ABSTRACT
This study considers online-marketing and -advertising effectiveness as well as consumer-segmentation research. Drawing from the value–attitude–behavior hierarchy, it compares different types of segmentation bases to identify which segmentation model best links to online-marketing click and order behavior plus brand-purchase behavior, and thus helps predict consumers' marketing responsiveness in the online-advertising context. Using real-life data from a German retailer plus factor and cluster analysis, the authors calculate three separate segmentation models based on human values and different attitudes. The results show that attitude-based segmentations are linked more closely to online-marketing response and brand-purchase behavior than the value-based segmentation.
MANAGEMENT SLANT
This study compares human values and attitudes, segmentation bases widely applied in practice, in terms of their predictive power on online-marketing responsiveness—that is, how they differentiate consumers' click and order behavior in different online-marketing channels and brand-purchase behavior.
Using an extensive real-life sample, the study identifies the most powerful of the segmentation bases for each type of marketing channel.
On the basis of these findings, practitioners will be able to custom tailor and fine-tune their segmentation models for their specific needs in online marketing to link them closely to actual click and order behavior of consumers.
INTRODUCTION
Online marketing has become a crucial pillar in the marketing arsenal in recent years and has grown immensely (Breuer, Brettel, and Engelen, 2011; Kim and McMillan, 2008; Li and Kannan, 2014). Technology-enabled data collection and interactivity in online marketing allow for radically new ways of communicating to and with consumers (Lambrecht and Tucker, 2013). The tool set in online marketing and especially online advertising comprises different instruments, such as display banners, search-engine optimization, and affiliate marketing (Li and Kannan, 2014).
The effectiveness of and response toward different marketing tools varies greatly across consumer groups, however (Bucklin, Gupta, and Siddarth, 1998). Online-marketing effectiveness therefore has become an important topic in research and practice over the last few years (Abou Nabout and Skiera, 2012; Breuer et al., 2011). Understanding the characteristics of consumer groups that differ in their responsiveness toward different online-marketing tools is of especially great interest (Cotte, Chowdhury, Ratneshwar, and Ricci, 2006; Forsythe, Liu, Shannon, and Gardner, 2006; Reimer, Rutz, and Pauwels, 2014).
In this context, consumer segmentation is an essential research instrument to examine and understand differences among consumer groups (Roberts, Kayande, and Stremersch, 2014; Wedel and Kamakura, 2000). Segmentation research, however, has not reached a state yet where definitive recommendations can be made on which model works best in which context, such as relating different consumer groups to varying degrees of online-marketing responsiveness (Liu, Ram, Lusch, and Brusco, 2010; Reimer et al., 2014; Steenkamp and Ter Hofstede, 2002). To differentiate between groups with high and low marketing responsiveness for different online-marketing tools, the authors' goal in the current study was to develop a segmentation model that produces segments with sufficiently differentiated click-behavior patterns, which is a prerequisite for the predictive validity of a segmentation model (Wedel and Kamakura, 2002).
In addition to the segmentation algorithm, the segmentation base is a crucial part of a segmentation model (Wedel and Kamakura, 2000). The segmentation base is the variable by which segments are formed (Steenkamp and Ter Hofstede, 2002). These variables can range from demographics (e.g., age and gender) to psychographics (e.g., personal values or attitudes) or behavior variables (e.g., purchase or usage behavior) and many more that have been developed in the past (Bock and Uncles, 2002; Chow and Amir, 2006). Today, however, researchers instead call for validation and comparison of segmentation bases and segmentation models to give guidance on which base to use in different settings, such as online-marketing responsiveness and effectiveness (Liu et al., 2010; Wedel and Kamakura, 2002).
In past research, many studies examined only single segmentation bases and the resulting segments, rather than comparing different segmentation bases and their segment solutions, as called for by researchers (Bhatnagar and Ghose, 2004; Rohm and Swaminathan, 2004). Only a few studies have had the possibility to link segments to real consumer behavior, such as click data. Rather, studies used questionnaires to ask about past behavior, intentions, or brand attitudes, with the known methodological downsides (Kahle, Beatty, and Homer, 1986; Kennedy, Best, and Kahle, 1988; Ko, Taylor, Sung, Lee, et al., 2012). To address this obvious research gap, the authors formulated the following research question:
RQ1: Which type of segmentation base produces segments with more differentiated marketing-responsiveness behavior?
To investigate this topic, the authors leveraged data from a leading European online apparel retailer. The study thus used psychographic data on the retailer's German customers as a segmentation base. The authors furthermore had access to the retailer's database for real-life online-marketing click and order data, plus brand-purchase data of the respective customers, to capture marketing responsiveness and purchase behavior.
THEORETICAL FRAMEWORK AND HYPOTHESES
For the comparison of segmentation bases, first a categorization of the different types of bases is necessary (Wedel and Kamakura, 2000). Several categorizations and classifications for segmentation bases exist, but one of the most frequently used (Frank, Massy, and Wind, 1972) has found wide application among segmentation researchers (Steenkamp and Ter Hofstede, 2002; Wedel and Kamakura, 2000). This classification structures segmentation bases in two dimensions:
One comprises the scope of the segmentation base, with the two categories general versus product or context specific, and
the other comprises the observability, with the two categories observable versus nonobservable (Frank et al., 1972; Wedel and Kamakura, 2000).
Although the latter is a rather technical differentiation that concerns data collection, the former categorizes segmentation bases in two basic classes:
segmentation bases that are independent of context and more general in nature, and
segmentation bases that are context specific and are linked to the product or market (Wedel and Kamakura, 2000).
Among the most important general segmentation bases are personal human values, and among the most frequently used context-specific segmentation bases are attitudes (Chow and Amir, 2006; McCarty and Shrum, 1993; Wedel and Kamakura, 2000).
A value can be defined as an ”enduring belief that a specific mode of conduct or end-state of existence is personally and socially preferable to alternative modes of conduct or end-states of existence” (Rokeach, 1968, p. 16), which emphasizes the value's general and long-term nature. An attitude can be defined as ”an organization of several beliefs focused on a specific object or situation, predisposing one to respond in some preferential manner” (Rokeach, 1968, p. 16), which points out the attitude's reference to specific situations and contexts.
A later study researched the relationship among values, attitudes, and behavior to develop the value–attitude–behavior hierarchy (Homer and Kahle, 1988). According to this theoretical framework, a person's general human values influence his or her attitudes regarding specific contexts and situations, such as fashion. A person's more conservative values, for example, may lead to attitudes toward fashion that are less open to new trends. The specific attitudes then influence the person's behavior in a given situation within the respective context (Homer and Kahle, 1988; Shim and Eastlick, 1998). In the fashion example, an attitudinal aversion toward fashion trends may, in a specific apparel-purchase situation, lead to the person buying clothes of the same style as in the previous season (Gutman and Mills, 1982).
This relationship implies that specific attitudes directly are linked to and hierarchically are closer to actual behavior than human values (Defever, Pandelaere, and Roe, 2011; Hansen, 2008; Homer and Kahle, 1988). So far, however, this relationship only has been proven for the concepts themselves, not for the segmentation solutions based on these concepts (Homer and Kahle, 1988; Shim and Eastlick, 1998). Applied to the research of segmentation bases, this implication means that the link between specific attitudes as segmentation base and consumer behavior, such as marketing-responsiveness behavior, should be stronger than the link between human values as segmentation base and consumer behavior. As a result, attitude-based segmentation should produce segments with more differentiated marketing-responsiveness behavior than value-based segmentation (Homer and Kahle, 1988; Vinson, Scott, and Lamont, 1977).
As mentioned before, this research used data from an online apparel retailer. In this setting, the attitudinal context is on the intersection of fashion attitudes and online-shopping attitudes.
H1: Fashion attitude-based segmentation produces consumer segments with more differentiated marketing responsiveness and brand-purchase behavior patterns than those consumer segments found through value-based segmentation.
H2: Online shopping attitude-based segmentation produces consumer segments with more differentiated marketing responsiveness and brand-purchase behavior patterns than those consumer segments found through value-based segmentation.
METHODOLOGY
The goal of this study was to compare fashion attitude-based and online shopping attitude-based segmentations with value-based segmentations regarding their ability to produce segments with differentiated online-marketing responsiveness and brand-purchase behavior. Several elements therefore were considered:
appropriate measurement scales for the respective attitudes and values,
a segmentation algorithm to form the segments,
appropriate variables to reflect online-marketing responsiveness and brand purchase, and
a statistical method to compare the degree of differentiation of marketing responsiveness between the segment solutions.
On the basis of extensive research of the relevant literature, the authors identified the most suitable scales for human values, fashion attitudes, and online-shopping attitudes for use in this study setting. They paid special attention to scales that have been reproduced and validated interculturally, if not developed for the German market.
For human values, this study follows the work of a particular researcher (Schwartz, 1992), which built on the work of another researcher (Rokeach, 1968), and advances it by developing an interculturally stable framework of human values in a circumplex structure of 10 value domains, such as achievement, universalism, security, and power. This structure has been validated in more than 200 studies and more than 60 countries (Schmidt, Bamberg, Davidov, Herrmann, et al., 2007). The portrait value questionnaire, (Schwartz, Melech, Lehmann, Burgess, et al., 2001) was used as a measuring scale in this research.
For a suitable fashion attitude scale, the authors used the fashion orientation scale (Gutman and Mills, 1982). This scale captures attitudes toward fashion, such as fashion interest, fashion leadership, or antifashion attitude. It was developed for the U.S. market but was reproduced and validated in the United Kingdom (Goldsmith, Freiden, and Kilsheimer, 1993).
To measure online-shopping attitudes, the current study builds on the work of prior researchers (Brengman, Geuens, Weijters, Smith, et al., 2005). Their Internet shopper lifestyle scale measures attitudes toward the online-shopping context, such as online-shopping distrust or online-shopping convenience. Although the measure is named a “lifestyle scale,” these constructs are, in fact, attitudes (per the definition of Rokeach, 1968), because they are based on beliefs toward the online-shopping context.
For the segmentation algorithm, several methods exist, ranging from simple cogrouping of all subjects with the same top choice from a list of items to using complex artificial neural networks (Kamakura and Novak, 1992; Wedel and Kamakura, 2000). One of the most frequently used algorithms, however, is factor analysis with subsequent cluster analysis, which the current study therefore adopted (Steenkamp and Ter Hofstede, 2002; Wedel and Kamakura, 2000). The same algorithm was applied to all three segmentations to make results comparable.
The goal of online marketing in e-commerce is to draw consumers eventually to the website to create sales (Dinner, Van Heerde, and Neslin, 2014). There are several marketing-relevant channels to access the website of an e-commerce company, including
direct type-in of the web address,
access by search engine, which can be through search engine optimization or embedded advertisements in search engines (i.e., search engine marketing [SEM]),
referral sites, such as affiliates or price-comparison sites,
emailings, including newsletters or product-recommendation mailings, and
display-banner advertisements (Li and Kannan, 2014).
For measurement of the marketing responsiveness of consumers across these channels, an appropriate variable is the number of clicks placed in each channel per customer (Dinner et al., 2014; Ha, 2008). The study also took the number of orders into account. The attribution of an order was made to all channels used by the customer before the order in a maximum time period of 30 days (Li and Kannan, 2014).
To also reflect the brand-purchase behavior, which follows the marketing responsiveness, the authors applied another metric: For each of the observed customers, brand-purchase behavior was captured as the percentage of their purchase among the top 100 brands (by turnover) of the collaborating online retailer within an observation period of 12 months. This approach to capture purchase and choice behavior within a substantial observation period follows the rationale of previous studies (Andrews and Currim, 2002; Sriram, Chintagunta, and Neelamegham, 2006; Van Kerckhove, Geuens, and Vermeir, 2012). To evaluate which of the segment solutions had a more differentiated marketing responsiveness and brand-purchase pattern, one-way analysis of variance (ANOVA) with F test was the most appropriate method in this setting, because it tests the significance of differences among group means (Aaker, Kumar, and Day, 2004; Backhaus, Erichson, Plinke, and Weiber, 2010; Malhotra, Birks, and Wills, 2012).
RESULTS
For the data collection, an online survey was conducted in spring 2014. Participants were German customers of the collaborating online apparel retailer who had made at least one purchase up to six months before the survey date. The click and purchase data for the marketing-responsiveness variables were obtained from the retailer's database, which directly captures the respective behavior information from all customers on the website. After elimination of incomplete and implausible data, 3,219 responses remained to be used in the analyses.
To analyze the underlying factor structure for each of the three segmentation scales, the authors calculated exploratory factor analysis (EFA) separately on each scale. To ensure that EFA was applicable to these specific scales, the authors first used two measures to test applicability: the Kaiser–Meyer–Olkin measure of sampling adequacy, and Bartlett's test of sphericity (Backhaus et al., 2010). The Kaiser–Meyer–Olkin measure turned out acceptable results for all three scales (0.84 for the value scale, 0.91 for the fashion scale, and 0.68 for the online-shopping scale). Bartlett's test results lay at the 0.000 significance level for all three scales. Both measures support the applicability of EFA for each of the three scales.
In the EFA, the authors performed principal-components analysis with subsequent varimax rotation using the IBM statistics software SPSS22. During the EFA for the value scale, the authors removed 12 of the 40 original items for unacceptable factor loadings or cross-loadings. From the remaining 28 items, they extracted seven factors with eigenvalue above 1.0, which explained 58.7 percent of the total variance. The factor loadings ranged between .546 and .822.
To evaluate the reliability of the factor structure, the authors calculated Cronbach's alpha. It ranged between .499 and .850. Although some factors had values for Cronbach's alpha below the popular threshold of .7, the study's solution is still accepted as reliable because of the low number of items in the respective factors (as per the rationale of Nunnally and Bernstein, 1978, and Peter, 1997). The seven factors were mostly in line with the original structure (see Schwartz, 1992). Two of the original 10 value domains, however, were not identified as factors in the analysis: Self-Direction and Security. Apart from that, the two value domains power and achievement fell into one common factor. The final value factors in the analysis were Achievement/Power, Hedonism, Universalism, Benevolence, Conformity, Stimulation, and Tradition. (See Table 1 for the results of the EFA.)
The original version of the Fashion Orientation Scale (Gutman and Mills, 1982) contains 17 items structured in four fashion attitude factors. During the EFA for the fashion scale, three items were dropped because of unacceptable factor loadings and cross-loadings. The 14 remaining items resulted in three factors with eigenvalue above 1.0. These factors explained 55.0 percent of the total variance. The factor loadings ranged from .496 to .815. Cronbach's alpha for the fashion factors ranged between .516 and .869.
As for the values EFA, the factors with low Cronbach's alpha had only a few items. The solution therefore was accepted as reliable (Nunnally and Bernstein, 1978; Peter, 1997). The three factors in the solution resemble the original construct closely. The two original factors, Fashion Leadership and Fashion Interest, however, were extracted as one common factor. The final solution therefore consists of the factors Fashion Leadership/Interest, Importance of Being Well Dressed, and Antifashion Attitude. (The results of the EFA are shown in Table 2.)
For the online-shopping attitude scale, in the EFA seven of the original 21 items of the Internet Shopper Lifestyle Scale (Brengman et al., 2005) were removed because of unacceptable factor loadings or cross-loadings. From the remaining items, five factors with eigenvalue above 1.0 and factor loadings between .539 and .862 were extracted. These factors explained 58.7 percent of the total variance. Cronbach's alpha ranged between .456 and .714. Again, the lower values were measured for those factors with only a few items. In keeping with previous research (Nunnally and Bernstein, 1978; Peter, 1997), the solution was accepted as sufficiently reliable.
The five factors of the EFA solution were mostly in line with the original factor structure (Brengman et al., 2005). Only the factor Internet Offer could not be identified in the sample. The final factors for online-shopping attitudes were Internet Convenience, Internet Self-Inefficacy, Internet Logistics, Internet Distrust, and Internet Window-Shopping (See Table 3). In a subsequent step, the authors performed confirmatory factor analysis (CFA) on each of the three factor constructs to evaluate convergent and discriminant validity and confirm the factor structures. They used the IBM statistics software SPSS AMOS (Version 22) to calculate CFA in this study. For each of the constructs, the items loaded acceptably high on their intended factors.
During CFA, the authors applied several measures to evaluate the constructs. The goodness-of-fit index (GFI) and adjusted goodness-of-fit index (AGFI) produced satisfactory results for the values construct (GFI = 0.878, AGFI = 0.850) and very high results for the fashion attitudes (GFI = 0.966, AGFI = 0.951) and the online-shopping attitudes constructs (GFI = 0.988, AGFI = 0.981). The root-mean-square error of approximation (RMSEA) produced good results for each of the three constructs (values scale, RMSEA = 0.072; fashion attitudes scale, RMSEA = 0.053; online-shopping attitudes scale, RMSEA = 0.031).
The comparative fit index (CFI) produced acceptable results for the values construct (CFI = 0.806) and very high values for the fashion-attitude construct (CFI = 0.952) and the online-shopping attitude construct (CFI = 0.965). The minimal chi-square standardized by its degrees of freedom (CMIN/DF) produced rather high values for each of the three constructs (value scale, CMIN/DF = 17.633; fashion-attitude scale, CMIN/DF = 10.179, online-shopping scale, CMIN/DF = 4.164). In keeping with the argumentation of previous researchers (Bagozzi, 1981; Jöreskog and Sörbom, 1982), however, these rather high values are acceptable, because this measure increases overproportionally with large sample sizes, such as the one used in this study.
Subsequent to the determination and confirmation of the factor structure for the three constructs, in the next stage, cluster analysis identified the underlying segments for each of the three segmentations. Drawbacks of cluster analysis raised by academics include its dependency on decisions taken by the researcher during the process (Iacobucci, 2015; Punj and Stewart, 1983). An advanced two-step approach (first introduced by Punj and Stewart, 1983) has been established combining the advantages of hierarchical and nonhierarchical clustering algorithms and limiting the decisions to be taken during the process (Mooi and Sarstedt, 2011; Punj and Stewart, 1983).
In the first step of this approach, hierarchical clustering with Ward's method determines the appropriate number of clusters using the elbow criterion and identifies the most promising starting point for the cluster centroids. These are fed into a nonhierarchical k-means clustering to determine the final cluster solutions for all three segmentations. ANOVA and Duncan's test serve to validate the cluster solutions (Backhaus et al., 2010; Iacobucci and Churchill, 2010; Ko et al., 2012).
The cluster solution for the values segmentation comprised eight clusters. The cluster sizes ranged from 10.5 percent to 15.8 percent of the sample. ANOVA and Duncan's test showed significant differences in cluster means across the factors, validating the cluster solution. In the original study (Schwartz, 1992), the goal was the exploration of value systems rather than a segmentation of persons. Accordingly, no original segments exist to which the current study's solution could be compared.
The segments in this study's solution therefore were named by the authors to reflect their profiles across the value factors. The resulting value clusters were:
Benevolent,
Demanding,
Achiever/Power Seeker,
Traditionalist/Conformist,
Universalist/Hedonist,
Stimulation Seeker,
Pragmatic, and
Egoist.
To increase applicability of these cluster solutions, the authors examined the demographic profiles of the identified value clusters (See Table 4).
When the authors explored the demographics profiles, the Benevolent segment showed an above-average share of female consumers, with a substantial shift toward the age cohorts of 40 or older. The Demanding segment consisted of a slightly above-average male share and an overproportional share of consumers aged 29 and younger. The Achiever/Power Seeker segment also showed a slightly overproportional share of male consumers and a slightly above-average share of 20–29-year-old individuals.
The Traditionalist/Conformist segment had the second-highest share of male consumers and a slightly above-average share of people younger than 29 years. In the Universalist/Hedonist segment, women were overrepresented, as well as consumers younger than 19 and older than 50. The gender split within the Stimulation-Seeker segment was very close to the overall average, with the age groups between 20 and 39 years being slightly overrepresented. The Pragmatic segment showed the highest share of women, with an above-average share in the age group of 40–49 years. Last, the Egoist segment had the highest share of men, with the age groups of 30 years or older being overrepresented.
The results of the fashion attitude cluster analysis showed six distinct clusters. Their sizes ranged from 14.9 percent to 18.1 percent of the study sample. Subsequent validation with ANOVA and Duncan's test showed significant differences in the cluster means on all factors.
Five of the seven clusters identified in the original study (Gutman and Mills, 1982) emerged in this analysis: Leaders, Independents, Neutrals, Uninvolveds, and Negatives. Additionally, one cluster with a completely new profile emerged from this analysis. To reflect its distinct scores across the fashion-attitude factors, the authors named it Instrumentalists. The clusters identified in this study thus were rather similar to those found in the original study, which suggests robustness of this cluster solution. Again, for illustration and applicability, the authors examined the demographic profiles of the identified fashion-attitude clusters (See Table 5).
With regard to the demographic profiles, the Leaders segment showed the highest share of women and a slightly above-average share in the age group of 29 years and younger. The Independents segment's gender split was very close to average, with slightly overproportional shares in the age groups of 19 years and younger as well as 30–39 years. The Neutrals segment had an above-average share of men as well as age groups older than 30 years. The segment of Uninvolveds showed a slightly above-average share of female consumers and age groups between 30 and 49. The new segment of Instrumentalists showed a gender split very close to the sample average and an above-average share of consumers aged 29 or younger. Last, the Negatives segment had the highest share of male consumers and an above-average share of the age group of 50 years and older.
In the cluster analysis for online-shopping attitudes, eight distinct clusters emerged, with sizes between 12.2 percent and 16.9 percent of the sample. ANOVA and Duncan's test resulted in significant differences of means across factors and thus validated this cluster solution. The original segmentation (Brengman et al., 2005) identified four clusters of online shoppers. Three of them were found in this analysis: Shopping Lovers, Tentative Shoppers, and Suspicious Learners.
In addition, four new cluster profiles emerged that were not identified in the original work. These clusters were named on the basis of their unique factor scores: Information Seekers, Direct Contact Seekers, Window-Shoppers, and Distrustful Shoppers. The identified clusters were rather similar to the original solution, which suggests a certain robustness and reliability of this segmentation model. (Table 6 illustrates the demographic profiles of the cluster solution.)
The segment of Shopping Lovers showed the second-highest share of male consumers and an above-average share in the age groups younger than 39 years. The Tentative Shoppers segment had a close to average gender split and a strongly above-average share of consumers ages 40 and older. The Information Seekers segment had the highest share of men, with the age groups of 40 and older overrepresented. The Direct Contact Seekers had an above-average share of male consumers, with the highest share in the age group between 20 and 29 years.
The segment of Window-Shoppers showed a close to average split of genders and an overproportional share in the age group of 29 years and younger. The Distrustful Shoppers segment had the highest share of female consumers and an above-average share of people older than 30. Last, the segment of Suspicious Learners had the second-highest share of women and a slightly above-average share of the age group between 30 and 39 years.
To test the hypotheses, the authors calculated one-way ANOVAs with F tests for each of the three segmentations on the online-marketing responsiveness variables and the brand-purchase behavior variables. The former were the clicks and orders in each of the access channels to the website:
direct type-in,
search engine optimization and SEM (for which the search string contained the retailer's brand name or no brand name, or the string itself was provided by the search engine),
referral through affiliate or price comparison sites,
display banners (reaching out to new customers or retargeting customers who had accessed the website before and using content from this visit), and
promotional emailings (Li and Kannan, 2014).
In the ANOVA on clicks for the value segmentation, the authors found significant differences among means across the channels at the 5 percent level or lower for two channels. For the fashion-attitude segmentation, they found significant differences for six channels, and for the online-shopping attitude segmentation they found significant differences for two channels. In the ANOVA on orders, the authors found significant differences for the value segmentation for only one channel, for the fashion-attitude segmentation for seven channels, and for the online-shopping attitude segmentation for three channels. The results of the one-way ANOVA on clicks (See Table 7) and on orders (See Table 8) are reported.
Regarding the brand-purchase behavior, the results of the ANOVA showed significant differences in means for four of the brands for the value segmentation, for 11 brands for the fashion-attitude segmentation, and for seven brands for the online-shopping segmentation. The results of the ANOVA on brand-purchase behavior are shown (See Table 9).
In summary, the ANOVAs for clicks and for orders as well as for brand purchase showed significant differences of means in more online-marketing channels for the fashion-attitude segmentation than for the value segmentation. Hypothesis 1 thus can be verified. The ANOVA for clicks showed significant differences across means in the same number of online-marketing channels for the online-shopping attitudes segmentation and the value segmentation. The ANOVA for orders and for brand purchase, however, showed significant differences in more channels for the online shopping segmentation than for the value segmentation. Although the differences were not as strong as between fashion-attitudes segmentation and value segmentation, Hypothesis 2 still can be verified.
Because both hypotheses can be verified in this study, it is interesting to further illustrate how differentiated actual consumer behavior was in the three segmentation models. To do so, the authors explored how each segment in each segmentation model differed in number of clicks (See Table 10) and number of orders (See Table 11) from the cross-segment average.
Although this serves as an illustration of the results of all three ANOVAs, it is still interesting to explore the data. As expected, the spread of the clusters' maximum differentiation from average was higher for the two attitude-based segmentations than for the segmentation based on values. When one looks at how segments reacted differently to these two marketing channels, it is notable that the value cluster Universalists/Hedonists reacted most positively in clicks to SEM advertising, whereas for orders through e-mail marketing Egoists reacted most positively.
For the fashion segmentation, the two segments Leaders and Independents reacted most positively in clicks to SEM advertising and in orders to e-mail marketing. For the online-shopping segmentation, Distrustful Shoppers reacted most positively in clicks to SEM advertising, whereas Tentative Shoppers showed the most positive reaction in orders to e-mail marketing. The latter segment, however, also showed a notably positive reaction to clicks in SEM advertising.
DISCUSSION
Implications for Research and Practice
The goal of this study was to find out whether attitude-based segmentations produce segments with more differentiated online-marketing responsiveness and brand-purchase behavior than value-based segmentations, as hypothesized on the basis of the value–attitude–behavior hierarchy by previous researchers (Homer and Kahle, 1988). This helps evaluate predictive validity of these bases.
This study therefore used survey-based data from 3,219 German customers of a leading European online apparel retailer plus click-behavior data with approximately 163,000 clicks from their customer database. This research used three different segmentation bases—human values, fashion attitudes, and online shopping attitudes—with combined EFA, CFA, and two-step cluster analysis to form segments for each of the three segmentation bases. This study used one-way ANOVA with F test to link the segments to the online-marketing responsiveness and brand-purchase behavior. The results of these analyses confirm the authors' hypotheses.
The segment profiles across demographics and behavior also allow an illustrative insight into the patterns that researchers may find on the basis of the respective segmentation models. For demographics, this shows certain differences across segments that can help target specific, more attractive groups of customers. Of even more interest are the differences in segment behavior that also form the basis of the ANOVAs that validate this study's hypotheses. In the illustrative profiles, one can see differences as high as
36.4 percent above average clicks through the SEM brand channel and 28.2 percent above average orders through the emailing channel for certain fashion-attitude segments, and
39.1 percent above average clicks through the SEM brand channel and 26.0 percent above average orders through the emailing channel for certain online-shopping attitude segments.
This illustrates the value of segmentation models with the respective segmentation bases.
These findings have a number of contributions in several fields. In terms of academic contributions, this study applies and validates the value–attitude–behavior hierarchy to segmentation research. This research thus confirms a theoretical framework to understand better the connection of values and attitudes as segmentation bases and actual consumer behavior.
This study also develops insights to give guidance in the choice of the most appropriate base for a segmentation analysis with the goal of examining online-marketing responsiveness behavior. In this case, attitudes appear to produce better results than human values. The results also show, however, that fashion attitudes produced much better results than online-shopping attitudes. This finding underlines the importance of choosing the most appropriate domain of attitudes, because several domains are possible for a given research context. As for this study, attitudes in the domain of fashion as well as in the domain of online shopping were applicable.
This research also confirms and validates three segmentation bases that, to the authors' best knowledge, have not been applied to the German fashion e-commerce market before. More than that, this study even discovered new segments, such as the Instrumentalists in the fashion segmentation or the Information Seekers for the online-shopping segmentation.
In terms of practical and managerial contribution, there is an immense value in the deeper understanding these findings create on how to configure the segmentation model in a market-segmentation project aimed at online-marketing responsiveness. As discussed in the introduction, a strong understanding of the differences in online-marketing responsiveness across consumer segments is crucial, especially in the e-commerce context. In practice, however, segmentation bases often are chosen ad hoc and based on mere data availability (Steenkamp and Ter Hofstede, 2002). Instead, practitioners should identify the expected findings and then choose the most appropriate base. If these findings are in the area of marketing responsiveness, then attitudes appear to be more promising than human values.
This study also helps practitioners to advance their knowledge and understanding of the German fashion e-commerce market. It not only produced one isolated segmentation of this very interesting and dynamic market but examined it with three segmentations from three different, insightful perspectives: human values for a better understanding of consumers' personalities, plus fashion and online shopping attitudes for a deeper knowledge of consumers' context-specific mindset. This will help practitioners to better target consumers and tailor communication strategies.
Limitations and Further Research
This study examines different segmentation bases and their link to online-marketing responsiveness behavior. This research therefore used data on German customers of a leading online apparel retailer to test the hypotheses. Although it has produced fascinating findings, this approach also incurred several limitations to be considered.
When a study only uses data from one single company, there are obvious limitations to the generalizability. Although this study used a large dataset, the persons in the sample may share common characteristics that might have moved them toward the retailer's brand. This possible over- or underrepresentation of certain characteristics may be the reason why, for example, the factors Self-Direction and Security did not emerge in the EFA for human values (Backhaus et al., 2010). Further research therefore should use this study as groundwork and extend the analysis to other company samples and even other industrial contexts beyond fashion.
The dataset also comprises only customers from one single country, Germany. It is obvious that these customers differ in their cultural characteristics from customers in other countries. It may be questionable whether the findings can be generalized to those other countries and cultures. Further researchers should take up the challenge and extend the research to other cultural environments and markets to understand cultural influences. Last, researchers should take a longitudinal view to understand whether values or attitudes produce more stable segments over time.
Building on the better understanding of which segmentation bases are linked more closely to actual customer behavior, further research should investigate whether certain actual value or attitude patterns (i.e., segment profiles) are linked consistently to higher or lower response behavior in certain marketing channels and brand-choice behavior. This even could involve exploratory studies of building cross-base segments. The ultimate result could be to predict marketing response and brand-choice behavior on the basis of value and attitude segment patterns.
ABOUT THE AUTHORS
Stefan Scheuffelen is a research assistant at RWTH Aachen University, Germany, and is project leader at the Boston Consulting Group in the Berlin area. His research interests are around consumer segmentation and consumer behavior in e-commerce.
Jan Kemper is an assistant professor in innovation management and entrepreneurship at RWTH Aachen University and managing director of Omio, an online travel agency. His research interests center on entrepreneurship, e-commerce, finance, media, and cross-cultural management. He has published in Psychology & Marketing and elsewhere.
Malte Brettel is professor for business administration at RWTH Aachen University, and adjunct professor for entrepreneurship at WHU Otto Beisheim School of Management, Germany. He has published in the Journal of Advertising Research, Journal of Business Venturing, Journal of Product Innovation Management, Strategic Management Journal, and elsewhere.
- Received January 2, 2018.
- Received (in revised form) May 18, 2018.
- Accepted August 9, 2018.
- Copyright© 2019 ARF. All rights reserved.