The impact of the sensory experience on scale and preference heterogeneity: The GMNL model approach applied to pig castration and meat quality

1 The impact of the sensory experience on scale and preference heterogeneity: The GMNL model approach applied to pig castration and meat quality Kallas, Zein 1 ; Borrisser-Pairó, Francesc 2 ; Martínez, Beatriz 3 ; Vieira, Ceferina 3 ; Rubio, Begonia 3 ; Panella, Nuria 2 ; Gil, Marta 2 ; Linares, M. Belén 4 ; Garrido, M. Dolores 4 ; Olivares, Álvaro 5 ; Miguel 5 Ibañez; M. Angels, Oliver 2 ; José María, Gil 1 . 1 Centre for Agro-food Economy and Development, Castelldefels, Barcelona, Spain. 2 IRTA-Monells, Product Quality program, Finca Cmps i Armet, e-17121 Monells, Girona, Spain. 3 Estación Tecnológica de la Carne, I. Tecnológico Agrario de Castilla y León, Guijuelo, Spain. 4 Facultad de Veterinaria, Tecnología de los alimentos, Universidad de Murcia, Murcia, Spain. 5 Facultad de Veterinaria, Producción Animal, Universidad Complutense, Madrid, Spain. Paper prepared for presentation at the EAAE-AAEA Joint Seminar ‘Consumer Behavior in a Changing World: Food, Culture, Society” March 25 to 27, 2015 Naples, Italy 1 The impact of the sensory experience on scale and preference heterogeneity: The GMNL model approach applied to pig castration and meat quality Topic: Consumer behavior: preference analysis Abstract The EU is considering a future ban on surgical pig castration by 2018 which may affect markets and consumers preferences. This study analysed consumers’ expected preference toward a masking strategy obtained from a mixture of spices and smoking of high level boar taint frankfurter sausages. In addition, we analysed the impact of the sensory experience on the nonobserved heterogeneity both at the scale and mean preferences. We carried out two NonHypothetical Discrete Choice Experiments (DCE) by creating a real shopping scenario before and after the hedonic sensory test for a sample of 150 consumers from the Metropolitan area of Madrid, Spain. Data used in this analysis were obtained from questionnaires completed in a controlled environment and estimated the recently developed Generalised Multinomial Logit Model (GMNL). Results showed the appropriateness of the proposed masking strategy of boar meat as a raw material for the processed meat industry in Spain. Consumers also declared their willingness to pay a premium for this flavours. The sensory experience have had impact on both the scale and preference heterogeneity. The degree of randomness and uncertainty of consumers decreased significantly in their final election and the source of unobserved heterogeneity obtained from the scale become more independent. Key Words: Consumers preference, Boar taint, Non-Hypothetical Choice Experiments, Generalized Multinomial Logit GMNL. 3 Introduction and objectives While the piglet castration is regulated by the CE Directive 93/2001 stating that it may only be performed using methods that do not involve tearing of the tissue, the EU is considering a future ban on surgical pig castration by 1 January 2018 (EC, 2010). This potential prohibition relies not only on its negative impact on pig welfare1 (Rault et al., 2011) but also because consumers demand lean meat with no off-odours and the production costs associated with entire males (Prunier et al., 2006) are lower. In a first pronouncement, from January 2012, in certified organic pig production, the surgical castration of piglets, if carried out, should be performed with prolonged analgesia and/or anaesthesia (Commission Regulation (EC) Nº 889/2008). In a transition period, surgical castration should only be performed with pain relief and efforts should be undertaken to facilitate the end of surgical castration (Heid & Hamm, 2013). The European changes in the animal welfare regulations and policies, in particular the pig sector, have been the results of an increasing societal pressure to seek for more humane production systems (EC, 2005 and EC, 2007). The pig welfare is one of the most studied sector as pork is the most produced and consumed meat in the EU (FAOSTAT, 2012). In this context, the pig production has received a special attention from the European authorities and several regulations have been approved. Recently, the EC directive 120/2008 banned the use of sow stalls since January 2012. Annually, approximately 100 million pigs or about 80% of the total population of male piglets are surgically castrated in the EU of which 48.7% are surgically castrated without anaesthesia (Fredriksen et al., 2009; Borrisser-Pairó et al., 2014). The practice of castration of male piglets is important to avoid boar taint, which is a distinctive and unpleasant odour and flavour perceived through a combination of Skatole and Androstenone compounds. In this context, within the EU, the Meat Hygiene Regulation 842/2004/EC states that meat manifesting a “pronounced sexual odour” is to be declared unfit for human consumption (Whittington et al., 2011). However, the castration practice increases the production costs and leads to fatter carcasses (Tuyttens et al., 2012). Raising entire male pigs is more profitable for farmers due to leaner carcasses and higher protein content (Lundström et al., 2009). In this context, the importance to implement alternatives for castration is gaining relevance (Heid & Hamm, 2012; Font-i-Furnols, 2012). 1 Behavioural changes observed after castration included reduced nursing and walking, and increased pain related behaviours suggesting that pigs experience acute pain in response to castration (Sutherland, 2015). 4 Many alternatives to castration have been explored: a) genetic selection and gender selection for ‘low-taint’ pigs (De Campos et al., 2015), b) different management and rearing strategies (Bonneau & Lebret 2010; Wesoly et al., 2015), c) slaughter at a younger age and lower weight (Von Borell et al., 2009), d) detection of boar taint at slaughter line (Vestergaard et al., 2006), e) mixing of tainted with untainted meat (Walstra, 1974) and f) masking unpleasant odours and flavours with the appropriate masking strategy such as spices, marinades or heat treatment (Valeeva et al., 2009; Mörlein et al., 2015) and if the castration is applied, the immunocastration is one of the recently most studied alternatives (Gamero-Negrón et al., 2015). An overwhelming majority of people are sensitive to the odour of skatole (Weiler et al., 2000), however less people is sensitive to androstenone (Blanch et al., 2012). In this context, the production of meat from entire males may appears to have a promising future (Bonneau & Chevillon, 2012) if is adequately commercialized within the processed pig meat sector (Bonneau, 1998; Gunn et al., 2004), since the boar tainted meat is better accepted in processed products than in retail cuts (Bañón et al., 2003; Valeeva et al., 2009). In this context, an effective and specific strategy of masking boar taint in the pork processed products became a worthy aspect to investigate. In the Spanish case, there is a need for updated knowledge about processing opportunities for boar meat, particularly if the castration is banned in a near future. As commented by Lunde et al. (2008), if castration is prohibited, it would be relevant to identify processing methods that could still provide quality products by allowing the processing industry to adjust their recipes to minimize such off-flavour. Several studies suggested different strategies to mask boar taint in processed meat. In readyto-eat marinades meat the use of liquid smoke and oregano extracts appeared to be a strategy for masking (Lunde et al., 2008). Likewise in fermented sausage the use of spices, liquid smoked or starter cultures have been proposed as potential solution to remove the perception of boar taint (Malmfors & Lundström, 1983; Stolzenbach et al., 2009). However, focusing only on the acceptance of the most effective masking strategy gives only a partial view of the consumers’ preferences. Liking or disliking a food product does not only depend on physicochemical properties but also on the consumers' expectations and attitudes toward extrinsic cues (Lange et al., 1998; Franchi, 2012; Meier-Dinkel et al., 2013; Asioli et al., 2014). The consumer final choice of a food product, and thus its willingness to pay for it (WTP), is a mixture between the sensory experience (measured by the intrinsic cues) and the other descriptors 5 of the products (commonly measured by the extrinsic cues) such as the price, the brand and the origin. The relevance of the sensory experience in food choices is not new and the hedonic experiences have a strong influence on the willingness-to pay (Heid & Hamm, 2013). Consumers’ experience modifies product quality perceptions and scoring behaviour, as well as it is likely to affect repurchase decisions (Poole et al., 2006). Even more, in some cases it is more important than label information (Combris et al., 2009). In this context, the objective of this paper is threefold: First, to analyse the consumers’ expected preference toward a masking strategy of high level boar taint of frankfurter sausages by a flavour developed from different herbs and natural smoking. Second to study the impact of the sensory experience of this masking strategy on such preferences through a hedonic test of four different frankfurter sausage; flavoured and original taste obtained from castrated pigs and boars2. Third, to assess how the non-observed heterogeneity both at the scale and mean preferences of the sausage’ attributes are affected by the experience in a real shopping scenario. To reach these objectives, we applied two Non-Hypothetical Discrete Choice Experiments (DCE) by creating a real shopping scenario before and after the hedonic sensory test for a sample of 150 consumers that carried out in Madrid, Spain. The Discrete Choice Experiments aims to identify the individual’s indirect utility function associated with attributes of products by examining the trade-offs they make when making choice decisions. Thus, several alternatives (products) that are described by several attributes with varying levels are presented to respondents in an array of choice sets. The respondent is then asked to select its preferred alternative within each choice set, thereby revealing his/her preference for certain attributes and levels. Subsequently, the relative importance or willingness to pay of the attributes can be indirectly recovered from respondents’ choices. To motivate the consumers in the recruitment process, they were rewarded by €15. In addition, before the DCE tasks, all participants were unexpectedly rewarded by an additional €5 while explaining that a real shopping scenario is to be created at the end of the experiment. Individuals who agree to participate were asked to purchase their selected product from a randomly chosen choice set obtained from the DCE and to pay its posted price. From one hand, at the empirical level, this study is the first paper that analyse consumers acceptance toward the proposed masking strategy in the frankfurter sausage obtained from boars 2 The sensory experiment details; how pigs were reared and slaughtered, the meat sampling, the sausage preparation, the masking flavour development, acceptance heterogeneity are the objective of another paper that is in progress as above commented. 6 meat in the Spanish market. On the other hand, at the methodological level, this paper contributes to the literature of the Discrete Choice Modelling (DCM) using the recently developed Generalised Multinomial Logit Model (GMNL) of Fiebig et al. (2010). In this context to our knowledge, this research is the first application, in the literature of boar tainted meat preferences studies that analyse the impact of sensory experience on consumers’ preferences and on both the scale and the preference heterogeneity. Materials and methods Consumer panel The impact of sensory experience on the expected preferences was analysed on a sample of 150 consumers selected from the Metropolitan area of Madrid province. Participants represent the Spanish consumers over 18 years of age who regularly purchase food and beverages and having purchased and consumed frankfurter sausage at least one time in the last month. Data used in this analysis were obtained from questionnaires completed in a controlled environment. A quota sampling procedure was used to guarantee a representative sample in terms of gender, postal districts (67 different districts) and age. Consumers were recruited from a specialized consumers study company (Silliker Ibérica, S.A.) and were economically compensated by €15 to participate in an experiment of about 1.5 hours. Table 1 summarize the main socio-demographic variables of the sample components. 7 Table 1: Summary of the socio-demographic variable of the sample. Sociodemographic variables (N) Age (150) (years) Gender (150) % female Study level (150) Family Income (€/month) (146) Primary Secondary University <1,500 1,500-3,000 >3,000 41.5 48.7 12.0 40.0 48.0 42.5 46.6 11.0 Employment situation (150) Housewife Retired Student unemployed Employee Part-time Employee full time Selfemployed part-time Selfemployed 5.3% 6.0% 15.3% 28.7% 10.7% 24.0% 2.0% 8.0% N 150 Experiment performance The applied methodological approach can be summarized into five main subsequent steps: i. First, participants were asked to answer in a short questionnaire their attitudes and consumption behaviour towards pork meat and pork frankfurter sausage in particular. Socioeconomic and life-style variables were also collected. ii. Second, participants were unexpectedly rewarded by an extra €5 and asked to select their preferred frankfurter sausage from different choice sets built within the DCE design. Consumers were warned that their selection will have a consequence as a real market will be created at the end of the experiment to exchange real money and real products. Thus, consumers who agree to participate in this market should purchase their selected product. No additional information about the products was given, except what appears in each choice set label. In this step we aimed to analyse the expected preferences of consumers on the basis of their past experiences and available information related to the characteristics of the product or to a similar one (Deliza & MacFie, 1996). iii. Third, a hedonic evaluation test was carried out to assess the impact of the chosen masking strategy for boar taint and to create a current sensory experience of the analysed product. Participants tasted four different frankfurter sausages with two different treatments: if the meat is obtained from castrated pig or boars and if the flavour is original or flavoured with spices and naturally smoked. iv. Fourth, consumers were informed about which type of sausages they tasted in order to associate their sensory experience with the specific products and characteristics. Then, the 8 same DCE was repeated and consumers turned to reselect their preferred products from the same choice sets and asked to take into consideration their sensory experience. This phase allow to analyse if the sensory experience have resulted in agreement or disagreement with what they expected. These changes play an important role in the final acceptance or rejection of the product (Font-i-Furnols & Guerrero, 2014) and may affect the final choice decision of the consumers. v. Fifth, a real market was created to exchange real product and money. Consumers who accepted to participate were obliged to purchase their chosen product from a randomly selected choice set. Attitude and consumption behaviour towards pork meat The survey collected information on the socio-economic backgrounds of the consumers along with their attitudes, preferences and opinions regarding pork consumption, especially frankfurter sausage. The short survey contains questions regarding the consumption frequency of fresh pork meat and frankfurter sausage, the brand they usually purchase and how much they usually pay when purchase frankfurter sausage. They were also asked to state if at the moment of the survey they have frankfurter sausage at home and if they are planning to purchase in future days. These questions were completed by life style questions about if they do sport or any physical activities not related to work, the last time they have made a blood test, if they are smoker, if they are alcohol drinker and in which frequency. The hedonic test The inclusion of a natural spices mixture (white pepper, paprika, mustard seed, nutmeg mace, coriander, sweet marjoram and small cardamom) combined with smoking process was the strategy selected to mask the perceived boar taint in frankfurter sausage. For consumer sensory analysis, four different batches of frankfurter sausages were prepared: 1) sausages with meat and fat from male castrated pig (original flavour), 2) smoked sausages with meat and fat from male castrated pig, manufactured including a natural spices mixture (flavoured), 3) sausages with meat and fat from entire pigs with androstenone concentrations 1.10-2.75 µg/g of fat (original flavour) and 4) smoked sausages with meat and fat from entire pigs manufactured including a natural spices mixture (flavoured). 9 The products were prepared in a pilot plant according to industrial procedures. The basic recipe contained lean meat (50%), pork fat (25%), ice/water (25%), potato starch (2.5%), soybean protein (2%), sodium chloride (2%), kappa carrageenan (0.5%), sodium polyphosphate (0.3%), dextrose (0.25%), pork flavour JBT-200 (0.2%), sodium ascorbate (0.05%) and sodium nitrite (150 ppm). To manufacture flavoured sausages a natural spices mixture (0.7%) was added to basic recipe. All additives and spices were provided by Proanda S.L. (Sevilla, Spain). For each batch of sausage, meat, fat, ice, water, and other ingredients were emulsified by using a bowl cutter (CM -41, Mainca. Barcelona, Spain). The batters obtained were vacuum stuffed (Tecmaq Microwat, Barcelona, Spain) into 20 mm collagen casing (NB300, Edicas, Ripoll, Spain), smoked by sawdust (in case of flavoured sausages) and cooked to 72º C core temperature (Verinox Junior 1100, Vigolo Vattaro, Italy). After cooking, sausages were cooled, vacuum-packaged and kept at 4ºC until the day of the analysis. All tests were performed in a room equipped with individual tasting booths according to ISO 8589 (2007). Sensory evaluation was carried out on two consecutive days, starting 7 days after the production of the sausages. Seven sensory sessions were conducted with approximately 22 consumers per session. For each session, samples were prepared shortly before analysis as described below. Sausages were heated for 3 min in a pot of boiling water on a ceramic glass cooktop (TEKA - IR 642). After cooking, the sausages were cut into pieces about 4.5 cm length and placed in aluminum souffle cups with lids, coded with a 3-digit random code. The cups were kept warm in an oven (Bosh-HB 22R251E) at 60 ºC for 10 min before being served. Serving temperature of the samples was 50 ± 2 °C. Ambient conditions (temperature and relative humidity) in the sensory room were 23-25ºC and 40-60% respectively, which were within the range of acceptable conditions. Consumers evaluated in a blind condition, the acceptability of the four different frankfurter sausage (meat obtained from castrated and boars pig and with original flavour or flavours with spices and naturally smoked) under white light in the order printed on the recording sheet, which was established to avoid the effect of sample order presentation, first-order or carry-over effects (Macfie et al., 1989). Consumers ate unsalted toasted bread and drank mineral water to rinse their palate between samples. Although castration might affect texture sensory parameters, such as tenderness and juiciness, Bañón et al. (2003) reported that these attributes were not related with boar taint when the meat is cooked, whereas the aroma and taste were strongly affect. Thus, in this study, each consumer rated odour, flavour and overall acceptability as the most important sensory attribute (Resurreccion, 2004) 16 The Mixed or heterogeneous logit models (MIXL) (also in the literature is referred to as Random Parameter Logit model, RPL) are currently quite popular. They extend the MNL introducing for unobserved heterogeneity by allowing random coefficients on attributes (Ben-Akiva et al., 1997). In MIXL the utility to person n from choosing alternative j in choice set t is given by: 1, , 1, , 1, , njt n njt njt n Nj JtUx n T      (4) Where, nn    and where () n  is the vector of person n specific deviations from the mean value of the  s. The n  is described by an underlying continuous distribution for the attributes defined by the researcher. In most applications the multivariate normal distribution is the most used, MVN (0,). In this case, n  is also assumed to be one for identification. For the MIXL model, the choice probability is: 1 exp[( ) ] () exp[( ) ] nnjt jnt J nnjt j x PX x         jT   (5) Recently Louviere & Mayer (2007), Louviere et al. (2008) argued that much of the preference heterogeneity captured by random parameters in MIXL can be better captured by the scale term; and thus known as “scale heterogeneity”. Besides, they stated that the normal distributions of the random attributed in the MIXL do not appear to be very close to it, as followed in almost MIXL applications. The MIXL turns to be likely a poor approximation to stated data if scale heterogeneity is not accounted for (Fiebig et al., 2010). The scale heterogeneity is the variation of the degree of randomness in the decision-making process over respondents and hence is the degree of individuals’ certainty. It is based on the differences of the variance of the error term ()  across individual. In this context, the analysis of the scale heterogeneity is important, especially for the stated preference studies (i.e. based on questionnaire). In this type of studies, consumers may interpret and process choice tasks and situations differently. They may have varying levels of attention paid to the task they are presented, as well as the level of certainty in their choice (Train & Weeks, 2005). Thus, it would be expected that the scale of the error term could be greater for some consumers than for others. 17  The Generalized Multinomial Logit Model (GMNL) On the basis of Keane (2006), Feibig et al. (2010) developed the Generalized Multinomial Logit model (GMNL). Within this approach, the n  is no longer set to be one, and a particular specification of this term is assumed. In this case, multiplying equation (4) through by n  , Feibig et al. (2010) identified that the utility to person n from choosing alternative j on choice set t is given by: [γ(1 γ)] njt n n n n njt njt UX        (6) where γ is a parameter between 0 and 1. n  is a scaling factor that proportionately scales the  up or down for each individual n. To impose that   γ 0,1 in estimation, Fiebeig et al., (2010) used a logistic transform ** γexp(γ) [1 exp( γ)]and estimate * γ . Thus γ is a mixing parameter, and its value determines the level of mixing or interaction between the scale heterogeneity coefficient n  and the parameter heterogeneity coefficient n  . Since the scale heterogeneity factor n  represents the person-specific scale of the idiosyncratic error, it should be positive. Fiebig et al. (2010) proposed that n  follows a log-normal distribution with mean 1 and standard deviation  , 2 LN (1, ) n   , with the estimated  capturing the scale heterogeneity across consumers. Thus, to ensure is poditive, Feibig et al. (2010) an exponential transformation of exp( ) nn    where N(0,1) n  . Because n  only enters the model as a product of n   (equation 6), some normalization on n  is required to identify  . Fiebig et al. (2010) recommend setting the mean of n  to 1 so  is the mean of the utility weights. Because the mean of the log-normally distributed n  is exp( ) nn    and the 2 E( ) exp( 2) n   , thus to ensure E( ) 1 n  , we need to set 2 (2)   . Let njt y be ‘1’ if respondent n choose alternative j in choice set t, and ‘0’ otherwise. The probability that consumer n choose alternative j in choice set t is: 1 exp([ (1 ) ] ) () exp([ (1 ) ] ) nn nnnjt jnt J nn nnnjt j X PX X            γγ γγ jT   (7) 18 Finally, regarding the Alternative Specific constant (ASC) which usually measure intangible aspects (not gathered by the attributes’ utilities) that some consumers like and others dislike. Fiebig et al. (2010) indicated that including it within the general GMNL specification may produce special estimation problems. Greene & Hensher (2010) proposed three possible strategies to deal with ASC: 1. Consider the ASC 0 () j  as fixed parameters, assuming homogenous preference for ASCs. In this case the equation (6) is specified as follows: 0 ()[ γ(1 γ)] njt j n n n n njt njt UX         2. Make them a part of the general specification of the GMNL model (i.e. to behave like the attributes), then the utility of the ASCs 0 () j  is scaled and considered to be random. The ASCs are considered as the  s components. In this case equation (6) is specified as follows: 00 0 [( )γ()(1γ)( )] njt nj njn nnjnnjtnjt UX          It is worth mentioning that Feibig et al. (2010) observed that this may cause estimator to fail. 3. Consider the ASC only as random parameter and force no special scaling for this variable, so the scaling parameter 1 n   and γ =0 . In this case, equation (6) is: 00 ()[γ(1 γ)] njt j nj n n n n njt njt UX          where 00 () jnj    are the heterogeneous intercepts (which do not have scale heterogeneity), with 0  being the mean vector and the 0n  being the stochastic component. The GMNL model is specified by default to consider the n  as uncorrelated. That is mean the covariance matrix of n  is constrained to be a diagonal matrix (a matrix in which all values above and to the right of the diagonal are equal to zero), thus, there are only variances estimated, no covariances. However, the GMNL can be specified to allow for correlated parameters. The presence of multiple observations on stated-choice responses for each sampled individual means that the potential for correlated responses across observations can be the product of many sources including the sequencing of offered choice situations that results in mixtures of learning and inertia effects, among other possible influences on choice response (Hensher et al., 2005). Thus, discrete choice data with repeated choice situations containing the same attributes and levels may have unobserved effects that are correlated among alternatives in a given choice situation. One way to recognize this 19 is to permit correlation of random parameters of attributes that are common across alternatives observation (Hensher et al., 2005). In the case of correlated random parameters, the set of random parameters has a full covariance matrix with estimated variances and covariances. Thus, when we have more than one random parameter the estimated standard deviation n  are no longer independent. To assess this, we have to decompose the estimated standard deviation parameters into their attribute specific and attributescorrelations standard deviations. The decomposition procedure is done following the Cholesky decomposition method. This method decouples the contribution to each standard deviation parameter made through correlation with other random parameter estimates and the actual contribution made solely through heterogeneity around the mean of each random parameter estimate (Hensher et al., 2005). The correlated parameters GMNL model reports both the “confounding” standard deviation and its Cholesky decomposing matrix. The diagonal value of the Cholesky matrix represent the true standard deviation for each random parameter once the cross-correlated parameter terms have been unconfounded. The below–diagonal elements in Cholesky decomposition matrix are the covariances (cross-correlation) among the random parameter estimates. From the abovementioned aspects of the GMNL model, in this case study, we used a GMNL model with correlated random parameters n  and considering the ASC 0 () j  as fixed parameters. This decision is because it showed to have the best goodness of fit compared to other specification in terms of Pseudo-R2, AIC and improvement in the Likelihood functions. We used the GMXLOGIT procedure in NLOGIT 5 with Once parameters are estimated, they represent the marginal utility of attributes and its contribution to the total utility function. Thus, the Marginal Rate of Substitution (MRS) between attributes can be obtained. As one of the attributes is expressed in monetary term (i.e. the price), it is possible to determine its “implicit price” (IP) or part-worth as follows: Product_attribute Product _attribute monetary_attribute IP       (8) As the attributes are codified with coding effect in order to avoid the base level to be confounded with the Alternative specific constant of the no–option, the implicit price to move from the base level of each attribute to the analysed level should be multiplied by 2. 20 Results and discussion Hedonic scores for the different meat tasted Before analysing the impact of the sensory experience on expected preferences, results of consumers’ acceptability for the different sausage types are first reported. Sensory parameters scores for sausages with and without the masking strategy from castrated pigs and boars are shown in Table 1. Table 1. Least squares means and standard error (SE) of sensory parameters evaluated in frankfurter sausages from castrated pigs and boars. Type of pork meat Odour Flavour Overall acceptability Original sausage from boar meat 5.40c (1.43) 5.62c (1.56) 5.46c (1.61) Flavoured1 sausage from boar meat 6.51a (1.27) 6.36b (1.26) 6.42a (1.18) Original sausage from castrated meat 5.69b (1.19) 6.20b (1.33) 5.91b (1.30) Flavoured1 sausage from castrated meat 6.62ª (1.16) 6.69a (1.18) 6.62a (1.29) a, b, c, Statistical differences among types of frankfurter sausage for all consumers at 95 %. 1with spices and naturally smoked. Comparing the overall acceptability of the four types of sausage, results showed significant differences. The flavoured sausage obtained from castrated and boar meat had the highest acceptability scores than the remaining type of sausage. This confirms that the applied masking strategy had a positive effect on frankfurter sausage acceptance (Lunde et al., 2008; Stolzenbach et al., 2009). In this context, the original sausage from boar meat had received the lowest valuation in all the attributes analysed which is clearly showing the negative impact of boar taint on acceptance (Bañon et al., 2003). Focusing on the flavour attribute, flavoured sausage obtained from boar meat exhibit a similar acceptance with the original sausage obtained from castrated meat. Finally, regarding the odour attribute, the masking strategy clearly show the non-significant difference between sausages obtained from boar and castrated meat which confirm the positive impact of the proposed ingredients. These results may represent the starting point of the processed pork meat for masking boar taint especially in sausages. 21 Impact of sensory evaluation on expected preferences Table 2 reports the marginal utilities of the attributes resulting from the G-MNL models with correlated random parameters for the pre and post sensory experiment. Both models showed the highest improvement in the likelihood and the best information criteria (AIC, BIC) over the MNL and MIXL models in line of the results of Fiebig et al. (2010), Greene & Hensher (2010) and Pancras & Dey (2011). As can be seen, at a 99% confidence level, we can reject the null hypothesis that all coefficients are jointly equal to zero with a Log-Likelihood ratio test highly significant. The goodness of fit is assessed through the McFadden’s pseudo-R2 (0.20 and 0.26 for pre and post sensory respectively) which is within the acceptable range for the discrete choice models. Table 2: Results from model estimations for consumer data with and without information. Pre sensory Post sensory Random Parameters in utility functions () Boar animal 0.28**a -0.01b Private Brand -0.21***a -0.26***a Flavoured1-0.47**b 0.50**a Non-Random Parameters in utility functions () Price -1.69***a -1.51***b Opt-Out -0.40***b 0.02a Diagonal values in Cholesky matrix Non-castrated animal 1.20*** 1.04*** Private brand 0.04 0.46** Flavoured 0.28 0.70*** Covariances of the Random parameters Private Brand : Non-castrated animal -0.11 -0.01 Flavoured : Non-castrated animal -0.20 -0.07 Flavoured : Private brand 0.07 -0.48* scale parameters Variance parameter tau in scale parameter 1.10*** 0.01 Weighting parameter Gamma 0.41*** 0.72*** Standard deviations of parameters distribution Std. Dev. Non-castrated animal 1.20*** 1.04*** Std. Dev. Private brand 0.10 0.47*** Std. Dev. Other Spanish origin 1.19*** 1.25*** Log-Likelihood (θ) -1,529.54 -1,418.39 Log-Likelihood (0) -1,931.32 -1,931.32 LL ratio test 803.55 (0.000) 1,025.87 (0.000) Pseudo R2 0.208 0.265 AIC/N 2.576 2.391 Significance levels: *** p<0.01; **p<0.05; * p< 0.10. Attributes with different superscript letters in columns (a, b) differ (P < 0.05). 1 with spices and naturally smoked. 22 The positive/negative sign of the coefficient implies higher/lower levels of utility associated with these attributes’ levels. In this context, the model estimates showed that all attribute coefficients are statistically significant except the boar animal level in the post sensory. This result revealed that before the hedonic valuation experience, consumers exhibit a preference for meat obtained from pigs reared in natural condition of life without any modification (i.e. castration) as a potential preference for a positive pig welfare. Previous studies have shown that Spanish consumers are not aware if castration is related to meat quality (Kallas et al., 2013). However, after testing the different sausage products, consumers were aware of the importance of castration on meat quality mainly on the odour off the product. Thus, the utility of meat obtained from boars pigs decreased significantly from 0.28 to -0.01. For the brand attribute preference, consumers showed a rejection of private brands with nonsignificant differences before and after the hedonic test. This result is confirmed by the descriptive results obtained from participants regarding their consumption behaviours of sausage. Almost a 70% of participants usually purchase manufacturers brand over the private one. Focusing on the flavour attribute, results showed an interesting pattern. While consumers before testing the products showed a negative expected preference for the flavoured sausage with spices and naturally smoked, after testing the product their utility become positive with highly significant difference. A fact that commented by Bredahl et al. (1998) who affirmed that the quality expectations and quality experience diverge widely. These results are in accordance to what commented also by Flores (1997) and Leistner (1995) that smoked meat products are not appealing to consumers in the Mediterranean. In addition, the results highlight the appropriateness of this masking strategy against boar taint as an effective masking strategy. In this line, as expected, the negative sign of the price implies that an increase in the levels of the price attribute, will decrease the utility of the alternatives presented to consumers. Comparing values before and after the experiments results showed slight significant differences. The utility of the price slightly decreased, showing that after eating experience some consumer shifted their election towards alternatives with higher prices that contains specific patterns of the other preferred levels such as flavoured, non-castrated and manufacturers’ brands. In this line, there is a clear indication against the no-choice option in both experiments, with highly significance and negative value in the pre sensory test. On average, consumers have a clear preference for the offered products. Interpreting the tau parameter (  ) which is the key parameter that captures the scale heterogeneity, results showed a substantial scale heterogeneity in the data with a highly significant value of 1.10 in the pre sensory test and a non-significant value of 0.01 in the post experiment. As the parameter tau increases, the degree of scale heterogeneity increases. This result shows that when consumers taste the different sausage 23 products accompanied by a posterior information about what they eat, the variation of the degree of randomness in their final decision and hence their degree of uncertainty decreased significantly. Regarding the unobserved taste (preference) heterogeneity (usually captured by the diagonal values in Cholesky matrix, but no more valid in the case of correlated parameters) is captured by the standard deviation of the random parameters (bottom of table 1) as detailed in the methodology section. The estimated models showed statistically significant results with the exception of the private brand in the pre sensory. Thus, as mentioned by Lenk (2011), when the estimated standard deviation of parameters distribution are close to zero, then the unobserved heterogeneity is mostly due to heterogeneity in the scale parameter and not taste. In our results the values are far from zero (i. e. 1.20 and 1.19 in the pre experiment and 0.47, 1.04 and 1.25 in the post experiment) and thus there is a mixture between both sources of heterogeneities. This can be verified analysing the gamma estimate. The main motivation of the G-MNL model is to separate the estimation of scale heterogeneity from taste heterogeneity (Lenk, 2011) which might be identified by the gamma parameter. The estimate of the gamma in both model are relatively far from zero. Which implies that preference heterogeneity is invariant to scale heterogeneity and thus both heterogeneity are independent. This independence is emphasized in the post experiment. Gamma parameter increased from 0.41 to 0.72 and thus both types of heterogeneity becomes more independently identified. Finally, the estimates of the covariance matrix of the random parameters showed non-significant values between attributes in both the pre and post experiment with the exception of the flavoured and private brand levels in the post experiment with a highly negative correlation (-0.82)3. In this context, consumers seem to negatively associate the new and well accepted flavour of sausage with spices and naturally smoked with the private label. For the economic interpretation, the implicit price of the attributes levels was calculated following equation 8. Since these estimates are stochastic, it is usual to calculate their confidence intervals. In this study we employ the method proposed by Krinsky & Robb (1986) through 1,000 random repetitions. Analysing preferences before and after the sensory test, results show significant modifications in the implicit prices of the levels as shown in Table 3. The obtained results, confirms the previously explained results about levels utilities in Table 2. To avoid any misinterpretation, results must be considered as the willingness to pay (€/product) to shift preference from the base level of the attribute to the evaluated one. In this line, in contrast to the castrated animal claim, consumers in the pre sensory were willingness to pay for the non-castrated animal (0.340€/product) while after the sensory experience this aspect was not statistically significant. 3 The correlation matrix estimated in both models was not reported in the paper due to their non-statistical significance (except the flavoured and private brand in the post experiment). 24 Moreover, compared to the manufacturer brand, the private brands were not valued with a negative implicit prices in the pre and post sensory experience. This results showed that consumers ask for a price discount to accept the private brand of about 0.252€/product and 0.342€/product in the pre and post sensory test respectively. Finally, the willingness to pay for the proposed flavour was negatively valuated before testing asking for about 0.558 discount by product and showing a significant change of preferences after testing the product, revealing a willingness to pay of 0.66€ for the masking strategy. Table 3: The willingness to pay from the GMNL Model. Pre Sensory Post sensory Non-castrated animal 0.340**a (0.06; 0.62) -0.022 b (-0.12; 0.08) Private Brand -0.252***a (-0.38; -0.1) -0.342**a (-0.62; -0.060) Flavoured1 -0.558**b (-0.99; -0.12) 0.660**a (0.12; 1.18) Significance levels: *** p<0.01; **p<0.05; * p< 0.10. a,b: Differences between preferences (pre sensory and post sensory) within each group at 95%. 1 with spices and naturally smoked. Conclusion Our evidence suggests the appropriateness of the proposed mixture of spices and smoking as a valid masking strategy of boar tainted meat to be used as a raw material for the production of frankfurter sausages. The results proves that this strategy was able to masking boar taint odour and flavour and does not diminish the consumer acceptability in these type of products. In addition, consumers showed their willingness to pay a premium for this type of flavours. In this context, this strategy may represent the starting point for the processed meat industry to adapt their receipt that allows to include the boar tainted meat for the production of frankfurter sausages. However, more research is needed to study other meat products and new cooking strategies. The sensory experience for the different frankfurter sausages with the main identified strategy of masking boar taint, have had impact on both the scale and preference heterogeneity by affecting the tau and the gamma mixing parameter. The hedonic test, decreased the degree of randomness and uncertainty of consumers in their final election and the source of unobserved heterogeneity obtained from the scale become more independent than the taste preference. These results may highlight the importance of the direct promotion of these types of product in the retail 25 point by giving potential consumers the opportunity to test the product. These advertising activities in the point of sales may represent a valid way to promote consumers to purchase the product by decreasing their level of uncertainty. In this context, more studies are needed to be done, especially to analyse the impact of sensory experience on the attributes non-attendance and on the internal and external validity of choices. These topics are proposed for future research. Acknowledgements This study was financially supported by National Institute for Agronomic Research of Spain (INIA): Market potential and quality of meat and meat products from entire males. European perspectives of banning pig castration. BOARMARKET; RTA-2011-00027-C02-01. References Asioli, D., Canavari, M., Pignatti, E., Obermowe, T., Sidali, K. L., Vogt, C., & Spiller, A. (2014). Sensory experiences and expectations of Italian and German organic consumers. Journal of International Food & Agribusiness Marketing, 26(1), 13-27. Bañón, S., Costa, E., Gil, M. D., & Garrido, M. D. (2003). A comparative study of boar taint in cooked and dry-cured meat. Meat Science, 63(3), 381-388. Bech, M., & Gyrd‐Hansen, D. (2005). Effects coding in discrete choice experiments. Health economics, 14(10), 1079-1083. Ben-Akiva, M., D. McFadden, M. Abe, U. Böckenholt, D. Bolduc, D. Gopinath, T. Morikawa et al. 1997. Modelling methods for discrete choice analysis. Marketing Letter, 8(3) 273–286. Blanch, M.; Panella-Riera, N.; Chevillon, P.; Font i Furnols, M.; Gil; M.; Gil, J.M.; Kallas, Z. and Oliver, M.A. (2012) Impact of consumer’s sensitivity to androstenone on acceptability of meat from entire males in three European countries: France, Spain and United Kingdom. Meat science, 90(3), 572–578. Bonneau, M. (1998). Use of entire males for pig meat in the European Union. Meat Science, 49, S257-S272. Bonneau, M., & Chevillon, P. (2012). Acceptability of entire male pork with various levels of androstenone and skatole by consumers according to their sensitivity to androstenone. Meat science, 90(2), 330-337. Bonneau, M., & Lebret, B. (2010). Production systems and influence on eating quality of pork. Meat science, 84(2), 293-300. Borrisser-Pairó, F., Kallas, Z., Panella-Riera, N., Avena, M., Ibáñez, M., Olivares, A. (2014). Analysis of the stakeholders’ attitudes for banning of castration of male pigs in Europe: focus group methodology. Paper presented at the 60 th International Congress of Meat Science and Technology, Punta Del Este, Uruguay. Brazell, J., Diener, C., Karniouchina, E., Moore, W., Séverin, V. & Uldry, P. (2006). The no-choice option and dual response choice designs. Marketing Letters, 17(4), 255-268. Bredahl, L., Grunert, K.G., & Fertin, C. (1998). Relating consumer perceptions of pork quality to physical product characteristics. Food Quality and Preference, 9(4), 273-281. Combris, P., Bazoche, P., Giraud-Héraud, E., & Issanchou, S. (2009). Food choices: What do we learn from combining sensory and economic experiments? Food Quality and Preference, 20(8), 550–557. De Campos, C. F., Lopes, M. S., e Silva, F. F., Veroneze, R., Knol, E. F., Lopes, P. S., & Guimarães, S. E. (2015). Genomic selection for boar taint compounds and carcass traits in a commercial pig population. Livestock Science. doi:10.1016/j.livsci.2015.01.018, In Press. De Oliveira Faria, M., Cipriano, T. M., da Cruz, A. G., dos Santos, B. A., Pollonio, M. A. R., & Campagnol, P. C. B. (2015). Properties of bologna-type sausages with pork back-fat replaced with pork skin and amorphous cellulose. Meat Science, in Press.