Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Using eye tracking to account for attribute non-attendance in choice experiments

Using eye tracking to account for attribute non-attendance in choice experiments Abstract This study uses eye-tracking measures to account for attribute non-attendance (ANA) in choice experiments. Using the case of sustainability labelling on coffee, we demonstrate various approaches to account for ANA based on the fixation count cut-offs, definitions for detecting ignored attributes, and methods for modelling ANA. Some of the sustainability attributes identified through eye-tracking measures as being ‘visually ignored’ were truly ignored, whereas in none of the tested approaches was price truly ignored. The adequacy of eye tracking as a visual ANA measure might thus depend on the type of attribute. Further, the study unveiled inconsistencies in identifying non-attenders using visual ANA and the coefficient of variation. Based on our results, we cannot conclude that eye tracking always adequately identifies ANA. However, we identified several major challenges that can assist in further optimising the use of eye tracking in the context of ANA. 1. Introduction A growing body of literature applies choice experiments (CEs) as a valuation method. In a CE, respondents select their preferred alternative from a choice set1 in which each alternative is described by attributes with varying levels. The analysis is based on the economic theory of consumer behaviour (Lancaster, 1966; McFadden, 1974), which assumes continuous preferences and thus unlimited substitutability between the attributes (Ryan and Bate, 2001). This implies that all the attributes presented as well as the trade-offs between attributes are considered (Hensher, Rose and Greene, 2005). However, several studies have questioned the assumption of compensatory behaviour because respondents may ignore some attributes in a choice task (Hensher, Rose and Greene, 2005, 2012; Hole, 2011; Hensher, 2006; Lancsar and Louviere, 2006; Campbell, Hutchinson and Scarpa, 2008; 2011; Scarpa et al. 2009; 2010; Carlsson, Kataria and Lampi, 2010; Hensher and Greene, 2010; Kragt, 2013). Respondents may not make the assumed trade-offs between all the attributes due to attribute non-attendance (ANA), resulting in a violation of the continuity axiom. This decision heuristic has gained increased attention in the CE literature (Hensher, 2014). Not accounting for ANA has been found to affect coefficient estimates and model performance (Campbell, Hutchinson and Scarpa, 2008, 2011; Hensher and Rose, 2009; Scarpa et al. 2009; 2010; Carlsson, Kataria and Lampi, 2010; Mariel, Hoyos and Meyerhoff, 2013). Two primary methods have been proposed to identify ANA in CEs: (i) asking respondents which attributes they ignored (i.e. stated ANA) and (ii) inferring ANA based on observed choices (i.e. inferred ANA). Respondents can be asked whether an attribute was ignored when making a decision at the end of the entire choice task sequence (i.e. serial stated ANA) (Hensher, Rose and Greene, 2005; Carlsson, Kataria and Lampi, 2010; Alemu et al., 2013; Kehlbacher, Balcombe and Bennett, 2013; Kragt, 2013; Scarpa et al., 2013) or after each individual choice task (i.e. choice task stated ANA) (Puckett and Hensher, 2008, 2009; Meyerhoff and Liebe, 2009; Scarpa, Thiene and Hensher, 2010; Caputo et al., 2017). The disadvantage of stated ANA is that these measures are self-reported, which raises concerns about reliability (Hensher and Rose, 2009). For example, responses may be influenced by how the question is asked or interpreted. Respondents may not recall how they chose, or may not answer the attendance statement truthfully (Kragt, 2013; Scarpa et al. 2013), or may bias their answer in a socially desirable manner (Mørkbak, Olsen and Campbell, 2014). Additionally, serial stated ANA questions may be difficult to be answered because respondents may have applied different attribute processing strategies for each choice task (Puckett and Hensher, 2009; Hess and Hensher, 2010). While asking these questions at the end of each choice task allows the respondents to indicate different ANA behaviour for each choice task, it also informs them about the researcher’s interest in their attribute attendance, which may influence their attribute attendance in subsequent choice tasks. Another drawback is the additional financial cost in terms of survey time of repeatedly asking these supplementary questions as well as the increased risk of respondent fatigue, which can lead to more random decision-making (Campbell et al., 2015). The standard modelling approach in stated ANA studies is to restrict the coefficients in the utility function of attributes that have been stated as ignored to zero (Hensher, Rose and Greene, 2005). However, the stated ANA literature suggests that even though respondents stated that they ignored an attribute, it is possible that they may in fact have still attended to it, and it may (perhaps to a lesser extent) have influenced their choice. This would lead to coefficients for these ignored attributes that are in fact significantly different from zero (Campbell and Lorimer, 2009; Hess and Hensher, 2010; Alemu et al. 2013). Thus, assuming attributes to be ignored based on stated ANA may lead to biased results. Rather than relying on self-reported ANA information, a second method infers ANA behaviour using analytical models such as equality constrained latent class (ECLC) models, which impose specific restrictions on the utility functions for each class by constraining some coefficients to zero for selected attributes in a certain class (Scarpa et al. 2009; 2013; Hensher and Greene, 2010; Campbell, Hensher and Scarpa, 2011; Caputo, Nayga and Scarpa, 2013; Kragt, 2013; Lagarde, 2013). Yet another method of inferring ANA is based on the coefficient of variation of individual-specific posterior means (Hess and Hensher, 2010; Scarpa et al. 2013; Mørkbak, Olsen and Campbell, 2014). More research has been called for on other methods to account for ANA (Scarpa et al. 2013; Caputo et al. 2017). In this study, we propose a third method based on visual ANA that is defined as visually ignoring information about attribute levels (Balcombe, Fraser and McSorly, 2015). The use of eye tracking has been widely applied in the fields of marketing and psychology; however, it is relatively new in the field of economics. While some researchers, such as Scarpa et al. (2013), have suggested the use of eye tracking to obtain information on ANA in CEs, limited studies have done so (Balcombe, Fraser and McSorly, 2015; Spinks and Mortimer, 2016; Krucien, Ryan and Hermens, 2017). In the present study, visual attention is measured by eye-tracking equipment during the CE, and eye-fixation counts are used as one of the indicators of visual attention. Based on the fixation counts for a particular attribute, we created a discrete measure of visual attendance by indicating whether a respondent visually attended an attribute or not. Similarly, as in stated ANA studies, we apply the standard approach and develop models in which the coefficients in the utility function are restricted to zero for the attributes identified as visually ignored. Next, these models incorporating visual ANA are compared to a CE model in which full attendance is assumed. We then assess the performance of the visual attendance measure to identify the ignored attributes by testing whether the attributes identified as visually ignored influenced respondents’ choices. To do so, we estimate two coefficients for each attribute (attended and ignored) and test whether the ignored coefficient is different from zero, similar to what has been done in stated ANA studies (Campbell and Lorimer, 2009; Hess and Hensher, 2010; Balcombe, Burton and Rigby, 2011; Alemu et al. 2013). The visual ANA study of Balcombe, Fraser and McSorly (2015) made a few assumptions, such as requiring at least two fixations to consider an attribute visually ‘attended to’ and not allowing ANA behaviour to vary across choice sets by modelling serial ANA. We advance the investigation of visual ANA and extend the work of Balcombe, Fraser and McSorly (2015) in three directions in relation to (i) the fixation count cut-off, (ii) the definition of visual ANA detection in a choice task and (iii) the ANA modelling. Based on these three aspects, a total of six approaches or combinations are applied to incorporate visual ANA in choice models (see the overview in Table 3). Table 3. Overview of the six combinations based on the two modelling approaches to account for visual ANA, the fixation count used as the cut-off, and the two definitions for detecting whether a specific attribute was ignored during a choice task Fixation count FC 2 FC 1 ANA modelling approach Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A* Definition B** Definition A Definition B Abbreviation defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Fixation count FC 2 FC 1 ANA modelling approach Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A* Definition B** Definition A Definition B Abbreviation defA-S defB-S defA-CT defB-CT FC1-S FC1-CT *Based on choice set. **Based on alternatives. Table 3. Overview of the six combinations based on the two modelling approaches to account for visual ANA, the fixation count used as the cut-off, and the two definitions for detecting whether a specific attribute was ignored during a choice task Fixation count FC 2 FC 1 ANA modelling approach Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A* Definition B** Definition A Definition B Abbreviation defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Fixation count FC 2 FC 1 ANA modelling approach Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A* Definition B** Definition A Definition B Abbreviation defA-S defB-S defA-CT defB-CT FC1-S FC1-CT *Based on choice set. **Based on alternatives. 2. Material and methods 2.1. Sustainability labels on coffee Sustainability aspects of food are credence attributes and are thus unobservable to consumers unless explicitly labelled. However, consumers may be overwhelmed with information in a shopping environment and may not pay attention to all food labels (Grunert, 2011). We apply our study to the case of sustainability labelling on coffee, as coffee is one of the most popular sustainability-labelled food products. Many US coffee products carry sustainability labels such as Fair Trade (e.g. Fair Trade USA), Rainforest Alliance and USDA Organic, which are all included in our study (see also Van Loo et al., 2015). Coffee producers are often certified for more than one type of label. For example, in 2013, 77 per cent of the Fair Trade certified producer organisations reported holding at least one additional certification (52 per cent Organic and 12 per cent Rainforest Alliance) (Fairtrade International, 2015, p. 59). For coffee specifically, approximately 37 per cent of Fair Trade coffee is also organic certified (Fairtrade International, 2015). Due to the proliferation of sustainability labels for coffee, coffee packages often carry several of these labels (Pierrot, Giovannucci and Kasterine, 2011). 2.2. Experimental design of the CE Participants were recruited from a consumer profile database (N = 6,500) of the University of Arkansas Sensory Service Center (Fayetteville, AR, USA) which includes area residents. In total, 81 consumers who had purchased coffee in the 2 months preceding the study (March and April 2013) and who had no history of eye disease or eye surgery participated in the study. While the number of participants could be considered low for a CE, it is relatively high for an eye-tracking study. Each participant was given a $20 gift card. Approximately half (53 per cent) of the participants were female (Table 1). Each age and income category is represented. The sample is slightly biased towards participants with higher education. Table 1. Socio-demographic characteristics of the sample (%, n = 81) Gender  Male 46.9  Female 53.1 Age group  18–24 years 17.3  25–34 years 37.0  35–44 years 21.0  45–54 years 14.8  55 and older 9.9 Children  Yes 54.3  No 45.7 Educational level completed  High school/GED or lower 7.4  Some college or 2-year college degree 27.2  4-year college degree (BA, BS) 39.5  Master’s or PhD degree 25.9 Annual household income  Less than $20,000 25.9  $20,000–$49,999 35.8  $50,000–$99,999 25.9  More than $100,000 12.4 Gender  Male 46.9  Female 53.1 Age group  18–24 years 17.3  25–34 years 37.0  35–44 years 21.0  45–54 years 14.8  55 and older 9.9 Children  Yes 54.3  No 45.7 Educational level completed  High school/GED or lower 7.4  Some college or 2-year college degree 27.2  4-year college degree (BA, BS) 39.5  Master’s or PhD degree 25.9 Annual household income  Less than $20,000 25.9  $20,000–$49,999 35.8  $50,000–$99,999 25.9  More than $100,000 12.4 Table 1. Socio-demographic characteristics of the sample (%, n = 81) Gender  Male 46.9  Female 53.1 Age group  18–24 years 17.3  25–34 years 37.0  35–44 years 21.0  45–54 years 14.8  55 and older 9.9 Children  Yes 54.3  No 45.7 Educational level completed  High school/GED or lower 7.4  Some college or 2-year college degree 27.2  4-year college degree (BA, BS) 39.5  Master’s or PhD degree 25.9 Annual household income  Less than $20,000 25.9  $20,000–$49,999 35.8  $50,000–$99,999 25.9  More than $100,000 12.4 Gender  Male 46.9  Female 53.1 Age group  18–24 years 17.3  25–34 years 37.0  35–44 years 21.0  45–54 years 14.8  55 and older 9.9 Children  Yes 54.3  No 45.7 Educational level completed  High school/GED or lower 7.4  Some college or 2-year college degree 27.2  4-year college degree (BA, BS) 39.5  Master’s or PhD degree 25.9 Annual household income  Less than $20,000 25.9  $20,000–$49,999 35.8  $50,000–$99,999 25.9  More than $100,000 12.4 All coffee products in the experiment were ground medium roast coffee, the most popular type of coffee in the USA (Mintel, 2012). The coffee products were described using a combination of five attributes: four sustainability labels and price (Table 2). Each of the sustainability labels had two levels (present or not present) and the price attribute had four levels based on a store check in food stores in Fayetteville (AR) USA in April 2013. Table 2. Attributes and levels used in the CE Attributes Level Fair Trade label USDA Organic label 0 = Not present Rainforest Alliance label 1 = Present Carbon Footprint label Price (per 12 ounces) $4.30 $6.30 $8.30 $10.30 Attributes Level Fair Trade label USDA Organic label 0 = Not present Rainforest Alliance label 1 = Present Carbon Footprint label Price (per 12 ounces) $4.30 $6.30 $8.30 $10.30 Table 2. Attributes and levels used in the CE Attributes Level Fair Trade label USDA Organic label 0 = Not present Rainforest Alliance label 1 = Present Carbon Footprint label Price (per 12 ounces) $4.30 $6.30 $8.30 $10.30 Attributes Level Fair Trade label USDA Organic label 0 = Not present Rainforest Alliance label 1 = Present Carbon Footprint label Price (per 12 ounces) $4.30 $6.30 $8.30 $10.30 Given the four sustainability attributes and their levels, a full factorial design would have resulted in 64 (24 × 4) alternatives resulting in 4,096 choice sets–each consisting of two alternatives. The CE design followed Street and Burgess (2007). We used a fractional factorial design and the generator ([1 1 1 1 1]) to obtain the design of eight choice sets (Burgess, 2007; Street and Burgess, 2007) with an efficiency of 97.6 per cent. To increase the similarity to a real shopping experience, a no-buy alternative was added. Hence, in each choice set, participants were presented with two types of ground medium roasted coffee as well as a no-buy alternative (Figure 1). Due to the hypothetical nature of our CE, a cheap talk script was presented prior to the choice tasks. The location of the labels (USDA Organic, Rainforest Alliance, Fair Trade, Carbon Footprint) on the package (from left to right) was randomised for each of the eight choice sets to avoid an order effect due to label location. In addition, the randomisation for each of the choice sets was repeated 10 times, resulting in 10 different surveys. Each respondent was randomly assigned to one of these 10 surveys. Additionally, within each survey, the eight choice sets were randomly presented to avoid order effects. Fig. 1. View largeDownload slide Example of a choice set with AOI. Frames indicate the AOI for the Carbon Footprint label, Rainforest Alliance label, Fair Trade label and USDA Organic label and price. Fig. 1. View largeDownload slide Example of a choice set with AOI. Frames indicate the AOI for the Carbon Footprint label, Rainforest Alliance label, Fair Trade label and USDA Organic label and price. 2.3. Experimental procedure for the eye-tracking experiment Participants’ visual attention was recorded using a contact-free eye-tracking device (model: RED, SensoMotoric Instruments GmbH (SMI), Teltow, Germany) located in a panel beneath a 56 cm computer screen (474 mm by 297 mm and screen resolution of 1680 px by 1050 px) (see Appendix A). The approximate distance between the display monitor and each participant’s head was 70 cm. The sampling rate and tracking resolution of the eye-tracking device were 120 Hz and 0.03°, respectively. Visual stimuli, which are the images presented on the screen representing the choice set, were randomly presented using stimulus presentation software (Experiment Suite 360°TM, SensoMotoric Instruments GmbH, Teltow, Germany). The eye-tracking device was individually calibrated using a five-point calibration method with a low mean tracking error (less than 0.4°). After successful calibration, two warm-up choice sets were presented to familiarise each participant with the experimental procedures. As in Balcombe, Fraser and McSorly (2015), participants knew that eye tracking was used; however, they were not aware of its purpose. Between the choice tasks (i.e. during the inter-stimulus intervals), participants were asked to maintain their fixation on a central black cross against a white background for approximately 8 s. Similar to the procedure used in Balcombe, Fraser and McSorly (2015), the participants viewed each choice set as long as they wanted before indicating their choice. For three choice sets, the choice was not indicated, resulting in a total of 645 choices. On an average, the participants spent 73 s on all eight choice tasks combined (without the inter-stimulus intervals), which equates to an average of less than 10 s per choice task. The average tracking ratio, i.e. ‘the number of non-zero gaze positions divided by the sampling frequency multiplied by run duration, expressed in per cent’ (SMI, 2016, p. 335), across our sample is 89.5 per cent. 2.4. Eye-tracking measures Areas of interest (AOI) were defined on the coffee packages (Figure 1) corresponding to the five attributes. Using the eye-tracking software (BeGazeTM, ver. 3.0, SensoMotoric Instruments GmbH, Teltow, Germany), fixation counts were calculated for the five AOIs in each of the eight choice sets. The fixation count is the number of times the participant fixated his or her gaze on the AOI. More fixations are an indication that an area is more noticeable or more important to the viewer than other areas (Poole, Ball and Phillips, 2005). The number of fixations within the AOI has been considered a reliable measure for the visual attention given to that AOI (Bialkova and van Trijp, 2011). The low speed event detection method (suggested for <200 Hz) was selected in BeGaze for the fixation detection. In this method, the fixation is the primary event, and other events are derived from it. The method uses two specific detection parameters: a minimum fixation duration (i.e. the minimum time window in which the gaze is analysed) of 80 ms and a maximum dispersion of 100 px. The low-speed event detection method uses a dispersion-based algorithm. For details on this algorithm, we refer to the BeGaze Manual 3.6 (SMI, 2016, p. 317). For each stimulus, the first fixation was excluded. According to Holmqvist et al. (2011), this approach is often used because the fixation position at stimulus onset has not been influenced by the stimulus content. 2.5. Accounting for visual ANA In this study, three aspects are taken into account when going from fixation counts to incorporating visual ANA in the choice modelling: (i) fixation count cut-off, (ii) definition of visual ANA to identify an attribute as ignored in a choice task and (iii) modelling approach for visual ANA. 2.5.1. Fixation count cut-offs Balcombe, Fraser and McSorly (2015) assumed that at least two fixation counts are required to consider an attribute ‘attended to’. However, even one fixation count can signify that the person fixated on the information and thus may have attended to the information. Hence, in addition to the arbitrary assumption by Balcombe, Fraser and McSorly (2015) that two fixation counts are required to consider an attribute visually attended to, we also used a fixation count of one as a less strict threshold to consider an attribute as being visually attended to. 2.5.2. Defining visual ANA To define visual ANA, we used the fixation count as a measure of visual attention and created the discrete measure ‘visual ANA’.2 We used two definitions to identify an attribute as being ‘ignored’ in a particular choice task. Based on visual attention to a specific attribute in the choice set as a whole (Definition A). This definition was used by Balcombe, Fraser and McSorly (2015), who considered an attribute to be ignored in a choice task if the fixation count for the attribute summated over the alternatives within one choice set was below the cut-off. Thus, the fixation count for one attribute is calculated for the choice set as a whole. Based on visual attention to a specific attribute in each of the alternatives within the choice set (Definition B). An attribute is judged to be ignored in a given choice task if the attribute was ignored (fixation count less than the cut-off) in both of the designed alternatives (if the attribute was present in both alternatives). Thus, in this second definition, visual attention to the attribute in each alternative is considered. Notably, the use of Definition A or B will only have an impact on the price attribute, as it is present in both of the designed alternatives, and the alternatives never share any of the quality attributes. We assumed that respondents, when looking at one package, did not infer which labels the other package had. 2.5.3. Modelling approaches for visual ANA Two modelling approaches were used to account for visual ANA, one at the respondent level (serial visual ANA) and one at the choice set level (choice task visual ANA). Serial ANA refers to classifying a respondent as an attender or non-attender for a particular attribute for the entire choice task sequence, while choice task ANA allows for differences in attendance across choice tasks. Serial and choice task visual ANA are similar to serial and choice task stated ANA but instead of stated attendance, visual attendance is used. Serial visual ANA: Following Balcombe, Fraser and McSorly (2015), we classified a respondent as having visually ignored an attribute over the whole sequence of choice tasks if the participant ignored a given attribute in more than half (i.e. more than four) of the choice tasks. Thus, using serial visual ANA, a person is either a non-attender or an attender of an attribute for the entire sequence of choice tasks in the CE. Choice task visual ANA: Several authors, however, have advocated that respondents’ processing strategies may change as they progress through a sequence of choice tasks, meaning that their tendency to ignore attributes may not be consistent throughout a panel of choices (Meyerhoff and Liebe, 2009; Puckett and Hensher, 2009; Hess and Hensher, 2010; Scarpa, Thiene and Hensher, 2010). Hence, it may be important to allow for varying ANA behaviour from one choice task to another. Choice task visual ANA allows visual ANA to vary across choice tasks. When a respondent visually ignored a given attribute in a choice task, this attribute is characterised as non-attended for that particular choice task. Therefore, for each choice set and for each attribute, a participant is classified as having attended or not attended to the attribute. For each of the two visual ANA definitions to detect ignored attributes (Definition A and Definition B), both modelling approaches – serial visual ANA (S) and choice task visual ANA (CT) – were applied. This leads to four combinations – defA-CT, defA-S, defB-CT and defB-S (Table 3) – when a fixation count of two is applied as the cut-off. Definition A combined with the serial ANA modelling method with a fixation count of two as the cut-off (defA-S) is the approach used by Balcombe, Fraser and McSorly (2015). For the cut-off fixation count of one, definitions A and B result in the exact same discrete measure of visual ANA; thus, no distinction is made between definitions A and B (see Table 3). Therefore, a fixation count of one as the cut-off results in two additional combinations or approaches (FC1-S and FC1-CT). 2.6. Discrete choice models While the multinomial model (MNL) assumes homogeneity in consumer preferences, we assume that heterogeneity may be an issue in analysing consumer preferences for food labelling (Bonnet and Simioni, 2001; Van Loo et al., 2014). Therefore, a random parameter logit (RPL) model was estimated (with 500 Halton draws) that allows for random taste variation and for the panel structure, given that each respondent made eight choices. This approach results in the estimation of a mean and standard deviation for each of the random taste parameters. For simplicity, we assume price to be a fixed coefficient, which is a widely practised specification in the field (Revelt and Train, 1998; Layton and Brown, 2000; Morey and Rossmann, 2003; Lusk and Schroeder, 2004; Caputo, Nayga and Scarpa, 2013). This restriction allows the distribution of the WTP to be easily calculated from the non-price coefficients. We further assume that the coefficients of the four sustainability labels follow a normal distribution (Lusk and Schroeder, 2004; Caputo, Nayga and Scarpa, 2013). Two additional modelling issues are taken into account – the correlations across taste parameters and across utilities – to make the estimates more robust and consistent with consumer choice behaviour (Barreiro-Hurle, Gracia and de-Magistris, 2010; Gracia, Barreiro-Hurle and Perez, 2012, 2014). To allow for dependence across tastes, no restrictions were applied to the correlations among the random parameters. Additionally, because the design consists of two designed alternatives and one no-buy alternative, correlations across utilities may exist (Scarpa, Ferrini and Willis, 2005). The no-buy alternative is truly experienced, while the designed alternatives can only be imagined. Therefore, the utilities of the buying alternatives are likely to be more correlated among themselves than with the no-buy alternative. To account for this correlation pattern, we employed an RPL model with an error component (RPL-EC) (Scarpa, Ferrini and Willis, 2005; 2007), whereby both designed alternatives share an extra error component that is a zero-mean normally distributed random parameter. Specifically, with our attributes, the utility that individual i obtains from alternative j at choice situation t takes the following form: Uijt=β0No_Buyijt+β1Organicijt+β2Rainforestijt+β3FairTradeijt+β4CarbonFootprintijt+β5Priceijt+ηij(1−No_Buyijt)+εijt, where j pertains to alternatives A, B and C. No_Buyijt is an indicator variable that takes the value of 1 when the no-buy alternative is chosen and 0 when either product profile A or B is selected. β0 is an alternative-specific constant representing the no-buy alternative. Price is the price (US$) of a package of 12 ounces of coffee. ηij is the zero-mean normal error term, or the error component term, which is only associated with alternatives that portray a purchase decision and is absent in the utility of the no-buy alternative. εijt is the unobserved random error term. The marginal WTP values are calculated as a negative ratio, where the numerator is the estimated mean values of the coefficients associated with a particular sustainability label and the denominator is the price coefficient. The data were analysed using NLOGIT 5.0 by Econometric Software Inc. (Greene, 1990). 2.7. Accounting for ANA The standard approach to account for stated ANA is to restrict the coefficient in the utility function to zero for the attributes that the respondents stated as ignored, which results in the removal of the attribute from the choice consideration (Hensher, Rose and Greene, 2005). This method has been incorporated into the NLOGIT 5.0 software by coding an attribute as −888 if it is not attended to and assigns a zero to the attribute coefficients rather than to the attribute levels (Greene, 2012). This approach has been applied in several studies on stated ANA (Hensher, Rose and Greene, 2005, 2009; Alemu et al. 2013; Kragt, 2013; Scarpa et al. 2013). We used the same approach, using a dummy variable to denote whether the attribute was visually ignored (visual ANA). This is defined at the choice set level (choice task ANA) or at the respondent level (serial ANA). We did not collect any data on stated ANA. For each of the six combinations (defA-CT, defA-S, defB-CT, defB-S, FC1-S, FC1-CT), we estimated a visual ANA model in which the coefficients of the visually ignored attributes are restricted to zero. In addition, a full attendance model in which all attributes are assumed to be attended to was estimated. 2.8. Are the attributes identified as visually non-attended actually ignored? We examine whether the attributes identified as visually non-attended are in fact also truly ignored3 when respondents are making the choice by estimating the coefficient of ignored attributes (Section 2.8.1). In addition, we compare the results of serial visual ANA with results from the coefficient of variation method (Section 2.8.2), an inferred method to identify non-attenders. 2.8.1. Coefficients of ignored attributes The coefficients of the ignored attributes are no longer restricted to zero; but are freely estimated. In the stated ANA literature, some studies have indicated that people reporting to have ignored a certain attribute may have a marginal utility for that attribute that differs from zero. Hence, respondents who stated that they ignored an attribute may have actually considered it (Carlsson, Kataria and Lampi, 2010). As a result, instead of restricting the coefficient of ignored attributes to zero, some stated ANA studies estimate two coefficients for each attribute: one for those who stated that they attended to the attribute and one for those who stated that they did not attend to the attribute (Campbell and Lorimer, 2009; Hess and Hensher, 2010; Alemu et al. 2013; Scarpa et al. 2013). Similarly, we estimate models with two coefficients for each attribute: one for visually ignored attributes and one for visually attended attributes. If the visually non-attended attributes were truly ignored, the corresponding coefficient should not differ statistically from zero. Because of the number of coefficients to be estimated and our limited sample size, MNL models were estimated. 2.8.2. Coefficient of variation of individual-specific coefficient distributions Using the serial visual ANA, a respondent is identified as an attender or a non-attender for a particular attribute. For each respondent, the allocation of having ignored an attribute or not based on serial visual ANA is compared with the allocation of this respondent as an attender or non-attender based on the inferred method using the coefficient of variation.4 Following Hess and Hensher (2010), we attempted to infer whether a respondent ignored a particular attribute or not (thus, inferred serial ANA) by analysing the individual-specific coefficient distributions that have been conditioned on observed choices. For additional details, we refer readers to the NLOGIT reference guide (Greene, 2012, section N29.8). Based on the RPL model in Table 5, the means and standard deviations for the conditional distributions were calculated for each coefficient of the random parameters. Rather than using the conditional mean to infer whether a respondent ignored an attribute or not, Hess and Hensher (2010) suggested using the coefficient of variation as an inferred method to identify ignored attributes. The coefficient of variation is the ratio between the standard deviation and the mean of the conditional distribution. This measure is used by Hess and Hensher (2010, p. 786) to incorporate uncertainty into the conditional distributions and gives an indication of whether the conditional mean is indistinguishable from zero. Hess and Hensher (2010) reported that this approach is better than using the conditional mean because a respondent may have a low sensitivity to an attribute without actually ignoring it. Hence, relying only on a low mean to allocate a respondent to the ignored group might be incorrect, and therefore using the coefficient of variation is suggested. Following Hess and Hensher (2010), we allocate respondents with a coefficient of variation of two or above for a certain attribute to the ignored group for that attribute. Subsequently, we evaluate whether the identification of respondents as having ignored or not ignored an attribute based on the serial visual ANA matches with the identification based on the coefficient of variation. Table 5. RPL model with error component (RPL-EC) parameter estimates (n = 645) Fixation count FC 2 FC 1 ANA modelling Full-AA Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Fair Trade Mean 0.63** (0.29) 0.73** (0.36) 0.75** (0.33) 0.76*** (0.27) 0.81*** (0.22) 0.65** (0.31) 0.65** (0.29) Standard deviation 0.83*** (0.28) 0.79** (0.32) 0.73** (0.37) 0.65 (0.51) 0.54 (0.47) 0.58* (0.29) 0.40 (0.30) USDA Organic Mean 1.02*** (0.35) 1.41*** (0.50) 1.35*** (0.47) 0.87*** (0.25) 0.84*** (0.24) 1.09*** (0.40) 0.77*** (0.26) Standard deviation 1.23*** (0.33) 0.85* (0.52) 0.89 (1.05) 0.39 (0.49) 0.18 (1.14) 1.13** (0.50) 0.84* (0.47) Rainforest Alliance Mean 0.74*** (0.26) 0.73*** (0.26) 0.68*** (0.23) 0.60*** (0.23) 0.60*** (0.21) 0.75*** (0.27) 0.59*** (0.20) Standard deviation 0.58 (0.37) 0.45 (0.41) 0.33 (0.44) 0.42 (0.73) 0.15 (1.60) 0.55 (0.52) 0.36 (0.52) Carbon Footprint Mean 0.30 (0.26) 0.38 (0.36) 0.40 (0.32) 0.50** (0.23) 0.58*** (0.22) 0.27 (0.31) 0.34 (0.22) Standard deviation 0.90* (0.48) 0.72 (0.74) 0.71 (0.98) 0.39 (0.77) 0.39 (1.27) 0.62 (0.50) 0.39 (1.03) Price −0.85*** (0.05) −0.81*** (0.05) −0.84*** (0.05) −0.61*** (0.05) −0.54*** (0.06) −0.77*** (0.05) −0.68*** (0.04) No_Buy −8.76*** (0.86) −8.91*** (0.85) −9.13*** (0.98) −7.35*** (0.90) −6.75*** (1.01) −8.01*** (0.81) −7.61*** (0.81) Error component Standard deviation 2.71*** (0.91) 3.15*** (1.15) 3.63*** (0.85) 3.03*** (0.96) 3.51*** (0.93) 2.55*** (0.96) 2.63*** (0.90) Log likelihood −349 −354 −370 −404 −433 −357 −375  AIC 731.2 741.2 774.1 842.4 900.6 747.9 783.6  BIC 807.2 817.2 850.0 918.4 976.6 823.9 859.6 Fixation count FC 2 FC 1 ANA modelling Full-AA Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Fair Trade Mean 0.63** (0.29) 0.73** (0.36) 0.75** (0.33) 0.76*** (0.27) 0.81*** (0.22) 0.65** (0.31) 0.65** (0.29) Standard deviation 0.83*** (0.28) 0.79** (0.32) 0.73** (0.37) 0.65 (0.51) 0.54 (0.47) 0.58* (0.29) 0.40 (0.30) USDA Organic Mean 1.02*** (0.35) 1.41*** (0.50) 1.35*** (0.47) 0.87*** (0.25) 0.84*** (0.24) 1.09*** (0.40) 0.77*** (0.26) Standard deviation 1.23*** (0.33) 0.85* (0.52) 0.89 (1.05) 0.39 (0.49) 0.18 (1.14) 1.13** (0.50) 0.84* (0.47) Rainforest Alliance Mean 0.74*** (0.26) 0.73*** (0.26) 0.68*** (0.23) 0.60*** (0.23) 0.60*** (0.21) 0.75*** (0.27) 0.59*** (0.20) Standard deviation 0.58 (0.37) 0.45 (0.41) 0.33 (0.44) 0.42 (0.73) 0.15 (1.60) 0.55 (0.52) 0.36 (0.52) Carbon Footprint Mean 0.30 (0.26) 0.38 (0.36) 0.40 (0.32) 0.50** (0.23) 0.58*** (0.22) 0.27 (0.31) 0.34 (0.22) Standard deviation 0.90* (0.48) 0.72 (0.74) 0.71 (0.98) 0.39 (0.77) 0.39 (1.27) 0.62 (0.50) 0.39 (1.03) Price −0.85*** (0.05) −0.81*** (0.05) −0.84*** (0.05) −0.61*** (0.05) −0.54*** (0.06) −0.77*** (0.05) −0.68*** (0.04) No_Buy −8.76*** (0.86) −8.91*** (0.85) −9.13*** (0.98) −7.35*** (0.90) −6.75*** (1.01) −8.01*** (0.81) −7.61*** (0.81) Error component Standard deviation 2.71*** (0.91) 3.15*** (1.15) 3.63*** (0.85) 3.03*** (0.96) 3.51*** (0.93) 2.55*** (0.96) 2.63*** (0.90) Log likelihood −349 −354 −370 −404 −433 −357 −375  AIC 731.2 741.2 774.1 842.4 900.6 747.9 783.6  BIC 807.2 817.2 850.0 918.4 976.6 823.9 859.6 Note: *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. Table 5. RPL model with error component (RPL-EC) parameter estimates (n = 645) Fixation count FC 2 FC 1 ANA modelling Full-AA Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Fair Trade Mean 0.63** (0.29) 0.73** (0.36) 0.75** (0.33) 0.76*** (0.27) 0.81*** (0.22) 0.65** (0.31) 0.65** (0.29) Standard deviation 0.83*** (0.28) 0.79** (0.32) 0.73** (0.37) 0.65 (0.51) 0.54 (0.47) 0.58* (0.29) 0.40 (0.30) USDA Organic Mean 1.02*** (0.35) 1.41*** (0.50) 1.35*** (0.47) 0.87*** (0.25) 0.84*** (0.24) 1.09*** (0.40) 0.77*** (0.26) Standard deviation 1.23*** (0.33) 0.85* (0.52) 0.89 (1.05) 0.39 (0.49) 0.18 (1.14) 1.13** (0.50) 0.84* (0.47) Rainforest Alliance Mean 0.74*** (0.26) 0.73*** (0.26) 0.68*** (0.23) 0.60*** (0.23) 0.60*** (0.21) 0.75*** (0.27) 0.59*** (0.20) Standard deviation 0.58 (0.37) 0.45 (0.41) 0.33 (0.44) 0.42 (0.73) 0.15 (1.60) 0.55 (0.52) 0.36 (0.52) Carbon Footprint Mean 0.30 (0.26) 0.38 (0.36) 0.40 (0.32) 0.50** (0.23) 0.58*** (0.22) 0.27 (0.31) 0.34 (0.22) Standard deviation 0.90* (0.48) 0.72 (0.74) 0.71 (0.98) 0.39 (0.77) 0.39 (1.27) 0.62 (0.50) 0.39 (1.03) Price −0.85*** (0.05) −0.81*** (0.05) −0.84*** (0.05) −0.61*** (0.05) −0.54*** (0.06) −0.77*** (0.05) −0.68*** (0.04) No_Buy −8.76*** (0.86) −8.91*** (0.85) −9.13*** (0.98) −7.35*** (0.90) −6.75*** (1.01) −8.01*** (0.81) −7.61*** (0.81) Error component Standard deviation 2.71*** (0.91) 3.15*** (1.15) 3.63*** (0.85) 3.03*** (0.96) 3.51*** (0.93) 2.55*** (0.96) 2.63*** (0.90) Log likelihood −349 −354 −370 −404 −433 −357 −375  AIC 731.2 741.2 774.1 842.4 900.6 747.9 783.6  BIC 807.2 817.2 850.0 918.4 976.6 823.9 859.6 Fixation count FC 2 FC 1 ANA modelling Full-AA Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Fair Trade Mean 0.63** (0.29) 0.73** (0.36) 0.75** (0.33) 0.76*** (0.27) 0.81*** (0.22) 0.65** (0.31) 0.65** (0.29) Standard deviation 0.83*** (0.28) 0.79** (0.32) 0.73** (0.37) 0.65 (0.51) 0.54 (0.47) 0.58* (0.29) 0.40 (0.30) USDA Organic Mean 1.02*** (0.35) 1.41*** (0.50) 1.35*** (0.47) 0.87*** (0.25) 0.84*** (0.24) 1.09*** (0.40) 0.77*** (0.26) Standard deviation 1.23*** (0.33) 0.85* (0.52) 0.89 (1.05) 0.39 (0.49) 0.18 (1.14) 1.13** (0.50) 0.84* (0.47) Rainforest Alliance Mean 0.74*** (0.26) 0.73*** (0.26) 0.68*** (0.23) 0.60*** (0.23) 0.60*** (0.21) 0.75*** (0.27) 0.59*** (0.20) Standard deviation 0.58 (0.37) 0.45 (0.41) 0.33 (0.44) 0.42 (0.73) 0.15 (1.60) 0.55 (0.52) 0.36 (0.52) Carbon Footprint Mean 0.30 (0.26) 0.38 (0.36) 0.40 (0.32) 0.50** (0.23) 0.58*** (0.22) 0.27 (0.31) 0.34 (0.22) Standard deviation 0.90* (0.48) 0.72 (0.74) 0.71 (0.98) 0.39 (0.77) 0.39 (1.27) 0.62 (0.50) 0.39 (1.03) Price −0.85*** (0.05) −0.81*** (0.05) −0.84*** (0.05) −0.61*** (0.05) −0.54*** (0.06) −0.77*** (0.05) −0.68*** (0.04) No_Buy −8.76*** (0.86) −8.91*** (0.85) −9.13*** (0.98) −7.35*** (0.90) −6.75*** (1.01) −8.01*** (0.81) −7.61*** (0.81) Error component Standard deviation 2.71*** (0.91) 3.15*** (1.15) 3.63*** (0.85) 3.03*** (0.96) 3.51*** (0.93) 2.55*** (0.96) 2.63*** (0.90) Log likelihood −349 −354 −370 −404 −433 −357 −375  AIC 731.2 741.2 774.1 842.4 900.6 747.9 783.6  BIC 807.2 817.2 850.0 918.4 976.6 823.9 859.6 Note: *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. 3. Results and discussion 3.1. Visual attribute ANA frequency In our specifications of visual ANA (Definition A or B), the modelling approach for visual ANA (serial or choice task) as well as the fixation count cut-off determine the ANA frequency. The proportions of ANA for each of the attributes and for the six different combinations are presented in Table 4. Table 4. Proportions (%) of choice task and serial visual ANA depending on the definition applied and fixation count (FC) (n = 645) FC 2 FC 1 Serial ANA Choice task ANA Serial ANA Choice task ANA Definition A Definition B Definition A Definition B Fair Trade 40.7 40.7 52.2 52.2 12.3 28.1 USDA Organic 55.6 55.6 59.6 59.6 23.5 30.4 Rainforest Alliance 43.2 43.2 54.9 54.9 17.3 30.1 Carbon Footprint 40.7 40.7 52.9 52.9 16.0 28.1 Price 12.3 23.4 24.5 40.7 4.9 9.6 FC 2 FC 1 Serial ANA Choice task ANA Serial ANA Choice task ANA Definition A Definition B Definition A Definition B Fair Trade 40.7 40.7 52.2 52.2 12.3 28.1 USDA Organic 55.6 55.6 59.6 59.6 23.5 30.4 Rainforest Alliance 43.2 43.2 54.9 54.9 17.3 30.1 Carbon Footprint 40.7 40.7 52.9 52.9 16.0 28.1 Price 12.3 23.4 24.5 40.7 4.9 9.6 Table 4. Proportions (%) of choice task and serial visual ANA depending on the definition applied and fixation count (FC) (n = 645) FC 2 FC 1 Serial ANA Choice task ANA Serial ANA Choice task ANA Definition A Definition B Definition A Definition B Fair Trade 40.7 40.7 52.2 52.2 12.3 28.1 USDA Organic 55.6 55.6 59.6 59.6 23.5 30.4 Rainforest Alliance 43.2 43.2 54.9 54.9 17.3 30.1 Carbon Footprint 40.7 40.7 52.9 52.9 16.0 28.1 Price 12.3 23.4 24.5 40.7 4.9 9.6 FC 2 FC 1 Serial ANA Choice task ANA Serial ANA Choice task ANA Definition A Definition B Definition A Definition B Fair Trade 40.7 40.7 52.2 52.2 12.3 28.1 USDA Organic 55.6 55.6 59.6 59.6 23.5 30.4 Rainforest Alliance 43.2 43.2 54.9 54.9 17.3 30.1 Carbon Footprint 40.7 40.7 52.9 52.9 16.0 28.1 Price 12.3 23.4 24.5 40.7 4.9 9.6 For serial visual ANA, the proportions of visual non-attenders for the sustainability labels range from 41 per cent to 56 per cent of the total number of participants for a fixation count of 2, while the proportions range from 12 per cent to 23 per cent for a fixation count of one. For both fixation counts, the number of respondents who ignore the price is lower than for the sustainability labels. Using a fixation count of 2 as the cut-off, only 12 per cent and 23 per cent of the respondents were classified as visual non-attenders for price for Definition A and Definition B, respectively.5 For a fixation count of one, only 5 per cent of the respondents were classified as ignoring the price. When applying choice task visual ANA, the number of choice tasks in which the sustainability labels were ignored ranges from 52 per cent to 60 per cent for a fixation count of two as the cut-off and from 28 per cent to 30 per cent for a fixation count of one as the cut-off. Price was ignored in 25 per cent or 41 per cent of the choice tasks depending on the ANA definition applied for the case of a fixation count of two as the cut-off and in 10 per cent of the choice tasks for a fixation count of one. 3.2. Standard ANA approach Similar to the standard approach in stated ANA, models are estimated with the parameters for the visually ignored attributes being constrained to zero; i.e. four models for the combinations with a fixation count of two: serial and choice task visual ANA based on Definition A (defA-CT, defA-S) and based on Definition B (defB-CT and defB-S) (Table 5) and two additional models for a fixation count of one: choice task and serial visual ANA modelling approaches (FC1-S and FC1-CT). The full attendance model (full-AA) pertains to the estimation assuming full attribute attendance and is included as a benchmark. To allow for heterogeneous preferences among the respondents and correlation across utilities, RPL-EC models were estimated (Table 5). The MNL model estimations are included in Appendix B. In all the models, the coefficient of the no-buy alternative is negative and statistically significant, indicating that participants increase their utility when choosing one of the proposed coffee alternatives compared with the no-buy alternative. In all of the models, the hypothesis of correlation across utilities is verified because the standard deviation of the error component (ηij) for the purchase alternatives is statistically significant. Correlations across the random parameters were also allowed. The Cholesky matrices are presented in Appendix C. The coefficients of the attributes have the expected signs. The price coefficient is negative, as expected, and statistically significant at the 0.01 level, indicating that consumers’ utility decreases with increasing price. In the full-AA model, the coefficients of Organic, Rainforest Alliance and Fair Trade are significant, implying that respondents’ utility increases when one of these labels is present on a coffee package. The results show that USDA Organic is the highest valued attribute, resulting in the strongest utility increase. The USDA Organic label is preferred over Rainforest Alliance and Fair Trade. The full-AA model has significant standard deviations of the random parameters (except for Rainforest Alliance), indicating the presence of considerable unobserved heterogeneity in taste preferences across the respondents. Turning to the standard ANA approach in which the parameters of the visually ignored attributes are restricted to zero (defA-CT, defA-S, defB-CT, defB-S, FC1-S and FC1-CT), we find that most of the parameters for the attended attributes are significant at the 5 per cent or 1 per cent level. In all six models, the coefficient of USDA Organic is the largest. Carbon Footprint is not significant for all models except for the choice task modelling approach with a fixation count of two as the cut-off. While the standard deviations of the random parameters for Fair Trade and USDA Organic of the full attendance model were significant at 1 per cent, this is no longer the case when accounting for visual ANA.6 While the full attendance model with significant standard deviations shows preference heterogeneity, accounting for visual ANA captures an important part of the heterogeneity across participants. This result illustrates that confounding between ANA and preference heterogeneity might be an issue (Hess et al., 2013), and thus preference heterogeneity may be incorrectly interpreted when ANA is not addressed, which further illustrates its importance (Hess et al. 2013). The full attendance model outperforms the model in which the coefficients of the ignored attributes are restricted to zero, as illustrated by the decrease in model fit for the visual ANA models (a decrease in log likelihood and an increase in the AIC and BIC statistics) (Table 5). A worse model fit could be due to treating attributes as having a zero parameter when some are not actually ignored. 3.3. Are the attributes identified as visually non-attended actually ignored? 3.3.1. Coefficients of ignored attributes We test whether the attributes identified as visually non-attended truly have coefficients that are equal to zero by estimating them freely, which leads to model estimations with two coefficients for each attribute (attended and ignored) that are referred to as defA-S2, defB-S2, defA-CT2, and defB-CT2 as well as FC1-S2 and FC1-CT2 for fixation counts two and one as the cut-off, respectively (Table 6). Of the 30 ignored coefficients,7 17 are not significantly different from zero. In the cases of the Rainforest Alliance, Fair Trade and Carbon Footprint labels, being identified as visually non-attended using one of the six approaches means that these attributes were truly ignored, with the exception of Rainforest Alliance in the choice task ANA modelling approach with a fixation count of two. Table 6. MNL parameter estimations with two coefficients (attended and ignored) (n = 645) Fixation count FC 2 FC 1 ANA modelling Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S2 defB-S2 defA-CT2 defB-CT2 FC1-S2 FC1-CT2 Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Attended  Fair Trade 0.54*** (0.16) 0.55*** (0.17) 0.72*** (0.18) 0.72*** (0.18) 0.43*** (0.14) 0.46*** (0.15)  USDA Organic 1.46*** (0.20) 1.45*** (0.20) 1.11*** (0.20) 1.12*** (0.20) 0.99*** (0.16) 0.94*** (0.16)  Rainforest Alliance 0.86*** (0.18) 0.87*** (0.18) 0.93*** (0.19) 0.93*** (0.19) 0.72*** (0.15) 0.73*** (0.16)  Carbon Footprint 0.16 (0.17) 0.18 (0.17) 0.37** (0.18) 0.38** (0.18) 0.06 (0.14) 0.17 (0.15)  Price −0.65*** (0.05) −0.65*** (0.05) −0.62*** (0.05) −0.63*** (0.05) −0.61*** (0.05) −0.62*** (0.05) Ignored  Fair Trade 0.10 (0.18) 0.08 (0.18) 0.03 (0.17) 0.03 (0.17) −0.41 (0.33) −0.03 (0.22)  USDA Organic 0.39** (0.17) 0.40** (0.17) 0.81*** (0.17) 0.80*** (0.17) 0.35 (0.24) 0.69*** (0.22)  Rainforest Alliance 0.30 (0.19) 0.30 (0.19) 0.38** (0.17) 0.38** (0.17) 0.08 (0.28) 0.34 (0.22)  Carbon Footprint −0.01 (0.19) −0.04 (0.19) −0.13 (0.17) −0.14 (0.17) 0.11 (0.28) −0.17 (0.22)  Price −0.40*** (0.08) −0.52*** (0.06) −0.51*** (0.06) −0.54*** (0.05) −0.46*** (0.12) −0.43*** (0.08)  No_Buy −5.47*** (0.40) −5.38*** (0.40) −5.09*** (0.39) −5.06*** (0.39) −5.28*** (0.39) −5.18*** (0.39)  Log likelihooda −376 −379 −386 −386 −388 −393  AIC 774.1 779.8 793.1 794.5 797.1 807.0  BIC 823.2 829.0 842.3 843.7 846.2 856.2 p-values for testing statistical differences between the coefficients attended versus ignored  Fair Trade 0.024** 0.016** 0.001*** 0.001*** 0.006*** 0.018**  USDA Organic <0.001*** <0.001*** 0.081* 0.074* 0.006*** 0.144  Rainforest Alliance 0.007*** 0.005*** 0.007*** 0.007*** 0.016** 0.051*  Carbon Footprint 0.214 0.165 0.013** 0.010*** 0.430 0.081*  Price 0.001*** 0.005*** 0.008*** 0.013** 0.088* 0.008*** Fixation count FC 2 FC 1 ANA modelling Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S2 defB-S2 defA-CT2 defB-CT2 FC1-S2 FC1-CT2 Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Attended  Fair Trade 0.54*** (0.16) 0.55*** (0.17) 0.72*** (0.18) 0.72*** (0.18) 0.43*** (0.14) 0.46*** (0.15)  USDA Organic 1.46*** (0.20) 1.45*** (0.20) 1.11*** (0.20) 1.12*** (0.20) 0.99*** (0.16) 0.94*** (0.16)  Rainforest Alliance 0.86*** (0.18) 0.87*** (0.18) 0.93*** (0.19) 0.93*** (0.19) 0.72*** (0.15) 0.73*** (0.16)  Carbon Footprint 0.16 (0.17) 0.18 (0.17) 0.37** (0.18) 0.38** (0.18) 0.06 (0.14) 0.17 (0.15)  Price −0.65*** (0.05) −0.65*** (0.05) −0.62*** (0.05) −0.63*** (0.05) −0.61*** (0.05) −0.62*** (0.05) Ignored  Fair Trade 0.10 (0.18) 0.08 (0.18) 0.03 (0.17) 0.03 (0.17) −0.41 (0.33) −0.03 (0.22)  USDA Organic 0.39** (0.17) 0.40** (0.17) 0.81*** (0.17) 0.80*** (0.17) 0.35 (0.24) 0.69*** (0.22)  Rainforest Alliance 0.30 (0.19) 0.30 (0.19) 0.38** (0.17) 0.38** (0.17) 0.08 (0.28) 0.34 (0.22)  Carbon Footprint −0.01 (0.19) −0.04 (0.19) −0.13 (0.17) −0.14 (0.17) 0.11 (0.28) −0.17 (0.22)  Price −0.40*** (0.08) −0.52*** (0.06) −0.51*** (0.06) −0.54*** (0.05) −0.46*** (0.12) −0.43*** (0.08)  No_Buy −5.47*** (0.40) −5.38*** (0.40) −5.09*** (0.39) −5.06*** (0.39) −5.28*** (0.39) −5.18*** (0.39)  Log likelihooda −376 −379 −386 −386 −388 −393  AIC 774.1 779.8 793.1 794.5 797.1 807.0  BIC 823.2 829.0 842.3 843.7 846.2 856.2 p-values for testing statistical differences between the coefficients attended versus ignored  Fair Trade 0.024** 0.016** 0.001*** 0.001*** 0.006*** 0.018**  USDA Organic <0.001*** <0.001*** 0.081* 0.074* 0.006*** 0.144  Rainforest Alliance 0.007*** 0.005*** 0.007*** 0.007*** 0.016** 0.051*  Carbon Footprint 0.214 0.165 0.013** 0.010*** 0.430 0.081*  Price 0.001*** 0.005*** 0.008*** 0.013** 0.088* 0.008*** Note: *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. aFor the Full-AA MNL model: log likelihood = −401.4, AIC = 814.8, BIC = 841.7 (see Appendix B). Table 6. MNL parameter estimations with two coefficients (attended and ignored) (n = 645) Fixation count FC 2 FC 1 ANA modelling Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S2 defB-S2 defA-CT2 defB-CT2 FC1-S2 FC1-CT2 Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Attended  Fair Trade 0.54*** (0.16) 0.55*** (0.17) 0.72*** (0.18) 0.72*** (0.18) 0.43*** (0.14) 0.46*** (0.15)  USDA Organic 1.46*** (0.20) 1.45*** (0.20) 1.11*** (0.20) 1.12*** (0.20) 0.99*** (0.16) 0.94*** (0.16)  Rainforest Alliance 0.86*** (0.18) 0.87*** (0.18) 0.93*** (0.19) 0.93*** (0.19) 0.72*** (0.15) 0.73*** (0.16)  Carbon Footprint 0.16 (0.17) 0.18 (0.17) 0.37** (0.18) 0.38** (0.18) 0.06 (0.14) 0.17 (0.15)  Price −0.65*** (0.05) −0.65*** (0.05) −0.62*** (0.05) −0.63*** (0.05) −0.61*** (0.05) −0.62*** (0.05) Ignored  Fair Trade 0.10 (0.18) 0.08 (0.18) 0.03 (0.17) 0.03 (0.17) −0.41 (0.33) −0.03 (0.22)  USDA Organic 0.39** (0.17) 0.40** (0.17) 0.81*** (0.17) 0.80*** (0.17) 0.35 (0.24) 0.69*** (0.22)  Rainforest Alliance 0.30 (0.19) 0.30 (0.19) 0.38** (0.17) 0.38** (0.17) 0.08 (0.28) 0.34 (0.22)  Carbon Footprint −0.01 (0.19) −0.04 (0.19) −0.13 (0.17) −0.14 (0.17) 0.11 (0.28) −0.17 (0.22)  Price −0.40*** (0.08) −0.52*** (0.06) −0.51*** (0.06) −0.54*** (0.05) −0.46*** (0.12) −0.43*** (0.08)  No_Buy −5.47*** (0.40) −5.38*** (0.40) −5.09*** (0.39) −5.06*** (0.39) −5.28*** (0.39) −5.18*** (0.39)  Log likelihooda −376 −379 −386 −386 −388 −393  AIC 774.1 779.8 793.1 794.5 797.1 807.0  BIC 823.2 829.0 842.3 843.7 846.2 856.2 p-values for testing statistical differences between the coefficients attended versus ignored  Fair Trade 0.024** 0.016** 0.001*** 0.001*** 0.006*** 0.018**  USDA Organic <0.001*** <0.001*** 0.081* 0.074* 0.006*** 0.144  Rainforest Alliance 0.007*** 0.005*** 0.007*** 0.007*** 0.016** 0.051*  Carbon Footprint 0.214 0.165 0.013** 0.010*** 0.430 0.081*  Price 0.001*** 0.005*** 0.008*** 0.013** 0.088* 0.008*** Fixation count FC 2 FC 1 ANA modelling Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S2 defB-S2 defA-CT2 defB-CT2 FC1-S2 FC1-CT2 Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Attended  Fair Trade 0.54*** (0.16) 0.55*** (0.17) 0.72*** (0.18) 0.72*** (0.18) 0.43*** (0.14) 0.46*** (0.15)  USDA Organic 1.46*** (0.20) 1.45*** (0.20) 1.11*** (0.20) 1.12*** (0.20) 0.99*** (0.16) 0.94*** (0.16)  Rainforest Alliance 0.86*** (0.18) 0.87*** (0.18) 0.93*** (0.19) 0.93*** (0.19) 0.72*** (0.15) 0.73*** (0.16)  Carbon Footprint 0.16 (0.17) 0.18 (0.17) 0.37** (0.18) 0.38** (0.18) 0.06 (0.14) 0.17 (0.15)  Price −0.65*** (0.05) −0.65*** (0.05) −0.62*** (0.05) −0.63*** (0.05) −0.61*** (0.05) −0.62*** (0.05) Ignored  Fair Trade 0.10 (0.18) 0.08 (0.18) 0.03 (0.17) 0.03 (0.17) −0.41 (0.33) −0.03 (0.22)  USDA Organic 0.39** (0.17) 0.40** (0.17) 0.81*** (0.17) 0.80*** (0.17) 0.35 (0.24) 0.69*** (0.22)  Rainforest Alliance 0.30 (0.19) 0.30 (0.19) 0.38** (0.17) 0.38** (0.17) 0.08 (0.28) 0.34 (0.22)  Carbon Footprint −0.01 (0.19) −0.04 (0.19) −0.13 (0.17) −0.14 (0.17) 0.11 (0.28) −0.17 (0.22)  Price −0.40*** (0.08) −0.52*** (0.06) −0.51*** (0.06) −0.54*** (0.05) −0.46*** (0.12) −0.43*** (0.08)  No_Buy −5.47*** (0.40) −5.38*** (0.40) −5.09*** (0.39) −5.06*** (0.39) −5.28*** (0.39) −5.18*** (0.39)  Log likelihooda −376 −379 −386 −386 −388 −393  AIC 774.1 779.8 793.1 794.5 797.1 807.0  BIC 823.2 829.0 842.3 843.7 846.2 856.2 p-values for testing statistical differences between the coefficients attended versus ignored  Fair Trade 0.024** 0.016** 0.001*** 0.001*** 0.006*** 0.018**  USDA Organic <0.001*** <0.001*** 0.081* 0.074* 0.006*** 0.144  Rainforest Alliance 0.007*** 0.005*** 0.007*** 0.007*** 0.016** 0.051*  Carbon Footprint 0.214 0.165 0.013** 0.010*** 0.430 0.081*  Price 0.001*** 0.005*** 0.008*** 0.013** 0.088* 0.008*** Note: *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. aFor the Full-AA MNL model: log likelihood = −401.4, AIC = 814.8, BIC = 841.7 (see Appendix B). For all model estimations using the serial ANA modelling approach, the coefficients of the ignored attributes indicate that respondents who were identified as visual non-attenders for the Rainforest Alliance, and Fair Trade labels truly ignored these attributes.8 For the serial ANA modelling approach with a fixation count of one (FC1-S2), the ignored coefficient of USDA Organic is also not significantly different from zero. Thus, for the serial ANA approach with a fixation count of one, respondents who were identified as visual non-attenders for one of the sustainability labels (USDA Organic, Fair Trade, or Rainforest Alliance) truly ignored these attributes. For all model estimations using the choice task ANA modelling approach, the coefficients of the ignored attributes indicate that in the choice tasks in which we considered Fair Trade and Carbon Footprint (and Rainforest Alliance for a fixation count of one) as visually ignored, they were indeed truly ignored. Thus, these choice tasks were answered as if the visually ignored attribute was not present in the choice set. For these attributes, restricting the coefficient to zero if it was visually ignored was appropriate and resulted in the removal of the attribute from the choice consideration. While 17 out of 30 ignored coefficients were not statistically different from zero, 13 out of the 30 estimated ignored parameters were significantly different from zero; thus, setting the coefficients of these parameters to zero may not be appropriate. In all six ANA models, the coefficient of ignored price was significant at the 1 per cent level. In all ANA models except FC1-S2, the coefficient of ignored USDA Organic was also significant. When using a fixation count of two as the cut-off, the choice task ANA model also had significant coefficients of the ignored Rainforest Alliance. Whereas assuming that visually ignored attributes were truly ignored was appropriate for all sustainability labels in the serial ANA approach with a fixation count of one and for all sustainability labels except USDA Organic in the serial ANA approach with a fixation count of two, this was not the case for price. Therefore, it is inappropriate to constrain the coefficients of the visually ignored price and also in some cases the sustainability labels USDA Organic and Rainforest Alliance to zero. This finding is important because it indicates that some attributes that were classified as visually non-attended based on the ANA definitions and modelling approaches were actually not ignored and could have influenced the choice. Importantly, there is a difference in the meaning of having coefficients for the ignored attributes significantly different from zero based on either stated ANA or visual ANA. When based on stated ANA, it refers to an attribute identified as ignored when in fact the attribute did impact one’s choice (Campbell and Lorimer, 2009); however, this attribute possibly had a reduced impact. For visual ANA, a coefficient for an ignored attribute significantly different from zero refers to an attribute being identified as not having looked at while the attribute impacts one’s choice. The statistical differences between the two coefficients (attended and ignored) were tested using the Krinsky and Robb method (Krinsky and Robb, 1986; 1990) and suggest some differences in behaviour (Table 6). For price, the coefficient of attended price was statistically lower (more negative) than the coefficient of ignored price (except for FC1-S2, where the difference was not statistically significant). The ignored coefficients of USDA Organic were significantly different from zero in five of the six models. Most of the ignored coefficients of USDA Organic were also significantly smaller than the attended coefficients of USDA Organic. For the choice task ANA models with a fixation count of two (defA-CT2 and defB-CT2), the Rainforest Alliance ignored coefficients were also significantly different from zero and statistically lower than the coefficient of attended Rainforest Alliance. This result indicates that a classification as visually ignored, on average, leads to a lower utility for the USDA Organic label (and also for the Rainforest Alliance label for two of the models) and a less negative utility for price. These less negative (price) or positive (USDA Organic, Rainforest Alliance) coefficients of the attributes identified as ignored may have two possible explanations. First, the Price, USDA Organic and Rainforest Alliance attributes that may be classified as visually non-attended may in fact not have been truly ignored when choosing the preferred alternative. Instead, respondents who paid less attention to these attributes received less negative (price) or less positive (USDA Organic) utility from these attributes. Second, it is possible that the smaller coefficient for the ignored subset might be a combination of truly ignored attributes (zeros) and attended attributes, which therefore results in a lower value for the coefficient. This could be an artefact of the experimental design or due to a number of reasons, including tracking error and/or wrongly identifying ANA due to tracking error and/or peripheral vision effects as explained in the next paragraph. Balcombe, Fraser and McSorly (2015) noted that people must look long enough at information for it to be processed. However, our results show that measuring whether an attribute is visually attended to might be attribute dependent. For example, the number of fixations needed to visually attend to attributes may differ depending on the attribute itself. Price and USDA Organic (and in some cases Rainforest Alliance) are attributes that were attended to when we identified them as being ignored. This result may be due to the respondents’ familiarity with these attributes. Therefore, even when we defined them as visually non-attended, they were not actually ignored. Wrongly assuming a coefficient to be zero also explains reduced model fit (see Section 3.4). In addition to familiarity, issues specifically related to the eye-tracking technology may provide an explanation for why there are significant coefficient estimates for some of the ignored parameters. While fixating on one attribute, another attribute might be viewed simultaneously and interpreted without truly fixating on it. If the participants did not fixate on information presented in an AOI, it does not mean that they were not aware that it was there (Bergstrom and Shall, 2014, p. 6), since fixations only report visual attention taking place in the foveal vision and not in the parafoveal and peripheral vision. While people’s primary attention is focused on what they see in the foveal vision, they might still grasp information presented in other parts of their visual field. Some authors (Henderson and Hollingworth, 1999; Henderson et al., 2003) have reported that a functional field of view can be 4°, while others (Holmqvist et al. 2011) have suggested using a margin of 1°–1.5°.9 We also cannot rule out the possibility of a measurement error due to the eye-tracker as the reason why the coefficients of the ignored parameters are significant, as it would be ‘unnatural’ for people to decouple visual attention and fixations in this manner.10 More information on the limitations and challenges of eye tracking are reported in a separate section (Section 5). 3.3.2. Coefficient of variation In this second method, we evaluate whether respondents who were identified as having ignored or not ignored an attribute based on the serial visual ANA have the same allocation (ignored versus not ignored) based on the coefficient of variation method (Table 7). Table 7. Comparison of allocation of respondents for serial visual ANA and serial inferred ANA (count) (n = 81) Serial inferred ANA Serial visual ANA FC 2 FC 1 NI I NI + I (%total) NI I NI + I (%total) Fair Trade NI 41 25 66 (81.5%) 60 6 66 (81.5%) I 7 8 15 (18.5%)a 11 4 15 (18.5%) NI + I (%total) 48 (59.3%) 33 (40.7%)a 71 (87.7%) 10 (12.3%) Matching allocations (%) 60.5%b 79.0% USDA Organic NI 34 35 69 (85.2%) 56 13 69 (85.2%) I 2 10 12 (14.8%) 6 6 12 (14.8%) NI + I (%total) 36 (44.4%) 45 (55.6%) 62 (76.5%) 19 (23.5%) Matching allocations (%) 54.3% 76.5% Rainforest Alliance NI 44 32 76 (93.8%) 63 13 76 (93.8%) I 2 3 5 (6.2%) 4 1 5 (6.2%) NI + I (%total) 46 (56.8%) 35 (43.2%) 67 (82.7%) 14 (17.3%) Matching allocations (%) 58.0% 79.0% Carbon Footprint NI 37 15 52 (64.2%) 47 5 52 (64.2%) I 11 18 29 (35.8%) 21 8 29 (35.8%) NI + I (%total) 48 (59.3%) 33 (40.7%) 68 (84.0%) 13 (16.0%) Matching allocations (%) 67.9% 67.9% Serial inferred ANA Serial visual ANA FC 2 FC 1 NI I NI + I (%total) NI I NI + I (%total) Fair Trade NI 41 25 66 (81.5%) 60 6 66 (81.5%) I 7 8 15 (18.5%)a 11 4 15 (18.5%) NI + I (%total) 48 (59.3%) 33 (40.7%)a 71 (87.7%) 10 (12.3%) Matching allocations (%) 60.5%b 79.0% USDA Organic NI 34 35 69 (85.2%) 56 13 69 (85.2%) I 2 10 12 (14.8%) 6 6 12 (14.8%) NI + I (%total) 36 (44.4%) 45 (55.6%) 62 (76.5%) 19 (23.5%) Matching allocations (%) 54.3% 76.5% Rainforest Alliance NI 44 32 76 (93.8%) 63 13 76 (93.8%) I 2 3 5 (6.2%) 4 1 5 (6.2%) NI + I (%total) 46 (56.8%) 35 (43.2%) 67 (82.7%) 14 (17.3%) Matching allocations (%) 58.0% 79.0% Carbon Footprint NI 37 15 52 (64.2%) 47 5 52 (64.2%) I 11 18 29 (35.8%) 21 8 29 (35.8%) NI + I (%total) 48 (59.3%) 33 (40.7%) 68 (84.0%) 13 (16.0%) Matching allocations (%) 67.9% 67.9% NI, not ignored; I, ignored; FC, fixation count. aOut of 81 respondents, 15 were considered to have ignored the Fair Trade attribute based on the inferred method, and 33 respondents were considered to have ignored the attribute based on serial visual ANA with fixation count 2. bFor 8 and 41 respondents, the classifications into ignoring and not ignoring, respectively, are the same between the inferred and visual methods, resulting in a matching allocation of 60% for Fair Trade ((41 + 8)/81). Table 7. Comparison of allocation of respondents for serial visual ANA and serial inferred ANA (count) (n = 81) Serial inferred ANA Serial visual ANA FC 2 FC 1 NI I NI + I (%total) NI I NI + I (%total) Fair Trade NI 41 25 66 (81.5%) 60 6 66 (81.5%) I 7 8 15 (18.5%)a 11 4 15 (18.5%) NI + I (%total) 48 (59.3%) 33 (40.7%)a 71 (87.7%) 10 (12.3%) Matching allocations (%) 60.5%b 79.0% USDA Organic NI 34 35 69 (85.2%) 56 13 69 (85.2%) I 2 10 12 (14.8%) 6 6 12 (14.8%) NI + I (%total) 36 (44.4%) 45 (55.6%) 62 (76.5%) 19 (23.5%) Matching allocations (%) 54.3% 76.5% Rainforest Alliance NI 44 32 76 (93.8%) 63 13 76 (93.8%) I 2 3 5 (6.2%) 4 1 5 (6.2%) NI + I (%total) 46 (56.8%) 35 (43.2%) 67 (82.7%) 14 (17.3%) Matching allocations (%) 58.0% 79.0% Carbon Footprint NI 37 15 52 (64.2%) 47 5 52 (64.2%) I 11 18 29 (35.8%) 21 8 29 (35.8%) NI + I (%total) 48 (59.3%) 33 (40.7%) 68 (84.0%) 13 (16.0%) Matching allocations (%) 67.9% 67.9% Serial inferred ANA Serial visual ANA FC 2 FC 1 NI I NI + I (%total) NI I NI + I (%total) Fair Trade NI 41 25 66 (81.5%) 60 6 66 (81.5%) I 7 8 15 (18.5%)a 11 4 15 (18.5%) NI + I (%total) 48 (59.3%) 33 (40.7%)a 71 (87.7%) 10 (12.3%) Matching allocations (%) 60.5%b 79.0% USDA Organic NI 34 35 69 (85.2%) 56 13 69 (85.2%) I 2 10 12 (14.8%) 6 6 12 (14.8%) NI + I (%total) 36 (44.4%) 45 (55.6%) 62 (76.5%) 19 (23.5%) Matching allocations (%) 54.3% 76.5% Rainforest Alliance NI 44 32 76 (93.8%) 63 13 76 (93.8%) I 2 3 5 (6.2%) 4 1 5 (6.2%) NI + I (%total) 46 (56.8%) 35 (43.2%) 67 (82.7%) 14 (17.3%) Matching allocations (%) 58.0% 79.0% Carbon Footprint NI 37 15 52 (64.2%) 47 5 52 (64.2%) I 11 18 29 (35.8%) 21 8 29 (35.8%) NI + I (%total) 48 (59.3%) 33 (40.7%) 68 (84.0%) 13 (16.0%) Matching allocations (%) 67.9% 67.9% NI, not ignored; I, ignored; FC, fixation count. aOut of 81 respondents, 15 were considered to have ignored the Fair Trade attribute based on the inferred method, and 33 respondents were considered to have ignored the attribute based on serial visual ANA with fixation count 2. bFor 8 and 41 respondents, the classifications into ignoring and not ignoring, respectively, are the same between the inferred and visual methods, resulting in a matching allocation of 60% for Fair Trade ((41 + 8)/81). First, the rates of ignoring the different attributes between visual and inferred ANA are compared (Table 7). The proportions of respondents ignoring Fair Trade, USDA Organic, Rainforest Alliance and Carbon Footprint using this inferred method11 are 18.5 per cent, 14.8 per cent, 6.2 per cent and 35.8 per cent, respectively. The inferred method has much lower rates of ignoring compared to the visual ANA with a fixation count of two (Table 7), except for Carbon Footprint, which is quite similar. For example, out of the 81 respondents, 15 (18.5 per cent) were considered to have ignored the Fair Trade attribute based on the inferred method, and 33 respondents (40.7 per cent) were considered to have ignored the attribute based on serial visual ANA with fixation count 2. The proportions obtained by the inferred method are more in line with those obtained through visual ANA with a fixation count of one as the cut-off, except for Carbon Footprint. In addition to the rate of ignoring attributes, a comparison of the allocation of the respondents identified as visually ignoring an attribute is of interest (Table 7). For example, for the Fair Trade attribute, 49 respondents are classified the same for both the inferred and the visual ANA with fixation count 2 (41 attenders and 8 non-attenders). A large proportion of the respondents who were identified as having ignored an attribute based on the serial visual ANA were not identified as ignoring this attribute by the inferred method. This shows that there is little agreement regarding the identification of serial non-attenders between the visual and inferred approach. This may suggest that a large proportion of the respondents who were identified as having ignored an attribute using the serial visual ANA did not actually ignore it. Clearly, there are differences in the allocations into the ignored group between the serial visual ANA and inferred ANA. However, it remains uncertain which method is the most accurate in identifying ignored attributes. Hess and Hensher (2010) compared this inferred ANA method (i.e. the coefficient of variation approach) with stated ANA and found large differences or inconsistencies between the two approaches in terms of rates of ignoring and in the allocation of respondents into the two groups (attenders versus non-attenders). The inferred method also has drawbacks such as the use of the arbitrary threshold value of two for the coefficient of variation (Hess and Hensher, 2010). As mentioned by Hess and Hensher (2010), more research is needed into how to allocate respondents to the ignored group based on the coefficient of variation, such as defining a less arbitrary threshold for the coefficient of variation method. 3.4. Model fit across estimations When comparing model fits, restricting the coefficients of the ignored attributes to zero results in a decrease in model fit (a decrease in log likelihood and an increase in the AIC and BIC statistics) compared with the full attendance model (Table 5). A worse model fit is likely due to treating attributes as having a zero parameter when some are not actually ignored (which was confirmed by the model estimations with two coefficients per attribute). For both fixation cut-offs, the serial ANA modelling approach results in a better model fit than the choice task ANA modelling approach. For models with a fixation count of two, defining an ignored attribute based on the whole choice set (Definition A) results in a better model fit than defining it based on the alternatives (Definition B). Models in which two separate coefficients were estimated (Table 6) have a better model fit compared with the model assuming full attendance (Appendix B) and also outperforms the model in which the ignored coefficients were set to zero (Appendix B). Results thus show that setting the ignored coefficients to zero when in fact they are statistically different from zero thus worsens the model fit. Though, several studies on stated ANA have reported an improvement in model fit when constraining ignored coefficients to zero (Hensher, Rose and Greene, 2005; Campbell, Hutchinson and Scarpa, 2008; Kragt, 2013). However, similar to our study, Alemu et al. (2013) reported a decrease in model fit when restricting the coefficients to zero. This decrease in model fit could also be attributed to the number of observations that are essentially excluded from contributing to the likelihood function (Alemu et al. 2013), which may also explain why we see a decrease in model fit when moving from serial to choice task ANA, as the percentage of ignored attributes also increases (Table 4). Since not all coefficients that were identified as ignored were, in fact, ignored based on estimates of these coefficients (Section 3.3.1) and based on the coefficient of variation (Section 3.3.2), we can conclude that eye tracking did not always successfully identify visual non-attenders for all attributes. This is also confirmed by the decrease in model fit for most models when constraining the ignored coefficients to zero compared with the full attendance models. We discuss the potential challenges and limitations of using eye tracking for ANA research in Section 5. 3.5. WTP We cannot always assume that the respondents did not look at the attribute just because they did not care about it; i.e. we do not know the reason why they did not consider the attribute when making their choices. For instance, some people may have visually ignored an attribute because the task was too complex, while others may have ignored it because they do not derive utility from that attribute. In the first case (too complex), a person who ignored an attribute may have the same preferences as someone who attended to it and thus may also have the same WTP. In this case, the WTP based on the attended attributes would be applicable for both groups. In the second case, where respondents ignore an attribute because they do not derive any utility from it, the WTP for the ignored group is zero. Because we do not know the actual reasons for visually ignoring the attributes, the WTP estimates are calculated only using the coefficient estimates for the attended attributes from Table 5, resulting in WTP estimates for sustainability labels of $1.7312 or lower. The Krinsky and Robb method (Krinsky and Robb, 1986; 1990) indicated no statistical differences in the mean estimates for the WTP distribution between the full attendance model and the models accounting for visual ANA. This result suggests that accounting for visual ANA, in our study, did not have significant implications in terms of mean WTP estimates. While coffee prices largely depend on the quality of the beans, their origin, the blend, and the brand, organic coffee prices are often higher than prices for coffee certified by the Rainforest Alliance or Fair Trade USA (FAO, 2009). Retail prices for ground coffee ranged from $3.00 to $9.99 per 12-ounce package (store check, Fayetteville, AR, 2013). Because the prices for both conventional and certified coffee vary considerably by the quality and origin as well as by the nature of the outlet and brand, it is very difficult to determine a premium that is solely attributable to a sustainability certification rather than these other factors (Consumers International, 2005; FAO, 2009). Based on a store check (Fayetteville, AR, 2013), the price premium for coffee with a sustainability label ranges from $1.50 to $2.30 per 12 ounces when comparing coffees with and without a label of the same brand. 4. Conclusion Given that CEs are commonly used to assess attribute valuation, there is an urgent need for and considerable research interest in methods to account for ANA. Stated and inferred approaches to address ANA do not directly measure whether attributes were ignored in a CE. Limitations associated with stated and inferred approaches to obtain ANA information at the choice task level justify studying other ways to address respondents’ true processing strategies and attendance at the choice task level. We contribute to this research area by using eye-tracking measures to evaluate visual attendance to the attributes in a CE. This method does not rely on self-reported ANA behaviour and does not attempt to infer ANA based on respondents’ choice behaviour. Instead, we measured participants’ visual attention to specify whether they visually attended to the information presented to them and extend the work of Balcombe, Fraser and McSorly (2015) on visual ANA in CEs using eye-tracking technology in three ways (alternative fixation count cut-offs, alternative definitions of visual ANA, and alternative modelling approaches). For each combination of fixation cut-off, visual ANA definition and modelling approach, we identified visually ignored attributes. Restricting the coefficients of ignored attributes to zero led to a decrease in model fit compared to the full attendance model, and accounting for visual ANA did not significantly influence the WTP estimates. Not all visually ignored attributes were truly ignored based on the respondents’ choice behaviours (price and USDA Organic labels in most cases and Rainforest Alliance in some), i.e. the respondents paid some attention to these attributes in their choice consideration. Hence, our results suggest that Fair Trade, Carbon Footprint and in most cases Rainforest Alliance were truly ignored when they were identified as ignored using visual ANA. However, the same does not hold for price and USDA Organic. This result shows that the best way to account for visual ANA might depend on the attribute itself. For price and USDA Organic, it is likely that people do not need to fixate their eyes on these attributes to the same extent as to the other attributes because it is possible that they are already acquainted or more familiar with these attributes. It is possible that for USDA Organic and price, they were aware of the information presented from other parts of their visual field (parafoveal and peripheral vision) not covered by eye fixations. The price, USDA Organic, and (in some cases) Rainforest Alliance attributes that may have been classified as visually non-attended were in fact not truly ignored. Most coefficient estimates for attributes that were classified as visually ignored were lower than the attended coefficients. This result might occur for two reasons. First, respondents who paid less attention to these attributes received less negative (price) or positive (USDA Organic, Rainforest Alliance) utility from these attributes. Second, the smaller coefficient of the ignored subset might be a combination of truly ignored attributes (zeros) and attended attributes and therefore result in a lower average. The true explanation is likely to be a combination of both reasons. Next, the allocation of attenders and non-attenders based on visual ANA was compared with the allocation based on the coefficient of variation, an inferred method. The method of the coefficient of variation indicated that a small proportion of those classified as having ignored an attribute using the eye-tracking’s visual ANA were also identified as having ignored it based on the inferred method. Given the insignificance of many of the coefficients of the ignored attributes, the mismatch between the allocation based on the inferred method and the serial visual ANA and the decrease in the model fit when using visual ANA, we can conclude, at least based on our study, that using eye tracking does not always provide an adequate measure that can guarantee that an attribute should be considered as having been ignored or not. Nevertheless, eye-tracking measures can provide useful information regarding respondents’ decision-making behaviour. It can help us to better understand their decision-making process and its relation with choice, preferences and attention (Grebitus, Roosen and Seitz, 2015). Previous studies illustrated that eye tracking is useful to quantify visual attention, a measure for the degree of attention, during food choices (Van Loo et al. 2015; Lewis, Grebitus and Nayga, 2016); however, Balcombe et al. (2017) report that the relationship between visual attention and preference may be weak. While eye tracking has been used to quantify visual attention and its effects during food choice, applying it to identify visual ANA is more challenging, as suggested by our findings. We discuss these challenges and limitations in the following sections. Specifically, we first discuss the limitations of eye-tracking research in general, followed by limitations specifically related to ANA research based on the insights from our study. 5. Limitations and challenges 5.1. The use of eye tracking Eye tracking provides measures of where participants fixate but not why; thus, the motivations and cognitions underlying these eye movements remain unknown (Graham, Orquin and Visschers, 2012). Familiarity might influence how extensively a piece of information is examined, as possibly observed with the price and USDA Organic attributes in our study (Pieters, Rosbergen and Hartog, 1996, 1999; Graham, Orquin and Visschers, 2012). Moreover, participants’ fixations do not necessarily imply understanding and do not reveal anything about the higher-level processes of attention and comprehension. As suggested by Graham, Orquin and Visschers (2012), conducting an interview after an eye-tracking task may provide additional insight into what respondents were thinking during the task. While eye-tracking studies might be less prone to social desirability compared to studies that ask respondents directly about the information to which they attend, knowing that their eye movements will be monitored could potentially influence their behaviour (Graham, Orquin and Visschers, 2012). Finally, eye tracking is a relatively expensive and time-consuming method. 5.2. The use of eye tracking to study ANA While eye tracking monitors consumers’ visual attention, this technology has some limitations that are particularly relevant when studying visual ANA. These issues relate to (i) the use of fixations, (ii) the size of the AOIs, (iii) the definitions for visual ANA, (iv) the influence of familiarity and (v) the tracking ratio. First, eye tracking uses fixations as an indication of visual attention, but it only reveals what happens in an individual’s foveal vision. While fixations take place in our foveal vision, which is where our primary attention is focused (Bergstrom and Shall, 2014), eye tracking does not allow us to measure the attention paid in one’s parafoveal and peripheral vision (Bergstrom and Shall, 2014; Orquin, Ashby and Clarke, 2016). If the participants did not fixate on information presented in an AOI, it does not necessarily mean that they were not aware that it was there (Bergstrom and Shall, 2014, p. 6). They may still have grasped information present in other parts of their visual field without focusing on it. While eye fixations within a certain AOI are believed to provide reliable measures for visual attention for that AOI (Bialkova and van Trijp, 2011), eye fixations do not necessarily represent everything that participants might have observed. Based on our study, we can conclude that fixations do not necessarily allow us to identify whether the information was ignored. Second, eye tracking uses AOIs for which specific metrics (such as fixation counts) are calculated, and thus the results depend on how these areas are defined. Unfortunately, there is no consensus in the decision-making research literature on the definition of AOIs. As mentioned by Orquin, Ashby and Clarke (2016, p. 103), ‘this lack of standardisation in AOI definition and reporting presents direct problems to the advancement of behavioural decision-making research’. Thus, there is a need to have a more standardised way to define AOIs, as the size of the AOI determines its fixation count as well as other eye-tracking metrics in use. Third, there are different ways to define visual ANA with respect to the number of fixation counts and the number of choice tasks. Balcombe, Fraser and McSorly (2015) used the threshold of at least two fixations to consider an attribute as attended. In addition, they identify a respondent as a non-attender when this person ignores an attribute in more than half of the choice tasks. These are two rather arbitrary values and will influence the discrete measure of visual ANA (ignored versus not ignored). Fourth, how familiar respondents are with the attribute or logo – as in our study – might matter. If they are very familiar with it, they may not need to fixate on it to recognise that the logo is present. If they are less familiar, they may need to look longer at and fixate more on the information for it to be processed. Thus, the decisions about the fixation cut-off as well as the size of the AOI for identifying information as being ignored might depend on the respondent’s familiarity. Fifth, not all eye movements are recorded. In our study, 89.5 per cent of eye movement data was recorded, and thus, we cannot guarantee if some attributes were ignored or unfixated. The five limitations discussed above represent major challenges, but they can hopefully assist researchers and future studies to further optimise the use of eye tracking in the context of ANA in CEs. We discuss several specific suggestions for future research in the following subsection. 5.3. Suggestions for future research on the use of eye tracking in CEs While we attempted to account for the limitations described above as much as possible in the present study, there are several suggestions based on our study that future studies could take into account. In our study, the logos were rather large and in close proximity to each other. In addition, the coffee packages did not contain much other information. Since packages never had the same labels, it is possible that respondents could infer the quality attributes on one package based on the labels on the other package in a choice set. We recommend that future studies include other information cues on the food packages (e.g. nutritional labels, ingredients lists, graphics) that compete for the respondent’s attention and make it more difficult for the participant to be aware of the presence of certain logos without fixating on them. In this study, we used a head-free, below-screen eye-tracker that provided us with a relatively high tracking ratio (89.5 per cent) across our respondents. Previous research has shown no significant difference between head-mounted and head-free types of eye-trackers with respect to the pupil size-based index of cognitive activity (Bartels and Marshall, 2012). Furthermore, the head-free, below-screen tracker allows participants to be less constrained and may promote more natural human behaviour. Nevertheless, future studies might consider using a chin-rest, since this could avoid movement of the head, possibly resulting in better tracking quality (in terms of reliability and accuracy) compared with the head-free, below-screen tracker that we used in this study. While we compared different visual ANA approaches, future studies could expand on comparing various ANA approaches (visual, stated and inferred) and evaluating which of these techniques or combinations of techniques is the most appropriate to account for ANA. Specifically, for visual ANA, future research should improve and standardise how to apply visual ANA in choice modelling based on eye tracking and remove any arbitrary steps. Studies could, for example, test other fixation count cut-offs and other visual ANA definitions. For example, for serial ANA, we used the approach used by Balcombe, Fraser and McSorly (2015), whereby a respondent is identified as a non-attender if he/she ignored an attribute in more than half of the choice tasks. However, ‘more than half’ is an arbitrary approach, and other approaches could be tested to further fine-tune the serial visual ANA method. In addition, other AOI sizes and margins could be tested. In addition to incorporating ANA or other decision-making strategies into choice models, future studies could use eye tracking as a tool to study how to optimise experimental designs that would discourage respondents from ignoring information, thereby avoiding violation of the assumption of rational utility maximisation (i.e. that the complete information presented is considered by respondents). We share the opinion of Balcombe, Fraser and McSorly (2015), who suggested the use of eye tracking to assist in the visual design and appearance of the CE instrument. Several eye-tracking studies have demonstrated that bottom-up factors can trigger attention (Visschers, Hess and Siegrist, 2010; Bialkova and van Trijp, 2011). Bottom-up or stimulus-driven forms of attention are caused by characteristics of the visual stimulus itself (colour, size, location, saliency) and occur without specifically searching for them (Wolfe, 1998). Further insights into these bottom-up factors can help in the visual design or layout of choice sets that can reduce visual ANA in CEs. Spinks and Mortimer (2016), for example, investigated the impact of CE complexity (in terms of number of attributes ranging from 3 to 8) on the prevalence of ANA. They reported higher levels of ANA when more attributes are present and provided information on how to improve the CE design, measuring health care preferences in their study. They suggest the use of eye tracking during the CE design and piloting phase, which is also suggested by Balcombe et al. (2017). Future studies could test other issues of the visual CE design and its impact on ANA such as the pictorial or verbal representation of attributes or alternatives, and other variations in representation of choice sets that could reduce ANA in CE studies. Finally, in addition to ANA research, eye tracking can be used to study other decision-making strategies and decision heuristics and can assist in explaining the decision-making process in relation to food choices (Uggeldahl et al., 2016) and preferences. As mentioned by Orquin and Loose (2013), more research on how to effectively combine attention research from eye tracking with decision-making research is warranted. Supplementary data Supplementary data are available at European Review of Agricultural Economics online. Acknowledgements The authors thank the editor (Iain Fraser) and three anonymous journal reviewers for their valuable comments and suggestions that helped us to significantly improve the quality of the paper. Footnotes 1 In the CE literature, the set of alternatives that an individual must consider to arrive at his/her choice is called the choice set (Hensher, Rose and Greene, 2015). The choice task refers to the action itself of selecting the preferred alternative in the choice set. 2 Visual attention to an attribute is ‘a continuous measure of the degree to which a respondent evaluates the attribute’, while attendance is ‘a discrete measure indicating whether respondents will be considered to have attended an attribute or not’ (Balcombe et al., 2015, p. 449). 3 ‘Truly ignored’ is related to the coefficient of ‘visually ignored’ attributes being zero. 4 We could not apply ECLC models because the limited sample size caused issues with them. 5 The visual attendance towards price depends on the definition applied because price was presented in each of the two buying alternatives. 6 As noted by a reviewer, this could also be due to the relatively small sample size. Only the observations in which the attribute was attended to are used for the coefficient estimation, while for the other observations, the coefficients are set to zero. 7 Five parameters were estimated (USDA Organic, Rainforest Alliance, Fair Trade, Carbon Footprint and Price) for each of the six approaches, resulting in a total of 30 parameter estimations. 8 For Carbon Footprint, the attended coefficients are insignificant. Therefore, being wrongly classified as having ignored Carbon Footprint may result in an insignificant coefficient. 9 As suggested by a reviewer, the functional field of view is likely different for illustration and text, with a smaller functional view for text (such as price) as compared to illustrations (such as the sustainability labels). Hughes et al. (2003) also reported differences in visual attention for text versus pictures. 10 We thank an anonymous reviewer for pointing out this possibility. 11 The inferred method cannot estimate the percentage of respondents ignoring price since price is considered fixed. 12 This corresponds with = 1.407/−0.812 for model defA-S. References Alemu , M. H. , Mørkbak , M. R. , Olsen , S. B. and Jensen , C. L. ( 2013 ). Attending to the reasons for attribute non-attendance in choice experiments . Environmental & Resource Economics 54 : 333 – 359 . Google Scholar CrossRef Search ADS Balcombe , K. , Burton , M. and Rigby , D. ( 2011 ). Skew and attribute non-attendance within the Bayesian mixed logit model . Journal of Environmental Economics and Management 62 : 446 – 461 . Google Scholar CrossRef Search ADS Balcombe , K. G. , Fraser , I. M. and McSorly , E. ( 2015 ). Visual attention and attribute attendance in multi-attribute choice experiments . Journal of Applied Econometrics 30 : 447 – 467 . Google Scholar CrossRef Search ADS Balcombe , K. , Fraser , I. , Williams , L. and McSorley , E. ( 2017 ). Examining the relationship between visual attention and stated preferences: a discrete choice experiment using eye-tracking . Journal of Economic Behavior & Organization 144 : 238 – 257 . Google Scholar CrossRef Search ADS Barreiro-Hurle , J. , Gracia , A. and de-Magistris , T. ( 2010 ). The effects of multiple health and nutrition labels on consumer food choices . Journal of Agricultural Economics 61 : 426 – 443 . Google Scholar CrossRef Search ADS Bartels , M. and Marshall , S. P. ( 2012 ). Measuring cognitive workload across different eye tracking hardware platforms. 2012 ACM Symposium on Eye Tracking and Applications, Santa Barbara, California, 28–30 March, 2012. Bergstrom , J. R. and Schall , A. ( 2014 ). Eye Tracking in User Experience Design . Waltham, MA : Elsevier . Bialkova , S. and van Trijp , H. C. M. ( 2011 ). An efficient methodology for assessing attention to and effect of nutrition information displayed front-of-pack . Food Quality and Preference 22 : 592 – 601 . Google Scholar CrossRef Search ADS Bonnet , C. and Simioni , M. ( 2001 ). Assessing consumer response to protected designation of origin labelling: a mixed multinomial logit approach . European Review of Agricultural Economics 28 : 433 – 449 . Google Scholar CrossRef Search ADS Burgess , L. ( 2007 ). Discrete Choice Experiments [computer software], Department of Mathematical Sciences, University of Technology, Sydney. http://maths.science.uts.edu.au/maths/wiki/SPExpts. Accessed 1 August 2016. Campbell , D. , Boeri , M. , Doherty , E. and Hutchinson , W. G. ( 2015 ). Learning, fatigue and preference formation in discrete choice experiments . Journal of Economic Behavior and Organization 119 : 345 – 363 . Google Scholar CrossRef Search ADS Campbell , D. , Hensher , D. A. and Scarpa , R. ( 2011 ). Non-attendance to attributes in environmental choice analysis: a latent class specification . Journal of Environmental Planning and Management 54 : 1061 – 1076 . Google Scholar CrossRef Search ADS Campbell , D. , Hutchinson , W. G. and Scarpa , R. ( 2008 ). Incorporating discontinuous preferences into the analysis of discrete choice experiments . Environmental & Resource Economics 41 : 401 – 417 . Google Scholar CrossRef Search ADS Campbell , D. and Lorimer , V. S. ( 2009 ). Accommodating attribute processing strategies in stated choice analysis: Do respondents do what they say they do? 17th Annual Conference of the European Association of Environmental and Resource Economics, Amsterdam. Caputo , V. , Nayga , R. M. and Scarpa , R. ( 2013 ). Food miles or carbon emissions? Exploring labelling preference for food transport footprint with a stated choice study . Australian Journal of Agricultural and Resource Economics 57 : 465 – 482 . Google Scholar CrossRef Search ADS Caputo , V. , Van Loo , E. J. , Scarpa , R. , Nayga , R. M. and Verbeke , W. ( 2017 ). Comparing serial, and choice task stated and inferred attribute non‐attendance methods in food choice experiments . Journal of Agricultural Economics DOI:10.1111/1477-9552.12246 . Carlsson , F. , Kataria , M. and Lampi , E. ( 2010 ). Dealing with ignored attributes in choice experiments on valuation of Sweden’s environmental quality objectives . Environmental & Resource Economics 47 : 65 – 89 . Google Scholar CrossRef Search ADS Consumers International ( 2005 ). From Bean to Cup: How Consumer Choice Impacts on Coffee Producers and the Environment . London : Consumers International . Fairtrade International ( 2015 ). Monitoring the Scope and Benefits of Fair Trade , 7th ed . Bonn : Fairtrade International . Food and Agriculture Organization of the United Nations (FAO) ( 2009 ). The Market for Organic and Fair-Trade Coffee. Study prepared in the framework of FAO project GCP/RAF/404/GER: ‘Increasing Incomes and Food Security of Small Farmers in West and Central Africa Through Exports of Organic and Fair-Trade Tropical Products’. Trade and Markets Division. Rome: FAO. Gracia , A. ( 2014 ). Consumers’ preferences for a local food product: a real choice experiment . Empirical Economics 47 : 111 – 128 . Google Scholar CrossRef Search ADS Gracia , A. , Barreiro-Hurle , J. and Perez , L. P. Y. ( 2012 ). Can renewable energy be financed with higher electricity prices? Evidence from a Spanish region . Energy Policy 50 : 784 – 794 . Google Scholar CrossRef Search ADS Graham , D. J. , Orquin , J. L. and Visschers , V. H. ( 2012 ). Eye tracking and nutrition label use: a review of the literature and recommendations for label enhancement . Food Policy 37 : 378 – 382 . Google Scholar CrossRef Search ADS Grebitus , C. , Roosen , J. and Seitz , C. C. ( 2015 ). Visual attention and choice: a behavioral economics perspective on food decisions . Journal of Agricultural & Food Industrial Organization 13 : 73 – 81 . Google Scholar CrossRef Search ADS Greene , W. H. ( 1990 ). LIMDEP Version 5.0 . New York, NY : Econometric Software. Inc . Greene , W. H. ( 2012 ). Nlogit Version 5.0: Reference Guide . Plainview, NY : Econometric Software Inc . Grunert , K. G. ( 2011 ). Sustainability in the food sector: a consumer behaviour perspective . International Journal on Food System Dynamics 2 : 207 – 218 . Henderson , J. M. and Hollingworth , A. ( 1999 ). The role of fixation position in detecting scene changes across saccades . Psychological Science 10 : 438 – 443 . Google Scholar CrossRef Search ADS Henderson , J. M. , Williams , C. C. , Castelhano , M. S. and Falk , R. J. ( 2003 ). Eye movements and picture processing during recognition . Perception & Psychophysics 65 : 725 – 734 . Google Scholar CrossRef Search ADS Hensher , D. A. ( 2006 ). How do respondents process stated choice experiments? Attribute consideration under varying information load . Journal of Applied Econometrics 21 : 861 – 878 . Google Scholar CrossRef Search ADS Hensher , D. A. ( 2014 ). Process heuristics in choice analysis: an editorial . Journal of Choice Modelling 11 : 1 – 3 . Google Scholar CrossRef Search ADS Hensher , D. A. and Greene , W. H. ( 2010 ). Non-attendance and dual processing of common-metric attributes in choice analysis: a latent class specification . Empirical Economics 39 : 413 – 426 . Google Scholar CrossRef Search ADS Hensher , D. A. and Rose , J. M. ( 2009 ). Simplifying choice through attribute preservation or non-attendance: implications for willingness to pay . Transportation Research Part E 45 : 583 – 590 . Google Scholar CrossRef Search ADS Hensher , D. A. , Rose , J. M. and Greene , W. H. ( 2005 ). The implications on willingness to pay of respondents ignoring specific attributes . Transportation 32 : 203 – 222 . Google Scholar CrossRef Search ADS Hensher , D. A. , Rose , J. M. and Greene , W. H. ( 2012 ). Inferring attribute non-attendance from stated choice data: implications for willingness to pay estimates and a warning for stated choice experiment design . Transportation 39 : 235 – 245 . Google Scholar CrossRef Search ADS Hensher , D. A. , Rose , J. M. and Greene , W. H. ( 2015 ). Applied Choice Analysis , 2nd edn . Cambridge, UK : Cambridge University Press . Google Scholar CrossRef Search ADS Hess , S. and Hensher , D. A. ( 2010 ). Using conditioning on observed choices to retrieve individual-specific attribute processing strategies . Transportation Research Part B 44 : 781 – 790 . Google Scholar CrossRef Search ADS Hess , S. , Stathopoulos , A. , Campbell , D. , O’Neill , V. and Caussade , S. ( 2013 ). It’s not that I don’t care, I just don’t care very much: confounding between attribute non-attendance and taste heterogeneity . Transportation 40 : 583 – 607 . Google Scholar CrossRef Search ADS Hole , A. R. ( 2011 ). A discrete choice model with endogenous attribute attendance . Economics Letters 110 : 203 – 205 . Google Scholar CrossRef Search ADS Holmqvist , K. , Nyström , M. , Andersson , R. , Dewhurst , R. , Jarodzka , H. and Van de Weijer , J. ( 2011 ) Eye Tracking: A Comprehensive Guide to Methods and Measures . Oxford , UK: Oxford University Press . Hughes , A. , Wilkens , T. , Wildemuth , B. M. and Marchionini , G. ( 2003 ). Text or pictures? An eyetracking study of how people view digital video surrogates. In: Proceedings of the International Conference on Image and Video Retrieval, 271–280. Kehlbacher , A. , Balcombe , K. and Bennett , R. ( 2013 ). Stated attribute non-attendance in successive choice experiments . Journal of Agricultural Economics 64 : 693 – 706 . Google Scholar CrossRef Search ADS Kragt , M. E. ( 2013 ). Stated and inferred attribute attendance models: a comparison with environmental choice experiments . Journal of Agricultural Economics 64 : 719 – 736 . Google Scholar CrossRef Search ADS Krinsky , I. and Robb , A. L. ( 1986 ). On approximating the statistical properties of elasticities . The Review of Economics and Statistics 68 : 715 – 719 . Google Scholar CrossRef Search ADS Krinsky , I. and Robb , A. L. ( 1990 ). On approximating the statistical properties of elasticities: a correction . The Review of Economics and Statistics 72 : 189 – 190 . Google Scholar CrossRef Search ADS Krucien , N. , Ryan , M. and Hermens , F. ( 2017 ). Visual attention in multi-attributes choices: what can eye-tracking tell us? Journal of Economic Behavior & Organization 135 : 251 – 267 . Google Scholar CrossRef Search ADS Lagarde , M. ( 2013 ). Investigating attribute non-attendance and its consequences in choice experiments with latent class models . Health Economics 22 : 554 – 567 . Google Scholar CrossRef Search ADS Lancaster , K. J. ( 1966 ). New approach to consumer theory . Journal of Political Economy 74 : 132 – 157 . Google Scholar CrossRef Search ADS Lancsar , E. and Louviere , J. ( 2006 ). Deleting ‘irrational’ responses from discrete choice experiments: a case of investigating or imposing preferences? Health Economics 15 : 797 – 811 . Google Scholar CrossRef Search ADS Layton , D. F. and Brown , G. ( 2000 ). Heterogeneous preferences regarding global climate change . The Review of Economics and Statistics 82 : 616 – 624 . Google Scholar CrossRef Search ADS Lewis , K. E. , Grebitus , C. and Nayga , R. M. ( 2016 ). The impact of brand and attention on consumers’ willingness to pay: evidence from an eye tracking experiment . Canadian Journal of Agricultural Economics/Revue canadienne d’agroeconomie 64 : 753 – 777 . Google Scholar CrossRef Search ADS Lusk , J. L. and Schroeder , T. C. ( 2004 ). Are choice experiments incentive compatible? A test with quality differentiated beef steaks . American Journal of Agricultural Economics 86 : 467 – 482 . Google Scholar CrossRef Search ADS Mariel , P. , Hoyos , D. and Meyerhoff , J. ( 2013 ). Stated or inferred attribute non-attendance? A simulation approach . Economía Agraria y Recursos Naturales (Agricultural and Resource Economics) 13 : 51 – 67 . Google Scholar CrossRef Search ADS McFadden , D. ( 1974 ). Conditional logit analysis of qualitative choice behavior. In: Zarembka P. (ed.) , Frontiers in Econometrics . New York : Academic Press , 105 – 142 . Meyerhoff , J. and Liebe , U. ( 2009 ). Discontinuous preferences in choice experiments: Evidence at the choice task level. 17th Annual Conference of the European Association of Environmental and Resource Economics, Amsterdam. Mintel ( 2012 ). Coffee–USA–October 2012 . London : Mintel Group Ltd . Morey , E. and Rossmann , K. ( 2003 ). Using stated-preference questions to investigate variations in willingness to pay for preserving marble monuments: classic heterogeneity, random parameters, and mixture models . Journal of Cultural Economics 27 : 215 – 229 . Google Scholar CrossRef Search ADS Mørkbak , M. R. , Olsen , S. B. and Campbell , D. ( 2014 ). Behavioral implications of providing real incentives in stated choice experiments . Journal of Economic Psychology 45 : 102 – 116 . Google Scholar CrossRef Search ADS Orquin , J. L. , Ashby , N. J. and Clarke , A. D. ( 2016 ). Areas of interest as a signal detection problem in behavioral eye-tracking research . Journal of Behavioral Decision Making 29 : 103 – 115 . Google Scholar CrossRef Search ADS Orquin , J. L. and Loose , S. M. ( 2013 ). Attention and choice: A review on eye movements in decision making . Acta psychological 144 : 190 – 206 . Google Scholar CrossRef Search ADS Pierrot , J. , Giovannucci , D. and Kasterine , A. ( 2011 ). Trends in the trade of certified coffees. Technical Paper . Geneva : International Trade Centre (ITC) . Pieters , R. G. M. , Rosbergen , E. and Hartog , M. ( 1996 ). Visual attention to advertising: the impact of motivation and repetition . Advances in Consumer Research 23 : 242 – 248 . Pieters , R. G. M. , Rosbergen , E. and Wedel , M. ( 1999 ). Visual attention to repeated print advertising: a test of scanpath theory . Journal of Marketing Research 36 : 424 – 438 . Google Scholar CrossRef Search ADS Poole , A. , Ball , L. and Phillips , P. ( 2005 ). In search of salience: a response-time and eye-movement analysis of bookmark recognition. In Fincher, S., Markopoulos, P., Moore, D. and Ruddle, R. (eds), People and Computers Xviii – Design for Life . London : Springer , 363 – 378 . Google Scholar CrossRef Search ADS Puckett , S. M. and Hensher , D. A. ( 2008 ). The role of attribute processing strategies in estimating the preferences of road freight stakeholders . Transportation Research Part E 44 : 379 – 395 . Google Scholar CrossRef Search ADS Puckett , S. M. and Hensher , D. A. ( 2009 ). Revealing the extent of process heterogeneity in choice analysis: an empirical assessment . Transportation Research Part A 43 : 117 – 126 . Revelt , D. and Train , K. ( 1998 ). Mixed logit with repeated choices: households’ choices of appliance efficiency level . The Review of Economics and Statistics 80 : 647 – 657 . Google Scholar CrossRef Search ADS Ryan , M. and Bate , A. ( 2001 ). Testing the assumptions of rationality, continuity and symmetry when applying discrete choice experiments in health care . Applied Economics Letters 8 : 59 – 63 . Google Scholar CrossRef Search ADS Scarpa , R. , Campbell , D. and Hutchinson , W. G. ( 2007 ). Benefit estimates for landscape improvements: sequential Bayesian design and respondents’ rationality in a choice experiment . Land Economics 83 : 617 – 634 . Google Scholar CrossRef Search ADS Scarpa , R. , Ferrini , S. and Willis , K. ( 2005 ). Performance of error component models for status-quo effects in choice experiments. In: R. Scarpa and A. Alberini (eds) , Applications of Simulation Methods in Environmental and Resource Economics . Dordrecht : Springer , 247 – 273 . Google Scholar CrossRef Search ADS Scarpa , R. , Gilbride , T. J. , Campbell , D. and Hensher , D. A. ( 2009 ). Modelling attribute non-attendance in choice experiments for rural landscape valuation . European Review of Agricultural Economics 36 : 151 – 174 . Google Scholar CrossRef Search ADS Scarpa , R. , Thiene , M. and Hensher , D. A. ( 2010 ). Monitoring choice task attribute attendance in nonmarket valuation of multiple park management services: does it matter? Land Economics 86 : 817 – 839 . Google Scholar CrossRef Search ADS Scarpa , R. , Zanoli , R. , Bruschi , V. and Naspetti , S. ( 2013 ). Inferred and stated attribute non-attendance in food choice experiments . American Journal of Agricultural Economics 95 : 165 – 180 . Google Scholar CrossRef Search ADS SMI (SensoMotoric Instruments) . ( 2016 ). BeGaze Manual. Version 3.6. February 2016. 416 pages. Spinks , J. and Mortimer , D. ( 2016 ). Lost in the crowd? Using eye-tracking to investigate the effect of complexity on attribute non-attendance in discrete choice experiments . BMC Medical Informatics and Decision Making 16 ( 1 ): 14 . Google Scholar CrossRef Search ADS Street , D. J. and Burgess , L. ( 2007 ). The Construction of Optimal Stated Choice Experiments . Hoboken, NJ : John Wiley & Sons, Inc . Google Scholar CrossRef Search ADS Uggeldahl , K. , Jacobsen , C. , Lundhede , T. H. and Olsen , S. B. ( 2016 ). Choice certainty in discrete choice experiments: Will eye tracking provide useful measures? Journal of Choice Modelling 20 : 35 – 48 . Google Scholar CrossRef Search ADS Van Loo , E. J. , Caputo , V. , Nayga , R. M. Jr. , Seo , H.-S. , Zhang , B. and Verbeke , W. ( 2015 ). Sustainability labels on coffee: consumer preferences, willingness-to-pay and visual attention to attributes . Ecological Economics 118 : 215 – 225 . Google Scholar CrossRef Search ADS Van Loo , E. J. , Caputo , V. , Nayga , R. M. Jr and Verbeke , W. ( 2014 ). Consumers’ valuation of sustainability labels on meat . Food Policy 49 : 137 – 150 . Google Scholar CrossRef Search ADS Visschers , V. H. M. , Hess , R. and Siegrist , M. ( 2010 ). Health motivation and product design determine consumers’ visual attention to nutrition information on food products . Public Health Nutrition 13 : 1099 – 1106 . Google Scholar CrossRef Search ADS Wolfe , J. M. ( 1998 ). Visual search. In: H. Pashler (ed.) , Attention . Hove, UK : Psychology Press , 13 – 22 . Author notes Review coordinated by Iain Fraser © Oxford University Press and Foundation for the European Review of Agricultural Economics 2018; all rights reserved. For permissions, please e-mail: [email protected] This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices) http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png European Review of Agricultural Economics Oxford University Press

Using eye tracking to account for attribute non-attendance in choice experiments

Download PDF
33 pages

Loading next page...
 
/lp/ou_press/using-eye-tracking-to-account-for-attribute-non-attendance-in-choice-JzBksjjA1r

References (69)

Publisher
Oxford University Press
Copyright
© Oxford University Press and Foundation for the European Review of Agricultural Economics 2018; all rights reserved. For permissions, please e-mail: [email protected]
ISSN
0165-1587
eISSN
1464-3618
DOI
10.1093/erae/jbx035
Publisher site
See Article on Publisher Site

Abstract

Abstract This study uses eye-tracking measures to account for attribute non-attendance (ANA) in choice experiments. Using the case of sustainability labelling on coffee, we demonstrate various approaches to account for ANA based on the fixation count cut-offs, definitions for detecting ignored attributes, and methods for modelling ANA. Some of the sustainability attributes identified through eye-tracking measures as being ‘visually ignored’ were truly ignored, whereas in none of the tested approaches was price truly ignored. The adequacy of eye tracking as a visual ANA measure might thus depend on the type of attribute. Further, the study unveiled inconsistencies in identifying non-attenders using visual ANA and the coefficient of variation. Based on our results, we cannot conclude that eye tracking always adequately identifies ANA. However, we identified several major challenges that can assist in further optimising the use of eye tracking in the context of ANA. 1. Introduction A growing body of literature applies choice experiments (CEs) as a valuation method. In a CE, respondents select their preferred alternative from a choice set1 in which each alternative is described by attributes with varying levels. The analysis is based on the economic theory of consumer behaviour (Lancaster, 1966; McFadden, 1974), which assumes continuous preferences and thus unlimited substitutability between the attributes (Ryan and Bate, 2001). This implies that all the attributes presented as well as the trade-offs between attributes are considered (Hensher, Rose and Greene, 2005). However, several studies have questioned the assumption of compensatory behaviour because respondents may ignore some attributes in a choice task (Hensher, Rose and Greene, 2005, 2012; Hole, 2011; Hensher, 2006; Lancsar and Louviere, 2006; Campbell, Hutchinson and Scarpa, 2008; 2011; Scarpa et al. 2009; 2010; Carlsson, Kataria and Lampi, 2010; Hensher and Greene, 2010; Kragt, 2013). Respondents may not make the assumed trade-offs between all the attributes due to attribute non-attendance (ANA), resulting in a violation of the continuity axiom. This decision heuristic has gained increased attention in the CE literature (Hensher, 2014). Not accounting for ANA has been found to affect coefficient estimates and model performance (Campbell, Hutchinson and Scarpa, 2008, 2011; Hensher and Rose, 2009; Scarpa et al. 2009; 2010; Carlsson, Kataria and Lampi, 2010; Mariel, Hoyos and Meyerhoff, 2013). Two primary methods have been proposed to identify ANA in CEs: (i) asking respondents which attributes they ignored (i.e. stated ANA) and (ii) inferring ANA based on observed choices (i.e. inferred ANA). Respondents can be asked whether an attribute was ignored when making a decision at the end of the entire choice task sequence (i.e. serial stated ANA) (Hensher, Rose and Greene, 2005; Carlsson, Kataria and Lampi, 2010; Alemu et al., 2013; Kehlbacher, Balcombe and Bennett, 2013; Kragt, 2013; Scarpa et al., 2013) or after each individual choice task (i.e. choice task stated ANA) (Puckett and Hensher, 2008, 2009; Meyerhoff and Liebe, 2009; Scarpa, Thiene and Hensher, 2010; Caputo et al., 2017). The disadvantage of stated ANA is that these measures are self-reported, which raises concerns about reliability (Hensher and Rose, 2009). For example, responses may be influenced by how the question is asked or interpreted. Respondents may not recall how they chose, or may not answer the attendance statement truthfully (Kragt, 2013; Scarpa et al. 2013), or may bias their answer in a socially desirable manner (Mørkbak, Olsen and Campbell, 2014). Additionally, serial stated ANA questions may be difficult to be answered because respondents may have applied different attribute processing strategies for each choice task (Puckett and Hensher, 2009; Hess and Hensher, 2010). While asking these questions at the end of each choice task allows the respondents to indicate different ANA behaviour for each choice task, it also informs them about the researcher’s interest in their attribute attendance, which may influence their attribute attendance in subsequent choice tasks. Another drawback is the additional financial cost in terms of survey time of repeatedly asking these supplementary questions as well as the increased risk of respondent fatigue, which can lead to more random decision-making (Campbell et al., 2015). The standard modelling approach in stated ANA studies is to restrict the coefficients in the utility function of attributes that have been stated as ignored to zero (Hensher, Rose and Greene, 2005). However, the stated ANA literature suggests that even though respondents stated that they ignored an attribute, it is possible that they may in fact have still attended to it, and it may (perhaps to a lesser extent) have influenced their choice. This would lead to coefficients for these ignored attributes that are in fact significantly different from zero (Campbell and Lorimer, 2009; Hess and Hensher, 2010; Alemu et al. 2013). Thus, assuming attributes to be ignored based on stated ANA may lead to biased results. Rather than relying on self-reported ANA information, a second method infers ANA behaviour using analytical models such as equality constrained latent class (ECLC) models, which impose specific restrictions on the utility functions for each class by constraining some coefficients to zero for selected attributes in a certain class (Scarpa et al. 2009; 2013; Hensher and Greene, 2010; Campbell, Hensher and Scarpa, 2011; Caputo, Nayga and Scarpa, 2013; Kragt, 2013; Lagarde, 2013). Yet another method of inferring ANA is based on the coefficient of variation of individual-specific posterior means (Hess and Hensher, 2010; Scarpa et al. 2013; Mørkbak, Olsen and Campbell, 2014). More research has been called for on other methods to account for ANA (Scarpa et al. 2013; Caputo et al. 2017). In this study, we propose a third method based on visual ANA that is defined as visually ignoring information about attribute levels (Balcombe, Fraser and McSorly, 2015). The use of eye tracking has been widely applied in the fields of marketing and psychology; however, it is relatively new in the field of economics. While some researchers, such as Scarpa et al. (2013), have suggested the use of eye tracking to obtain information on ANA in CEs, limited studies have done so (Balcombe, Fraser and McSorly, 2015; Spinks and Mortimer, 2016; Krucien, Ryan and Hermens, 2017). In the present study, visual attention is measured by eye-tracking equipment during the CE, and eye-fixation counts are used as one of the indicators of visual attention. Based on the fixation counts for a particular attribute, we created a discrete measure of visual attendance by indicating whether a respondent visually attended an attribute or not. Similarly, as in stated ANA studies, we apply the standard approach and develop models in which the coefficients in the utility function are restricted to zero for the attributes identified as visually ignored. Next, these models incorporating visual ANA are compared to a CE model in which full attendance is assumed. We then assess the performance of the visual attendance measure to identify the ignored attributes by testing whether the attributes identified as visually ignored influenced respondents’ choices. To do so, we estimate two coefficients for each attribute (attended and ignored) and test whether the ignored coefficient is different from zero, similar to what has been done in stated ANA studies (Campbell and Lorimer, 2009; Hess and Hensher, 2010; Balcombe, Burton and Rigby, 2011; Alemu et al. 2013). The visual ANA study of Balcombe, Fraser and McSorly (2015) made a few assumptions, such as requiring at least two fixations to consider an attribute visually ‘attended to’ and not allowing ANA behaviour to vary across choice sets by modelling serial ANA. We advance the investigation of visual ANA and extend the work of Balcombe, Fraser and McSorly (2015) in three directions in relation to (i) the fixation count cut-off, (ii) the definition of visual ANA detection in a choice task and (iii) the ANA modelling. Based on these three aspects, a total of six approaches or combinations are applied to incorporate visual ANA in choice models (see the overview in Table 3). Table 3. Overview of the six combinations based on the two modelling approaches to account for visual ANA, the fixation count used as the cut-off, and the two definitions for detecting whether a specific attribute was ignored during a choice task Fixation count FC 2 FC 1 ANA modelling approach Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A* Definition B** Definition A Definition B Abbreviation defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Fixation count FC 2 FC 1 ANA modelling approach Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A* Definition B** Definition A Definition B Abbreviation defA-S defB-S defA-CT defB-CT FC1-S FC1-CT *Based on choice set. **Based on alternatives. Table 3. Overview of the six combinations based on the two modelling approaches to account for visual ANA, the fixation count used as the cut-off, and the two definitions for detecting whether a specific attribute was ignored during a choice task Fixation count FC 2 FC 1 ANA modelling approach Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A* Definition B** Definition A Definition B Abbreviation defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Fixation count FC 2 FC 1 ANA modelling approach Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A* Definition B** Definition A Definition B Abbreviation defA-S defB-S defA-CT defB-CT FC1-S FC1-CT *Based on choice set. **Based on alternatives. 2. Material and methods 2.1. Sustainability labels on coffee Sustainability aspects of food are credence attributes and are thus unobservable to consumers unless explicitly labelled. However, consumers may be overwhelmed with information in a shopping environment and may not pay attention to all food labels (Grunert, 2011). We apply our study to the case of sustainability labelling on coffee, as coffee is one of the most popular sustainability-labelled food products. Many US coffee products carry sustainability labels such as Fair Trade (e.g. Fair Trade USA), Rainforest Alliance and USDA Organic, which are all included in our study (see also Van Loo et al., 2015). Coffee producers are often certified for more than one type of label. For example, in 2013, 77 per cent of the Fair Trade certified producer organisations reported holding at least one additional certification (52 per cent Organic and 12 per cent Rainforest Alliance) (Fairtrade International, 2015, p. 59). For coffee specifically, approximately 37 per cent of Fair Trade coffee is also organic certified (Fairtrade International, 2015). Due to the proliferation of sustainability labels for coffee, coffee packages often carry several of these labels (Pierrot, Giovannucci and Kasterine, 2011). 2.2. Experimental design of the CE Participants were recruited from a consumer profile database (N = 6,500) of the University of Arkansas Sensory Service Center (Fayetteville, AR, USA) which includes area residents. In total, 81 consumers who had purchased coffee in the 2 months preceding the study (March and April 2013) and who had no history of eye disease or eye surgery participated in the study. While the number of participants could be considered low for a CE, it is relatively high for an eye-tracking study. Each participant was given a $20 gift card. Approximately half (53 per cent) of the participants were female (Table 1). Each age and income category is represented. The sample is slightly biased towards participants with higher education. Table 1. Socio-demographic characteristics of the sample (%, n = 81) Gender  Male 46.9  Female 53.1 Age group  18–24 years 17.3  25–34 years 37.0  35–44 years 21.0  45–54 years 14.8  55 and older 9.9 Children  Yes 54.3  No 45.7 Educational level completed  High school/GED or lower 7.4  Some college or 2-year college degree 27.2  4-year college degree (BA, BS) 39.5  Master’s or PhD degree 25.9 Annual household income  Less than $20,000 25.9  $20,000–$49,999 35.8  $50,000–$99,999 25.9  More than $100,000 12.4 Gender  Male 46.9  Female 53.1 Age group  18–24 years 17.3  25–34 years 37.0  35–44 years 21.0  45–54 years 14.8  55 and older 9.9 Children  Yes 54.3  No 45.7 Educational level completed  High school/GED or lower 7.4  Some college or 2-year college degree 27.2  4-year college degree (BA, BS) 39.5  Master’s or PhD degree 25.9 Annual household income  Less than $20,000 25.9  $20,000–$49,999 35.8  $50,000–$99,999 25.9  More than $100,000 12.4 Table 1. Socio-demographic characteristics of the sample (%, n = 81) Gender  Male 46.9  Female 53.1 Age group  18–24 years 17.3  25–34 years 37.0  35–44 years 21.0  45–54 years 14.8  55 and older 9.9 Children  Yes 54.3  No 45.7 Educational level completed  High school/GED or lower 7.4  Some college or 2-year college degree 27.2  4-year college degree (BA, BS) 39.5  Master’s or PhD degree 25.9 Annual household income  Less than $20,000 25.9  $20,000–$49,999 35.8  $50,000–$99,999 25.9  More than $100,000 12.4 Gender  Male 46.9  Female 53.1 Age group  18–24 years 17.3  25–34 years 37.0  35–44 years 21.0  45–54 years 14.8  55 and older 9.9 Children  Yes 54.3  No 45.7 Educational level completed  High school/GED or lower 7.4  Some college or 2-year college degree 27.2  4-year college degree (BA, BS) 39.5  Master’s or PhD degree 25.9 Annual household income  Less than $20,000 25.9  $20,000–$49,999 35.8  $50,000–$99,999 25.9  More than $100,000 12.4 All coffee products in the experiment were ground medium roast coffee, the most popular type of coffee in the USA (Mintel, 2012). The coffee products were described using a combination of five attributes: four sustainability labels and price (Table 2). Each of the sustainability labels had two levels (present or not present) and the price attribute had four levels based on a store check in food stores in Fayetteville (AR) USA in April 2013. Table 2. Attributes and levels used in the CE Attributes Level Fair Trade label USDA Organic label 0 = Not present Rainforest Alliance label 1 = Present Carbon Footprint label Price (per 12 ounces) $4.30 $6.30 $8.30 $10.30 Attributes Level Fair Trade label USDA Organic label 0 = Not present Rainforest Alliance label 1 = Present Carbon Footprint label Price (per 12 ounces) $4.30 $6.30 $8.30 $10.30 Table 2. Attributes and levels used in the CE Attributes Level Fair Trade label USDA Organic label 0 = Not present Rainforest Alliance label 1 = Present Carbon Footprint label Price (per 12 ounces) $4.30 $6.30 $8.30 $10.30 Attributes Level Fair Trade label USDA Organic label 0 = Not present Rainforest Alliance label 1 = Present Carbon Footprint label Price (per 12 ounces) $4.30 $6.30 $8.30 $10.30 Given the four sustainability attributes and their levels, a full factorial design would have resulted in 64 (24 × 4) alternatives resulting in 4,096 choice sets–each consisting of two alternatives. The CE design followed Street and Burgess (2007). We used a fractional factorial design and the generator ([1 1 1 1 1]) to obtain the design of eight choice sets (Burgess, 2007; Street and Burgess, 2007) with an efficiency of 97.6 per cent. To increase the similarity to a real shopping experience, a no-buy alternative was added. Hence, in each choice set, participants were presented with two types of ground medium roasted coffee as well as a no-buy alternative (Figure 1). Due to the hypothetical nature of our CE, a cheap talk script was presented prior to the choice tasks. The location of the labels (USDA Organic, Rainforest Alliance, Fair Trade, Carbon Footprint) on the package (from left to right) was randomised for each of the eight choice sets to avoid an order effect due to label location. In addition, the randomisation for each of the choice sets was repeated 10 times, resulting in 10 different surveys. Each respondent was randomly assigned to one of these 10 surveys. Additionally, within each survey, the eight choice sets were randomly presented to avoid order effects. Fig. 1. View largeDownload slide Example of a choice set with AOI. Frames indicate the AOI for the Carbon Footprint label, Rainforest Alliance label, Fair Trade label and USDA Organic label and price. Fig. 1. View largeDownload slide Example of a choice set with AOI. Frames indicate the AOI for the Carbon Footprint label, Rainforest Alliance label, Fair Trade label and USDA Organic label and price. 2.3. Experimental procedure for the eye-tracking experiment Participants’ visual attention was recorded using a contact-free eye-tracking device (model: RED, SensoMotoric Instruments GmbH (SMI), Teltow, Germany) located in a panel beneath a 56 cm computer screen (474 mm by 297 mm and screen resolution of 1680 px by 1050 px) (see Appendix A). The approximate distance between the display monitor and each participant’s head was 70 cm. The sampling rate and tracking resolution of the eye-tracking device were 120 Hz and 0.03°, respectively. Visual stimuli, which are the images presented on the screen representing the choice set, were randomly presented using stimulus presentation software (Experiment Suite 360°TM, SensoMotoric Instruments GmbH, Teltow, Germany). The eye-tracking device was individually calibrated using a five-point calibration method with a low mean tracking error (less than 0.4°). After successful calibration, two warm-up choice sets were presented to familiarise each participant with the experimental procedures. As in Balcombe, Fraser and McSorly (2015), participants knew that eye tracking was used; however, they were not aware of its purpose. Between the choice tasks (i.e. during the inter-stimulus intervals), participants were asked to maintain their fixation on a central black cross against a white background for approximately 8 s. Similar to the procedure used in Balcombe, Fraser and McSorly (2015), the participants viewed each choice set as long as they wanted before indicating their choice. For three choice sets, the choice was not indicated, resulting in a total of 645 choices. On an average, the participants spent 73 s on all eight choice tasks combined (without the inter-stimulus intervals), which equates to an average of less than 10 s per choice task. The average tracking ratio, i.e. ‘the number of non-zero gaze positions divided by the sampling frequency multiplied by run duration, expressed in per cent’ (SMI, 2016, p. 335), across our sample is 89.5 per cent. 2.4. Eye-tracking measures Areas of interest (AOI) were defined on the coffee packages (Figure 1) corresponding to the five attributes. Using the eye-tracking software (BeGazeTM, ver. 3.0, SensoMotoric Instruments GmbH, Teltow, Germany), fixation counts were calculated for the five AOIs in each of the eight choice sets. The fixation count is the number of times the participant fixated his or her gaze on the AOI. More fixations are an indication that an area is more noticeable or more important to the viewer than other areas (Poole, Ball and Phillips, 2005). The number of fixations within the AOI has been considered a reliable measure for the visual attention given to that AOI (Bialkova and van Trijp, 2011). The low speed event detection method (suggested for <200 Hz) was selected in BeGaze for the fixation detection. In this method, the fixation is the primary event, and other events are derived from it. The method uses two specific detection parameters: a minimum fixation duration (i.e. the minimum time window in which the gaze is analysed) of 80 ms and a maximum dispersion of 100 px. The low-speed event detection method uses a dispersion-based algorithm. For details on this algorithm, we refer to the BeGaze Manual 3.6 (SMI, 2016, p. 317). For each stimulus, the first fixation was excluded. According to Holmqvist et al. (2011), this approach is often used because the fixation position at stimulus onset has not been influenced by the stimulus content. 2.5. Accounting for visual ANA In this study, three aspects are taken into account when going from fixation counts to incorporating visual ANA in the choice modelling: (i) fixation count cut-off, (ii) definition of visual ANA to identify an attribute as ignored in a choice task and (iii) modelling approach for visual ANA. 2.5.1. Fixation count cut-offs Balcombe, Fraser and McSorly (2015) assumed that at least two fixation counts are required to consider an attribute ‘attended to’. However, even one fixation count can signify that the person fixated on the information and thus may have attended to the information. Hence, in addition to the arbitrary assumption by Balcombe, Fraser and McSorly (2015) that two fixation counts are required to consider an attribute visually attended to, we also used a fixation count of one as a less strict threshold to consider an attribute as being visually attended to. 2.5.2. Defining visual ANA To define visual ANA, we used the fixation count as a measure of visual attention and created the discrete measure ‘visual ANA’.2 We used two definitions to identify an attribute as being ‘ignored’ in a particular choice task. Based on visual attention to a specific attribute in the choice set as a whole (Definition A). This definition was used by Balcombe, Fraser and McSorly (2015), who considered an attribute to be ignored in a choice task if the fixation count for the attribute summated over the alternatives within one choice set was below the cut-off. Thus, the fixation count for one attribute is calculated for the choice set as a whole. Based on visual attention to a specific attribute in each of the alternatives within the choice set (Definition B). An attribute is judged to be ignored in a given choice task if the attribute was ignored (fixation count less than the cut-off) in both of the designed alternatives (if the attribute was present in both alternatives). Thus, in this second definition, visual attention to the attribute in each alternative is considered. Notably, the use of Definition A or B will only have an impact on the price attribute, as it is present in both of the designed alternatives, and the alternatives never share any of the quality attributes. We assumed that respondents, when looking at one package, did not infer which labels the other package had. 2.5.3. Modelling approaches for visual ANA Two modelling approaches were used to account for visual ANA, one at the respondent level (serial visual ANA) and one at the choice set level (choice task visual ANA). Serial ANA refers to classifying a respondent as an attender or non-attender for a particular attribute for the entire choice task sequence, while choice task ANA allows for differences in attendance across choice tasks. Serial and choice task visual ANA are similar to serial and choice task stated ANA but instead of stated attendance, visual attendance is used. Serial visual ANA: Following Balcombe, Fraser and McSorly (2015), we classified a respondent as having visually ignored an attribute over the whole sequence of choice tasks if the participant ignored a given attribute in more than half (i.e. more than four) of the choice tasks. Thus, using serial visual ANA, a person is either a non-attender or an attender of an attribute for the entire sequence of choice tasks in the CE. Choice task visual ANA: Several authors, however, have advocated that respondents’ processing strategies may change as they progress through a sequence of choice tasks, meaning that their tendency to ignore attributes may not be consistent throughout a panel of choices (Meyerhoff and Liebe, 2009; Puckett and Hensher, 2009; Hess and Hensher, 2010; Scarpa, Thiene and Hensher, 2010). Hence, it may be important to allow for varying ANA behaviour from one choice task to another. Choice task visual ANA allows visual ANA to vary across choice tasks. When a respondent visually ignored a given attribute in a choice task, this attribute is characterised as non-attended for that particular choice task. Therefore, for each choice set and for each attribute, a participant is classified as having attended or not attended to the attribute. For each of the two visual ANA definitions to detect ignored attributes (Definition A and Definition B), both modelling approaches – serial visual ANA (S) and choice task visual ANA (CT) – were applied. This leads to four combinations – defA-CT, defA-S, defB-CT and defB-S (Table 3) – when a fixation count of two is applied as the cut-off. Definition A combined with the serial ANA modelling method with a fixation count of two as the cut-off (defA-S) is the approach used by Balcombe, Fraser and McSorly (2015). For the cut-off fixation count of one, definitions A and B result in the exact same discrete measure of visual ANA; thus, no distinction is made between definitions A and B (see Table 3). Therefore, a fixation count of one as the cut-off results in two additional combinations or approaches (FC1-S and FC1-CT). 2.6. Discrete choice models While the multinomial model (MNL) assumes homogeneity in consumer preferences, we assume that heterogeneity may be an issue in analysing consumer preferences for food labelling (Bonnet and Simioni, 2001; Van Loo et al., 2014). Therefore, a random parameter logit (RPL) model was estimated (with 500 Halton draws) that allows for random taste variation and for the panel structure, given that each respondent made eight choices. This approach results in the estimation of a mean and standard deviation for each of the random taste parameters. For simplicity, we assume price to be a fixed coefficient, which is a widely practised specification in the field (Revelt and Train, 1998; Layton and Brown, 2000; Morey and Rossmann, 2003; Lusk and Schroeder, 2004; Caputo, Nayga and Scarpa, 2013). This restriction allows the distribution of the WTP to be easily calculated from the non-price coefficients. We further assume that the coefficients of the four sustainability labels follow a normal distribution (Lusk and Schroeder, 2004; Caputo, Nayga and Scarpa, 2013). Two additional modelling issues are taken into account – the correlations across taste parameters and across utilities – to make the estimates more robust and consistent with consumer choice behaviour (Barreiro-Hurle, Gracia and de-Magistris, 2010; Gracia, Barreiro-Hurle and Perez, 2012, 2014). To allow for dependence across tastes, no restrictions were applied to the correlations among the random parameters. Additionally, because the design consists of two designed alternatives and one no-buy alternative, correlations across utilities may exist (Scarpa, Ferrini and Willis, 2005). The no-buy alternative is truly experienced, while the designed alternatives can only be imagined. Therefore, the utilities of the buying alternatives are likely to be more correlated among themselves than with the no-buy alternative. To account for this correlation pattern, we employed an RPL model with an error component (RPL-EC) (Scarpa, Ferrini and Willis, 2005; 2007), whereby both designed alternatives share an extra error component that is a zero-mean normally distributed random parameter. Specifically, with our attributes, the utility that individual i obtains from alternative j at choice situation t takes the following form: Uijt=β0No_Buyijt+β1Organicijt+β2Rainforestijt+β3FairTradeijt+β4CarbonFootprintijt+β5Priceijt+ηij(1−No_Buyijt)+εijt, where j pertains to alternatives A, B and C. No_Buyijt is an indicator variable that takes the value of 1 when the no-buy alternative is chosen and 0 when either product profile A or B is selected. β0 is an alternative-specific constant representing the no-buy alternative. Price is the price (US$) of a package of 12 ounces of coffee. ηij is the zero-mean normal error term, or the error component term, which is only associated with alternatives that portray a purchase decision and is absent in the utility of the no-buy alternative. εijt is the unobserved random error term. The marginal WTP values are calculated as a negative ratio, where the numerator is the estimated mean values of the coefficients associated with a particular sustainability label and the denominator is the price coefficient. The data were analysed using NLOGIT 5.0 by Econometric Software Inc. (Greene, 1990). 2.7. Accounting for ANA The standard approach to account for stated ANA is to restrict the coefficient in the utility function to zero for the attributes that the respondents stated as ignored, which results in the removal of the attribute from the choice consideration (Hensher, Rose and Greene, 2005). This method has been incorporated into the NLOGIT 5.0 software by coding an attribute as −888 if it is not attended to and assigns a zero to the attribute coefficients rather than to the attribute levels (Greene, 2012). This approach has been applied in several studies on stated ANA (Hensher, Rose and Greene, 2005, 2009; Alemu et al. 2013; Kragt, 2013; Scarpa et al. 2013). We used the same approach, using a dummy variable to denote whether the attribute was visually ignored (visual ANA). This is defined at the choice set level (choice task ANA) or at the respondent level (serial ANA). We did not collect any data on stated ANA. For each of the six combinations (defA-CT, defA-S, defB-CT, defB-S, FC1-S, FC1-CT), we estimated a visual ANA model in which the coefficients of the visually ignored attributes are restricted to zero. In addition, a full attendance model in which all attributes are assumed to be attended to was estimated. 2.8. Are the attributes identified as visually non-attended actually ignored? We examine whether the attributes identified as visually non-attended are in fact also truly ignored3 when respondents are making the choice by estimating the coefficient of ignored attributes (Section 2.8.1). In addition, we compare the results of serial visual ANA with results from the coefficient of variation method (Section 2.8.2), an inferred method to identify non-attenders. 2.8.1. Coefficients of ignored attributes The coefficients of the ignored attributes are no longer restricted to zero; but are freely estimated. In the stated ANA literature, some studies have indicated that people reporting to have ignored a certain attribute may have a marginal utility for that attribute that differs from zero. Hence, respondents who stated that they ignored an attribute may have actually considered it (Carlsson, Kataria and Lampi, 2010). As a result, instead of restricting the coefficient of ignored attributes to zero, some stated ANA studies estimate two coefficients for each attribute: one for those who stated that they attended to the attribute and one for those who stated that they did not attend to the attribute (Campbell and Lorimer, 2009; Hess and Hensher, 2010; Alemu et al. 2013; Scarpa et al. 2013). Similarly, we estimate models with two coefficients for each attribute: one for visually ignored attributes and one for visually attended attributes. If the visually non-attended attributes were truly ignored, the corresponding coefficient should not differ statistically from zero. Because of the number of coefficients to be estimated and our limited sample size, MNL models were estimated. 2.8.2. Coefficient of variation of individual-specific coefficient distributions Using the serial visual ANA, a respondent is identified as an attender or a non-attender for a particular attribute. For each respondent, the allocation of having ignored an attribute or not based on serial visual ANA is compared with the allocation of this respondent as an attender or non-attender based on the inferred method using the coefficient of variation.4 Following Hess and Hensher (2010), we attempted to infer whether a respondent ignored a particular attribute or not (thus, inferred serial ANA) by analysing the individual-specific coefficient distributions that have been conditioned on observed choices. For additional details, we refer readers to the NLOGIT reference guide (Greene, 2012, section N29.8). Based on the RPL model in Table 5, the means and standard deviations for the conditional distributions were calculated for each coefficient of the random parameters. Rather than using the conditional mean to infer whether a respondent ignored an attribute or not, Hess and Hensher (2010) suggested using the coefficient of variation as an inferred method to identify ignored attributes. The coefficient of variation is the ratio between the standard deviation and the mean of the conditional distribution. This measure is used by Hess and Hensher (2010, p. 786) to incorporate uncertainty into the conditional distributions and gives an indication of whether the conditional mean is indistinguishable from zero. Hess and Hensher (2010) reported that this approach is better than using the conditional mean because a respondent may have a low sensitivity to an attribute without actually ignoring it. Hence, relying only on a low mean to allocate a respondent to the ignored group might be incorrect, and therefore using the coefficient of variation is suggested. Following Hess and Hensher (2010), we allocate respondents with a coefficient of variation of two or above for a certain attribute to the ignored group for that attribute. Subsequently, we evaluate whether the identification of respondents as having ignored or not ignored an attribute based on the serial visual ANA matches with the identification based on the coefficient of variation. Table 5. RPL model with error component (RPL-EC) parameter estimates (n = 645) Fixation count FC 2 FC 1 ANA modelling Full-AA Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Fair Trade Mean 0.63** (0.29) 0.73** (0.36) 0.75** (0.33) 0.76*** (0.27) 0.81*** (0.22) 0.65** (0.31) 0.65** (0.29) Standard deviation 0.83*** (0.28) 0.79** (0.32) 0.73** (0.37) 0.65 (0.51) 0.54 (0.47) 0.58* (0.29) 0.40 (0.30) USDA Organic Mean 1.02*** (0.35) 1.41*** (0.50) 1.35*** (0.47) 0.87*** (0.25) 0.84*** (0.24) 1.09*** (0.40) 0.77*** (0.26) Standard deviation 1.23*** (0.33) 0.85* (0.52) 0.89 (1.05) 0.39 (0.49) 0.18 (1.14) 1.13** (0.50) 0.84* (0.47) Rainforest Alliance Mean 0.74*** (0.26) 0.73*** (0.26) 0.68*** (0.23) 0.60*** (0.23) 0.60*** (0.21) 0.75*** (0.27) 0.59*** (0.20) Standard deviation 0.58 (0.37) 0.45 (0.41) 0.33 (0.44) 0.42 (0.73) 0.15 (1.60) 0.55 (0.52) 0.36 (0.52) Carbon Footprint Mean 0.30 (0.26) 0.38 (0.36) 0.40 (0.32) 0.50** (0.23) 0.58*** (0.22) 0.27 (0.31) 0.34 (0.22) Standard deviation 0.90* (0.48) 0.72 (0.74) 0.71 (0.98) 0.39 (0.77) 0.39 (1.27) 0.62 (0.50) 0.39 (1.03) Price −0.85*** (0.05) −0.81*** (0.05) −0.84*** (0.05) −0.61*** (0.05) −0.54*** (0.06) −0.77*** (0.05) −0.68*** (0.04) No_Buy −8.76*** (0.86) −8.91*** (0.85) −9.13*** (0.98) −7.35*** (0.90) −6.75*** (1.01) −8.01*** (0.81) −7.61*** (0.81) Error component Standard deviation 2.71*** (0.91) 3.15*** (1.15) 3.63*** (0.85) 3.03*** (0.96) 3.51*** (0.93) 2.55*** (0.96) 2.63*** (0.90) Log likelihood −349 −354 −370 −404 −433 −357 −375  AIC 731.2 741.2 774.1 842.4 900.6 747.9 783.6  BIC 807.2 817.2 850.0 918.4 976.6 823.9 859.6 Fixation count FC 2 FC 1 ANA modelling Full-AA Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Fair Trade Mean 0.63** (0.29) 0.73** (0.36) 0.75** (0.33) 0.76*** (0.27) 0.81*** (0.22) 0.65** (0.31) 0.65** (0.29) Standard deviation 0.83*** (0.28) 0.79** (0.32) 0.73** (0.37) 0.65 (0.51) 0.54 (0.47) 0.58* (0.29) 0.40 (0.30) USDA Organic Mean 1.02*** (0.35) 1.41*** (0.50) 1.35*** (0.47) 0.87*** (0.25) 0.84*** (0.24) 1.09*** (0.40) 0.77*** (0.26) Standard deviation 1.23*** (0.33) 0.85* (0.52) 0.89 (1.05) 0.39 (0.49) 0.18 (1.14) 1.13** (0.50) 0.84* (0.47) Rainforest Alliance Mean 0.74*** (0.26) 0.73*** (0.26) 0.68*** (0.23) 0.60*** (0.23) 0.60*** (0.21) 0.75*** (0.27) 0.59*** (0.20) Standard deviation 0.58 (0.37) 0.45 (0.41) 0.33 (0.44) 0.42 (0.73) 0.15 (1.60) 0.55 (0.52) 0.36 (0.52) Carbon Footprint Mean 0.30 (0.26) 0.38 (0.36) 0.40 (0.32) 0.50** (0.23) 0.58*** (0.22) 0.27 (0.31) 0.34 (0.22) Standard deviation 0.90* (0.48) 0.72 (0.74) 0.71 (0.98) 0.39 (0.77) 0.39 (1.27) 0.62 (0.50) 0.39 (1.03) Price −0.85*** (0.05) −0.81*** (0.05) −0.84*** (0.05) −0.61*** (0.05) −0.54*** (0.06) −0.77*** (0.05) −0.68*** (0.04) No_Buy −8.76*** (0.86) −8.91*** (0.85) −9.13*** (0.98) −7.35*** (0.90) −6.75*** (1.01) −8.01*** (0.81) −7.61*** (0.81) Error component Standard deviation 2.71*** (0.91) 3.15*** (1.15) 3.63*** (0.85) 3.03*** (0.96) 3.51*** (0.93) 2.55*** (0.96) 2.63*** (0.90) Log likelihood −349 −354 −370 −404 −433 −357 −375  AIC 731.2 741.2 774.1 842.4 900.6 747.9 783.6  BIC 807.2 817.2 850.0 918.4 976.6 823.9 859.6 Note: *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. Table 5. RPL model with error component (RPL-EC) parameter estimates (n = 645) Fixation count FC 2 FC 1 ANA modelling Full-AA Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Fair Trade Mean 0.63** (0.29) 0.73** (0.36) 0.75** (0.33) 0.76*** (0.27) 0.81*** (0.22) 0.65** (0.31) 0.65** (0.29) Standard deviation 0.83*** (0.28) 0.79** (0.32) 0.73** (0.37) 0.65 (0.51) 0.54 (0.47) 0.58* (0.29) 0.40 (0.30) USDA Organic Mean 1.02*** (0.35) 1.41*** (0.50) 1.35*** (0.47) 0.87*** (0.25) 0.84*** (0.24) 1.09*** (0.40) 0.77*** (0.26) Standard deviation 1.23*** (0.33) 0.85* (0.52) 0.89 (1.05) 0.39 (0.49) 0.18 (1.14) 1.13** (0.50) 0.84* (0.47) Rainforest Alliance Mean 0.74*** (0.26) 0.73*** (0.26) 0.68*** (0.23) 0.60*** (0.23) 0.60*** (0.21) 0.75*** (0.27) 0.59*** (0.20) Standard deviation 0.58 (0.37) 0.45 (0.41) 0.33 (0.44) 0.42 (0.73) 0.15 (1.60) 0.55 (0.52) 0.36 (0.52) Carbon Footprint Mean 0.30 (0.26) 0.38 (0.36) 0.40 (0.32) 0.50** (0.23) 0.58*** (0.22) 0.27 (0.31) 0.34 (0.22) Standard deviation 0.90* (0.48) 0.72 (0.74) 0.71 (0.98) 0.39 (0.77) 0.39 (1.27) 0.62 (0.50) 0.39 (1.03) Price −0.85*** (0.05) −0.81*** (0.05) −0.84*** (0.05) −0.61*** (0.05) −0.54*** (0.06) −0.77*** (0.05) −0.68*** (0.04) No_Buy −8.76*** (0.86) −8.91*** (0.85) −9.13*** (0.98) −7.35*** (0.90) −6.75*** (1.01) −8.01*** (0.81) −7.61*** (0.81) Error component Standard deviation 2.71*** (0.91) 3.15*** (1.15) 3.63*** (0.85) 3.03*** (0.96) 3.51*** (0.93) 2.55*** (0.96) 2.63*** (0.90) Log likelihood −349 −354 −370 −404 −433 −357 −375  AIC 731.2 741.2 774.1 842.4 900.6 747.9 783.6  BIC 807.2 817.2 850.0 918.4 976.6 823.9 859.6 Fixation count FC 2 FC 1 ANA modelling Full-AA Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S defB-S defA-CT defB-CT FC1-S FC1-CT Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Fair Trade Mean 0.63** (0.29) 0.73** (0.36) 0.75** (0.33) 0.76*** (0.27) 0.81*** (0.22) 0.65** (0.31) 0.65** (0.29) Standard deviation 0.83*** (0.28) 0.79** (0.32) 0.73** (0.37) 0.65 (0.51) 0.54 (0.47) 0.58* (0.29) 0.40 (0.30) USDA Organic Mean 1.02*** (0.35) 1.41*** (0.50) 1.35*** (0.47) 0.87*** (0.25) 0.84*** (0.24) 1.09*** (0.40) 0.77*** (0.26) Standard deviation 1.23*** (0.33) 0.85* (0.52) 0.89 (1.05) 0.39 (0.49) 0.18 (1.14) 1.13** (0.50) 0.84* (0.47) Rainforest Alliance Mean 0.74*** (0.26) 0.73*** (0.26) 0.68*** (0.23) 0.60*** (0.23) 0.60*** (0.21) 0.75*** (0.27) 0.59*** (0.20) Standard deviation 0.58 (0.37) 0.45 (0.41) 0.33 (0.44) 0.42 (0.73) 0.15 (1.60) 0.55 (0.52) 0.36 (0.52) Carbon Footprint Mean 0.30 (0.26) 0.38 (0.36) 0.40 (0.32) 0.50** (0.23) 0.58*** (0.22) 0.27 (0.31) 0.34 (0.22) Standard deviation 0.90* (0.48) 0.72 (0.74) 0.71 (0.98) 0.39 (0.77) 0.39 (1.27) 0.62 (0.50) 0.39 (1.03) Price −0.85*** (0.05) −0.81*** (0.05) −0.84*** (0.05) −0.61*** (0.05) −0.54*** (0.06) −0.77*** (0.05) −0.68*** (0.04) No_Buy −8.76*** (0.86) −8.91*** (0.85) −9.13*** (0.98) −7.35*** (0.90) −6.75*** (1.01) −8.01*** (0.81) −7.61*** (0.81) Error component Standard deviation 2.71*** (0.91) 3.15*** (1.15) 3.63*** (0.85) 3.03*** (0.96) 3.51*** (0.93) 2.55*** (0.96) 2.63*** (0.90) Log likelihood −349 −354 −370 −404 −433 −357 −375  AIC 731.2 741.2 774.1 842.4 900.6 747.9 783.6  BIC 807.2 817.2 850.0 918.4 976.6 823.9 859.6 Note: *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. 3. Results and discussion 3.1. Visual attribute ANA frequency In our specifications of visual ANA (Definition A or B), the modelling approach for visual ANA (serial or choice task) as well as the fixation count cut-off determine the ANA frequency. The proportions of ANA for each of the attributes and for the six different combinations are presented in Table 4. Table 4. Proportions (%) of choice task and serial visual ANA depending on the definition applied and fixation count (FC) (n = 645) FC 2 FC 1 Serial ANA Choice task ANA Serial ANA Choice task ANA Definition A Definition B Definition A Definition B Fair Trade 40.7 40.7 52.2 52.2 12.3 28.1 USDA Organic 55.6 55.6 59.6 59.6 23.5 30.4 Rainforest Alliance 43.2 43.2 54.9 54.9 17.3 30.1 Carbon Footprint 40.7 40.7 52.9 52.9 16.0 28.1 Price 12.3 23.4 24.5 40.7 4.9 9.6 FC 2 FC 1 Serial ANA Choice task ANA Serial ANA Choice task ANA Definition A Definition B Definition A Definition B Fair Trade 40.7 40.7 52.2 52.2 12.3 28.1 USDA Organic 55.6 55.6 59.6 59.6 23.5 30.4 Rainforest Alliance 43.2 43.2 54.9 54.9 17.3 30.1 Carbon Footprint 40.7 40.7 52.9 52.9 16.0 28.1 Price 12.3 23.4 24.5 40.7 4.9 9.6 Table 4. Proportions (%) of choice task and serial visual ANA depending on the definition applied and fixation count (FC) (n = 645) FC 2 FC 1 Serial ANA Choice task ANA Serial ANA Choice task ANA Definition A Definition B Definition A Definition B Fair Trade 40.7 40.7 52.2 52.2 12.3 28.1 USDA Organic 55.6 55.6 59.6 59.6 23.5 30.4 Rainforest Alliance 43.2 43.2 54.9 54.9 17.3 30.1 Carbon Footprint 40.7 40.7 52.9 52.9 16.0 28.1 Price 12.3 23.4 24.5 40.7 4.9 9.6 FC 2 FC 1 Serial ANA Choice task ANA Serial ANA Choice task ANA Definition A Definition B Definition A Definition B Fair Trade 40.7 40.7 52.2 52.2 12.3 28.1 USDA Organic 55.6 55.6 59.6 59.6 23.5 30.4 Rainforest Alliance 43.2 43.2 54.9 54.9 17.3 30.1 Carbon Footprint 40.7 40.7 52.9 52.9 16.0 28.1 Price 12.3 23.4 24.5 40.7 4.9 9.6 For serial visual ANA, the proportions of visual non-attenders for the sustainability labels range from 41 per cent to 56 per cent of the total number of participants for a fixation count of 2, while the proportions range from 12 per cent to 23 per cent for a fixation count of one. For both fixation counts, the number of respondents who ignore the price is lower than for the sustainability labels. Using a fixation count of 2 as the cut-off, only 12 per cent and 23 per cent of the respondents were classified as visual non-attenders for price for Definition A and Definition B, respectively.5 For a fixation count of one, only 5 per cent of the respondents were classified as ignoring the price. When applying choice task visual ANA, the number of choice tasks in which the sustainability labels were ignored ranges from 52 per cent to 60 per cent for a fixation count of two as the cut-off and from 28 per cent to 30 per cent for a fixation count of one as the cut-off. Price was ignored in 25 per cent or 41 per cent of the choice tasks depending on the ANA definition applied for the case of a fixation count of two as the cut-off and in 10 per cent of the choice tasks for a fixation count of one. 3.2. Standard ANA approach Similar to the standard approach in stated ANA, models are estimated with the parameters for the visually ignored attributes being constrained to zero; i.e. four models for the combinations with a fixation count of two: serial and choice task visual ANA based on Definition A (defA-CT, defA-S) and based on Definition B (defB-CT and defB-S) (Table 5) and two additional models for a fixation count of one: choice task and serial visual ANA modelling approaches (FC1-S and FC1-CT). The full attendance model (full-AA) pertains to the estimation assuming full attribute attendance and is included as a benchmark. To allow for heterogeneous preferences among the respondents and correlation across utilities, RPL-EC models were estimated (Table 5). The MNL model estimations are included in Appendix B. In all the models, the coefficient of the no-buy alternative is negative and statistically significant, indicating that participants increase their utility when choosing one of the proposed coffee alternatives compared with the no-buy alternative. In all of the models, the hypothesis of correlation across utilities is verified because the standard deviation of the error component (ηij) for the purchase alternatives is statistically significant. Correlations across the random parameters were also allowed. The Cholesky matrices are presented in Appendix C. The coefficients of the attributes have the expected signs. The price coefficient is negative, as expected, and statistically significant at the 0.01 level, indicating that consumers’ utility decreases with increasing price. In the full-AA model, the coefficients of Organic, Rainforest Alliance and Fair Trade are significant, implying that respondents’ utility increases when one of these labels is present on a coffee package. The results show that USDA Organic is the highest valued attribute, resulting in the strongest utility increase. The USDA Organic label is preferred over Rainforest Alliance and Fair Trade. The full-AA model has significant standard deviations of the random parameters (except for Rainforest Alliance), indicating the presence of considerable unobserved heterogeneity in taste preferences across the respondents. Turning to the standard ANA approach in which the parameters of the visually ignored attributes are restricted to zero (defA-CT, defA-S, defB-CT, defB-S, FC1-S and FC1-CT), we find that most of the parameters for the attended attributes are significant at the 5 per cent or 1 per cent level. In all six models, the coefficient of USDA Organic is the largest. Carbon Footprint is not significant for all models except for the choice task modelling approach with a fixation count of two as the cut-off. While the standard deviations of the random parameters for Fair Trade and USDA Organic of the full attendance model were significant at 1 per cent, this is no longer the case when accounting for visual ANA.6 While the full attendance model with significant standard deviations shows preference heterogeneity, accounting for visual ANA captures an important part of the heterogeneity across participants. This result illustrates that confounding between ANA and preference heterogeneity might be an issue (Hess et al., 2013), and thus preference heterogeneity may be incorrectly interpreted when ANA is not addressed, which further illustrates its importance (Hess et al. 2013). The full attendance model outperforms the model in which the coefficients of the ignored attributes are restricted to zero, as illustrated by the decrease in model fit for the visual ANA models (a decrease in log likelihood and an increase in the AIC and BIC statistics) (Table 5). A worse model fit could be due to treating attributes as having a zero parameter when some are not actually ignored. 3.3. Are the attributes identified as visually non-attended actually ignored? 3.3.1. Coefficients of ignored attributes We test whether the attributes identified as visually non-attended truly have coefficients that are equal to zero by estimating them freely, which leads to model estimations with two coefficients for each attribute (attended and ignored) that are referred to as defA-S2, defB-S2, defA-CT2, and defB-CT2 as well as FC1-S2 and FC1-CT2 for fixation counts two and one as the cut-off, respectively (Table 6). Of the 30 ignored coefficients,7 17 are not significantly different from zero. In the cases of the Rainforest Alliance, Fair Trade and Carbon Footprint labels, being identified as visually non-attended using one of the six approaches means that these attributes were truly ignored, with the exception of Rainforest Alliance in the choice task ANA modelling approach with a fixation count of two. Table 6. MNL parameter estimations with two coefficients (attended and ignored) (n = 645) Fixation count FC 2 FC 1 ANA modelling Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S2 defB-S2 defA-CT2 defB-CT2 FC1-S2 FC1-CT2 Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Attended  Fair Trade 0.54*** (0.16) 0.55*** (0.17) 0.72*** (0.18) 0.72*** (0.18) 0.43*** (0.14) 0.46*** (0.15)  USDA Organic 1.46*** (0.20) 1.45*** (0.20) 1.11*** (0.20) 1.12*** (0.20) 0.99*** (0.16) 0.94*** (0.16)  Rainforest Alliance 0.86*** (0.18) 0.87*** (0.18) 0.93*** (0.19) 0.93*** (0.19) 0.72*** (0.15) 0.73*** (0.16)  Carbon Footprint 0.16 (0.17) 0.18 (0.17) 0.37** (0.18) 0.38** (0.18) 0.06 (0.14) 0.17 (0.15)  Price −0.65*** (0.05) −0.65*** (0.05) −0.62*** (0.05) −0.63*** (0.05) −0.61*** (0.05) −0.62*** (0.05) Ignored  Fair Trade 0.10 (0.18) 0.08 (0.18) 0.03 (0.17) 0.03 (0.17) −0.41 (0.33) −0.03 (0.22)  USDA Organic 0.39** (0.17) 0.40** (0.17) 0.81*** (0.17) 0.80*** (0.17) 0.35 (0.24) 0.69*** (0.22)  Rainforest Alliance 0.30 (0.19) 0.30 (0.19) 0.38** (0.17) 0.38** (0.17) 0.08 (0.28) 0.34 (0.22)  Carbon Footprint −0.01 (0.19) −0.04 (0.19) −0.13 (0.17) −0.14 (0.17) 0.11 (0.28) −0.17 (0.22)  Price −0.40*** (0.08) −0.52*** (0.06) −0.51*** (0.06) −0.54*** (0.05) −0.46*** (0.12) −0.43*** (0.08)  No_Buy −5.47*** (0.40) −5.38*** (0.40) −5.09*** (0.39) −5.06*** (0.39) −5.28*** (0.39) −5.18*** (0.39)  Log likelihooda −376 −379 −386 −386 −388 −393  AIC 774.1 779.8 793.1 794.5 797.1 807.0  BIC 823.2 829.0 842.3 843.7 846.2 856.2 p-values for testing statistical differences between the coefficients attended versus ignored  Fair Trade 0.024** 0.016** 0.001*** 0.001*** 0.006*** 0.018**  USDA Organic <0.001*** <0.001*** 0.081* 0.074* 0.006*** 0.144  Rainforest Alliance 0.007*** 0.005*** 0.007*** 0.007*** 0.016** 0.051*  Carbon Footprint 0.214 0.165 0.013** 0.010*** 0.430 0.081*  Price 0.001*** 0.005*** 0.008*** 0.013** 0.088* 0.008*** Fixation count FC 2 FC 1 ANA modelling Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S2 defB-S2 defA-CT2 defB-CT2 FC1-S2 FC1-CT2 Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Attended  Fair Trade 0.54*** (0.16) 0.55*** (0.17) 0.72*** (0.18) 0.72*** (0.18) 0.43*** (0.14) 0.46*** (0.15)  USDA Organic 1.46*** (0.20) 1.45*** (0.20) 1.11*** (0.20) 1.12*** (0.20) 0.99*** (0.16) 0.94*** (0.16)  Rainforest Alliance 0.86*** (0.18) 0.87*** (0.18) 0.93*** (0.19) 0.93*** (0.19) 0.72*** (0.15) 0.73*** (0.16)  Carbon Footprint 0.16 (0.17) 0.18 (0.17) 0.37** (0.18) 0.38** (0.18) 0.06 (0.14) 0.17 (0.15)  Price −0.65*** (0.05) −0.65*** (0.05) −0.62*** (0.05) −0.63*** (0.05) −0.61*** (0.05) −0.62*** (0.05) Ignored  Fair Trade 0.10 (0.18) 0.08 (0.18) 0.03 (0.17) 0.03 (0.17) −0.41 (0.33) −0.03 (0.22)  USDA Organic 0.39** (0.17) 0.40** (0.17) 0.81*** (0.17) 0.80*** (0.17) 0.35 (0.24) 0.69*** (0.22)  Rainforest Alliance 0.30 (0.19) 0.30 (0.19) 0.38** (0.17) 0.38** (0.17) 0.08 (0.28) 0.34 (0.22)  Carbon Footprint −0.01 (0.19) −0.04 (0.19) −0.13 (0.17) −0.14 (0.17) 0.11 (0.28) −0.17 (0.22)  Price −0.40*** (0.08) −0.52*** (0.06) −0.51*** (0.06) −0.54*** (0.05) −0.46*** (0.12) −0.43*** (0.08)  No_Buy −5.47*** (0.40) −5.38*** (0.40) −5.09*** (0.39) −5.06*** (0.39) −5.28*** (0.39) −5.18*** (0.39)  Log likelihooda −376 −379 −386 −386 −388 −393  AIC 774.1 779.8 793.1 794.5 797.1 807.0  BIC 823.2 829.0 842.3 843.7 846.2 856.2 p-values for testing statistical differences between the coefficients attended versus ignored  Fair Trade 0.024** 0.016** 0.001*** 0.001*** 0.006*** 0.018**  USDA Organic <0.001*** <0.001*** 0.081* 0.074* 0.006*** 0.144  Rainforest Alliance 0.007*** 0.005*** 0.007*** 0.007*** 0.016** 0.051*  Carbon Footprint 0.214 0.165 0.013** 0.010*** 0.430 0.081*  Price 0.001*** 0.005*** 0.008*** 0.013** 0.088* 0.008*** Note: *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. aFor the Full-AA MNL model: log likelihood = −401.4, AIC = 814.8, BIC = 841.7 (see Appendix B). Table 6. MNL parameter estimations with two coefficients (attended and ignored) (n = 645) Fixation count FC 2 FC 1 ANA modelling Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S2 defB-S2 defA-CT2 defB-CT2 FC1-S2 FC1-CT2 Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Attended  Fair Trade 0.54*** (0.16) 0.55*** (0.17) 0.72*** (0.18) 0.72*** (0.18) 0.43*** (0.14) 0.46*** (0.15)  USDA Organic 1.46*** (0.20) 1.45*** (0.20) 1.11*** (0.20) 1.12*** (0.20) 0.99*** (0.16) 0.94*** (0.16)  Rainforest Alliance 0.86*** (0.18) 0.87*** (0.18) 0.93*** (0.19) 0.93*** (0.19) 0.72*** (0.15) 0.73*** (0.16)  Carbon Footprint 0.16 (0.17) 0.18 (0.17) 0.37** (0.18) 0.38** (0.18) 0.06 (0.14) 0.17 (0.15)  Price −0.65*** (0.05) −0.65*** (0.05) −0.62*** (0.05) −0.63*** (0.05) −0.61*** (0.05) −0.62*** (0.05) Ignored  Fair Trade 0.10 (0.18) 0.08 (0.18) 0.03 (0.17) 0.03 (0.17) −0.41 (0.33) −0.03 (0.22)  USDA Organic 0.39** (0.17) 0.40** (0.17) 0.81*** (0.17) 0.80*** (0.17) 0.35 (0.24) 0.69*** (0.22)  Rainforest Alliance 0.30 (0.19) 0.30 (0.19) 0.38** (0.17) 0.38** (0.17) 0.08 (0.28) 0.34 (0.22)  Carbon Footprint −0.01 (0.19) −0.04 (0.19) −0.13 (0.17) −0.14 (0.17) 0.11 (0.28) −0.17 (0.22)  Price −0.40*** (0.08) −0.52*** (0.06) −0.51*** (0.06) −0.54*** (0.05) −0.46*** (0.12) −0.43*** (0.08)  No_Buy −5.47*** (0.40) −5.38*** (0.40) −5.09*** (0.39) −5.06*** (0.39) −5.28*** (0.39) −5.18*** (0.39)  Log likelihooda −376 −379 −386 −386 −388 −393  AIC 774.1 779.8 793.1 794.5 797.1 807.0  BIC 823.2 829.0 842.3 843.7 846.2 856.2 p-values for testing statistical differences between the coefficients attended versus ignored  Fair Trade 0.024** 0.016** 0.001*** 0.001*** 0.006*** 0.018**  USDA Organic <0.001*** <0.001*** 0.081* 0.074* 0.006*** 0.144  Rainforest Alliance 0.007*** 0.005*** 0.007*** 0.007*** 0.016** 0.051*  Carbon Footprint 0.214 0.165 0.013** 0.010*** 0.430 0.081*  Price 0.001*** 0.005*** 0.008*** 0.013** 0.088* 0.008*** Fixation count FC 2 FC 1 ANA modelling Serial ANA Choice task ANA Serial ANA Choice task ANA Definition ignored attribute Definition A Definition B Definition A Definition B defA-S2 defB-S2 defA-CT2 defB-CT2 FC1-S2 FC1-CT2 Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Coefficient Standard error Attended  Fair Trade 0.54*** (0.16) 0.55*** (0.17) 0.72*** (0.18) 0.72*** (0.18) 0.43*** (0.14) 0.46*** (0.15)  USDA Organic 1.46*** (0.20) 1.45*** (0.20) 1.11*** (0.20) 1.12*** (0.20) 0.99*** (0.16) 0.94*** (0.16)  Rainforest Alliance 0.86*** (0.18) 0.87*** (0.18) 0.93*** (0.19) 0.93*** (0.19) 0.72*** (0.15) 0.73*** (0.16)  Carbon Footprint 0.16 (0.17) 0.18 (0.17) 0.37** (0.18) 0.38** (0.18) 0.06 (0.14) 0.17 (0.15)  Price −0.65*** (0.05) −0.65*** (0.05) −0.62*** (0.05) −0.63*** (0.05) −0.61*** (0.05) −0.62*** (0.05) Ignored  Fair Trade 0.10 (0.18) 0.08 (0.18) 0.03 (0.17) 0.03 (0.17) −0.41 (0.33) −0.03 (0.22)  USDA Organic 0.39** (0.17) 0.40** (0.17) 0.81*** (0.17) 0.80*** (0.17) 0.35 (0.24) 0.69*** (0.22)  Rainforest Alliance 0.30 (0.19) 0.30 (0.19) 0.38** (0.17) 0.38** (0.17) 0.08 (0.28) 0.34 (0.22)  Carbon Footprint −0.01 (0.19) −0.04 (0.19) −0.13 (0.17) −0.14 (0.17) 0.11 (0.28) −0.17 (0.22)  Price −0.40*** (0.08) −0.52*** (0.06) −0.51*** (0.06) −0.54*** (0.05) −0.46*** (0.12) −0.43*** (0.08)  No_Buy −5.47*** (0.40) −5.38*** (0.40) −5.09*** (0.39) −5.06*** (0.39) −5.28*** (0.39) −5.18*** (0.39)  Log likelihooda −376 −379 −386 −386 −388 −393  AIC 774.1 779.8 793.1 794.5 797.1 807.0  BIC 823.2 829.0 842.3 843.7 846.2 856.2 p-values for testing statistical differences between the coefficients attended versus ignored  Fair Trade 0.024** 0.016** 0.001*** 0.001*** 0.006*** 0.018**  USDA Organic <0.001*** <0.001*** 0.081* 0.074* 0.006*** 0.144  Rainforest Alliance 0.007*** 0.005*** 0.007*** 0.007*** 0.016** 0.051*  Carbon Footprint 0.214 0.165 0.013** 0.010*** 0.430 0.081*  Price 0.001*** 0.005*** 0.008*** 0.013** 0.088* 0.008*** Note: *, ** and *** indicate statistical significance at the 10%, 5% and 1% levels, respectively. aFor the Full-AA MNL model: log likelihood = −401.4, AIC = 814.8, BIC = 841.7 (see Appendix B). For all model estimations using the serial ANA modelling approach, the coefficients of the ignored attributes indicate that respondents who were identified as visual non-attenders for the Rainforest Alliance, and Fair Trade labels truly ignored these attributes.8 For the serial ANA modelling approach with a fixation count of one (FC1-S2), the ignored coefficient of USDA Organic is also not significantly different from zero. Thus, for the serial ANA approach with a fixation count of one, respondents who were identified as visual non-attenders for one of the sustainability labels (USDA Organic, Fair Trade, or Rainforest Alliance) truly ignored these attributes. For all model estimations using the choice task ANA modelling approach, the coefficients of the ignored attributes indicate that in the choice tasks in which we considered Fair Trade and Carbon Footprint (and Rainforest Alliance for a fixation count of one) as visually ignored, they were indeed truly ignored. Thus, these choice tasks were answered as if the visually ignored attribute was not present in the choice set. For these attributes, restricting the coefficient to zero if it was visually ignored was appropriate and resulted in the removal of the attribute from the choice consideration. While 17 out of 30 ignored coefficients were not statistically different from zero, 13 out of the 30 estimated ignored parameters were significantly different from zero; thus, setting the coefficients of these parameters to zero may not be appropriate. In all six ANA models, the coefficient of ignored price was significant at the 1 per cent level. In all ANA models except FC1-S2, the coefficient of ignored USDA Organic was also significant. When using a fixation count of two as the cut-off, the choice task ANA model also had significant coefficients of the ignored Rainforest Alliance. Whereas assuming that visually ignored attributes were truly ignored was appropriate for all sustainability labels in the serial ANA approach with a fixation count of one and for all sustainability labels except USDA Organic in the serial ANA approach with a fixation count of two, this was not the case for price. Therefore, it is inappropriate to constrain the coefficients of the visually ignored price and also in some cases the sustainability labels USDA Organic and Rainforest Alliance to zero. This finding is important because it indicates that some attributes that were classified as visually non-attended based on the ANA definitions and modelling approaches were actually not ignored and could have influenced the choice. Importantly, there is a difference in the meaning of having coefficients for the ignored attributes significantly different from zero based on either stated ANA or visual ANA. When based on stated ANA, it refers to an attribute identified as ignored when in fact the attribute did impact one’s choice (Campbell and Lorimer, 2009); however, this attribute possibly had a reduced impact. For visual ANA, a coefficient for an ignored attribute significantly different from zero refers to an attribute being identified as not having looked at while the attribute impacts one’s choice. The statistical differences between the two coefficients (attended and ignored) were tested using the Krinsky and Robb method (Krinsky and Robb, 1986; 1990) and suggest some differences in behaviour (Table 6). For price, the coefficient of attended price was statistically lower (more negative) than the coefficient of ignored price (except for FC1-S2, where the difference was not statistically significant). The ignored coefficients of USDA Organic were significantly different from zero in five of the six models. Most of the ignored coefficients of USDA Organic were also significantly smaller than the attended coefficients of USDA Organic. For the choice task ANA models with a fixation count of two (defA-CT2 and defB-CT2), the Rainforest Alliance ignored coefficients were also significantly different from zero and statistically lower than the coefficient of attended Rainforest Alliance. This result indicates that a classification as visually ignored, on average, leads to a lower utility for the USDA Organic label (and also for the Rainforest Alliance label for two of the models) and a less negative utility for price. These less negative (price) or positive (USDA Organic, Rainforest Alliance) coefficients of the attributes identified as ignored may have two possible explanations. First, the Price, USDA Organic and Rainforest Alliance attributes that may be classified as visually non-attended may in fact not have been truly ignored when choosing the preferred alternative. Instead, respondents who paid less attention to these attributes received less negative (price) or less positive (USDA Organic) utility from these attributes. Second, it is possible that the smaller coefficient for the ignored subset might be a combination of truly ignored attributes (zeros) and attended attributes, which therefore results in a lower value for the coefficient. This could be an artefact of the experimental design or due to a number of reasons, including tracking error and/or wrongly identifying ANA due to tracking error and/or peripheral vision effects as explained in the next paragraph. Balcombe, Fraser and McSorly (2015) noted that people must look long enough at information for it to be processed. However, our results show that measuring whether an attribute is visually attended to might be attribute dependent. For example, the number of fixations needed to visually attend to attributes may differ depending on the attribute itself. Price and USDA Organic (and in some cases Rainforest Alliance) are attributes that were attended to when we identified them as being ignored. This result may be due to the respondents’ familiarity with these attributes. Therefore, even when we defined them as visually non-attended, they were not actually ignored. Wrongly assuming a coefficient to be zero also explains reduced model fit (see Section 3.4). In addition to familiarity, issues specifically related to the eye-tracking technology may provide an explanation for why there are significant coefficient estimates for some of the ignored parameters. While fixating on one attribute, another attribute might be viewed simultaneously and interpreted without truly fixating on it. If the participants did not fixate on information presented in an AOI, it does not mean that they were not aware that it was there (Bergstrom and Shall, 2014, p. 6), since fixations only report visual attention taking place in the foveal vision and not in the parafoveal and peripheral vision. While people’s primary attention is focused on what they see in the foveal vision, they might still grasp information presented in other parts of their visual field. Some authors (Henderson and Hollingworth, 1999; Henderson et al., 2003) have reported that a functional field of view can be 4°, while others (Holmqvist et al. 2011) have suggested using a margin of 1°–1.5°.9 We also cannot rule out the possibility of a measurement error due to the eye-tracker as the reason why the coefficients of the ignored parameters are significant, as it would be ‘unnatural’ for people to decouple visual attention and fixations in this manner.10 More information on the limitations and challenges of eye tracking are reported in a separate section (Section 5). 3.3.2. Coefficient of variation In this second method, we evaluate whether respondents who were identified as having ignored or not ignored an attribute based on the serial visual ANA have the same allocation (ignored versus not ignored) based on the coefficient of variation method (Table 7). Table 7. Comparison of allocation of respondents for serial visual ANA and serial inferred ANA (count) (n = 81) Serial inferred ANA Serial visual ANA FC 2 FC 1 NI I NI + I (%total) NI I NI + I (%total) Fair Trade NI 41 25 66 (81.5%) 60 6 66 (81.5%) I 7 8 15 (18.5%)a 11 4 15 (18.5%) NI + I (%total) 48 (59.3%) 33 (40.7%)a 71 (87.7%) 10 (12.3%) Matching allocations (%) 60.5%b 79.0% USDA Organic NI 34 35 69 (85.2%) 56 13 69 (85.2%) I 2 10 12 (14.8%) 6 6 12 (14.8%) NI + I (%total) 36 (44.4%) 45 (55.6%) 62 (76.5%) 19 (23.5%) Matching allocations (%) 54.3% 76.5% Rainforest Alliance NI 44 32 76 (93.8%) 63 13 76 (93.8%) I 2 3 5 (6.2%) 4 1 5 (6.2%) NI + I (%total) 46 (56.8%) 35 (43.2%) 67 (82.7%) 14 (17.3%) Matching allocations (%) 58.0% 79.0% Carbon Footprint NI 37 15 52 (64.2%) 47 5 52 (64.2%) I 11 18 29 (35.8%) 21 8 29 (35.8%) NI + I (%total) 48 (59.3%) 33 (40.7%) 68 (84.0%) 13 (16.0%) Matching allocations (%) 67.9% 67.9% Serial inferred ANA Serial visual ANA FC 2 FC 1 NI I NI + I (%total) NI I NI + I (%total) Fair Trade NI 41 25 66 (81.5%) 60 6 66 (81.5%) I 7 8 15 (18.5%)a 11 4 15 (18.5%) NI + I (%total) 48 (59.3%) 33 (40.7%)a 71 (87.7%) 10 (12.3%) Matching allocations (%) 60.5%b 79.0% USDA Organic NI 34 35 69 (85.2%) 56 13 69 (85.2%) I 2 10 12 (14.8%) 6 6 12 (14.8%) NI + I (%total) 36 (44.4%) 45 (55.6%) 62 (76.5%) 19 (23.5%) Matching allocations (%) 54.3% 76.5% Rainforest Alliance NI 44 32 76 (93.8%) 63 13 76 (93.8%) I 2 3 5 (6.2%) 4 1 5 (6.2%) NI + I (%total) 46 (56.8%) 35 (43.2%) 67 (82.7%) 14 (17.3%) Matching allocations (%) 58.0% 79.0% Carbon Footprint NI 37 15 52 (64.2%) 47 5 52 (64.2%) I 11 18 29 (35.8%) 21 8 29 (35.8%) NI + I (%total) 48 (59.3%) 33 (40.7%) 68 (84.0%) 13 (16.0%) Matching allocations (%) 67.9% 67.9% NI, not ignored; I, ignored; FC, fixation count. aOut of 81 respondents, 15 were considered to have ignored the Fair Trade attribute based on the inferred method, and 33 respondents were considered to have ignored the attribute based on serial visual ANA with fixation count 2. bFor 8 and 41 respondents, the classifications into ignoring and not ignoring, respectively, are the same between the inferred and visual methods, resulting in a matching allocation of 60% for Fair Trade ((41 + 8)/81). Table 7. Comparison of allocation of respondents for serial visual ANA and serial inferred ANA (count) (n = 81) Serial inferred ANA Serial visual ANA FC 2 FC 1 NI I NI + I (%total) NI I NI + I (%total) Fair Trade NI 41 25 66 (81.5%) 60 6 66 (81.5%) I 7 8 15 (18.5%)a 11 4 15 (18.5%) NI + I (%total) 48 (59.3%) 33 (40.7%)a 71 (87.7%) 10 (12.3%) Matching allocations (%) 60.5%b 79.0% USDA Organic NI 34 35 69 (85.2%) 56 13 69 (85.2%) I 2 10 12 (14.8%) 6 6 12 (14.8%) NI + I (%total) 36 (44.4%) 45 (55.6%) 62 (76.5%) 19 (23.5%) Matching allocations (%) 54.3% 76.5% Rainforest Alliance NI 44 32 76 (93.8%) 63 13 76 (93.8%) I 2 3 5 (6.2%) 4 1 5 (6.2%) NI + I (%total) 46 (56.8%) 35 (43.2%) 67 (82.7%) 14 (17.3%) Matching allocations (%) 58.0% 79.0% Carbon Footprint NI 37 15 52 (64.2%) 47 5 52 (64.2%) I 11 18 29 (35.8%) 21 8 29 (35.8%) NI + I (%total) 48 (59.3%) 33 (40.7%) 68 (84.0%) 13 (16.0%) Matching allocations (%) 67.9% 67.9% Serial inferred ANA Serial visual ANA FC 2 FC 1 NI I NI + I (%total) NI I NI + I (%total) Fair Trade NI 41 25 66 (81.5%) 60 6 66 (81.5%) I 7 8 15 (18.5%)a 11 4 15 (18.5%) NI + I (%total) 48 (59.3%) 33 (40.7%)a 71 (87.7%) 10 (12.3%) Matching allocations (%) 60.5%b 79.0% USDA Organic NI 34 35 69 (85.2%) 56 13 69 (85.2%) I 2 10 12 (14.8%) 6 6 12 (14.8%) NI + I (%total) 36 (44.4%) 45 (55.6%) 62 (76.5%) 19 (23.5%) Matching allocations (%) 54.3% 76.5% Rainforest Alliance NI 44 32 76 (93.8%) 63 13 76 (93.8%) I 2 3 5 (6.2%) 4 1 5 (6.2%) NI + I (%total) 46 (56.8%) 35 (43.2%) 67 (82.7%) 14 (17.3%) Matching allocations (%) 58.0% 79.0% Carbon Footprint NI 37 15 52 (64.2%) 47 5 52 (64.2%) I 11 18 29 (35.8%) 21 8 29 (35.8%) NI + I (%total) 48 (59.3%) 33 (40.7%) 68 (84.0%) 13 (16.0%) Matching allocations (%) 67.9% 67.9% NI, not ignored; I, ignored; FC, fixation count. aOut of 81 respondents, 15 were considered to have ignored the Fair Trade attribute based on the inferred method, and 33 respondents were considered to have ignored the attribute based on serial visual ANA with fixation count 2. bFor 8 and 41 respondents, the classifications into ignoring and not ignoring, respectively, are the same between the inferred and visual methods, resulting in a matching allocation of 60% for Fair Trade ((41 + 8)/81). First, the rates of ignoring the different attributes between visual and inferred ANA are compared (Table 7). The proportions of respondents ignoring Fair Trade, USDA Organic, Rainforest Alliance and Carbon Footprint using this inferred method11 are 18.5 per cent, 14.8 per cent, 6.2 per cent and 35.8 per cent, respectively. The inferred method has much lower rates of ignoring compared to the visual ANA with a fixation count of two (Table 7), except for Carbon Footprint, which is quite similar. For example, out of the 81 respondents, 15 (18.5 per cent) were considered to have ignored the Fair Trade attribute based on the inferred method, and 33 respondents (40.7 per cent) were considered to have ignored the attribute based on serial visual ANA with fixation count 2. The proportions obtained by the inferred method are more in line with those obtained through visual ANA with a fixation count of one as the cut-off, except for Carbon Footprint. In addition to the rate of ignoring attributes, a comparison of the allocation of the respondents identified as visually ignoring an attribute is of interest (Table 7). For example, for the Fair Trade attribute, 49 respondents are classified the same for both the inferred and the visual ANA with fixation count 2 (41 attenders and 8 non-attenders). A large proportion of the respondents who were identified as having ignored an attribute based on the serial visual ANA were not identified as ignoring this attribute by the inferred method. This shows that there is little agreement regarding the identification of serial non-attenders between the visual and inferred approach. This may suggest that a large proportion of the respondents who were identified as having ignored an attribute using the serial visual ANA did not actually ignore it. Clearly, there are differences in the allocations into the ignored group between the serial visual ANA and inferred ANA. However, it remains uncertain which method is the most accurate in identifying ignored attributes. Hess and Hensher (2010) compared this inferred ANA method (i.e. the coefficient of variation approach) with stated ANA and found large differences or inconsistencies between the two approaches in terms of rates of ignoring and in the allocation of respondents into the two groups (attenders versus non-attenders). The inferred method also has drawbacks such as the use of the arbitrary threshold value of two for the coefficient of variation (Hess and Hensher, 2010). As mentioned by Hess and Hensher (2010), more research is needed into how to allocate respondents to the ignored group based on the coefficient of variation, such as defining a less arbitrary threshold for the coefficient of variation method. 3.4. Model fit across estimations When comparing model fits, restricting the coefficients of the ignored attributes to zero results in a decrease in model fit (a decrease in log likelihood and an increase in the AIC and BIC statistics) compared with the full attendance model (Table 5). A worse model fit is likely due to treating attributes as having a zero parameter when some are not actually ignored (which was confirmed by the model estimations with two coefficients per attribute). For both fixation cut-offs, the serial ANA modelling approach results in a better model fit than the choice task ANA modelling approach. For models with a fixation count of two, defining an ignored attribute based on the whole choice set (Definition A) results in a better model fit than defining it based on the alternatives (Definition B). Models in which two separate coefficients were estimated (Table 6) have a better model fit compared with the model assuming full attendance (Appendix B) and also outperforms the model in which the ignored coefficients were set to zero (Appendix B). Results thus show that setting the ignored coefficients to zero when in fact they are statistically different from zero thus worsens the model fit. Though, several studies on stated ANA have reported an improvement in model fit when constraining ignored coefficients to zero (Hensher, Rose and Greene, 2005; Campbell, Hutchinson and Scarpa, 2008; Kragt, 2013). However, similar to our study, Alemu et al. (2013) reported a decrease in model fit when restricting the coefficients to zero. This decrease in model fit could also be attributed to the number of observations that are essentially excluded from contributing to the likelihood function (Alemu et al. 2013), which may also explain why we see a decrease in model fit when moving from serial to choice task ANA, as the percentage of ignored attributes also increases (Table 4). Since not all coefficients that were identified as ignored were, in fact, ignored based on estimates of these coefficients (Section 3.3.1) and based on the coefficient of variation (Section 3.3.2), we can conclude that eye tracking did not always successfully identify visual non-attenders for all attributes. This is also confirmed by the decrease in model fit for most models when constraining the ignored coefficients to zero compared with the full attendance models. We discuss the potential challenges and limitations of using eye tracking for ANA research in Section 5. 3.5. WTP We cannot always assume that the respondents did not look at the attribute just because they did not care about it; i.e. we do not know the reason why they did not consider the attribute when making their choices. For instance, some people may have visually ignored an attribute because the task was too complex, while others may have ignored it because they do not derive utility from that attribute. In the first case (too complex), a person who ignored an attribute may have the same preferences as someone who attended to it and thus may also have the same WTP. In this case, the WTP based on the attended attributes would be applicable for both groups. In the second case, where respondents ignore an attribute because they do not derive any utility from it, the WTP for the ignored group is zero. Because we do not know the actual reasons for visually ignoring the attributes, the WTP estimates are calculated only using the coefficient estimates for the attended attributes from Table 5, resulting in WTP estimates for sustainability labels of $1.7312 or lower. The Krinsky and Robb method (Krinsky and Robb, 1986; 1990) indicated no statistical differences in the mean estimates for the WTP distribution between the full attendance model and the models accounting for visual ANA. This result suggests that accounting for visual ANA, in our study, did not have significant implications in terms of mean WTP estimates. While coffee prices largely depend on the quality of the beans, their origin, the blend, and the brand, organic coffee prices are often higher than prices for coffee certified by the Rainforest Alliance or Fair Trade USA (FAO, 2009). Retail prices for ground coffee ranged from $3.00 to $9.99 per 12-ounce package (store check, Fayetteville, AR, 2013). Because the prices for both conventional and certified coffee vary considerably by the quality and origin as well as by the nature of the outlet and brand, it is very difficult to determine a premium that is solely attributable to a sustainability certification rather than these other factors (Consumers International, 2005; FAO, 2009). Based on a store check (Fayetteville, AR, 2013), the price premium for coffee with a sustainability label ranges from $1.50 to $2.30 per 12 ounces when comparing coffees with and without a label of the same brand. 4. Conclusion Given that CEs are commonly used to assess attribute valuation, there is an urgent need for and considerable research interest in methods to account for ANA. Stated and inferred approaches to address ANA do not directly measure whether attributes were ignored in a CE. Limitations associated with stated and inferred approaches to obtain ANA information at the choice task level justify studying other ways to address respondents’ true processing strategies and attendance at the choice task level. We contribute to this research area by using eye-tracking measures to evaluate visual attendance to the attributes in a CE. This method does not rely on self-reported ANA behaviour and does not attempt to infer ANA based on respondents’ choice behaviour. Instead, we measured participants’ visual attention to specify whether they visually attended to the information presented to them and extend the work of Balcombe, Fraser and McSorly (2015) on visual ANA in CEs using eye-tracking technology in three ways (alternative fixation count cut-offs, alternative definitions of visual ANA, and alternative modelling approaches). For each combination of fixation cut-off, visual ANA definition and modelling approach, we identified visually ignored attributes. Restricting the coefficients of ignored attributes to zero led to a decrease in model fit compared to the full attendance model, and accounting for visual ANA did not significantly influence the WTP estimates. Not all visually ignored attributes were truly ignored based on the respondents’ choice behaviours (price and USDA Organic labels in most cases and Rainforest Alliance in some), i.e. the respondents paid some attention to these attributes in their choice consideration. Hence, our results suggest that Fair Trade, Carbon Footprint and in most cases Rainforest Alliance were truly ignored when they were identified as ignored using visual ANA. However, the same does not hold for price and USDA Organic. This result shows that the best way to account for visual ANA might depend on the attribute itself. For price and USDA Organic, it is likely that people do not need to fixate their eyes on these attributes to the same extent as to the other attributes because it is possible that they are already acquainted or more familiar with these attributes. It is possible that for USDA Organic and price, they were aware of the information presented from other parts of their visual field (parafoveal and peripheral vision) not covered by eye fixations. The price, USDA Organic, and (in some cases) Rainforest Alliance attributes that may have been classified as visually non-attended were in fact not truly ignored. Most coefficient estimates for attributes that were classified as visually ignored were lower than the attended coefficients. This result might occur for two reasons. First, respondents who paid less attention to these attributes received less negative (price) or positive (USDA Organic, Rainforest Alliance) utility from these attributes. Second, the smaller coefficient of the ignored subset might be a combination of truly ignored attributes (zeros) and attended attributes and therefore result in a lower average. The true explanation is likely to be a combination of both reasons. Next, the allocation of attenders and non-attenders based on visual ANA was compared with the allocation based on the coefficient of variation, an inferred method. The method of the coefficient of variation indicated that a small proportion of those classified as having ignored an attribute using the eye-tracking’s visual ANA were also identified as having ignored it based on the inferred method. Given the insignificance of many of the coefficients of the ignored attributes, the mismatch between the allocation based on the inferred method and the serial visual ANA and the decrease in the model fit when using visual ANA, we can conclude, at least based on our study, that using eye tracking does not always provide an adequate measure that can guarantee that an attribute should be considered as having been ignored or not. Nevertheless, eye-tracking measures can provide useful information regarding respondents’ decision-making behaviour. It can help us to better understand their decision-making process and its relation with choice, preferences and attention (Grebitus, Roosen and Seitz, 2015). Previous studies illustrated that eye tracking is useful to quantify visual attention, a measure for the degree of attention, during food choices (Van Loo et al. 2015; Lewis, Grebitus and Nayga, 2016); however, Balcombe et al. (2017) report that the relationship between visual attention and preference may be weak. While eye tracking has been used to quantify visual attention and its effects during food choice, applying it to identify visual ANA is more challenging, as suggested by our findings. We discuss these challenges and limitations in the following sections. Specifically, we first discuss the limitations of eye-tracking research in general, followed by limitations specifically related to ANA research based on the insights from our study. 5. Limitations and challenges 5.1. The use of eye tracking Eye tracking provides measures of where participants fixate but not why; thus, the motivations and cognitions underlying these eye movements remain unknown (Graham, Orquin and Visschers, 2012). Familiarity might influence how extensively a piece of information is examined, as possibly observed with the price and USDA Organic attributes in our study (Pieters, Rosbergen and Hartog, 1996, 1999; Graham, Orquin and Visschers, 2012). Moreover, participants’ fixations do not necessarily imply understanding and do not reveal anything about the higher-level processes of attention and comprehension. As suggested by Graham, Orquin and Visschers (2012), conducting an interview after an eye-tracking task may provide additional insight into what respondents were thinking during the task. While eye-tracking studies might be less prone to social desirability compared to studies that ask respondents directly about the information to which they attend, knowing that their eye movements will be monitored could potentially influence their behaviour (Graham, Orquin and Visschers, 2012). Finally, eye tracking is a relatively expensive and time-consuming method. 5.2. The use of eye tracking to study ANA While eye tracking monitors consumers’ visual attention, this technology has some limitations that are particularly relevant when studying visual ANA. These issues relate to (i) the use of fixations, (ii) the size of the AOIs, (iii) the definitions for visual ANA, (iv) the influence of familiarity and (v) the tracking ratio. First, eye tracking uses fixations as an indication of visual attention, but it only reveals what happens in an individual’s foveal vision. While fixations take place in our foveal vision, which is where our primary attention is focused (Bergstrom and Shall, 2014), eye tracking does not allow us to measure the attention paid in one’s parafoveal and peripheral vision (Bergstrom and Shall, 2014; Orquin, Ashby and Clarke, 2016). If the participants did not fixate on information presented in an AOI, it does not necessarily mean that they were not aware that it was there (Bergstrom and Shall, 2014, p. 6). They may still have grasped information present in other parts of their visual field without focusing on it. While eye fixations within a certain AOI are believed to provide reliable measures for visual attention for that AOI (Bialkova and van Trijp, 2011), eye fixations do not necessarily represent everything that participants might have observed. Based on our study, we can conclude that fixations do not necessarily allow us to identify whether the information was ignored. Second, eye tracking uses AOIs for which specific metrics (such as fixation counts) are calculated, and thus the results depend on how these areas are defined. Unfortunately, there is no consensus in the decision-making research literature on the definition of AOIs. As mentioned by Orquin, Ashby and Clarke (2016, p. 103), ‘this lack of standardisation in AOI definition and reporting presents direct problems to the advancement of behavioural decision-making research’. Thus, there is a need to have a more standardised way to define AOIs, as the size of the AOI determines its fixation count as well as other eye-tracking metrics in use. Third, there are different ways to define visual ANA with respect to the number of fixation counts and the number of choice tasks. Balcombe, Fraser and McSorly (2015) used the threshold of at least two fixations to consider an attribute as attended. In addition, they identify a respondent as a non-attender when this person ignores an attribute in more than half of the choice tasks. These are two rather arbitrary values and will influence the discrete measure of visual ANA (ignored versus not ignored). Fourth, how familiar respondents are with the attribute or logo – as in our study – might matter. If they are very familiar with it, they may not need to fixate on it to recognise that the logo is present. If they are less familiar, they may need to look longer at and fixate more on the information for it to be processed. Thus, the decisions about the fixation cut-off as well as the size of the AOI for identifying information as being ignored might depend on the respondent’s familiarity. Fifth, not all eye movements are recorded. In our study, 89.5 per cent of eye movement data was recorded, and thus, we cannot guarantee if some attributes were ignored or unfixated. The five limitations discussed above represent major challenges, but they can hopefully assist researchers and future studies to further optimise the use of eye tracking in the context of ANA in CEs. We discuss several specific suggestions for future research in the following subsection. 5.3. Suggestions for future research on the use of eye tracking in CEs While we attempted to account for the limitations described above as much as possible in the present study, there are several suggestions based on our study that future studies could take into account. In our study, the logos were rather large and in close proximity to each other. In addition, the coffee packages did not contain much other information. Since packages never had the same labels, it is possible that respondents could infer the quality attributes on one package based on the labels on the other package in a choice set. We recommend that future studies include other information cues on the food packages (e.g. nutritional labels, ingredients lists, graphics) that compete for the respondent’s attention and make it more difficult for the participant to be aware of the presence of certain logos without fixating on them. In this study, we used a head-free, below-screen eye-tracker that provided us with a relatively high tracking ratio (89.5 per cent) across our respondents. Previous research has shown no significant difference between head-mounted and head-free types of eye-trackers with respect to the pupil size-based index of cognitive activity (Bartels and Marshall, 2012). Furthermore, the head-free, below-screen tracker allows participants to be less constrained and may promote more natural human behaviour. Nevertheless, future studies might consider using a chin-rest, since this could avoid movement of the head, possibly resulting in better tracking quality (in terms of reliability and accuracy) compared with the head-free, below-screen tracker that we used in this study. While we compared different visual ANA approaches, future studies could expand on comparing various ANA approaches (visual, stated and inferred) and evaluating which of these techniques or combinations of techniques is the most appropriate to account for ANA. Specifically, for visual ANA, future research should improve and standardise how to apply visual ANA in choice modelling based on eye tracking and remove any arbitrary steps. Studies could, for example, test other fixation count cut-offs and other visual ANA definitions. For example, for serial ANA, we used the approach used by Balcombe, Fraser and McSorly (2015), whereby a respondent is identified as a non-attender if he/she ignored an attribute in more than half of the choice tasks. However, ‘more than half’ is an arbitrary approach, and other approaches could be tested to further fine-tune the serial visual ANA method. In addition, other AOI sizes and margins could be tested. In addition to incorporating ANA or other decision-making strategies into choice models, future studies could use eye tracking as a tool to study how to optimise experimental designs that would discourage respondents from ignoring information, thereby avoiding violation of the assumption of rational utility maximisation (i.e. that the complete information presented is considered by respondents). We share the opinion of Balcombe, Fraser and McSorly (2015), who suggested the use of eye tracking to assist in the visual design and appearance of the CE instrument. Several eye-tracking studies have demonstrated that bottom-up factors can trigger attention (Visschers, Hess and Siegrist, 2010; Bialkova and van Trijp, 2011). Bottom-up or stimulus-driven forms of attention are caused by characteristics of the visual stimulus itself (colour, size, location, saliency) and occur without specifically searching for them (Wolfe, 1998). Further insights into these bottom-up factors can help in the visual design or layout of choice sets that can reduce visual ANA in CEs. Spinks and Mortimer (2016), for example, investigated the impact of CE complexity (in terms of number of attributes ranging from 3 to 8) on the prevalence of ANA. They reported higher levels of ANA when more attributes are present and provided information on how to improve the CE design, measuring health care preferences in their study. They suggest the use of eye tracking during the CE design and piloting phase, which is also suggested by Balcombe et al. (2017). Future studies could test other issues of the visual CE design and its impact on ANA such as the pictorial or verbal representation of attributes or alternatives, and other variations in representation of choice sets that could reduce ANA in CE studies. Finally, in addition to ANA research, eye tracking can be used to study other decision-making strategies and decision heuristics and can assist in explaining the decision-making process in relation to food choices (Uggeldahl et al., 2016) and preferences. As mentioned by Orquin and Loose (2013), more research on how to effectively combine attention research from eye tracking with decision-making research is warranted. Supplementary data Supplementary data are available at European Review of Agricultural Economics online. Acknowledgements The authors thank the editor (Iain Fraser) and three anonymous journal reviewers for their valuable comments and suggestions that helped us to significantly improve the quality of the paper. Footnotes 1 In the CE literature, the set of alternatives that an individual must consider to arrive at his/her choice is called the choice set (Hensher, Rose and Greene, 2015). The choice task refers to the action itself of selecting the preferred alternative in the choice set. 2 Visual attention to an attribute is ‘a continuous measure of the degree to which a respondent evaluates the attribute’, while attendance is ‘a discrete measure indicating whether respondents will be considered to have attended an attribute or not’ (Balcombe et al., 2015, p. 449). 3 ‘Truly ignored’ is related to the coefficient of ‘visually ignored’ attributes being zero. 4 We could not apply ECLC models because the limited sample size caused issues with them. 5 The visual attendance towards price depends on the definition applied because price was presented in each of the two buying alternatives. 6 As noted by a reviewer, this could also be due to the relatively small sample size. Only the observations in which the attribute was attended to are used for the coefficient estimation, while for the other observations, the coefficients are set to zero. 7 Five parameters were estimated (USDA Organic, Rainforest Alliance, Fair Trade, Carbon Footprint and Price) for each of the six approaches, resulting in a total of 30 parameter estimations. 8 For Carbon Footprint, the attended coefficients are insignificant. Therefore, being wrongly classified as having ignored Carbon Footprint may result in an insignificant coefficient. 9 As suggested by a reviewer, the functional field of view is likely different for illustration and text, with a smaller functional view for text (such as price) as compared to illustrations (such as the sustainability labels). Hughes et al. (2003) also reported differences in visual attention for text versus pictures. 10 We thank an anonymous reviewer for pointing out this possibility. 11 The inferred method cannot estimate the percentage of respondents ignoring price since price is considered fixed. 12 This corresponds with = 1.407/−0.812 for model defA-S. References Alemu , M. H. , Mørkbak , M. R. , Olsen , S. B. and Jensen , C. L. ( 2013 ). Attending to the reasons for attribute non-attendance in choice experiments . Environmental & Resource Economics 54 : 333 – 359 . Google Scholar CrossRef Search ADS Balcombe , K. , Burton , M. and Rigby , D. ( 2011 ). Skew and attribute non-attendance within the Bayesian mixed logit model . Journal of Environmental Economics and Management 62 : 446 – 461 . Google Scholar CrossRef Search ADS Balcombe , K. G. , Fraser , I. M. and McSorly , E. ( 2015 ). Visual attention and attribute attendance in multi-attribute choice experiments . Journal of Applied Econometrics 30 : 447 – 467 . Google Scholar CrossRef Search ADS Balcombe , K. , Fraser , I. , Williams , L. and McSorley , E. ( 2017 ). Examining the relationship between visual attention and stated preferences: a discrete choice experiment using eye-tracking . Journal of Economic Behavior & Organization 144 : 238 – 257 . Google Scholar CrossRef Search ADS Barreiro-Hurle , J. , Gracia , A. and de-Magistris , T. ( 2010 ). The effects of multiple health and nutrition labels on consumer food choices . Journal of Agricultural Economics 61 : 426 – 443 . Google Scholar CrossRef Search ADS Bartels , M. and Marshall , S. P. ( 2012 ). Measuring cognitive workload across different eye tracking hardware platforms. 2012 ACM Symposium on Eye Tracking and Applications, Santa Barbara, California, 28–30 March, 2012. Bergstrom , J. R. and Schall , A. ( 2014 ). Eye Tracking in User Experience Design . Waltham, MA : Elsevier . Bialkova , S. and van Trijp , H. C. M. ( 2011 ). An efficient methodology for assessing attention to and effect of nutrition information displayed front-of-pack . Food Quality and Preference 22 : 592 – 601 . Google Scholar CrossRef Search ADS Bonnet , C. and Simioni , M. ( 2001 ). Assessing consumer response to protected designation of origin labelling: a mixed multinomial logit approach . European Review of Agricultural Economics 28 : 433 – 449 . Google Scholar CrossRef Search ADS Burgess , L. ( 2007 ). Discrete Choice Experiments [computer software], Department of Mathematical Sciences, University of Technology, Sydney. http://maths.science.uts.edu.au/maths/wiki/SPExpts. Accessed 1 August 2016. Campbell , D. , Boeri , M. , Doherty , E. and Hutchinson , W. G. ( 2015 ). Learning, fatigue and preference formation in discrete choice experiments . Journal of Economic Behavior and Organization 119 : 345 – 363 . Google Scholar CrossRef Search ADS Campbell , D. , Hensher , D. A. and Scarpa , R. ( 2011 ). Non-attendance to attributes in environmental choice analysis: a latent class specification . Journal of Environmental Planning and Management 54 : 1061 – 1076 . Google Scholar CrossRef Search ADS Campbell , D. , Hutchinson , W. G. and Scarpa , R. ( 2008 ). Incorporating discontinuous preferences into the analysis of discrete choice experiments . Environmental & Resource Economics 41 : 401 – 417 . Google Scholar CrossRef Search ADS Campbell , D. and Lorimer , V. S. ( 2009 ). Accommodating attribute processing strategies in stated choice analysis: Do respondents do what they say they do? 17th Annual Conference of the European Association of Environmental and Resource Economics, Amsterdam. Caputo , V. , Nayga , R. M. and Scarpa , R. ( 2013 ). Food miles or carbon emissions? Exploring labelling preference for food transport footprint with a stated choice study . Australian Journal of Agricultural and Resource Economics 57 : 465 – 482 . Google Scholar CrossRef Search ADS Caputo , V. , Van Loo , E. J. , Scarpa , R. , Nayga , R. M. and Verbeke , W. ( 2017 ). Comparing serial, and choice task stated and inferred attribute non‐attendance methods in food choice experiments . Journal of Agricultural Economics DOI:10.1111/1477-9552.12246 . Carlsson , F. , Kataria , M. and Lampi , E. ( 2010 ). Dealing with ignored attributes in choice experiments on valuation of Sweden’s environmental quality objectives . Environmental & Resource Economics 47 : 65 – 89 . Google Scholar CrossRef Search ADS Consumers International ( 2005 ). From Bean to Cup: How Consumer Choice Impacts on Coffee Producers and the Environment . London : Consumers International . Fairtrade International ( 2015 ). Monitoring the Scope and Benefits of Fair Trade , 7th ed . Bonn : Fairtrade International . Food and Agriculture Organization of the United Nations (FAO) ( 2009 ). The Market for Organic and Fair-Trade Coffee. Study prepared in the framework of FAO project GCP/RAF/404/GER: ‘Increasing Incomes and Food Security of Small Farmers in West and Central Africa Through Exports of Organic and Fair-Trade Tropical Products’. Trade and Markets Division. Rome: FAO. Gracia , A. ( 2014 ). Consumers’ preferences for a local food product: a real choice experiment . Empirical Economics 47 : 111 – 128 . Google Scholar CrossRef Search ADS Gracia , A. , Barreiro-Hurle , J. and Perez , L. P. Y. ( 2012 ). Can renewable energy be financed with higher electricity prices? Evidence from a Spanish region . Energy Policy 50 : 784 – 794 . Google Scholar CrossRef Search ADS Graham , D. J. , Orquin , J. L. and Visschers , V. H. ( 2012 ). Eye tracking and nutrition label use: a review of the literature and recommendations for label enhancement . Food Policy 37 : 378 – 382 . Google Scholar CrossRef Search ADS Grebitus , C. , Roosen , J. and Seitz , C. C. ( 2015 ). Visual attention and choice: a behavioral economics perspective on food decisions . Journal of Agricultural & Food Industrial Organization 13 : 73 – 81 . Google Scholar CrossRef Search ADS Greene , W. H. ( 1990 ). LIMDEP Version 5.0 . New York, NY : Econometric Software. Inc . Greene , W. H. ( 2012 ). Nlogit Version 5.0: Reference Guide . Plainview, NY : Econometric Software Inc . Grunert , K. G. ( 2011 ). Sustainability in the food sector: a consumer behaviour perspective . International Journal on Food System Dynamics 2 : 207 – 218 . Henderson , J. M. and Hollingworth , A. ( 1999 ). The role of fixation position in detecting scene changes across saccades . Psychological Science 10 : 438 – 443 . Google Scholar CrossRef Search ADS Henderson , J. M. , Williams , C. C. , Castelhano , M. S. and Falk , R. J. ( 2003 ). Eye movements and picture processing during recognition . Perception & Psychophysics 65 : 725 – 734 . Google Scholar CrossRef Search ADS Hensher , D. A. ( 2006 ). How do respondents process stated choice experiments? Attribute consideration under varying information load . Journal of Applied Econometrics 21 : 861 – 878 . Google Scholar CrossRef Search ADS Hensher , D. A. ( 2014 ). Process heuristics in choice analysis: an editorial . Journal of Choice Modelling 11 : 1 – 3 . Google Scholar CrossRef Search ADS Hensher , D. A. and Greene , W. H. ( 2010 ). Non-attendance and dual processing of common-metric attributes in choice analysis: a latent class specification . Empirical Economics 39 : 413 – 426 . Google Scholar CrossRef Search ADS Hensher , D. A. and Rose , J. M. ( 2009 ). Simplifying choice through attribute preservation or non-attendance: implications for willingness to pay . Transportation Research Part E 45 : 583 – 590 . Google Scholar CrossRef Search ADS Hensher , D. A. , Rose , J. M. and Greene , W. H. ( 2005 ). The implications on willingness to pay of respondents ignoring specific attributes . Transportation 32 : 203 – 222 . Google Scholar CrossRef Search ADS Hensher , D. A. , Rose , J. M. and Greene , W. H. ( 2012 ). Inferring attribute non-attendance from stated choice data: implications for willingness to pay estimates and a warning for stated choice experiment design . Transportation 39 : 235 – 245 . Google Scholar CrossRef Search ADS Hensher , D. A. , Rose , J. M. and Greene , W. H. ( 2015 ). Applied Choice Analysis , 2nd edn . Cambridge, UK : Cambridge University Press . Google Scholar CrossRef Search ADS Hess , S. and Hensher , D. A. ( 2010 ). Using conditioning on observed choices to retrieve individual-specific attribute processing strategies . Transportation Research Part B 44 : 781 – 790 . Google Scholar CrossRef Search ADS Hess , S. , Stathopoulos , A. , Campbell , D. , O’Neill , V. and Caussade , S. ( 2013 ). It’s not that I don’t care, I just don’t care very much: confounding between attribute non-attendance and taste heterogeneity . Transportation 40 : 583 – 607 . Google Scholar CrossRef Search ADS Hole , A. R. ( 2011 ). A discrete choice model with endogenous attribute attendance . Economics Letters 110 : 203 – 205 . Google Scholar CrossRef Search ADS Holmqvist , K. , Nyström , M. , Andersson , R. , Dewhurst , R. , Jarodzka , H. and Van de Weijer , J. ( 2011 ) Eye Tracking: A Comprehensive Guide to Methods and Measures . Oxford , UK: Oxford University Press . Hughes , A. , Wilkens , T. , Wildemuth , B. M. and Marchionini , G. ( 2003 ). Text or pictures? An eyetracking study of how people view digital video surrogates. In: Proceedings of the International Conference on Image and Video Retrieval, 271–280. Kehlbacher , A. , Balcombe , K. and Bennett , R. ( 2013 ). Stated attribute non-attendance in successive choice experiments . Journal of Agricultural Economics 64 : 693 – 706 . Google Scholar CrossRef Search ADS Kragt , M. E. ( 2013 ). Stated and inferred attribute attendance models: a comparison with environmental choice experiments . Journal of Agricultural Economics 64 : 719 – 736 . Google Scholar CrossRef Search ADS Krinsky , I. and Robb , A. L. ( 1986 ). On approximating the statistical properties of elasticities . The Review of Economics and Statistics 68 : 715 – 719 . Google Scholar CrossRef Search ADS Krinsky , I. and Robb , A. L. ( 1990 ). On approximating the statistical properties of elasticities: a correction . The Review of Economics and Statistics 72 : 189 – 190 . Google Scholar CrossRef Search ADS Krucien , N. , Ryan , M. and Hermens , F. ( 2017 ). Visual attention in multi-attributes choices: what can eye-tracking tell us? Journal of Economic Behavior & Organization 135 : 251 – 267 . Google Scholar CrossRef Search ADS Lagarde , M. ( 2013 ). Investigating attribute non-attendance and its consequences in choice experiments with latent class models . Health Economics 22 : 554 – 567 . Google Scholar CrossRef Search ADS Lancaster , K. J. ( 1966 ). New approach to consumer theory . Journal of Political Economy 74 : 132 – 157 . Google Scholar CrossRef Search ADS Lancsar , E. and Louviere , J. ( 2006 ). Deleting ‘irrational’ responses from discrete choice experiments: a case of investigating or imposing preferences? Health Economics 15 : 797 – 811 . Google Scholar CrossRef Search ADS Layton , D. F. and Brown , G. ( 2000 ). Heterogeneous preferences regarding global climate change . The Review of Economics and Statistics 82 : 616 – 624 . Google Scholar CrossRef Search ADS Lewis , K. E. , Grebitus , C. and Nayga , R. M. ( 2016 ). The impact of brand and attention on consumers’ willingness to pay: evidence from an eye tracking experiment . Canadian Journal of Agricultural Economics/Revue canadienne d’agroeconomie 64 : 753 – 777 . Google Scholar CrossRef Search ADS Lusk , J. L. and Schroeder , T. C. ( 2004 ). Are choice experiments incentive compatible? A test with quality differentiated beef steaks . American Journal of Agricultural Economics 86 : 467 – 482 . Google Scholar CrossRef Search ADS Mariel , P. , Hoyos , D. and Meyerhoff , J. ( 2013 ). Stated or inferred attribute non-attendance? A simulation approach . Economía Agraria y Recursos Naturales (Agricultural and Resource Economics) 13 : 51 – 67 . Google Scholar CrossRef Search ADS McFadden , D. ( 1974 ). Conditional logit analysis of qualitative choice behavior. In: Zarembka P. (ed.) , Frontiers in Econometrics . New York : Academic Press , 105 – 142 . Meyerhoff , J. and Liebe , U. ( 2009 ). Discontinuous preferences in choice experiments: Evidence at the choice task level. 17th Annual Conference of the European Association of Environmental and Resource Economics, Amsterdam. Mintel ( 2012 ). Coffee–USA–October 2012 . London : Mintel Group Ltd . Morey , E. and Rossmann , K. ( 2003 ). Using stated-preference questions to investigate variations in willingness to pay for preserving marble monuments: classic heterogeneity, random parameters, and mixture models . Journal of Cultural Economics 27 : 215 – 229 . Google Scholar CrossRef Search ADS Mørkbak , M. R. , Olsen , S. B. and Campbell , D. ( 2014 ). Behavioral implications of providing real incentives in stated choice experiments . Journal of Economic Psychology 45 : 102 – 116 . Google Scholar CrossRef Search ADS Orquin , J. L. , Ashby , N. J. and Clarke , A. D. ( 2016 ). Areas of interest as a signal detection problem in behavioral eye-tracking research . Journal of Behavioral Decision Making 29 : 103 – 115 . Google Scholar CrossRef Search ADS Orquin , J. L. and Loose , S. M. ( 2013 ). Attention and choice: A review on eye movements in decision making . Acta psychological 144 : 190 – 206 . Google Scholar CrossRef Search ADS Pierrot , J. , Giovannucci , D. and Kasterine , A. ( 2011 ). Trends in the trade of certified coffees. Technical Paper . Geneva : International Trade Centre (ITC) . Pieters , R. G. M. , Rosbergen , E. and Hartog , M. ( 1996 ). Visual attention to advertising: the impact of motivation and repetition . Advances in Consumer Research 23 : 242 – 248 . Pieters , R. G. M. , Rosbergen , E. and Wedel , M. ( 1999 ). Visual attention to repeated print advertising: a test of scanpath theory . Journal of Marketing Research 36 : 424 – 438 . Google Scholar CrossRef Search ADS Poole , A. , Ball , L. and Phillips , P. ( 2005 ). In search of salience: a response-time and eye-movement analysis of bookmark recognition. In Fincher, S., Markopoulos, P., Moore, D. and Ruddle, R. (eds), People and Computers Xviii – Design for Life . London : Springer , 363 – 378 . Google Scholar CrossRef Search ADS Puckett , S. M. and Hensher , D. A. ( 2008 ). The role of attribute processing strategies in estimating the preferences of road freight stakeholders . Transportation Research Part E 44 : 379 – 395 . Google Scholar CrossRef Search ADS Puckett , S. M. and Hensher , D. A. ( 2009 ). Revealing the extent of process heterogeneity in choice analysis: an empirical assessment . Transportation Research Part A 43 : 117 – 126 . Revelt , D. and Train , K. ( 1998 ). Mixed logit with repeated choices: households’ choices of appliance efficiency level . The Review of Economics and Statistics 80 : 647 – 657 . Google Scholar CrossRef Search ADS Ryan , M. and Bate , A. ( 2001 ). Testing the assumptions of rationality, continuity and symmetry when applying discrete choice experiments in health care . Applied Economics Letters 8 : 59 – 63 . Google Scholar CrossRef Search ADS Scarpa , R. , Campbell , D. and Hutchinson , W. G. ( 2007 ). Benefit estimates for landscape improvements: sequential Bayesian design and respondents’ rationality in a choice experiment . Land Economics 83 : 617 – 634 . Google Scholar CrossRef Search ADS Scarpa , R. , Ferrini , S. and Willis , K. ( 2005 ). Performance of error component models for status-quo effects in choice experiments. In: R. Scarpa and A. Alberini (eds) , Applications of Simulation Methods in Environmental and Resource Economics . Dordrecht : Springer , 247 – 273 . Google Scholar CrossRef Search ADS Scarpa , R. , Gilbride , T. J. , Campbell , D. and Hensher , D. A. ( 2009 ). Modelling attribute non-attendance in choice experiments for rural landscape valuation . European Review of Agricultural Economics 36 : 151 – 174 . Google Scholar CrossRef Search ADS Scarpa , R. , Thiene , M. and Hensher , D. A. ( 2010 ). Monitoring choice task attribute attendance in nonmarket valuation of multiple park management services: does it matter? Land Economics 86 : 817 – 839 . Google Scholar CrossRef Search ADS Scarpa , R. , Zanoli , R. , Bruschi , V. and Naspetti , S. ( 2013 ). Inferred and stated attribute non-attendance in food choice experiments . American Journal of Agricultural Economics 95 : 165 – 180 . Google Scholar CrossRef Search ADS SMI (SensoMotoric Instruments) . ( 2016 ). BeGaze Manual. Version 3.6. February 2016. 416 pages. Spinks , J. and Mortimer , D. ( 2016 ). Lost in the crowd? Using eye-tracking to investigate the effect of complexity on attribute non-attendance in discrete choice experiments . BMC Medical Informatics and Decision Making 16 ( 1 ): 14 . Google Scholar CrossRef Search ADS Street , D. J. and Burgess , L. ( 2007 ). The Construction of Optimal Stated Choice Experiments . Hoboken, NJ : John Wiley & Sons, Inc . Google Scholar CrossRef Search ADS Uggeldahl , K. , Jacobsen , C. , Lundhede , T. H. and Olsen , S. B. ( 2016 ). Choice certainty in discrete choice experiments: Will eye tracking provide useful measures? Journal of Choice Modelling 20 : 35 – 48 . Google Scholar CrossRef Search ADS Van Loo , E. J. , Caputo , V. , Nayga , R. M. Jr. , Seo , H.-S. , Zhang , B. and Verbeke , W. ( 2015 ). Sustainability labels on coffee: consumer preferences, willingness-to-pay and visual attention to attributes . Ecological Economics 118 : 215 – 225 . Google Scholar CrossRef Search ADS Van Loo , E. J. , Caputo , V. , Nayga , R. M. Jr and Verbeke , W. ( 2014 ). Consumers’ valuation of sustainability labels on meat . Food Policy 49 : 137 – 150 . Google Scholar CrossRef Search ADS Visschers , V. H. M. , Hess , R. and Siegrist , M. ( 2010 ). Health motivation and product design determine consumers’ visual attention to nutrition information on food products . Public Health Nutrition 13 : 1099 – 1106 . Google Scholar CrossRef Search ADS Wolfe , J. M. ( 1998 ). Visual search. In: H. Pashler (ed.) , Attention . Hove, UK : Psychology Press , 13 – 22 . Author notes Review coordinated by Iain Fraser © Oxford University Press and Foundation for the European Review of Agricultural Economics 2018; all rights reserved. For permissions, please e-mail: [email protected] This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)

Journal

European Review of Agricultural Economics – Oxford University Press

Published: Feb 15, 2018

There are no references for this article.