Menu
Log in
The Sensometric Society
 



Abstracts - Keynotes

K1 - Protecting Data from Users

Paul Holtzman
Consultant, Chicago

Through the years, I have worked with an assortment of companies essentially developing software for platforms. These platforms serve as the base for a process from which the company’s clients or customers could address research questions... a process including the steps of creating and deploying surveys and then analyzing and reporting on data captured by those surveys, as appropriate to the research questions. Much of my work has been oriented towards achieving the goal of automating this process as much as possible, to assist users of the platform. Back in the old days, the process culminated in a report, provided with a good bit of researcher intervention. Today, in the age of democratization of software, the trend is towards users executing steps along the process by themselves, with little to no intervention.

A point of great concern, amplified by democratization, has always been generating, and then handing off, results to users who may or may not fully appreciate what they were receiving. Some of the concern could be addressed by direct contact with users, for example via training. But this has been replaced to some extent by reliance on annotation and result-triggered comments (with the occasion link to a “white paper” or blog note) added to reports automatically generated on the platform. (Just to be sure, I am a strong supporter of annotation, triggering and offering links.) These can be helpful but only if the user reads and understands them… often doubtful. A more subtle way of addressing this concern, for me, has been building protections into the process, offered at least as options. This is the focus of my monologue. And just to be sure, by “protections”, I mean “protecting data from the user”. 

Data are meant to be innocent (themselves having no agenda) pieces of information, evidence, then reduced via analysis to some insight or result upon which a decision can be made regarding “next steps”. At issue then is how can data be captured, analyzed and reported protecting this information, ensuring at least a reasonably clear direction for the required decision, reported with an equally clear sense of risk assessment.

I’ll focus on a small set of guidelines: (1) simplicity, (2) transparency, (3) robust and unbreakable analyses all providing results that are (4) transportable beyond the specific data set from which they were extracted. These guidelines are then applied to three parts of the process mentioned above: (1) survey construction, specifically measurements, and sampling plan (a good bit of pressure to needs to be applied here to capture those innocent data), (2) analysis, (3) reporting results.


K2 - Visualising and Exploring Multi-Dimensional Sensory Data

Sugnet Lubbe
MuViSU, Department of Statistics and Actuarial Science, Stellenbosch University, South Africa

The biplot, as introduced by Gabriel in 1971, is a very useful tool for visually exploring multi-dimensional data sets. Here we will take a different point of view, considering the biplot as a multi-dimensional scatterplot. Different data set from the literature and sensory analyses of wine will be used to illustrate the use and characteristics of biplots. Attention will be given to incorporating ordinal categorical data, mixtures of numerical and categorical data, multi-way data as well as grouped data set.


K3 - AI - wonder weapon or hype

Andrea Ahlemeyer-Stubbe
servicepro GmbH and Data Mining & More

In the field of mass data processing, Industry 4.0, text and image processing and in marketing, especially in dealing with customer behaviour and customer needs, AI is a key to success. The advantages of automation, speed and a high degree of individualisation contribute to this. A special aspect is the creation of artificially generated but seemingly real statements/images. The application possibilities are manifold and will generate new insights now and in the future, especially through simulations.

But unfortunately, AI is often only used as a prestige-promoting buzzword and the actual processes behind it are neither new nor particularly spectacular. The keynote offers possibilities for classification and demarcation, which arise from the area of tension between the expert-based, classical, experimental design and statistical methods and the approach of AI. Particularly in areas of application that in the past often worked with very small data derived from controlled surveys and used the large spectrum of static methods, it is necessary to weigh up how and whether AI can currently make an additional contribution to success.


Abstracts - Oral presentations

O1 - Beyond Liking: Innovative approach using CATA data to better understand consumer’s associations to products

Thierry Worch, Feline Heussen, Jonathan Rason
FrieslandCampina, Amersfoort, Netherlands

To ensure the success of food products on the market, there is a strong need to deepen the understanding of our products by going ‘beyond liking’. This can be done by measuring the intrinsic (e.g. sensory perception) and extrinsic (e.g. emotional, abstract, and/or functional associations) properties of products. 

In recent studies, such product descriptions (intrinsic and extrinsic properties) are often measured using the Check-All-That-Apply (CATA) task, as it is a simple, fast, and reliable task. However, in those studies, the data collected are often analyzed separately and/or in an aggregated way. Although such approaches provide some relevant information regarding the products, it ignores the individual responses provided by the respondents themselves.

In this presentation, a new methodology that looks at the individual CATA responses through the co-ticking pattern of consumers for pairs of terms is proposed. By combining the intrinsic and extrinsic terms in the same analysis, one can then evaluate how these different types of data connect. For instance, which sensory attributes affect the perception of quality? Additionally, a statistical test through simulations was implemented to assess the strength of the connection between pairs of terms. These connections are then visualized through network graphs. 

For illustration, this methodology is applied to a set of 7 samples: from this study, different connections between the sensory characteristics and the extrinsic properties were found across products. For example, Sample E has a strong association to a golden brown color and crispiness which provides an impression of quality that Sample A does not have. This translates in a higher average liking score for Sample E than for Sample A.


O2 - Exploring food pairing through a cross-cultural projective mapping

Araceli Arellano-Covarrubias1Carlos Gómez-Corona2Paula Varela3Mara Galmarini4Héctor B. Escalona-Buendía1
1Universidad Autónoma Metropolitana, Mexico City, Mexico2XOC Estudio, Mexico City, Mexico3NOFIMA, AS, Norway4Pontificia Universidad Católica Argentina (UCA), Argentina

Culture is a driver of food choices and, therefore, food pairing preferences. Food pairing has been widely studied where food researchers intend to find successful food and beverage combinations; however, little cross-cultural research can be found. In this research, food pairing was explored from a cross-cultural perspective through projective mapping. Four countries (Mexico, Argentina, France, and Norway), thirty foods and six beverages were selected. Projective mapping methodology was applied online to 100 consumers for each country, in order to create maps that represent successful food pairings among consumers. Participants were instructed to position each food item in a limited area close to a food or beverage that represented a good combination, according to consumer preferences. For each country, the positions of each product (coordinates) were analyzed through Multiple Factorial Analysis (MFA). An Agglomerative Hierarchical Clustering (AHC) was performed to the participants’ coordinates of the MFA, in order to make a segmentation of the participants. Four clusters for each country were selected and the participants from the cluster with the least number of participants (outliers) were discarded. A Generalized Procrustes Analysis (GPA) was performed to the original coordinates of each country (without outliers) and for each one of the participants’ clusters, in order to create consensus maps. The two factors of each GPA were used to perform multiple RV coefficient analyses to test the similarities between the countries’ coordinates and between the clusters’ coordinates. The two factors of each GPA were used to perform k-means clustering, using the coordinates of the six beverages as a starting clustering point; and therefore, to find the foods that were paired with each beverage. Results of RV coefficients showed that the coordinates were not similar, and consequently, the food and beverage pairings were not similar either (RV<0.66; p-value<0.0001 for all comparisons).  Some differences and similarities in food-beverage pairing were found between countries and between participants’ clusters in each country. For example, in Mexico and Norway, blond beer was clustered with salty snacks and peanuts. Both red and white wines were clustered with red meat in Argentina, Mexico, and France. In Norway, red wine was clustered with red meat and cheese, while white wine was clustered with some other foods, such as fruits like mango, apple, orange, and pineapple. In general, food and beverage pairing was effectively explored using projective mapping, from which several differences and similarities were found between cultures and between clusters of participants.


O3 - The pain of accurate texture profiling and magnitude estimation

Jason Cohen1Benedicte Bimont2
1Gastrograph AI, New York City, USA2Gastrograph AI, Hong Kong, Hong Kong

While the majority of sensory research focuses on flavor and aroma profiling and intensity scaling, texture, which plays a major role in the acceptance and preference of products, remains under studied. Profiling texture is more difficult and complicated than profiling flavor and aroma as texture terms change inherent meaning across contexts (for example, crunchy in a corn chip, crunchy in a potato chip, and crunchy in an apple), change intensity scales across those contexts, and for unstructured sets of overlapping meaning in different contexts (for example: gritty, sandy, and sharp). There is a larger lexical space as well when talking about texture, that varies depending on the language. This is often due to a great variation in vocabulary richness of the languages and leads to a lack of consistent categorization of sensory results across languages (a word for one texture in one language can sometimes be used to describe multiple texture attributes that would be described by distinct terms in another language: for example, the French word gluant can be used to describe gooey, gluey, sticky, or slimy in English). 

The current methodologies have major limitations: using a restricted selection of attributes (a constrained lexicon) for evaluation provides an incomplete profile of the product’s texture, loses inherent data structure from the natural co-occurrence of terms, and artificially inflates the occurrence of textures present in the lexicon; in contrast, free-choice profiling, where panelists use their own words to describe the product (an unconstrained lexicon), faces the opposite problem of a large low-probability set of infrequent terms due to inconsistencies between panelists vocabulary which are not replicable across panelists, products, and categories. 

Finally, using traditional methods to generate textural descriptors does not result in accurate magnitude estimations because of the overlap in inherent meaning between the terms; all standard methods of summation, averaging, or minimum weights would result in a inaccurate estimate of the overall intensity of the product or relative intensity between attributes. 

To solve the problems of contemporary methods in analyzing texture profiles, we propose a novel approach using Artificial Intelligence to analyze data from lexicon-free modified CATA, where panelists can select the textures they perceive from a lexicon and add manually any additional textures they can perceive. Textures can be profiled accurately, magnitudes of each attribute-cluster can be predicted, and local ontologies via synonyms can be clustered.


O4 - Application of SO-PLS Regression in Handling Multiblock L-shape Data

Quoc Cuong Nguyen1,2Daniele Asioli3Paula Varela4Tormod Næs4
1Ho Chi Minh City University of Technology, Ho Chi Minh City, Vietnam2Vietnam National University Ho Chi Minh City, Ho Chi Minh City, Vietnam3Department of Agri-Food Economics and Marketing, School of Agriculture Policy and Development, University of Reading, Reading, United Kingdom4Nofima AS, Ås, Norway

L-shape data includes three blocks of data such as sensory properties, consumer degree of liking, consumer attributes. The analysis of these three blocks of data can provide useful information for food product development and marketing and communication strategies. To investigate L-shape data, either L-Partial Least Square (L-PLS) regression or a Two-step PLS based Procedure (TSP) can be used. In two of the main blocks, i.e., sensory properties and consumer attributes, can consist of variables representing different aspects. This study proposes a new way to deal with this combination of L-shape and multiblock data sets using the Sequential and Orthogonalized - Partial Least Square (SO-PLS) regression in each step of the TSP approach. One of the main advantages of SO-PLS is the ability to handle blocks of different complexity and type (e.g., number of variables and underlying dimension in each block) explicitly.

Eight yoghurts varying in three intrinsic attributes (i.e., viscosity, particle size of oat flakes, flavour intensity) were tasted by a trained sensory panel (description via QDA and TCATA) and by Norwegian consumers who rated their liking using Labelled Affective Magnitude (LAM) scales. Consumers’ attitudes to health and taste were also collected. The multiblock L-shape data then included two blocks of sensory properties (static and dynamic sensory description), one block of consumer liking ratings and two blocks of consumer attitudes.

Consumers were segmented into three segments based on visual segmentation using Principal Component Analysis score and loading plots. Consumers in segment 1 (G1) liked the thin yoghurts while segment 2 (G2) liked flour yoghurts and segment 3 (G3) liked thick-flakes yoghurts.

In step 1, the consumer liking ratings are regressed onto static and dynamic sensory descriptors using SO-PLS regression. Results indicates that (i) segment G1 preferred products with thin, gritty textures and sweet flavour, (ii) segment G2 preferred products with dry, sandy perception, (iii) segment G3 preferred products with thick texture.

In step 2, consumer dummy variables (0/1 representing the three segments) are regressed onto attitudes to healthiness and taste of food using SO-PLS regression. We found that (i) segment G1 was linked to taste attitudes, (ii) segment G2 did not relate clearly to either health or taste attitudes, (iii) segment G3 was associated with attitudes to healthiness.

These findings demonstrate that Sequential and Orthogonalized - Two Step Procedure (SO-TSP) is a useful tool to understand the relation between consumer segments and different types of sensory properties and consumer attributes.


O5 - Application of multi-group multiple correspondence analysis in sensory data

Aida Eslami1,2Lauren Faye Toogood3Soudeh A. Khoubrouy4Hervé Abdi5
1Institut Universitaire de Cardiologie et de Pneumologie de Québec, Quebec, Canada2Department of Social and Preventive Medicine, Université Laval, Quebec, Canada3University of Exeter, Exeter, United Kingdom4The University of Texas at Dallas Erik Jonsson School of Engineering and Computer Science, Richardson, USA5School of Behavioral and Brain Sciences, The University of Texas at Dallas, Richardson, USA

In multivariate analysis, when variables are categorical the standard descriptive exploratory method is multiple correspondence analysis (MCA). MCA assumes that observations are independent and originate from a homogeneous population. However, the observations often comprise several groups known a priori (e.g., sex, ethnicity), a configuration known as multi-group data structure. In multi-group data, individuals from the same group are likely to be more similar to each other than to individuals from other groups. Several methods have been developed to consider a group structure in continuous multivariate analysis approaches, such as multigroup principal component analysis. To take into account the group structure in MCA, we recently developed a new method called dual STATIS-MCA. This method lets us create both global and group components and loadings. In addition, to measure the similarity between these global and group components and loadings we used different approaches based on (1) the vector correlation (RV) coefficients—a multivariate generalization of the squared Pearson correlation coefficient—and (2) Tucker’s congruence coefficient.  We illustrate this new procedure with a real case study in sensory data.


O6 - Evaluation of complementary numerical and visual approaches for investigating pairwise comparisons after principal component analysis

John Castura1Paula Varela2Tormod Naes2
1Compusense Inc., Guelph, Canada2Nofima A/S, Ås, Norway

We propose and evaluate numerical and visual methods for investigating paired comparisons after principal component analysis (PCA). PCA results can be visualized to facilitate an understanding of the relationships between the products and the sensory attributes. But identifying and visualizing significant product differences in multiple PCs simultaneously is not straightforward. A benefit of the proposed methods is that they provide a screening tool for evaluating PCA results rapidly. We begin with a real data set which is analyzed and submitted to the truncated total bootstrap (TTB) procedure. This TTB procedure simulates and analyzes results from virtual panels. The TTB-derived results form clouds of uncertainty around each product and paired comparison. These clouds can be visualized directly or by plotting the contours that enclose the highest 95% of their kernel-estimated densities. We find that these density regions tend to be smaller and mostly fit inside the 95% confidence ellipsoids that we propose in this manuscript. We show how to calculate the volumes of these confidence ellipsoids, which quantify uncertainty. We also show how to calculate P values to evaluate whether pairs of products are discriminated in the PCA subspace. The interpretation of these P values coincides with the visual interpretation of the confidence ellipsoids. We illustrate the methods with two real data sets, one a sensory quantitative-descriptive sensory data set from a trained panel, the other a consumer check-all-that-apply (CATA) data set. We also conduct a simulation study based on each of these data sets. The results from these simulation studies show that under repetition, approximately 95% of the ellipsoids covered the true result. This indicates that the proposed ellipsoids have an approximately frequentist interpretation. The complementary numerical and visual approaches can be applied to a wide range of data sets from sensory evaluation and to data from other domains.


O7 - Text mining as a tool in food sensory and consumer science. Hype or help?

Ingunn Berget1Charo E. Hodgins2Monique M. Raats2Lada Timotijevic2Ellen Van Kleef3Paula Varela1Martina Galler1
1Nofima, Ås, Norway2Food, Consumer Behaviour and Health Research Centre, School of Psychology, Faculty of Health & Medical Sciences, University of Surrey, Surrey, United Kingdom3Wageningen University and Research, Marketing and Consumer Behaviour Group, Wageningen, Netherlands

Text mining is used as a data led approach to explore both information online as well as consumer generated texts. Text mining is an interdisciplinary field comprising elements from mathematics, statistics, computer science and linguistics. Currently text mining is emerging as a potential tool in a range of scientific disciplines, including food sensory and consumer science. However, food sensory and consumer science literature shows that many studies tend to confirm findings gained using traditional research methods. Moreover, validation and interpretation of results can still be extremely time consuming. The potential for text mining to contribute to substantive new insights needs to still be established.

Papers on applications of text mining show a large variation not only in data sources, but also in data set size both regarding number and length of texts. Finally, there is also considerable variation in how data collection procedures, pre-processing and analyses are documented.  There is a need to establish rigorous procedures to incorporate text mining into food sensory and consumer science. To ensure these methods are used to their full potential there is a need for tools and systematic studies comparing different methods and data sources to develop a knowledge base for efficient use of text mining tools.

One type of analysis often performed for exploratory research on texts is topic modelling (TM). This refers to a statistical model that extracts a number of topics in a collection of texts. Typically, different types of coherence scores are used to select the number of topics in a heuristic way. To our knowledge the impact on the number of texts and the length of texts on the coherence measure has not been investigated thoroughly.  In this work we aim to explore the impact of varying the number and length of texts on the coherence score in topic modelling on a dataset extracted from web-harvesting using a third-party company (webz.io) based on key words related to meat reduction (January 2020). We will present typical challenges experienced when working with this type of data.


O8 - Reinventing "sensory data science" learning through digital technology: the JAR case study, a play in five acts

Sébastien Lê
L'Institut Agro Rennes Angers, Rennes, France

Digital pedagogy is not simply about using digital technologies to teach. It's about using them wisely: it's as much about using digital tools thoughtfully as it is about deciding when not to use digital tools, and paying attention to the impact of digital tools on learning.

In other words, digital pedagogy is the use of digital tools to enhance the educational experience.

The aim of this presentation is to introduce a new platform dedicated to the teaching of sensory data science. Beyond the classical digital tools that this platform integrates (quizzes, slides, videos), the learner has the possibility to run R code to analyze their data.

A detailed example will be presented around the analysis of data integrating JAR scales: from standard JAR data related to liking, to free JAR data (Luc et al.), without forgetting JAR data associated with the adequacy towards a given concept.

We will also present how this platform was built and provide feedback on both the construction of the platform and its use.


O9 - Analyzing Temporal Dominance of Sensations data with Categorical Functional Data Analysis

Caroline Peltier1,2Michel Visalli3,2Pascal Schlich3,2Hervé Cardot4
1Centre des Sciences du Goût et de l’Alimentation, CNRS, INRAE, Institut Agro, University of Bourgogne Franche-Comté,, Dijon, France2CNRS, INRAE, PROBE research infrastructure, ChemoSens facility, Dijon, France3Centre des Sciences du Goût et de l’Alimentation, CNRS, INRAE, Institut Agro, University of Bourgogne Franche-Comté, Dijon, France4Institut de Mathématiques de Bourgogne, UMR CNRS 5584, Université de Bourgogne, Dijon, France

In the last years, Temporal Dominance of Sensations (TDS, Pineau et al., 2009) was extensively used for obtaining temporal sensory profiles of food products. Classical analysis of TDS data relies on either dominance rate (proportion of subjects having chosen a given descriptor at a given point in time) or dominance duration (time during which a descriptor was perceived as dominant). None of these two approaches explicitly considers the sequence of descriptors (probability to go from to another descriptor). To overcome this limitation, the use of homogeneous Semi-Markov processes was advocated by Lecuelle et al. (2018). However, this approach seems to have not been used by other sensory scientists so far.

Recently, Preda et al. (2021) presented an R package including categorical functional data analysis (CFDA). This statistical approach, based on the seminal work by Deville (1982), extends the usual functional data analysis to temporal categorical data, and as such could be particularly relevant for TDS data.

This paper illustrates the relevance of CFDA for TDS data by applying it on both simulated and real TDS data from a study on 4 products assessed by 70 subjects on 12 descriptors. 

CFDA produces a PCA-like map of the sensory evaluations (subject x product) based on the sequences of the sensations. Each axis represents leading temporal patterns and the individual coordinates on these axes describe their main temporal characteristics.

CFDA outputs can also be used as inputs for further statistical analyses (clustering, discriminant analysis, etc.). Thus, important topics in sensory evaluation, such as segmentation of consumers and product discrimination can be addressed in this framework taking temporality of perception into account.

References:

Deville, J.-C.. (1982). Analyses de données chronologiques qualitatives: Comment analyser des calendriers? Annales de l’inséé, 45(45), 45. https://doi.org/10.2307/20076433
Lecuelle, G., Visalli, M., Cardot, H., Schlich, P., (2018) Modeling Temporal Dominance of Sensations with semi-Markov chains, Food Quality and Preference,67, 59-66,
Pineau, N., Schlich, P., Cordelle, S., Mathonnière, C., Issanchou, S., Imbert, A., Rogeaux, M., Etiévant, P., & Köster, E. (2009). Temporal Dominance of Sensations: Construction of the TDS curves and comparison with time-intensity. Food Quality and Preference. https://doi.org/10.1016/j.foodqual.2009.04.005
Preda, C., Grimonprez, Q., & Vandewalle, V. (2021). Categorical functional data analysis. The cfda r package. Mathematics, 9(23), 1–31. https://doi.org/10.3390/math9233074


O10 - Incomplete-block designs for rapid, holistic sensory methods: Comparing sorting and linking methods with a novel resampling approach

Jacob Lahne1Marlon Ac-Pangan1Marino Tejedor-Romero2David Orden2
1Virginia Tech, Blacksburg, VA, USA2Universidad de Alcalá, Madrid, Spain

A key limitation on sorting and projective-mapping tasks for sensory evaluation is the requirement that all samples be presented and sorted simultaneously by every panelist.  For food samples, this imposes an upper limit of around 25 samples per study because incomplete blocks are not compatible with the underlying cognitive task of groupwise similarity.  Recently, the free-linking task was developed as an alternative to sorting; briefly, free-linking asks subjects to draw a set of binary connections among samples in an individual “similarity graph”, from which it is possible to produce analyses of similarity equivalent to sorting-task results.  Because the cognitive task underlying free-linking is pairwise rather than groupwise similarity, it should be possible to use incomplete-blocks designs in free-linking, increasing the number of samples that can be analyzed.  Here, we present the results of a study in which we evaluated the feasibility of incomplete-block designs in similarity tasks.  We chose the 62 terms of the Chocolate Flavor Wheel as our sample set in order to have a known “ground truth” grouping of samples.  Subjects were assigned to one of three experimental conditions: a complete free-sorting of all 62 samples (N = 53), or either an incomplete free-sorting (N = 58) or free-linking (N = 30) of 3 blocks of 16/62 samples.  All results were treated as graph dissimilarities, and consensus groupings for each condition were determined using additive-tree splitting.  Incomplete sorting and linking were evaluated for graph parameters (density, degree, transitivity, cophenetic measures), and similarity to “gold standard” complete sorting and “ground truth” groupings (Rand index).  In order to determine the stability and variation in these results, we present a novel, graph-derived resampling approach to the incomplete blocks that avoid the propagation of missing cooccurrences.  We find that overall there are significant qualitative and statistical differences in groupings between incomplete linking and incomplete sorting, with linking producing results that are more discriminating, empirically reasonable, and which remain close to the gold standard and ground truth groupings.


O11 - Imputing Consumer Liking Data using XLSTAT: How Much Missing Data Can I Have? Which Method Should I Use? Can I Cluster Based on Imputed Values?

Joshua Brain1Anne Hasted1,2Ian Wakeling1
1Qi Statistics, London, United Kingdom2University of Nottingham, Nottingham, United Kingdom

Balanced incomplete designs in consumer research are a common occurrence, affording more samples to be evaluated across a pool of consumers than a complete design. The limitation from an analysis perspective is what to do with the missing data for the samples not evaluated per consumer, as each consumer only sees a subset of the total number of samples. 

The missing data of primary interest to consumer scientists is a liking-based metric and leads to several practical issues:
(i) why should I impute missing data and what is the preferred imputation method offered for missing liking data?
(ii) what is the maximum amount of missing data one can have before the quality of results is impacted?
(iii) can adding more consumers to the sample improve the quality of imputations?
(iv) how valid is clustering liking data that contains varying amounts of imputed values?

Our research aims to tackle the above. We test different imputation methods by converting full consumer data sets (FMCG) to balanced incomplete designs with varying amounts of imputed data. Imputations can therefore be compared to the respective full data set. The data sets we consider vary in the number of samples evaluated, in order to be reflective of the most frequent designs that used in consumer science We also evaluate the performance of selected clustering algorithms based on imputed data versus data complete, balanced designs.

The audience for this research: consumer scientists who analyse their own data yet do not have the time to code. To be relevant for consumer scientists who desire applicability to their everyday work, we’ve chosen to work in XLSTAT, as this is one of the most popular choices in the industry, across different product categories. Subsequently, we will be able to provide useful and practical recommendations to a wide consumer science audience. No prior knowledge of imputing data is required as this is the first step of our research into the applicability of imputation algorithms in popular software.


O12 - Equivalence and non-inferiority tests using replicated discrimination and preference data

Michael Meyners1B. Thomas Carr2,3Joachim Kunert4
1Procter & Gamble Service GmbH, Kronberg, Germany2Carr Consulting, Wilmette, IL, USA3Charles Sturt University, Wagga Wagga, Australia4Department of Statistics, TU Dortmund University, Dortmund, Germany

Using replications in discrimination tests is becoming more common in times of strict budgetary and time constraints. The analysis of replicated test data is well-understood for decades now; the standard analysis using the naïve binomial test can be used. However, it is no longer applicable if the objective is to show equivalence or non-inferiority, as potential differences between assessor (assessor heterogeneity, or overdispersion) might invalidate the binomial test. We reapply ideas described earlier for the development of a confidence interval for such data in order to derive a direct test for equivalence and non-inferiority using replicated discrimination and preference data. We derive an asymptotic test both for the case of equal number of replications for all assessors, and for the situation in which the number of replications varies. We provide tables with critical values and a power assessment which indicate that the loss in power is moderate only if the number of replications is not excessive. We exemplify its use on three small examples using data published earlier. This work closes a significant gap in the tool kit for the analysis of replicated discrimination and preference data and is expected to help abandoning the flawed approach of concluding for equivalence when a test for differences does not turn out to be statistically significant.


Abstracts - Poster presentations

P1 - Quality grading of pre-treated yams

Cecille AboagyeSeloame NyakuMaame AdjeiPriscilla AhadziEmmanuel Addo-Preko
University of Ghana, Legon, Ghana

Yam is a very important staple in Ghana. One major problem usually encountered with yams in Ghana is nematode infestation. Nematode infestation causes the yams to have a scaly appearance, develop cracks on the surface and have dark surface patches. This reduces their visual appeal causing financial loss to both farmers and yam traders. To control this, a technology was developed with banana paper and treated with abamectin to prevent the growth of nematodes during the growth of the yams. Three yam-growing communities in Ghana were selected for this study. In each yam community, the most popular yam variety was used for the test. For each yam community, four farms were selected for cultivating the yams. The yam seeds were treated in three different ways before planting. The treatments included banana paper treated with abamectin, untreated banana paper and farmer’s own practice. To determine if the sensory properties of the yams had undergone changes due to the treatments, yams were harvested and cooked and the sensory properties were discussed.

Three different activities were conducted to understand the sensory properties of the yams. The activities included cooking the yams, grading the quality of the yams, and discussing the properties of the yam and the technology used in a focus group discussion. In each community, between 16-18 participants took part in the test. Participants were to be either persons who consumed yam at least twice a week, or yam distributors, yam farmers, yam processors or extension officers. Participants were put into six groups for the cooking session. Two groups prepared the same dish. The dishes included boiled yam, roasted yam and pounded yam.  Participants were provided with a 3-point (good, okay, bad) scorecard to grade the qualities of the yam before and after cooking. Then a focus group discussion followed to understand their perception and liking of the yams, and their readiness to embrace the new technology. 

Both fresh and cooked products for all treatments were graded as good or okay and none was graded as bad. Most participants were willing to try the new technology if the trial was repeated. The sensory qualities of the yams in the fresh and cooked form did not change much when compared to the farmer’s practice.



P2 - Perception of yoghurt quality among a cross-section of Ghanaian consumers

Elijah ArhinfulEdmond ArhinMaame AdjeiCecille AboagyePriscilla Ahadzi
University of Ghana, Legon, Ghana

Yoghurt is an important dairy product and has numerous health benefits. Yoghurt quality is not only dependent on meeting standards and specifications but also dependent on consumer acceptability and preference. This project aimed at studying what consumers perceive quality yoghurt to be. The study was divided into two parts. 

Part one was a consumer survey to determine what consumers thought about product quality and yoghurt quality through the use of a pretested electronic questionnaire. Responses were obtained from 253 participants over a span of two weeks. Information on demographics, yoghurt quality perception, consumption patterns of consumers and the sensory properties that influence their choice of yoghurt was obtained.

Part two of the study involved the use of quantitative descriptive analysis to develop sensory profiles for a few selected yoghurt products using the most consumed yoghurts as indicated by the consumers. Eight (8) experienced and well-trained assessors from the University of Ghana Department of Nutrition and Food Science Sensory Evaluation Laboratory descriptive panel database were selected for the test. They developed sensory descriptors for six yoghurt samples of which 2 were frozen. The frozen yoghurts were thawed and served in liquid form like the others. A consensus list of all descriptors was compiled for all sensory modalities of the yoghurt samples. Assessors scored the intensities of the different attributes in triplicate using a ranking scale. Tukey pairwise comparison was used to determine the statistical differences in the sensory attributes derived during the descriptive analysis. Spearman Correlation Test was run on the data obtained from the questionnaire to find the correlation between the ranks of the data obtained.

Results showed that Fanyogo was the most preferred yoghurt by consumers followed by Hollandia. According to the responses, the most normally consumed yoghurt included Fanyogo, Hollandia, Fanmaxx, Yomi, Dolait, Varsity, Gogurt and Goyo. Terms that described the sensory properties of the yoghurt samples included creamy, particulate, milky, viscous etc. For most consumers, a quality product should be safe for consumption, should be consistent in meeting requirements and should also comply with all regulatory requirements. Also, quality yoghurt should be thick, have uniform flavour and colour distribution, have low-fat (health-wise) should be sweet and smooth in the mouth, be highly available, should be cheap and have good packaging.  


P3 - Using A Consumer-Led Product Development Approach to Develop Communion Wine Drink from Roselle Calyx (Sobolo) to be used in Ghanaian Churches

Emmanuel Addo-PrekoMaame Adjei
University of Ghana, Legon, Accra, Ghana

Consumer input in product development contributes to product success by assuring product appeal and acceptability. The consumer-led product development process is thus instrumental to product developers when formulating or optimizing new products. Incorporating consumer insights in product development for small-scale manufacturers can be challenging due to the skills gap and resources required. Communion wine is a religious product and is consumed in small quantities, in spite of these attributes of the product, consumers still demand a good sensory experience when consuming it.   In this study, we used focus group discussion (FDG) as a tool to guide the optimization of a communion wine drink (CWD) product being processed by a small-scale processor by exploring how religious sentiments may influence consumer appeal for communion wine products.  Two different FGDs were conducted to discuss consumer expectations and appeal for commercialized CWD products, and optimized prototypes of an existing product from the small-scale processor. The groups differed in age; a young group with participants aged between 18 and 25 years and a mature group with participants aged between 35and 65 years. As part of the discussion, participants tasted nine (9) CWD products. The FGD results highlighted a clear distinction between what the younger and mature adult groups preferred. Both groups agreed on the same products that were clearly not liked as a CWD. The reasons for the difference in products that were liked were based both on sensory differences between the products and how those differences affect the sentimentality of the religious act of taking communion. The FGD was insightful to understand how religious sentimentality influences product appeal. The sensory properties that make consumers like CWD were also understood. The findings provided guidelines for the small-scale processor on product positioning and for the product development team, insights to guide the direction for product optimization were achieved.


P4 - Perceptual Mapping of Beer in Ghanaian Consumers

Portia Owusu1Maame Yaakwaah Adjei1Carlos Gomez-Corona2Esther Sakyi-Dawson1Cecille Wendy Aboagye1Emmanuel Addo-Preko1Elijah Arhinful1
1University of Ghana, Accra, Ghana. 2Firmenich SA, Satigny, Switzerland

One of the oldest alcoholic drinks is beer, which is widely consumed. Multiple beer varieties have been developed as a result of variations in its processing. The three main types of beer made natively in Ghana are lagers, stouts, and ‘pito’ (a traditional Ghanaian beer) with a few craft beers slowly gaining popularity. Preference of one beer over the other can be influenced by the attributes of the beer; taste, foam, color, aroma, lacing, viscosity etc. Diversities in consumer preferences provides a wide range of standpoints for the production of creative beer styles with specific characteristics that are mostly appreciated by these consumers. However, there is a lack of information on beer consumer preferences in Ghana. In order to create distinct beer styles targeted at a certain group, it is important to understand the main factors that influence consumer preferences for beer. This will help Ghanaian brewery industries make inventive beers and boost their competitiveness with imported beers. This study sought to identify the styles of beer consumed, as well as the beer consumption pattern of consumers in Ghana. To achieve this, an observational study and online questionnaire were used. A total of 10 bars, pubs and clubs were randomly chosen using the balloting approach, for the observational study. Observations were done for a minimum of 2 hours at each location. The questionnaire used was divided into three main components; demographics of the participants, their consumption patterns, and their preferences for beer. Frequencies, means, and counts were used to analyze the survey data. There were 203 complete responses, majority of which were males (71%). The beer styles that were mostly consumed were lagers and ‘pito’ with Club lager beer being the most consumed beer brand (86%) in Ghana. Except for people aged 35 to 44, who ranked past experiences as the most significant factor, popularity was ranked by all age groups as the most important reason for drinking one's favorite beer. Quality attributes like aroma, foam, and carbonation were the least regarded when choosing their preferred beer, however taste, alcoholic content and color were the qualities that were most valued. In conclusion, popularity is the major driving force for purchase of beer by Ghanaian consumers, and beer liking is influenced by taste, alcoholic content and color.


P5 - Exploring the role of carrot juice in enhancing the sensory properties of bissap juice extract

Beatrice Tagoe1Dr. Maame Yaakwaah Adjei1Dr. Idolo Ifie2Cecille Aboagye1Emmanuel Addo-Preko1Elijah Arhinful1
1University of Ghana, Accra, Ghana. 2University of Leeds, Lead, United Kingdom

Hibiscus calyces over the years have been used to produce a variety of products such as beverages known in Ghana as “bissap” or “sobolo” due to its brilliant red colour and characteristic sour taste. Hibiscus-based juices are known for their instability during processing and storage. A study conducted showed that addition of carrot juice to bissap could improve its sensory properties. Therefore, the aim of this study was to produce different bissap-carrot blends and monitor how the sensory attributes of the samples changed with time during storage using a difference from control test. 

The samples produced included Bissap +Carrot with preservative (BCSB), Bissap +Blanched Carrot with preservative (BBCSB), and Bissap +Carrot +Sugar +Preservative +Flavour (SB) which were stored at ambient temperature and analyzed on 7, 21 and 35 days of storage. Each product had its own fresh sample as control. For each test day, assessors determined how different the fresh controls were from the ambient stored samples in terms of overall differences as well as differences of appearance, aroma, flavour, mouthfeel, and aftertaste. Fifteen screened panelists used a 6-point category scale with 0 indicating no difference and 6 indicating very large difference. Paired-t test and Dunnett’s test for multiple comparisons were used to analyze the data. 

Results showed that all samples at day 7 of storage were not significantly different from the control samples for all attributes as well as the overall difference. Significant differences for the attributes started to show after 21 days of storage for all samples. Only mouthfeel of BCSB and SB samples showed no significant difference when compared to the control samples. At day 35 of storage, all stored samples had all sensory attributes as well as overall difference being significantly different from that of their control samples. 

It was observed that at day 35 of storage, the aroma, flavour, mouthfeel, and aftertaste of all the samples (BBCSB, BCSB, and SB) had a degree of difference of 1 (very slight difference) when compared to the control samples with only appearance showing a degree of difference of 2 (slight difference). Only BCSB had a degree of difference of 2 (slight difference) in flavour when compared to the control. 

The samples showed changes with time, but the differences seen were slight, giving an indication that the samples may be able to stay longer at ambient condition especially when the appearance is worked on. 



Abstracts - Joint Sensometrics/SSP workshop:
New Technologies in Application: Case Studies & Data Analysis Strategies

W1 - Exploring the usage of smart-speaker surveys in sensory and consumer research

Jen Grady1, Rafael Drabek2
1P&K. 2ACCE

The application of voice-activated, hands-free technology will be explored in the context of consumer sensory research. Two different studies (one in US and one in Canada) were conducted among consumers utilizing Smart-Speaker technology to assess its application, practicality and added value in sensory research. Results from the research will be discussed, along with challenges and application for future use.


W2 - Unlocking the potential: Key industry learnings & experiences with emerging digital technologies

Chrisly Philip, Maria Elena Lozano
PepsiCo

We will share industry perspectives on piloting digital technologies that specialize in virtual product development and sensor systems. First, we will discuss how these technologies complement rather than replace established sensory methods and that when leveraged together have the potential to transform ways of working. Next, we will share best practices for defining research objectives, establishing success criteria, and managing team expectations. Finally, we will attempt to demystify the domain knowledge required to pilot these technologies by demonstrating how sensory science fundamentals and traditional analysis approaches are still applicable and valuable when assessing these technologies.


W3 - How Web3 surveys enrich machine learning models for sensory and consumer science

Tian Yu, Vanessa Rios de Souza
Aigora

Machine learning is becoming increasingly important in understanding sensory and consumer data. The quality of consumer data is crucial in building reliable and robust machine learning models. Using Web3 technology, consumers can safely store their taste and preference data in blockchain-backed digital wallets. These consumer-owned data can be used to build quality models which can predict the consumer liking for specific products or recommend products to a specific consumer.

© 2010 - 2024 Copyright The Sensometric Society.
Powered by Wild Apricot Membership Software