Using Known Patterns in Image Data to Determine Brand Positioning

INTRODUCTION

The effect of usage on image responses is a well–documented phenomenon (Bird & Ehrenberg 1966; Bird et al. 1970; Bird & Ehrenberg 1970; Barwise & Ehrenberg 1985; Barwise & Ehrenberg 1987; Riquier et al. 1996). We know that users of a brand are more likely to give an image response than former users or those who have never used the brands. Thus, bigger brands (with more users) will get more image responses than smaller brands, almost regardless of the image attribute.

This paper identifies a second pattern in image responses, one that is based on the attribute that is given to respondents. It appears that attributes obtain a certain level of responses, regardless of the number of users of the brand. This level is generally stable over time and appears to be associated with the prototypicality level of the attribute.

PATTERN 1: USAGE BIAS

Current users of the brand are more likely to give a response to image/ perceptual/attitude questions about the brand (Bird et al. 1970; Bird & Ehrenberg 1970). This applies for neutral, descriptive brand attributes (e.g. ‘has a red logo’) as well as positive, evaluative attributes (e.g. ‘provides good service’) (Barwise & Ehrenberg 1985). Consequently, brands with more users tend to score higher on image response questions.

Current users, former users, those who had tried it once, and those who had never used the brand all show different levels of aggregate response in the expected direction, i.e. a decreasing level of response from the first group to the last. This has been tested over a number of individual attributes across a series of attributes over time and found to be a stable pattern. This usage influence explains why big brands receive more responses than small brands, regardless of the attributes used for analysis.

ATTRIBUTE PATTERNS

Knowing that usage increases the likelihood of image responses helps interpret the data received (Bound & Ehrenberg 1998). It allows us to order image tables according to usage to improve our ability to make sense of image results (as advocated in Ehrenberg 1977). Doing this provides us with a table as shown in Table 1. The data was collected as part of a commercial image study with 600 customers in a services market conducted in 1997. The figures shown are the percentage of respondents who associated that brand with that attribute.

TABLE 1: IMAGE RESPONSES WITH THE BRANDS ORDERED BY NUMBER OF USERS

MEASURING BRAND IMAGE

Brand image was collected using a free choice, pick any format where both brands and perceptions are provided to respondents (see Holbrook et al. 1982). That is, respondents were presented with an image attribute (e.g. ‘tastes good’) and asked which, if any, of the listed brands they associated with this attribute. This has been found to be a valid and reliable method of collecting perceptual data (Barnard & Ehrenberg 1990) and is a technique commonly used in both commercial market research and research in marketing (see Brown 1985). Alternative brand image measures include forced choice measures such as ranking and rating of brands on a series of attributes. However, the seminal research in the area has shown that forced choice methods, while reducing non–response levels, result in overly positive responses being generated (Joyce, 1963)

In Table 1 we can clearly identify the usage effect for the different attributes. Brand 1 (which has the highest number of users and the highest number of responses overall) generally gains the highest or second highest proportion of responses whatever the attribute. This response level generally declines across the brands until Brand 8, with the lowest number of users, which generally receives the lowest number of responses for each of the attributes, as we would expect. This allows us to identify that, for example, Brand 7 is considered ‘business minded’ more than we would expect given its number of users. Without knowing this we might simply interpret an unordered table as saying that Brand 7 is not perceived as ‘business minded’ because it receives fewer mentions than all but two of the other brands.

Thus, knowledge of the usage pattern provides an interpretive framework and helps prevent drawing some misleading conclusions from the raw data. However, only undertaking this first step leads us to miss another very important aspect of this table. There is another pattern whereby the number of responses also varies according to the attribute. Looking at Table 1 we can get some sense of this pattern. For example, it is possible to see that the attribute ‘safe’ gains more responses for every brand than for example ‘good price’, which in turn gains more responses than ‘growing’. A simple ordering of the table by the number of responses for each attribute makes the pattern more evident. This is shown in Table 2.²

TABLE 2: TABLE ORDERED BY BRAND USAGE AND ATTRIBUTE BASED ON NUMBER OF RESPONSES

Now there are two patterns evident in the table. The first is the ‘usage effect’ previously evident in Table 1 whereby bigger brands generally gain more responses than smaller brands. The second is that the attribute ‘safe’ generally has more responses than the attribute ‘business minded’ which generally has more than the attribute ‘competent, and so on. It is important to note that at this stage the ordering is based on the total number of responses, as we have yet to explain the source of this pattern. In the next section we present one hypothesis: prototypicality.

PATTERN 2: PROTOTYPICALITY/FAMILY RESEMBLANCE

Prototypicality is a concept that describes the degree of category membership of any object. It was originally thought that all concepts elicited for a particular category could be considered equally representative of that category. However, it has been established empirically that there are different degrees of category membership (Rosch and Mervis 1975; Nedungadi and Hutchinson 1985), for example, a chair is more representative of the category ‘furniture’ than a radio. The determinants of prototypicality, as described by Nedungadi & Hutchinson (1985) are as follows:

• Objects which are frequently encountered as instances of that category will be perceived to be representative of that category.

• Objects that possess attributes that occur frequently within that category (that is, they have a higher family resemblance) are also perceived to be representative.

In the context of brand image, certain attributes can be said to be contributing more to category membership than others, based on the degree to which they represent the essential qualities that are needed to be part of that category or market. For example, in the banking industry ‘offering home loans’ could be classed as a highly prototypical attribute as it would be expected that almost every brand that was considered a bank would be offering home loans (even though some unusual banks do not, such as merchant banks or private ‘Swiss’ banks). In contrast ‘has low fees and charges’ is a less prototypical attribute as not all (indeed probably few) brands would be expected to have this quality, and not having ‘low fees and charges’ is unlikely to lead someone to question if that brand was indeed a bank.

The measurement of family resemblance (or prototypicality), established by Rosch and Mervis (1975), is based on the frequency of elicitation of an attribute for a series of items in that category. In this computation, each attribute is given a weight based on the number of items that share that attribute. So an attribute that is associated with six items in a category is considered more prototypical than an attribute that is mentioned for three items.

Our ability to detect this prototypicality in brand image data is made possible by knowing the usage influence on perceptual responses. For brand image data, we know that users are more likely to mention that brand regardless of the attribute, but that this influence is constant across all attributes. That is, the usage influence on responses for Brand 1 is the same regardless of the attribute. Therefore it has the same influence on each attribute. While this accounts for differences between brands in a single attribute, it does not account for differences between attributes. This difference (once brand size has been controlled for) could be related to the degree of prototypicality of the attribute, which, in line with Rosch and Mervis (1975) is the number of responses that each attribute gains across all of the brands/items.

Thus, it is a valid extension of the prototypicality construct to consider number of responses for a single attribute within a series of attributes to be an indicator of that attribute’s level of family resemblance or prototypicality.

UTILISING THE TWO PATTERNS: IDENTIFYING THE EXPECTED

Regardless of the source of the second attribute pattern, its identification provides us with some immediate benefits when interpreting image data. As the usage effect is reflected in the number of responses for each brand and the ‘prototypicality effect’ is evident in the number of responses for each attribute, a chi–squared formula can be employed to work out the expected number of responses in each brand-attribute cell. That is done using the following formula:³

Expected = Row total x column total
                        Total for all cells

We can calculate a table of expected values. Table 3 shows the table of expected values for the data used in this paper. In Table 3 we can see the pattern we get in our overall image data, with the higher responses in the top left hand corner (the larger brands on the more prototypical attributes), tapering to the lower responses in the bottom right hand corner.

TABLE 3: TABLE OF EXPECTED IMAGE DATA VALUES

Having established a benchmark, we are now able to identify the deviations from expected by simply subtracting the expected from the actual proportion of responses.⁴ The deviations are shown in Table 4.

TABLE 4: TABLE SHOWING DEVIATIONS FROM EXPECTED

Thus now we can see where the image responses deviate from the expected values that are based on the two patterns in image data that have been discussed in this paper.

This calculation is similar to that which drives the perceptual distance calculation in correspondence analysis based perceptual maps. This technique has advantages over correspondence analysis as it clearly shows where a brand scores lower than expected as well as scores higher than expected. This information is not as evident in perceptual maps.

FINAL OUTPUT AND MANAGERIAL APPLICATION

Now these deviations have been identified it is possible to identify:

• The positioning of each brand.

• Which attributes are ‘owned’ by particular brands and to what extent.

This can lead to conclusions concerning priorities for image development activities.

BRAND POSITIONING

Examining the deviations we can see that each of the brands has attributes upon which they score more, or less, than expected. Table 5 shows the attributes on which individual brands score higher than expected, that is, had the largest positive deviations from expected. Table 5 also shows the attributes with the largest negative deviations for each individual brand. For each brand, up to the three highest positive and lowest negative deviations that were 5 percentage points or greater have been listed.⁵ These reveal the distinctive image ‘positions’.

TABLE 5: SHOWING THE VARIOUS DEVIATIONS FROM THE EXPECTED RESULTS FOR EACH INDIVIDUAL BRAND

Thus, we get 2 pieces of important information – the attributes which the brand scores above or below expectations and a relative numerical value for that deviation. This gives a manager an indication of how their brand is performing on specific attributes, and identifies for which attributes remedial action (e.g. through marketing communications) perhaps need to be taken.

A deviation comes about when there is greater agreement between users and non–users of the brand. A positive or ‘upward’ deviation is typically when non–users also show a strong tendency to associate this attribute with the brand. And a negative or ‘downward’ deviation is typically when users behave more like non–users in that they do not tend to associate this attribute with the brand. For example, both users and non–users are likely to associate Coca–cola with ‘American’, hence this attribute will show an upward deviation for this brand, and both users and non–users are not likely to associate Coca–cola with ‘a traditional Chinese drink’, hence this attribute will show a downward deviation for this brand.

Such deviating attributes tend to be more descriptive or factual rather than evaluative or attitudinal. The more descriptive an attribute is, the more likely it is to show a deviation for a brand.

WHICH ATTRIBUTES ARE 'OWNED' BY PARTICULAR BRANDS AND TO WHAT EXTENT

By examining the deviations across brands for particular attributes it is possible to determine which attributes to use to develop a brand’s image. For example, a manager may choose to avoid attributes that are highly associated with another brand and avoid going head to head with a competitor. By looking at the deviations for each attribute it is possible to determine which are more commonly linked to competitor brands (or indeed the manager’s own brand). Table 6 shows the attributes that are strongly associated with a specific brand. It is important to note the influence of the negative deviations. Sometimes a ‘strength’ for a particular brand will appear because there is a specific ‘weakness’ in another brand (and as such all other brands appear stronger on that attribute). Thus, it is possible, and potentially useful, to single out ‘owned’ attributes which we have identified as when a brand has a positive deviation that is more than 50% of the largest negative deviation that is 5 percentage points or greater.⁶ The attributes are ordered by the centre column, which shows the size of 50% of the largest negative deviation.

TABLE 6: 'OWNED ATTRIBUTES'

It is also possible to identify ‘free’ attributes: those that are performing generally as we would expect across all brands (that is, with positive deviations less than 5 percentage points or 5 percentage points but not fitting into the ‘positive deviation’ category for any brand). These are shown below.

• Free attributes
• Responsive
• Growing
• Responsible
• Supports business

These attributes may represent opportunities to add to the image of a brand as they are not ‘owned’ by any brand in particular.

MANAGERIAL IMPLICATIONS

The patterns we have discussed in this paper provide an interpretive framework, allowing managers to turn image response data sets into information that can potentially guide decision making.

Identifying a brand’s distinctive position is potentially useful for devising salience enhancing marketing communication that is more likely to be clearly identified with the particular brand, that is, advertising which experiences low levels of ‘brand ambiguity’.⁷ For this reason, knowing what a brand or company stands for in the eyes of customers is generally seen as useful information to have prior to devising an advertising campaign.

Identifying the position of brands and the ‘owned’ and ‘free’ attributes also potentially allows marketing managers to determine attributes to develop as part of their brand image. A variety of strategies are possible:

• Work on the brand’s relative ‘weaknesses’. There is an argument that priority should be given to those attributes that are higher up the table, as well as those that have the greater negative deviations, as highly prototypical attributes probably play a more important role in establishing brand salience. The only caveat to this suggestion is when other brands ‘own’ the attribute in which case stressing this attribute may be counterproductive, effectively functioning as advertising for a competitor.
• Develop the association with a ‘free’ attribute – this is to take advantage of an attribute that is not easily identifiable with another brand. By incorporating it into marketing communication activities, it may be possible to develop that attribute as a specific strength for that brand. In this instance, however, it may be that those attributes lower on the table (those that gained a lower number of responses) may be a more beneficial target as they have an overall low awareness in the market place. This may make the image building exercise easier as few customers have pre–conceived ideas about the attribute. However, it is important to make sure that the attribute is not just totally irrelevant to customers in that market. This can be established through specific research into that attribute and testing any marketing communications prior to launch.

However, it must be noted that whether or not it is possible to make substantive changes to a brand’s image through marketing communication is an interesting empirical question that deserves investigation. While it is widely assumed in marketing practice, there seems to be a dearth of supporting evidence.

CONCLUSIONS AND IMPLICATIONS

This paper has built on knowledge about the existing pattern in image data, namely usage bias. It identifies a complementary, attribute–based pattern that appears to be related to the extent to which the attribute defines category membership. Knowledge of these two patterns allows us to identify the expected level of response for each brand on each attribute. This knowledge can then be used to identify deviations from the expected patterns.

These deviations help managers to identify their brand’s image position in a way not possible with statistical perceptual mapping approaches. This image can provide a useful guide for marketing communications decisions. The approach we have outlined also identifies which attributes are owned (more commonly associated with a particular brand than would be expected) or free (responses for all brands and attributes as would be expected). This is probably useful information for marketers seeking to alter their brand’s image.

END NOTES

1) The authors would like to thank the anonymous referees whose comments led to substantial improvements to this paper.

2) All calculations are based on unrounded figures.

3) This calculation should not be used when there appears to be a large negative deviation for the biggest brand on the most prototypical attribute (which usually coincides when there is a single dominant brand in the market.) In this case a logistic equation should be used to calculate the expected values as it ensures that the resultant figures are bounded at 100%.

4) Again, the figures are unrounded.

5) The reason for this decision rule is that it highlights the major deviations for each brand whilst avoiding highlighting minor deviations that may be attributable to factors such as sampling variation.

6) Although not strictly qualifying under this rule, Brand 2 on 'helps rural community' has been included given its extremely large positive deviations which makes it a reasonable conclusion to draw that it 'owns' this tribute.

7) Brand ambiguity refers to viewers mistakenly attributing an advertisement to another brand (Miller et al. 1971). High levels of ambiguity presumably severely hamper advertising effectiveness.

REFERENCES