs between the remaining attributes presented and the specific origin. These alternatives correspond to existing labelling requirements. This was intended to establish a link to random utility theory and avoid the unfeasibility problem. Then there was one alternative with the EU/non-EU denomination of origin. The design provided a constrained balanced approach (i.e., equal occurrence of each attribute, except for the origin attribute, which was present for each alternative). The relative d-efficiency of the partial profile design was 0.93.
Each respondent was then faced with 22 choice tasks, which, drawing on the extant literature following Bradley and Daly [
32], represents a point at which most analysts would expect fatigue effects to have set in. However, more recent findings by Hess, Hensher, and Daly [
33], who reviewed and tested for scale differences due to the number of choice tasks in datasets within five stated preference studies from a number of disciplines, while controlling for contexts of familiarity with the market in question, showed an absence of fatigue effects. Similarly, Louviere et al. [
34] show that there is little loss of reliability and validity when using larger and more complex choice tasks. In fact, the literature suggests that considerable gains can be achieved by increasing the number of choice tasks per respondent such as the generation of learning effects, which increase model structure reliability and precision [
35,
36,
37]. It has been reported that a similar increase in model precision can be obtained by increasing the number of tasks as by proportionally increasing the number of respondents [
36]. In addition, increasing the number of choice tasks has been reported to establish a learning effect whereby respondents have been found to learn to draw finer distinctions between attributes as they progress through the choice tasks. The respondents have thereby been reported to focus on brand or COO (as a proxy for other attributes) over price in the first task, while this effect diminishes in subsequent choice tasks [
35,
37].
Moreover, a complete heterogeneous design [
38], rather than a blocked design, was used to increase the statistical efficiency by providing more variation across respondents and to reduce the problems of scale effects (i.e., variations in preferences due to the block of the design from which data were generated). The heterogeneous design meant that the respondents were randomly assigned one of 100 versions of the full design. On completion of the DCE, the respondents were asked to rate (on a scale from 1 = agree to 5 = disagree) their understanding of the task assigned to them in the DCE. This included three statements referring to (i) the ease of understanding of how to provide responses; (ii) an understanding of the labelling attributes; and (iii) the ability to express what is important concerning the labelling of beef. Furthermore, the respondents were asked (on a scale from 1 = very easy to 5 = very difficult) to rate the perceived difficulty in expressing which type of beef labelling information was important. This was to infer the cognitive burden related to the responses.
Given that the standard information is provided on the label or on the package, which of the following three beef products would you prefer? (The country flags are only illustrative).
Mark your choice by using the buttons, and please bear in mind the price that is associated with your choice:
3.3. Statistical Analysis
Random utility theory (RUT) provides a family of probabilistic choice models that describe how choice probabilities relate to changes in choice tasks (i.e., attributes and their levels) and to individual choosers. In accordance with RUT, Equation (1) then describes the probability of individual
n choosing alternative
i from the choice task
Cn equalling the probability of the systematic (
) and random components (
) of the latent unobservable utility associated with alternative
i for individual
n being larger than the systematic and random components of all other alternatives competing with alternative
i within the choice task Equation (2) [
39].
where for all j options in choice set C
n.
The analysis of the DCE data was adapted to the study’s purpose of examining the choice probability for a product alternative with an EU/non-EU denomination of origin conditional upon the depth and content of the consideration set available (i.e., credence quality cues). Therefore, a mixed logit model approach was developed. The model takes the nested nature between choice of denomination (individual preferences) and the (exogenous) explanatory variables into account within mixed effect estimation, thus allowing a random error component so as to capture individual heterogeneity in responses within and across choice sets.
In modelling the nested data structure of
i persons who completed
j choice tasks, with each task including
k choice concepts, the general structure of the mixed logit model used was:
where the
r × n matrix
X is a representation of the
r explanatory variables;
β is a
r × 1 matrix of the parametric coefficients to be estimated on
X; the
q × n matrix
Z is a representation of the
q random effects;
ζ is a
q × 1 matrix of the random effect coefficients to be estimated on
Z so that it captures parts of the unobserved heterogeneity of the respondents; and
ε is the idiosyncratic error of unexplained variance in the dependent variable. However,
y is specified as a variable that follows a binomial distribution; for each respondent and for each choice concept in each choice task, this dependent variable takes a value of one for all those observations under which a respondent has chosen the EU/non-EU denomination of origin rather than the ‘specific country’ and a value of zero otherwise. The binomial distribution of
y is consistent with the design of the choice sets. Therefore, we employed a logistic link function
such that the model in Equation (2) became:
This type of model is less common in econometrics but is widely used for experimental data obtained, e.g., in crop sciences, medicine, or psychology. In these disciplines, the models are known as ‘linear mixed models’ (e.g., [
40]). In contrast to the conditional logit model that is commonly used in the context of choice experiments, our approach has much more flexibly, which allows it to capture unobserved heterogeneity within the data through the random effects ζ.
The model in Equation (4) was estimated using a Restricted Maximum Likelihood (REML) model, as implemented in the lme4 package [
40,
41] from the R network software [
42].
The model specification and the selection of the final model were based on the following steps:
The model was set to explain the choice of the dependent variable ‘EU/non-EU origin’ as a function of price level and the number of additional attributes provided as explanatory variables X.
Alternative specifications of Z were estimated as random effects; the selection of the best random effects specification was based on Likelihood Ratio tests for model selection.
The model was tested under alternative specifications of the explanatory variables, treating ‘Price level’ and ‘Number of information items provided’ as either discrete or continuous variables or as a combination thereof.
In a second set of regressions, the variable containing the number of information attributes was replaced by dummy variables for the actual information categories that were provided.
The marginal effects were computed according to procedures outlined by Fernihough [
43], but, with the code provided, they were revised because it simulates only one standard error for all marginal effects. The marginal effects in the present study were averages of the sample marginal effects (rather than average marginal effects) and were computed by multiplying each coefficient
estimated from Equation (4) by the transformed values from the logistic probability density function of the predicted values [
43].