Designing the experiment

Given that this research applies experiments to answer the research questions, it is important to discuss intricacies related to how to appropriately design choice experiments.

101 5.4.1Selection of attributes and their levels

The first step in designing choice experiments is to select attributes and their associated levels (Cattin & Wittink, 1982; Louviere et al., 2000). There are several ways of developing a list of suitable attributes and associated levels. Cattin and Wittink (1982) propose that in generating attributes for CA (the same approach can be applied to DCEs and BWS), the choice should be based on consumer input. Therefore, the attributes should include those most relevant to consumers. To generate the list of attributes, the researcher may conduct qualitative work applying techniques such as depth interviews, focus groups, and triadic sorting (repertory grids) or more quantitative methods, for example free sorting and picking from a list of attributes (Bech-Larsen & Nielsen, 1999; Walley et al., 1999). Other researchers, for example Jaeger et al. (2001), Hall and Lockshin (2000), Lockshin et al. (2006) and E. Cohen (2009), propose conducting an extensive literature review to determine the list of attributes. Additionally, the choice of attributes may depend on the specific purpose of the research, i.e. a certain attribute must be included in the design as the research aims to examine consumer preferences for this attribute and its associated levels.

5.4.2Factorial designs

Once the attributes and their levels are chosen, the next stage is to generate factorial designs. Factorial designs are combinations of attributes’ levels in which each level of each attribute is combined with every level of all other attributes (Louviere et al., 2000). Factorial designs can be full factorials or fractional factorial designs (see e.g. Ryan & Morgan, 2007).

5.4.2.1Full factorials

Full factorials are the factorial enumerations of all possible levels of all attributes (Louviere et al., 2000). A significant advantage of full factorial designs is that they guarantee that all effects of interest, for example means, variances, and regression parameters, are independent (Louviere et al., 2000). Therefore, if any interaction effects5 _{between attribute levels are to be}

tested, full factorials should be applied (Louviere et al., 2015).

5 _{Interactions occur when consumer preferences for one attribute’s level depend on the levels of another}

102 However, the usefulness of full factorials is often limited. Full factorial designs can be used if the number of all possible combinations of attributes levels is relatively low. The researcher must remember that this number increases exponentially with the increase in the number of attributes and/or levels, thus it is often difficult to apply full factorial (Louviere et al., 2015). For example, in ranking-based CA it is impractical to present more than 20 profiles to respondents for rank ordering as respondents are likely to pay less attention to the ranking task due to the cognitive burden. In DCEs and BWS, a large number of profiles may result in the problem of how to present these profiles to respondents for evaluation. For example, in BWS responders choose the best and worst alternatives out of a few alternatives, but it is impractical to use more than five or six alternatives in each set, particularly if stimuli are visual and data collected online (see further discussion in Chapter 7).

One way to tackle this problem might be blocking profiles into versions of scenarios and assigning participants to each block (see e.g. Johnson et al., 2013). For example, in the 2×4×4×5×8 design with five attributes of two, four, four, five and eight levels of the first, second, third, fourth and fifth attribute, there are 1280 possible unique profiles. The researcher may create, for example 80 versions of 16 scenarios and, if they wish to have three profiles in each set, they make three copies of each version. Then they randomly assign profiles from each copy without replacement to make 16 choice sets of size three. Finally, the researcher randomly assigns participants to each block (e.g. 20 participants to each of 80 blocks if the sample size is 1,600).

5.4.2.2Fractional factorials – Orthogonal Main Effects Plans (OMEPs)

An alternative to the full factorial designs are fractional factorial designs called orthogonal main effects plans or OMEPs, which comprise a subset of a full factorial (D. Street & Burgess, 2004). However, it is important to remember that OMEPs can be applied only if the researcher wants to examine the main effects of independent variables (Louviere, 1988), as orthogonality assumes no collinearity between attribute levels. Moreover, OMEPs should not be used if there are any interactions between levels of one attribute and levels of another attribute (see e.g. Johnson et al., 2013). Applying OMEPs in experiments, in which interactions may occur, is likely to bias the parameter estimates (Louviere, 1988). OMEPs can be generated using

103 statistical packages, for example SPSS and SAS (Mühlbacher, Kaczynski, Zweifel, & Johnson, 2016).

5.4.2.3Problems of duplicates and dominance

In designing BWS applying full factorial and OMEPs, it is important to ensure that there are no duplicates per set and no dominance in sets (Huber & Zwerina, 1996; Louviere et al., 2000). Duplicating refers to two or more identical profiles occurring in a given set, while dominance occurs if one profile dominates other profiles in a given set. For example, profiles are the same in respect to all attributes levels except for the price level, and the profile with the lowest price level may dominate other profiles in the set. This would result in overstated preferences for these levels of attributes that are bundled with the lowest price. To deal with the issues of duplicates and dominance, Louviere et al. (2000) propose to fold over the original OMEP and/or transposition (shift) columns in it. Folding over means mirror imaging the original design, i.e. replacing each ‘dummy’ coded attribute level 1 with 4, 2 with 3, 3 with 2 and 4 with 1 (in the four-level attribute), while shifting columns occurs through column 1 becoming column 3, column 2 becoming column 1 and column 3 becoming column 2 (in the three options per set design; Louviere et al., 2000). This results in, for example the lowest price becoming highest (fold over) or a certain price level becoming a certain level of other attribute (shifting columns).

5.4.2.4Balanced Incomplete Block Designs

As previously noted, one way of creating choice sets in DCEs and BWS is the application of full factorials and a random assignment of profiles to sets. An alternative to this is the use of OMEPs and so called Balanced Incomplete Block Designs (BIBDs), or their certain types, e.g. Youden designs (see in Gupta, 2005; Raghavarao, 1988). BIBD is a design where all pairs of profiles occur together within a block an equal number of times (Louviere et al., 2015). BIBDs can be found in various sources, for example A. Street and Street (1987). Catalogues of BIBDs comprise designs with no duplicates per set issue, which makes it easier for researchers to design a flawless experiment.

104 5.4.3Final steps in designing experiments

The remaining steps in conducting experiments comprise laying out choice sets and formatting a response task, recruiting a suitable sample of participants, and administering experiments to obtain rank order of profiles (CA) and best and worst choices (BWS). The final step is to analyse the choices using a valid method (Louviere et al., 2015).