Using Best–Worst Scaling Methodology to Investigate Consumer Ethical Beliefs Across Countries

Abstract

This study uses best–worst scaling experiments to examine differences across six countries in the attitudes of consumers towards social and ethical issues that included both product related issues (such as recycled packaging) and general social factors (such as human rights). The experiments were conducted using over 600 respondents from Germany, Spain, Turkey, USA, India, and Korea. The results show that there is indeed some variation in the attitudes towards social and ethical issues across these six countries. However, what is more telling are the similarities seen and the extent to which individual variation dominates observable demographics and country-based variables.

Appendices

Appendix A: A simple model for BW judgments

BWS is a fairly general scaling method that extends Thurstone’s (1927) Random Utility Theory-based model for paired comparison judgments to judgments of the largest/smallest, best/worst, most/least, etc., items, objects or cues in a set of three or more multiple items. Specifically, BWS assumes that there is some underlying subjective dimension, such as “degree of importance,” “degree of concern,” “degree of interest,” etc., and the researcher wishes to measure the location or position of some set of objects, items, etc., on that underlying dimension. We refer to the process of assigning numerical values that reflect the positions of the items on the underlying scale as “scaling.” The BWS approach is based on the view that such measurement arises from theory, and that theory and associated measurement are inseparable. Thus, the scale values derived from BWS are those that best satisfy a theory about the way in which individuals make BW judgments.

To begin, we assume that there is a master set of K items to be scaled, {I ₁, I ₂, ..., I _K }. The items are to be placed in c subsets, {i ₁, i ₂, ... i _C }, and some sample of individuals of interest is asked to identify, respectively, the best and worst items in each of the subsets (or in each of some subset of the subsets). If there are K total items to be scaled; then the total number of subsets that could be presented to the individuals is “K, pick 3”, which grows exponentially with K. Thus, one needs some systematic way to pick the subsets that makes sense, and as noted by Finn and Louviere (1992), constructing the sets from a 2^K orthogonal main effects design or some higher resolution design in the 2^K family of designs is a good approach, which coincides nicely with previous design theory for the case of only “best” choices (Louviere and Woodworth, 1983). There are other ways to construct appropriate sets, such as balanced incomplete block designs (BIBDs), and we illustrate the use of such designs in this paper.

Thus, BWS assumes that there is some underlying dimension of interest, and one wants to assign scale values to the K items on that single underlying dimension. It assumes that the choice of a pair of items from any subset is an indicator of that pair of items in that subset that are the farthest apart on the underlying dimension. That is, in any subset, say the cth subset, there are K(c−1)/2 pairs of items that could be chosen best and worst, and an additional K(c−1)/2 pairs of items that could be chosen worst and best. Thus, for any given subset presented to an individual like the cth subset, the individual implicitly chooses from 2 × K(c−1)/2 pairs. Let us denote the quantity 2 × K(c−1)/2 as M, and for ease of exposition (and because it reflects the case in this paper), we assume that P is constant in every subset (e.g., balanced incomplete block designs lead to subsets of fixed size, M). Now, we can formulate this choice process as a random utility model as follows:

$$D_{ij}=\delta_{ij}+\varepsilon_{ij} $$

(1)

where, D _ij is the latent or unobservable true difference in items i and j on the underlying dimension;

• δ _ij is an observable component of the latent difference that can be observed and measured; and _ij is an error component associated with each ij pair.

Due to the presence of the $\varepsilon_{ij}$ component, the choice process of any individual is stochastic when viewed by the researcher because we cannot know what the individual is thinking. Thus, we can formulate the model as a probability model to capture the probability that the individual chooses the ij pair in each subset:

$$P(ij|C)=P[(\delta_{ij}+\varepsilon_{ij})> \hbox{ all other }M-1 (\delta_{ik}+\varepsilon_{ik})\hbox{ pairs}], $$

(2)

where all terms are as previously defined. This problem can be solved by making assumptions about the distribution and properties of $\varepsilon_{ij}$. A simple assumption that leads to a tractable model form that has seen many applications in the social and business sciences is that $\varepsilon_{ij}$ is distributed independently and identically as an extreme value type 1 random variate (equivalently, as a Gumbel, Weibull or double exponential). It is well known that these assumptions lead to the multinomial logit (MNL) model (e.g., Ben-Akiva and Lerman, 1985; Louviere et al., 2000; Louviere and Woodworth, 1983), which is the form of analysis used in this paper. That is, the choice probabilities can be expressed as:

$$P(ij|C)=\exp(\delta_{ij})/\Sigma_{ik}, \exp(\delta_{ik}), \quad \hbox{ for all }M\delta_{ik} \hbox{ in } i_{C}. $$

(3)

We can express δ _ij a difference in two scale values, say s _i and s _j , or s _i −s _j . Hence, we can rewrite the model as:

$$P(ij|C)=\exp(s_{i}-s_{j})/ \Sigma_{ik}, \exp(s_{i}-s_{j}), \quad \hbox{ for all }M \{s_{i}, s_{k}\}\hbox{ pairs in } i_{C}. $$

(4)

Thus, the scale values of interest are s _i and s _j , which reflect the location of each item on the underlying scale.

If the subsets are constructed in such a way that the joint probability of choosing items i and j across all subsets can be estimated independently of the marginal probabilities (e.g., by using a 2^k orthogonal main effects design + its foldover, or a BIRD + its complement), then the model implied by Equation (4) can be estimated directly from the observed counts associated with each best–worst, worst–best pair summed over all subsets in the experiment. If the experiment does not allow one to calculate the total choices of all implied best–worst, worst–best pairs across the subsets (e.g., if one only uses the orthogonal main effects design or only the BIBD as discussed by Finn and Louviere, 1992), one can approximate the desired scale values by taking differences in the marginal best and worst counts for each item. That is, the simple score δ(b _i w _i ) = total best i − total worst i, approximates the unknown difference s _i −s _j for each individual or subset of individuals who exhibit the same underlying ordering of the items (apart from judgmental errors). We state this without proof, but note that one can easily see that this must be true by constructing an experiment that permits the joint choice probabilities for all the implied pairs to be estimated independently of the marginal probabilities, assuming an ordering of the items in that experiment, and simulating choices of the items with the highest and lowest rank in the order in each subset. It is easy to show that the total choices over all subsets for the implied pairs will be consistent with MNL, and once one obtains the MNL estimates, one can easily see that the best _i  − worst _i differences are perfectly proportional to the MNL estimates.

Appendix B: BW experiment

In this section, we will present you with 16 social and ethical issues. These will be organized in groups of four over the next two pages (a total of 20 groups or questions). For each group, select the one issue among the four that is least important to you and the one issue that is most important to you. Please make sure that you select only one least important and one most important for each group of four issues. We have included a description of the issues below; please keep them in mind throughout the rest of this section.

• Animal rights – describes the general treatment of animals for commercial purposes such as the use of animals for product testing, the displacement or killing of animals for natural resource exploitation (e.g., logging), or the cruel use of animals for entertainment.

• Animal by-products used – indicates that the product is made using animal by-products such as animal fat or lard.

• Product biodegradability – indicates that the materials used to make a product can be broken down naturally and hence are safer for the environment.

• Products made from recyclables – indicates that some or all of the materials used to make a product were obtained from recycled sources.

• Product safety information provided – means that information about the safe use of a product and/or potential dangers from using a product is/are included with the product.

• Human rights – describes the basic rights of all people as stated in the Universal Declaration of Human Rights such as the right to food, clothing, housing, education, etc.

• Packaging recyclability – indicates that part or all packaging materials can be recycled for future use (e.g., product packages, food containers, shipping boxes, etc.).

• Product disposability – indicates that a product can be disposed of without causing undue damage to the environment.

• Paying minimum wages – signifies that companies adhere to the minimum wage standards of the country(ies) in which they are operating.

• Unions allowed – indicates that unionization is legal within a country and that companies producing in that country do not attempt to prevent or curtail the unionization of their workers.

• Minimum living conditions met – means that companies supply their employees with basic and acceptable living accommodations when required.

• Sexual rights – indicates that discrimination against individuals based on their sexual orientation is not allowed.

• Safe working conditions – signifies that companies follow a set of procedures to create a safe working environment for their workers.

• Child labor not used – means that companies do not use workers under the minimum working age in the country(ies) in which they are operating.

• GM material used – indicates that the use of GM materials are allowed within a country and that companies use GM materials in their products.

• Gender, religious, racial rights – indicates that discrimination based on gender, religion, or race is not allowed.

Example

In this example, sexual rights are least important and human rights are most important. Please notice that only one issue was selected in each column (Least Important and Most Important).