Anchored vs. relative best–worst scaling and latent class vs. hierarchical Bayesian analysis of best–worst choice data: Investigating the importance of food quality attributes in a developing country

https://doi.org/10.1016/j.foodqual.2012.01.002Get rights and content

Abstract

Applying best–worst (BW) scaling to a multifaceted feature, e.g. food quality, is challenging as attribute non-attendance or lack of attribute discrimination risks invalidating the transformation of choice data to unidimensional scale. The relativism of BW scaling also typically prevents distinction of respondents or groups of respondents based on similarities to the study object. A dual-response BW scaling method employed here to obtain an anchored scale allowed comparisons of importance ratings across individuals. Attribute importance ratings and rankings obtained were compared with those from relative BW scaling. Latent class (LC) and hierarchical Bayesian (HB) analyses of individual specific BW choice data were also compared for ability to consider within- and between-respondent choice heterogeneity. Personal interviews with 449 consumers provided data on the importance of 16 food quality attributes of kale produced in peri-urban farming in Kenya. Major findings were that the anchoring model improved individual choice predictions compared with conventional relativistic BW scaling, i.e. was more reliable in measuring consumer preferences, and that HB analysis fitted the data better than LC analysis. HB analysis also successfully obtained individual parameter estimates from sparse data and is thus a promising tool for analysis of BW choices in sensory and consumer-orientated research.

Highlights

► Consumer perceptions of food quality were studied in a developing country context. ► Anchored best–worst scaling (BWS) was developed and compared with relative BWS. ► Latent class and hierarchical Bayesian estimations provided importance weightings. ► Food quality related to vegetables was shown to be a multi-dimensional phenomenon. ► Intrinsic credibility attributes were predominantly perceived as important.

Introduction

There is growing use of best–worst (BW) scaling to investigate preferences for products and programmes in consumer research (e.g. Jaeger and Cardello, 2009, Jaeger et al., 2009, Jaeger et al., 2008) and in several other disciplines, including health treatment (e.g. Lancasar, Louvierie, & Flynn, 2007), social care (Potolglu et al., 2011) and value research (Lee, Soutar, & Louviere, 2008). Among methods for the measurement of consumer preferences and product attribute importance, BW scaling can generate greater consumer discrimination in sensory testing, although it is viewed by panellists as the most demanding method (Hein, Jaeger, Carr, & Delahunty, 2008).

In a standard BW scaling approach, a number of attributes take a particular level and for a series of choice sets, respondents are asked to evaluate and select the attributes perceived to be farthest apart on an underlying latent scale. Fig. 1 illustrates the BW choice set used to collect data for this study. Methods to analyse BW data are well developed (e.g. Jaeger et al., 2008) and typically relate all estimated differences to provide a form with ratios of scale values (attribute importance) for each attribute relative to a single (arbitrary) attribute for which the importance cannot be estimated to avoid linear dependency. This then acts as the reference attribute.

The BW scaling approach to a multifaceted feature such as food quality can be challenging, as it results in a unidimensional scale. Food quality and choice are dynamic, multi-dimensional aspects linked to the concept of consumer acceptability (Issanchou, 1996). In addition, food quality includes a consumer perception-related subjective dimension (Grunert, 2005). Presence of attribute non-attendance or lack of discrimination among attributes risks invalidating the transformation of choice data into the unidimensional scale. The relativism of BW scaling is also considered a limitation (Jaeger et al., 2008) and typically excludes the possibility to distinguish respondents or groups of respondents based on similarities to the subject under measurement. Against this background, this study sought to: (1) introduce the dual-response anchored BW scaling method and compare it with the relative BW scaling approach in order to determine whether the anchored approach warrants inclusion as a tool in consumer research due to its internal validity and respondent segmentation qualities, (2) compare latent class and hierarchical Bayesian analyses of individual specific BW choice data based on their capacity to consider within- and between-respondent choice heterogeneity, and (3) present a case study investigating the importance to consumers of a range of quality issues for leafy vegetables produced by farmers in peri-urban areas (within 20 km of the city centre) in a developing country context.

Sections 1.2 Relative versus anchored dual-format best–worst scaling, 1.3 Analysing individual choices from best–worst data introduce the methodology of anchored BW scaling, considering in particular the theoretical and applied aspects of introducing a dual-response choice format and a no-choice option, and the use of a Bayesian approach to estimate BW data. Sections 1.4 Aggregate multinomial logit model, 1.5 Measures of performance introduce the measures used to assess the performance of the models in terms of reliability, fit and predictive accuracy. The empirical study examining the importance to consumers of food quality attributes of fresh vegetables is introduced in Section 1.6. The methodology is described in Section 2 and the results are presented and discussed in Section 3, which also includes a food policy-orientated discussion of the findings.

BW scaling as a technique was originally devised to allow respondents to make trade-offs in their choices between the relevant food safety issues that compete for their concern (Finn & Louviere, 1992). Being based on the theory of paired comparisons (Thurstone, 1927), BW scaling is related to a generally accepted theory of human decision-making and may thus overcome the potential response biases typical of rating scales (e.g. social desirability bias, acquiescence bias, extreme response bias), as identified by Paulhus (1991). The method has gained popularity and has been used e.g. to measure the importance of product benefits (Cohen & Markowitz, 2002), and to identify the relative importance of 11 general food values (Lusk & Briggeman, 2009).

The standard (relative) BW approach has several advantages over conventional rating scales (Flynn et al., 2007, Jaeger et al., 2008, Lusk and Briggeman, 2009), which together with its benefits with respect to potential response biases typical of rating scales motivated its use in this study. The methodology is generally easily understood by the respondents, while at the same time stressing the trade-offs between the importances of issues. In addition, compared with rating techniques, use of BW scaling may overcome cultural equivalence problems (e.g. those relating to terms of gradation).

However, there are two methodological remedies of the standard BW scaling approach that are relevant to address in sensory and consumer-orientated research.

Firstly, relating to the multi-faceted nature of consumers’ perception of food quality, a basic axiom in choice modelling is that of unlimited substitutability between the attributes tapping the latent construct. In a technical sense, continuity is needed to ensure that for any choice set in a B–W study, there exists some probability such that the respondent is indifferent between the “best” and the “worst” attributes, while substitutability requires that respondents would be indifferent between two choice sets which offer these attributes with equal probabilities, if the choice sets are identical in every other way. Continuity is a key issue in transforming BW choice data to a unidimensional probability scale, but the transformation has external validity if, and only if, individual respondents have made accurate and continuous trade-offs between all attributes.

In studying food decision-making this may represent too strong an assumption, since consumers are known to be highly influenced by habitual and emotional aspects, as well as being characterised by low levels of involvement for routine purchases (see e.g. Costa et al., 2003, Helsop, 2007). Mantel and Kardes (1999) found that product attributes (qualities such as price, size, nutritional value, durability, etc.) are often compared disproportionately, i.e. one is the subject of more intense comparison, thus eliciting more consideration when the consumer decides which attribute is the best. Mantel and Kardes also found that purchase decisions are not always made on the basis of an attribute-by-attribute comparison (attribute-based processing). Instead, consumers also make decisions based on an overall evaluation of their impressions, intuition and knowledge based on past experience, or attitude-based processing.

A growing body of literature is emerging around how individuals evaluate a set of attributes associated with ordered or unordered choice alternatives. The findings suggest that individuals make use of a number of attribute processing rules to arrive at choice outcome (Hensher & Greene, 2010). The effect of information processing strategies and attribute processing rules on choice outcomes is relevant to consider in a BW context, especially what if a respondent were to cancel, ignore or exclude a certain attribute, or a set of attributes (attribute aggregation), for example the washing and sprinkling attribute displayed in Fig. 1 in choice sets where this attribute occurs? This could be due to e.g. non-interest or lack of attribute attractiveness/relevance and, depending on the choice context, could originate from reference dependence, through lack of knowledge or uncertainty (Fenichel, Lupi, Hoehn, & Kaplowitz, 2009). Effectively, this would qualify as attribute non-attendance. However, what if a respondent were to lack the ability to fully discriminate between attributes because of task similarity (Tversky & Shafir, 1992), e.g. in responding to a set in which the respondent cannot make a sufficient distinction between the attributes. For example, the respondent(s) could be unable to disentangle how the two food safety attributes in Fig. 1 (washing and sprinkling versus use of pesticides) differ with respect to food quality. In addition, lack of discrimination could exist because of task attractiveness or complexity (Hensher, 2006) (e.g. in responding to a set in which the cognitive burden of deciding upon importance becomes too difficult when all attributes are considered vital).

Attribute non-attendance or lack of discrimination implies that respondents do not make trade-offs between all attributes included, potentially invoking attribute discontinuity and non-substitutability. Persistently disregarding from one, or several, attributes means there exists no probability that the respondent will be indifferent between the best and worst attributes within choice sets in which these attributes appear, or indifferent between any pair of choice sets where these attributes appear. Attribute discontinuity is of particular relevance in a BW scaling context, since the method aims to induce respondents to discriminate between attributes. Thus, if some respondents consider all attributes while others consider only a subset, leading to discontinuous preference orderings, the pooling of such choice-based data into an aggregated model for estimation of attribute importances may generate biased estimates.

The issue of attribute non-attendance or lack of discrimination and its relationship to attribute discontinuity in BW scaling remains largely unexplored within the sensory and consumer-orientated research related to food quality. In order to address the influence of non-attendance, past research within discrete choice-based modelling has either used self-reporting in collecting data on the attributes given non-attention by respondents (e.g. Hensher, 2008) through asking questions such as: ‘Did you disregard any (and if so which) attribute/level in choosing between these alternatives?’ either after each choice set, or at the end of the questionnaire. The benefit to the analyst of such control questions has recently been challenged. Evidence suggests that respondents claiming to have ignored a certain attribute have actually assigned it less importance (Hess & Hensher, 2010). Another strand of the literature has developed specific entropy estimation methods to account for attribute non-attendance when data on self-reported attribute non-attendance are not available, or sought (Scarpa, Gilbride, Campell, & Hensher, 2009). In the latter approach there is no requirement to have the respondent report back on attribute ignorance. Instead, the estimation technique is able to detect and adjust for this aspect.

Secondly, standard BW scaling measures relative, rather than absolute, preferences. Relating differences for each attribute to a single reference attribute removes the common origin of individual scales. The basic problem then, like that in traditional rating approaches, is that the concept of ‘best’ or ‘worst’ is only comparable within a respondent, but not across respondents. Bacon, Lenk, Seryakova, and Veccia (2008) argue that the relativism within discrete, choice-based methods excludes meaningful information that distinguishes respondents or groups of respondents.

In order to overcome these methodological shortcomings of the standard BW approach, here we adopted an anchored dual-response format, in line with suggestions within business marketing research (Sawtooth Software, 2009a, Sawtooth Software, 2009b). While being based on a self-reporting approach the dual-response approach brings additional important information features. First, in comparison with the non-attention approach, the level of information obtained is higher, since the respondent indicates the discrimination (i.e. the trade-offs being considered) within each choice set along the full range of best-to-worst. Secondly, asking for an additional discrimination among the attributes within choice sets relates to the discovered preference hypothesis (Plott, 1996) in the assessment of attribute importance. In this approach, people gain successive experience with the particular context, and the dual-response question may make respondents consider the choices within each choice set in a more analytical decision-making mode, because of the request to indicate three alternative relationships among attributes within sets. Proceeding along the choice sets, respondents may then become more observant of the connections between attributes. To our knowledge, no previous study has applied the anchored dual response format methodology of BW scaling in a food quality context. The novel contribution of this work lies in the evaluation and comparison of relative and anchored dual-format BW scaling.

To implement the anchored dual-response format, as shown in Fig. 1, respondents were first asked to choose the BW attribute combination and then choose between the available attributes and an elaborated no-choice alternative. This dual response question was used as a fifth choice option in the subsequent analysis. The response ‘Some are important, some are not’ represented the anchor between important and unimportant, and indicated that the anchor was perceived as either ‘best’ or ‘worst’. From this question a common scale origin exists (theoretically) with the utility of the anchor being zero. Moreover, a response ‘None of these four is important’ was taken to indicate that this alternative was viewed as being preferable to the other attributes within the choice set, i.e. representing the ‘best’ alternative (to indicate that the offered attributes failed to meet some importance threshold of the respondent). Thus, instead of ignoring the choice set and leaving it unanswered, or through lack of other way of expression providing a choice within a set of irrelevant attributes, the respondent could indicate this lack of importance directly. This anchoring alternative is therefore the same as a ‘None of these’ option in the choice literature. On the other hand, the response ‘All four are important’ was taken to indicate that all four attributes were preferred to the anchor and that the fifth attribute was the ‘worst’ alternative. This indicates a potential lack of attribute discrimination. Importantly for the analysis of the B–W data, use of this alternative takes out choice of the attribute that is the least important.

Although this anchoring approach generates degrees of importance that are scaled relative to a common anchor, there is still the possibility of heterogeneity among respondents in their conception of the anchor, as the definition of importance or unimportance can differ between individuals. The benefit of this approach is the potential information gain at the individual level allowing for discontinuity in attribute attendance to be detected and for further segmenting of respondents depending on their responses to the dual-response questions. An important caveat to the anchored dual-format approach could be if a new bias were to be introduced to the choice setting by the inclusion of this additional choice alternative, i.e. if the responses under the dual-response format violated the independence from irrelevant alternatives (IIA). This would be the case if the addition of the dual response task drew proportionately from each of the available alternatives.

Fig. 2a, Fig. 2b illustrate the coding of responses of the anchored BW choice task as presented in Fig. 1. Data were coded in relation to the design matrix format (see Appendix A) with number of columns (16) equal to the total number of attributes included. The presence of a given attribute in the choice set was dummy-coded. There was one data matrix for each choice of most important (with dummies as +1) (Fig. 2a) and one matrix for each choice of least important (dummies as −1) (Fig. 2b). Each matrix had seven rows, with rows two to five being associated with the presence of the quality attribute in the BW part of the choice set, while row six indicated the presence of the additional anchoring question. The illustration displays the coding when the respondent first chose alternative 3 (attribute 6: ‘Pesticides have been applied using the recommended dosage for a given symptom’) as ‘Most important’ and then alternative 2 (attribute 14: ‘The product bought is washed and sprinkled with running tap water at the market’) as ‘Least important’, while in the anchoring question selecting the alternative ‘All four are important’.

A general problem in analysing choice data in general, including data from BW scaling, is the relative scarcity of data, i.e. frequency counts for individuals, which has motivated the use of aggregated methods for analysis. For BW data, little is known concerning the risk of declining attention or cognitive burden on respondents from including more choice sets in a study. Cohen and Orme (2004) reported that more time was needed to complete a BW choice experiment than a rating experiment, while Hein et al. (2008) found BW scaling to be more demanding for respondents than other methods.

When the aim of the analysis is to generate individual specific information for, say, segmenting purposes, the potential conflict between more extensive data needed to obtain accuracy on one hand, and study design on the other, is still challenging. Marley and Louviere (2005) recently formulated probabilistic models for BW choice data that can be used to obtain distributions of individual BW choices: Formally, let the q-dimensional vector xkj represent the level of attribute j in choice set k and let the q-dimensional vector β contain the coefficients for degree of importance of the q food quality attributes to be elicited. In a situation with N respondents deciding simultaneously and independently, where pBWkii denotes the joint probability of choosing which alternative i is the best and which alternative i′ is the worse in a choice set k out of K choice sets, the log-likelihood function can be expressed as (Vermulen, Goos, & Vandebroek, 2010, p. 1428):lnLk(β)=n=1Nk=1Ki=1Ji=1JiiynkiilnpBWkiiwhere ynkii is a dummy variable which equals one if respondent n chooses attributes i and i′ as the best and worst alternative, respectively, and zero otherwise. Eq. (1) can then be used to obtain the maximum likelihood estimate of the degree of attribute importance β. Conventional estimation methods to obtain individuals’ importance parameters from Eq. (1) include the random parameter model (RPL) (Train, 2003). Lusk and Briggeman (2009) provided a comparison between RPL and MNL for BW data. Although the RPL model allows preferences to vary due to unknown characteristics across respondents, it requires conditioning individual heterogeneity on a pre-specified functional form. The results obtained relate to the degree to which the data are consistent with the imposed constraint on parameter space. There is thus a risk that the quality of choice data is lost in estimating the model. Furthermore, the RPL model does not reveal any information about the source of heterogeneity (Boxall & Adamowicz, 2002). In addition, in cases where there is a strong pattern of homogeneity within groups and heterogeneity between groups, the use of an RPL model is not recommended, as identification of parameters with discrete support is difficult. Exogenous segmentation criteria have been used to resolve this identification issue (Rigby & Burton, 2005), but such criteria typically lack theoretical support.

In the presence of between-respondent variation, a latent class (LC) model can deal with individual heterogeneity by implicitly grouping individuals that exhibit similar (unobservable to the analyst) preferences into specific classes. Latent class methods allow each respondent to belong to more than one segment, and solve simultaneously for the elements of β for each segment and for each respondent’s probability of belonging to each segment. The elements of β are, however, just obtained as point estimates (averages) for each segment. In comparing several methods for segmenting respondents based on choice data Moore, Gray-Lee, and Louviere (1996) found that the LC method generated more accurate results than aggregated or clustering methods. When latent constructs are assumed to affect the choices made by respondents, exogenous indicator variables can be used to determine class membership to improve the LC estimation. The problems with such approaches are similar to those for the RPL model.

Hierarchical Bayesian (HB) models have recently been used to handle the presence of within- and between-respondent choice heterogeneity. Note that both the LC and the HB approach investigate the probability distribution of the parameters given the data, instead of the opposite in the RPL model, which means that data quality is not lost in estimating either the LC or HB model. A further advantage of HB is the ability to generate individual specific data from sparse data sets. Duinevald and Meyners (2008) analyse discrimination rates in replicated triangle sets and review the hierarchical approach to Bayesian estimation. In addition, business marketing research has developed a HB approach that considers within- and between-respondent variations in choice data (Sawtooth Software, 2009a, Sawtooth Software, 2009b). This HB framework introduces two levels of analysis: (1) at an overall level the prior probabilities of an individual’s degrees of attribute importance (means and standard deviations) are drawn from an overall multivariate normal distribution of means and standard deviations. Using Gibbs sampling and a Metropolis Hastings Algorithm, iterative draws are made of the individual parameters of β until convergence is obtained. The first step conditions the individual parameters to the population mean and thus relates to the between-respondent variation. (2) After convergence, successive draws of β for each individual are continued while considering the relative density of the individual βs and while assuming the logit rule for individual choices providing the likelihoods. This step recognises within-respondent variation in β. Taken together, by using information concerning β pooled across respondents, individual estimates of β can be obtained even from sparse choice data.

To our knowledge, no previous study has compared the LC and HB methods for analysing the degrees of food quality attribute importance. The second novel contribution of this work lies in the evaluation and comparison of LC and HB for relative and anchored dual-format BW scaling data.

Because of the aim of this work was to compare latent class and hierarchical Bayesian estimation of BW data, the data were also analysed using multinomial logit (MNL) regression as a benchmark for ranking quality attributes and model fit. Jaeger et al. (2008) reported that MNL parameter estimates (i.e. aggregate analysis) are proportional to results obtained from analysing BW scores. Such standard BW scores, by which counting the number of times an attribute is chosen as most important is related to the number of times it is considered least important, were not directly applicable as a benchmark here, since the design meant that attributes were presented a different number of times. Instead, a joint MNL model was estimated for choices of most and least important attributes. The benchmark of a MNL model can only be used with respect to the results generated by the relative BW scaling models, since in each of these models parameters can be estimated relative to the same reference attribute.

We used several measures to assess the performance of the models in terms of attribute importance ranking, goodness of fit and predictive accuracy. Attribute ranking was compared to control for reversals in attribute importances between the anchored and the relative BW scaling method. Stability in ranking is an indicator of comparative reliability of the different methodological variants of the BW procedure. In comparing the LC and HB models, the average percentage certainty measure (Hauser, 1978), obtained as the difference between the log likelihood of each model and the log likelihood of a chance model, was used to assess model fit. In addition, a chance ratio measure (average percentage certainty divided by the chance) was used to compare the predictive accuracy between relative B–W scaling and anchored B–W scaling. In relative B–W scaling, a chance model has a predictive power of 25% (one out of four choice options) whereas in anchored B–W scaling it has a predictive power of 20% (one out of five choice options).

Throughout the developing world, food quality and safety have become increasingly important attributes, driven by middle and high income consumers changing their lifestyles and their ability to pay more for food (Mergenthaler, Weinberger, & Qaim, 2009). Food produce safety, aesthetics, appearance and presentation and personal trust in vendor-consumer interactions have been identified as important issues in this change (Rheinländer et al., 2008). However, consumer concerns about the quality and safety of products such as vegetables have arisen owing to an increase in food-borne illnesses (Huda, Muzaffar, & Ahmed, 2009) and problems with heavy metal contamination and the presence of pesticide and fertiliser residues in food (Ikeda et al., 2000). Intake of fresh vegetables is being promoted in developing countries to improve public health, and an increasing proportion of middle and high income groups is consuming fresh produce (especially leafy vegetables) in the form of raw salads, or as part of prepared meals. The middle and high income consumers generally, but not exclusively, tend to purchase their fresh vegetables from supermarkets or specialist retail outlets. Meanwhile, many low income households depend on leafy vegetables, which they purchase in informal markets because they are generally less expensive.

In this study, kale (Brassica oleracea O.) from peri-urban farming for consumption within Nairobi was selected as the study object. Kale is the most important green leafy vegetable consumed by households in Nairobi, and peri-urban farming supplies the full range of market outlets from traditional to more recent speciality stores. Kale grown for domestic consumption does not undergo the same type of production certification as kale for export. Kale plays an important role in the nutritional balance in developing countries (World Health Organization, 2004). Analyses were carried out for four major food retail market categories in order to encompass potential variations in the importance of food quality attributes among consumers. A metropolitan area (Nairobi) was chosen as the study site, since such areas usually play a leading role in transformation of the food system of a country (Pingali, 2007).

Section snippets

Food quality attributes

Food quality was defined here in line with Grunert (2005) to include all the desirable properties a food product is perceived to possess according to the consumer. Food safety was considered to be part of food quality, because health-related and process-related quality attributes typically represent credence qualities (Grunert, 2005). Previous studies suggest that safety, sensory quality, price, environmental friendliness, convenience, hygiene and handling and nutrition can all affect the level

Sample descriptions

A series of Kruskal–Wallis tests of independent samples was performed to test for differences in socio-demographic classification between respondents per market outlet. Table 1 presents a description of the sample of kale consumers surveyed and variables for which equality across market types was rejected. The frequency of Kale consumption scale (Table 1) had the alternatives: daily (5), 2–3 times per week (4), at least once per week (2), seldom (1), or never (0). Product use frequency is

Conclusions

Consumer quality perception and evaluation of food is generally carried out using intrinsic and/or extrinsic quality cues regarding the produce. However, rising consumer concerns about safety, health, convenience, etc. have resulted in increasing focus on credence quality attributes relating to the quality of the production process. Greater multi-dimensionality of the food quality construct stresses the importance of adopting elicitation methods capable of detecting attribute non-attendance

Acknowledgements

We thank the editor and anonymous reviewers for insightful comments. Their efforts have significantly improved this manuscript.

References (53)

  • D.L. Paulhus

    Measurement and control of response bias

  • R.L. Andrews et al.

    An empirical comparison of logit choice models with discrete versus continuous representation of heterogeneity

    Journal of Marketing Research

    (2002)
  • L. Bacon et al.

    Comparing apples to oranges

    Marketing Research

    (2008)
  • K. Balcombe et al.

    A general treatment of ‘don’t know’ responses from choice experiments

    European Review of Agricultural Economics

    (2011)
  • P. Boxall et al.

    Understanding heterogeneous preferences in random utility models: A latent class approach

    Environmental and Resource Economics

    (2002)
  • Brunsø, K, Fjord, T.A., & Grunert, K.G. (2002). Consumers’ food choice and quality perception. MAPP working paper 77....
  • S. Cohen et al.

    What’s your preference?

    Marketing Research

    (2004)
  • Cohen, S. H., & Markowitz, P. (2002). Renewing market segmentation: Some new tools to correct old problems. ESOMAR 2002...
  • K. Duinevald et al.

    Hierarchical Bayesian analysis of true discrimination rates in replicated triangle tests

    Food Quality and Preference

    (2008)
  • FAO and WHO (2004). Human Vitamin and Mineral Requirements. Report of a joint FAO/WHO expert consultation Bangkok,...
  • E.P. Fenichel et al.

    Split-sample tests of ‘no opinion’ responses in an attribute based choice model

    Land Economics

    (2009)
  • A. Finn et al.

    Determining the appropriate response to evidence of public concern: The case of food safety

    Journal of Public Policy and Marketing

    (1992)
  • K.G. Grunert

    Food quality and safety: Consumer perception and demand

    European Review of Agricultural Economics

    (2005)
  • J.R. Hauser

    Testing and accuracy, usefulness and significance of probabilistic choice models: An information-theoretic approach

    Operations Research

    (1978)
  • Helsop, L. A. (2007). Literature review of Canadian consumer attitudes and perceptions. Ottawa, Canada. Consumer...
  • D. Hensher

    How do respondents process stated choice experiments? Attribute consideration under varying information load

    Journal of Applied Econometrics

    (2006)
  • Cited by (30)

    • Importance of food choice motives vary with degree of food neophobia

      2021, Appetite
      Citation Excerpt :

      It mitigates the risk often associated with rating scales that all items are quite similarly rated. Where concerns regarding relativity in FCQ factor importance exist, an anchored variant of BWS could be used (Lagerkvist, Okello, & Karanja, 2012). The expectation would be that differences seen in the present research would be magnified when measuring FCQ factor importance using BWS.

    • Dairy farmers' use and non-use values in animal welfare: Determining the empirical content and structure with anchored best-worst scaling

      2016, Journal of Dairy Science
      Citation Excerpt :

      Similarly, the alternative “none of these 5 are important to me” indicated that each of the 5 motivational attributes shown had a utility lower than the anchor. The presence of the utility anchor then allows for an absolute ranking of attributes across respondents, and the anchor defines the zero level on the common interval scale (Lagerkvist et al., 2012). The coding of responses of the anchored BWS choices was based on the complete coding approach outlined by Lattery (2011).

    • Solar cooking in Senegalese villages: An application of best-worst scaling

      2014, Energy Policy
      Citation Excerpt :

      Due to the heuristics, respondents do not always trade-off all attributes and it is important to know what influence this has on results. Unfortunately, this issue is under-examined (Lagerkvist et al., 2012). Finally, even though we emphasized the importance of PSPs in this study, they should not be seen as a means to an end.

    View all citing articles on Scopus
    View full text