Margarine: The Drivers of Liking and Image
Howard R. Moskowitz
Moskowitz Jacobs Inc.
White Plains, New York
A developing area of product research deals with the relation among product variables when these product variables are rated by the same panelist or by two or more groups of panelists. This paper presents a study of the attributes of margarine, showing the depth of information about consumer perceptions and drivers of liking that emerges from a detailed analysis of relations among attributes. The paper considers three analyses of attribute relations: principal components analysis to identify basic dimensions of perception, linear functions relating overall liking to attribute liking or to image ratings, respectively, and quadratic functions relating overall liking or image ratings to sensory attribute levels. The analyses show how consumer data can be tapped for learning about the consumer perceptions on the one hand, and for product development guidance on the other.
The Language of Food
Descriptive analysis uses many terms to profile the attributes of products. In the case of food, these terms include sensory attributes (amount of), liking (overall, of specific attributes), image (more complex aspects of a product involving cognitive factors beyond sensory ones), performance (person-product interaction), and so-called “directionals” (too much versus too little of an attribute) (ASTM, 1996; Moskowitz, 1994). Often, the focus is on the process of describing the characteristics and not on the subsequent analysis of the interrelations among these characteristics. The latter analysis, properly done, provides a wealth of both scientific knowledge and pragmatic direction for development.
This paper deals with an in-depth analysis of margarine as a case in point. The aim is to understand the interrelation among sensory, liking and image attributes for a single product. The results of the analysis are important for two reasons. First, the approach shows a variety of analytical approaches to more deeply understand the language of food. Second, the data provide substantive results for a particular product -- margarine.
Drivers of Liking - Simplistic Analyses
Drivers of liking are sensory attributes that are responsible for acceptance. Although it is conventional wisdom that overall acceptance is a function of many attributes and their interactions, it is still instructive to identify which particular attributes are the key to moving acceptance. Armed with that specific knowledge the product developer begins to understand what must be done in order to increase product acceptance, even without systematic product manipulation.
The most fundamental analysis considers a set of products in a category, classifies the products into “winning” versus “losing” products (typically on the basis of the overall liking rating), and then searches for those attributes which co-vary with acceptance. The analysis may be quite simple -- sort the products into winners or losers to discover which attributes appear more often with the winners than the losers. The correlation may be more statistical - e.g., calculate the Pearson correlation between overall liking and a sensory attribute. At this basic level no formalized function relates sensory attribute level (a description of the product) to acceptance level (an evaluation of the product). Drivers of Liking - Attribute Liking vs. Overall Liking
Another way to assess drivers of liking considers how individual liking attributes drive overall liking. Typically, attribute liking (e.g., liking of appearance) positively correlates with overall liking. It is not the correlation itself (which only measures the strength of a linear relation), but the actual function relation between overall liking and attribute liking which is important. By expressing the relation between overall liking and attribute liking as a linear equation, the researcher estimates the unit increase of overall liking to be expected from a unit increase in attribute liking. The equation is written as:
Overall Liking = k + k(Attribute Liking). 01
The higher the coefficient k, the more important the attribute liking. In previous 1
publications the author has shown that each sensory input has a differential level of importance (Moskowitz & Krieger, 1995). Typically, flavor is first, followed by texture and then by appearance. The question here is the value of these coefficients for margarine. Drivers of Liking - Sensory Intensity vs. Acceptability
The second analysis fits a parabolic equation to the curve relating overall liking to sensory attribute level. Figure 1 shows the typical sensory-liking curve (in an idealized form). Although many sensory attributes interact to generate overall liking, it is easiest to understand relative importance and discover drivers by doing the analysis on an attribute by attribute basis (Moskowitz, 1981). The analyses involve the nature of the curve, the optimal point and the area under the curve.
Margarine is an especially interesting product to assess because margarine is used in conjunction with other products, unlike other products that have been analyzed in this fashion (e.g., gravies, sauces, pies, etc.). These other products stand by themselves, rather than complement a product. One of the key questions is whether or not a clear sensory-
liking curve really can be developed, given the use of margarine as an adjunct to other food products.
Figure 1 - How sensory attribute level drives overall liking. The prototypical function describes an inverted U shaped curve. The points represent products, the curve represents the fitted quadratic function.
Relative Importance of Sensory Inputs - Area under the Liking Curve
Given the fundamental sensory-liking curve shown in Figure 1, one key index for relative importance is the area under the curve (viz., the area subtended by the curve). To the degree that there is a large area under the curve, the attribute is important. Integrating the quadratic function, and then subtracting the rectangular areas not subtended by the curve (Figure 2) obtain the area. The curve can appear either fairly flat with little area subtended, wide with a great deal of area subtended, or even narrow, but with a great deal of area subtended. Of course, relative area can be large either because the sensory attribute spans a wide range of the scale (giving more area, simply by the sheer expanse of the X-axis) or because the sensory attribute spans a narrow range. However strong drives overall liking or purchase intent, so that liking changes dramatically, even within the narrow range.
Figure 2 - A sensory-liking curve subtending a relatively large area (viz., an important sensory attribute)
Sensory Preference Segmentation and Relative Importance of Attributes
Consumers differ in their sensory likes and dislikes. This inter-individual variation can be reduced to organizing rules in the population by the method of sensory preference segmentation (Moskowitz, Jacobs & Lazar, 1985). The segmentation algorithm creates individual curves relating overall liking to sensory attribute level, and divides consumers by the sensory optimal level (viz., that sensory level which corresponds to optimal liking). Sensory preference segmentation generates groups of consumers who show dramatically different sensory-liking curves. These curves differ from each other far more different for the segments than they do for other subgroups in the population (e.g., users of Brand A Vs users of Brand B, respectively). When the sensory-liking curves are created for the total panel versus for segments quite often the total panel curve looks flat, whereas the sensory segment curves look quite steep (and have different shapes). This difference between sensory segments means that the relative importance of the same attribute may differ by segment, with one segment finding the attribute quite important the other segment finding the attribute less important. Both segments may, in turn, find the attribute very important but exhibit different curves, so that the optimal sensory level for each segment differs from that of the other segment.
Image Attributes -- The Same Dynamics as Liking Attributes?
There are two key questions here. The first concerns the dimensionality of image attributes, and second concerns the relation between image attributes and sensory attributes.
Researchers working with food products recognize that there are classes of attributes - sensory, liking, and image, respectively. In various studies the author has found that many image attributes correlate positively with each other, and with overall liking, suggesting that consumers do not differentiate between overall acceptability and image (e.g., nutritious, etc.), at least when these attributes are used with products. For margarine does the same inter-correlation occur among image attributes? Does the high intercorrelation of image attributes and liking attributes occur?
A second set of issues concerns the sensory-liking functions, extended to image attributes in place of liking attributes. Moskowitz (1997b) has recently suggested that the dynamics of image attributes may be different from the dynamics of overall liking. Image attributes for one product category, soap, cannot be easily related to sensory characteristics, whereas overall liking can be so related (especially when the relation is uni-variate, sensory attribute Vs overall liking, of the form shown in Figure 1). Does this same failure to create image-sensory functions occur with margarine as well?
The stimuli comprised 49 in-market margarines, varying in brand and type. The products comprised stick and tub products. [Stick margarines are hard; tub margarines are soft]. All stimuli were tested “blind”, on both unsalted soda crackers (for cold product evaluation), and on white toast (for warm product evaluation).
Panelists rated each product on a set of attributes pertaining to crackers, then to toast, using anchored 0-100 point scales. The anchoring reduces ambiguity. All attributes were anchored at both the upper and lower ends. Most the attributes were repeated for cracker and for toast, except attributes relating to rate and ease of melting (which were appropriate for toast only). For the analysis, therefore, average data will be considered, from the combination of the toast and cracker ratings. Table 1 shows examples of the attributes and the anchors.
Examples of attribute questions and anchors
How much do you like the appearance of this margarine?
0=Hate the appearance ---> 100=Love the appearance
Describe the color of the margarine
0=Very light --->100=Very dark
How much do you like the margarine overall?
0=Hate the margarine ---> 100=Love the margarine
Describe the sweetness of the margarine
0=Not at all sweet ---> 100=Very sweet
Describe the creamy texture of the margarine
0=No creamy texture at all ---> 100=Very creamy texture
How healthful is this margarine?
0=Not at all healthful ---> 100=Extremely healthful
How caloric is this margarine?
0=Low calorie ----> 100=High calorie
Rate the fat level of this margarine
0=Low fat -----> 100=High fat
Rate the quality of this margarine
0=Low quality --> 100=High quality
Rate the similarity to butter
0=Very different from butter --> 100=Identical to butter
The attributes were rated in order of appearance, with visual attributes preceding aroma, aroma preceding taste, and taste preceding texture. For specific attributes, the liking rating was asked first, followed by the sensory attribute (e.g., liking of appearance preceded darkness). The image attributes were rated last. A computer screen presented the question and the anchors. The panelist could rate the margarine only within the range allowed (viz., 0-100). The computer also controlled the rate at which the panelist answered, so that the panelist could not rush through the evaluation.
Each panelist rated a randomized 10 of the 49 margarines in a three-hour session. The session was pre-recruit, and followed the sequence recommended by Moskowitz (1985). The session began with a short orientation about the test, and then comprised 10 ratings. Each panelist rated a randomized 10 of the products, with a new product rated every 15 minutes (approximately).
Panelists participated in groups of 25. There were a total of 150 panelists, yielding 1500 “product tastings”. These tastings were allocated across the 49 margarines, to yield approximately 30 ratings per margarine. In previous studies of a similar nature a base size of 30 has been found to yield stable data (although from the point of view of representativeness of the population more panelists would have been preferred, but a higher base size would have exceeded budgets for the project; Moskowitz, 1997a).
The ratings began with a set of ratings on toast, followed by ratings of the same margarine on the unsalted cracker. Panelists found this task to be relatively straightforward after the first product. At the end of the evaluation the panelists completed an attitude and usage questionnaire to record the type and brands of margarines the panelists purchased and consumed, respectively.
The panelists comprised margarine users (at least once per week), divided approximately evenly between users of soft/tub margarine, and users of hard/stick margarine. Efforts were made to balance the panel composition based upon the brands and forms used, but due to the small base size and the different types of products within each brand it was impossible maintain an equal balance of all brand users.
The 150 panelists came from three geographically dispersed U.S markets (50 panelists per market). The panel comprised 100 females and 50 males. The panelists were pre-recruited from lists available to the local field services, and invited to participate for the test session. The panelists were paid after participation.
The analysis of the data followed these steps:
Step 1 - Average the ratings on equivalent attributes (viz., same attribute rated on two carriers; “on cracker” and “on toast”, respectively) in order to create a single attribute
rating. [E.g., “strength of butter flavor” on cracker and “strength of butter flavor” toast were averaged to create a single attribute, “strength of butter flavor”].
Step 2 - Determine the structure underlying each type of attribute. Step 2 used principal components to identify the basic structure of the image attributes and the liking attributes (Systat, 1994)
Step 3 - Determine the relative importance of image and liking attributes by means of the linear equation: Overall Liking = k + k(Image Or Liking Attribute). The coefficient k 011
shows the relative importance of the attribute.
Step 4- Create curves relating evaluative attributes (liking, image) to sensory attributes, using quadratic regression. The equation is: Evaluative Attribute = k+ k(Sensory 0 12 . Attribute) + k(Sensory Attribute)The quadratic equation shows the degree to which a 2
sensory attribute can be said to “drive” an evaluative attribute. The greater the area under 2the curve the more important the attribute (see Figure 2). The higher the Pearson R, the
more the curve fits the data. A great deal of area, and a high R2, indicates that the attribute is a key driver of liking.
Step 5 - Identify key subgroups of panelists, based upon the type of margarine used, frequency of use, sensory preference segment, etc. Then create the sensory-liking curves and compute area under the curve as a key measure by which to compare groups.
Liking and Image Attributes - Do They Load Onto the Same Factors?
In numerous commercially funded studies (unpublished) the author discovered that many image attributes really re-state “overall liking”. For instance, in the case of
carbonated beverages the attributes of “overall quality” and “refreshing” so highly
correlate with overall liking as to be virtually equivalent. For margarines there appear to be two types of image attributes - those which re-state liking (e.g., quality) and those which represent other, more complex perceptions which don’t appear to be liking, nor
even the inverse of liking (e.g., high in calories).
One way to better understand the differences among these two types of attributes is through factor analysis (principal components), in order to discover the number of factors underlying the image and liking attributes. [The factor analysis was done by principal components, followed by a quartimax rotation, and extraction of factors whose eigenvalues exceed 1.0. The quartimax rotation provides a relatively simple, easy to interpret factor structure].
Table 2A shows the results of the factor analysis, and suggests that there are two major factors -- the first being the liking factor, and the second being the fat/calorie factor. The order of attributes is determined by the loading on the two factors, respectively. Factor 1 comprises liking, quality, and similarity to butter and healthfulness. Factor 2 comprises high in fat and high in calories, respectively. These two factors suggest that the consumers treat the positive image attributes similarly, but the consumers also differentiate between positive image factors and negative image factors. Furthermore, the negative image factors of “fat” and “calories” are not the opposite of “liking” and
“health”, but rather independent of the positive image attributes.
Factor structure for overall liking and image attributes. (Numbers in the body of the table are loadings on the two factors)
Attribute Factor 1 Factor 2
First Factor = Liking
Like Overall 0.95 -0.06
Quality 0.93 0.05
Similar To Butter 0.85 0.27
Healthful 0.82 0.33
Second Factor = Fat/Calories
High In Fat 0.29 0.94
High In Calories 0.29 0.93
Percent Variation Account For 55% 32%
The Dimensionality of Liking Attributes
In previous studies, the author has reported that although panelists can be instructed to rate the liking of different attributes, in actuality the attributes turn out to be so highly correlated with each other that the set of liking attributes can be really treated as one major attribute (Moskowitz, 1995; Moskowitz & Krieger, 1995). If this finding continues to be the case, then researchers may waste a lot of time, interview cost and analysis by instructing panelists to rate attribute liking, because all attribute liking ratings correlate with each other. We can test this hypothesis with the margarine data by running a principal components analysis on the different liking ratings. Table 2B shows the results of this analysis of liking ratings, and suggests that there are two clear factors.
Factor 1 comprises color, aftertaste, taste, and texture. Factor 2 comprises spreadability, aroma, and overall appearance.
The first conclusion is that the major liking attributes which are clearly understood (color, taste and texture) again load together, indicating that the key liking attributes all refer to the same basic underlying primary (viz., restatement of liking). The second conclusion is that other liking attributes such as aroma and spreadability load onto a different dimension. Putting the phrase “liking” in front of an attribute does not automatically relegate that attribute to a single common liking factor. Rather, the main liking attributes (color, taste, texture) all load together for margarine (and for other foods as well), for reasons still unknown.
Factor structure for attribute liking. (Numbers in the body of the table are loadings on the two factors)
Attribute Factor 1 Factor 2
First factor = simple attribute liking
Color 0.98 0.05
Aftertaste 0.96 -0.09
Taste 0.96 0.15
Texture 0.80 0.45
Second factor = less concrete attribute liking
Spreadability 0.08 0.91
Aroma 0.43 0.73
Appearance 0.47 0.51
Percent Variation Accounted For 54% 27%
The Dimensionality of Sensory Attributes
Quite often sensory attributes correlate with each other. A principal components analysis reveals the degree to which panelists separate the attributes from the different senses. Moskowitz (1996) reported that in comparison to consumers, expert panelists