A comparison of the effects of online
social interactions on the consumers
decision making process between regular
durable goods and annuities
Date:
January 16, 2016
A comparison of the effects of online social
interactions on the consumers decision
making process between regular durable
goods and annuities
Master thesis, MSc Marketing,Marketing Intelligence and Marketing Management University of Groningen, Faculty of Economics and Business
Abstract
This paper analyzes the effects of online social interaction (i.e. observational learning and electronic word of mouth) on the decision making process of two different product types. These product types are regular durable goods (i.e. digital cameras) and annuities (i.e. health insurances). Based on a choice based conjoint (CBC) analysis, this paper analyzes various valences of online social interaction, and for each valence there is a calculated utility which helps to see the preferences of consumers for each product type. To see the differences between the two product types the relative attribute importance and its ranking are compared. Next, for all the analyzed valences of social interaction, the interactions were created to see the moderating effect of eWOM on OL. The aim of these interactions is to predict the effect of observational learning when electronic word of mouth also becomes available and to test whether the accessibility-diagnosticity model holds for both product types. The analysis shows that the two product types do differ on its characteristics and their eventual decision, since the main effects of social interaction slightly differs for annuities compared to regular durable goods. However, for both product types the moderating effects turned out to be only marginally significant for a couple of social interaction valences.
Preface and acknowledgements
First of all, I would like to thank my first supervisor Dr. Hans Risselada for his time, guidance, feedback and useful meetings during this semester. I really appreciate the working procedure we have used, where I also had the opportunity to make mistakes and to learn from new insights. This brings me to the members of my thesis group; Marin Teinsma, Fyne van Breevoort, Albert Koller, Erick Jansema and Pascuál Ab, who also provided me with new insights, tools and support. I also need to thank Felix Eggers, my second supervisor, for his time and for helping me out with difficulties according to choice based conjoints and the related software. Also worth mentioning are the respondents who did spend some of their time to fill in one of my surveys. Finally, I would also like to thank my family and friends, which supported me during all my years of studying.
Table of content
Introduction ... 5
Theoretical Framework ... 12
The conceptual model ... 12
Literature Review ... 12 Hypotheses ... 26 Methodology ... 28 Research setting ... 28 Research design ... 29 Experimental design ... 31 Measurement items ... 31 Data collection ... 32 Modeling approach ... 32 Model Diagnostics ... 33 Results ... 35 Sample statistics ... 35
CBC Analysis - the main effects ... 36
Covariates ... 38
Validity ... 39
Parameter interpretation of the main effects ... 40
CBC Analysis - The moderating effect of eWOM valence ... 43
Parameter interpretation of the interaction effects ... 44
Discussion ... 49
Theoretical implications ... 49
Managerial Implications ... 52
Limitations and Recommendation ... 52
References ... 53
Articles ... 54
Books ... 57
Electronic documents ... 57
Introduction
People within the Netherlands are one of the most insured people of the world (Calma, 2007). Compared to the Dutch people only the Swiss citizens are more insured (Laan, 2014). Dutch people do not like risks. However, an insurance is not always necessary.
Most of the time customers spend too much money, if you would compare it to what they eventually get out of it. The insurance premium someone must pay is often vigorous, while the coverage is limited due to all sorts of restrictions. Research from a Dutch price comparison site, called verzekeringssite.nl, demonstrates that almost half of the citizens within the Netherlands have contracted a superfluous insurance in the past five years (Laan, 2014). In other words, almost half of the consumers did not make their decision in an optimal way. It seems that consumers not necessarily have or use the right information at the moment of decision making about this kind of products, which belong to the product category called
"annuities". An annuity (e.g. a health insurance or retirement saving plan) is a financial
product sold by financial institutions that are designed to accept and grow money from a customer and then, pay out a steady amount of cash at later points in time. At their core, annuities are contracts between a financial institution and a customer. An annuity’s goal is to provide a stable and long term source of income for the customer. Therefore it seems important to make a deliberate and well considerate decision about the choice you make according to these products.
of annuities it can be expected that the decision making process of such a product differs from the decision making process of a regular durable product. Within this paper I explain why this is the case by means of positioning the two product categories in the FCB-grid. Based on this theory there can be concluded that the consumer’s decision behavior for annuities and general durable goods are based on another sequence of steps (Fennis & Stroebe, 2016), and can be triggered in a different way. Furthermore, the Elaboration likelihood model (Petty, Cacioppo and Schumann, 1983) states that persuasive communications can cause changes in attitude through two different routes of processing. Based on the differences in characteristics and different positions the regular durable goods consumers seems to be triggered differently by different kind of information. Nowadays this information is widely available by the use of online platforms which provide information from all kind of sources like companies, the government but also to a large extent from other customers. Within this paper I will focus on consumer-generated information that is called online social interaction.
Previous research (Chen et al., 2011) already focused on the effect of the two kinds of social interaction on purchase decisions of more simple products like cameras. Their longitudinal studies leaded to significant results for the main effect social interaction (i.e. observational learning and electronic word of mouth) on the decision making process. Moreover, the results show that OL and eWOM differ in their impacts on the decision making process. Negative eWOM has a bigger influence on product choice than positive eWOM. In addition, they found that the opposite holds for OL. The researchers found the interaction effect between OL and eWOM valence to be non-significant.
Less is known about the effect of social interaction on more complex products. While prior research focused more on regular durable products (Chen et al. 2011), this paper is also aiming at annuities to find possible differences between them. Annuities are way different in their origin than the regular durable goods consumers buy on a more regular base.
For most of the durable products it is clear what you will get before buying the product. For an annuity this is way harder to predict upfront. According to the specific attributes of an annuity, a customer’s decision making process of annuities involves another sequence of steps
Within this research I will research the effects of observational learning and eWOM on the online decision making process. As mentioned before, these effects are also examined in the paper of Chen and his colleagues (2011), but only for regular durable goods. However, this research tries to elaborate the existing knowledge about the strength of these social interactions by means of a comparison between regular durable goods and annuities. So, within this paper the effects of OL and eWOM on the consumers decision making process will be tested and compared for both product types. According to the first type of social interaction the following research question is formulated:
“To what extent do consumers rely on observational learning (OL) cues
in an online decision making process for annuities, compared to regular
durable goods?”
According to the effect of electronic word of mouth (eWOM), the second research question is:
“To what extent do consumers rely on electronic word of mouth (eWOM)
in an online decision making situation for annuities, compared to regular
durable goods?”
“To what extent does electronic word of mouth (eWOM) influence the
relationship of observational learning (OL) cues in an online decision
making situation for annuities, compared to regular durable goods?”
To address the gap in the literature, I will focus on the two different kind of social interactions between customers and their effect on two different types of product categories. To find out if social interactions will have different effects for regular durable goods and annuities, both product categories are compared with each other. Then, some important psychological theories are explained. Based on the comparison the product types are categorized in the FCB-grid. From this theory we can see that both product types have some similarities but also let the consumer differ in their decision making process. A different place on the Elaboration continuum will show that consumers of the product categories will be susceptible to different types of information (i.e. heuristic cues and information quality). As mentioned in the beginning, consumers are not always making the right decision when it comes to products like annuities (Laan, 2104). Within this research I investigate the decision making process of regular durable goods and annuities, and I will check if this decision making process can be steered by the presentation of different kind of information (OL cues and eWOM).
Data was collected by means of two different online surveys in late 2016. The primary data that resulted from merging the two surveys into one are analysed by means of a Choice-Based conjoint analysis with predefined segmentation. To research combinations of the different valences of OL and eWOM, 160 respondents were faced with choice sets that had different combinations on observational learning cues (represented by the percentage of bought
products), eWOM information (represented by customer review quotes) and two additional
product they preferred most. The data of these choice tasks were than analyzed to investigate the differences between general durable goods and annuities.
The results show that observational learning and eWOM valence are evaluated slightly different between regular durable goods and annuities. Within this research setting eWOM and observational learning seems to be relatively more important for regular durable goods than for annuities. Moreover, both types of social interactions are ranked very high based on their relative importance compared to other attributes; eWOM is ranked first, observational learning is ranked second, followed by price and brand age seems to be evaluated as least important. Furthermore, this research also provide some evidence that the interactions between eWOM and OL is only marginally significant for some combinations of valences. Since these interactions are only marginally significant one cannot conclude much on them. The non-significance between observational learning and e-WOM was also founded earlier for a digital cameras (Chen et al., 2007).
From the theoretical framework and the results we can conclude that regular durable goods and annuities do differ on some important aspects. Therefore the impact of different kinds of social interactions also have slightly different effects on the decision making process for both products. The effects of OL cues turns out have a stronger impact on the decision making process, compared to an annuity. However, the effects of eWOM cues did not have a stronger impact on an annuity in comparison with a regular durable good.
making process. Since the other possible interactions are insignificant these effects of eWOM will have no significant influence on the effect of OL.
The findings of this research has implications for several interested parties; information providers, retailers and researches interested in the decision behavior of consumers. Based on the fact that little is known about the effects of consumer generated information on annuities this paper can make a contribution to the literature. In addition, this research will also assist marketers to better organize online information and design a more effective website. Showing information in an effective way can create an big advantage for companies, especially nowadays since there is a lot of information available through the internet. In addition, for companies and marketers it can be highly relevant to know whether behavior-based- and preference-based social interaction between customers has an effect on the decision making process of complex products like annuities. Based on this new knowledge these firms and marketers can create different and more effective strategies, by means of a good allocation of different kinds of information according to the different kind of products.
Theoretical Framework
In this section we will take a closer look at the conceptual model of this research, followed by the definitions of the different concepts within this model. In the first place the reader will be provided with an extensive description and explanation of annuities. In addition, I also make the important distinction between the decision making process of annuities and general durable goods and the ways these decision making processes can be influenced by different kind of information. After that, I will elaborate on the concept of social interaction, and two of its forms; observational learning and electronic word of mouth. The expected effects on purchase decisions are described and explained, and as a result different hypotheses are formulated.
The conceptual model
Figure 1: Conceptual model with hypotheses
Literature Review
complex and risky products, like an annuity. Some examples of these annuities are health insurances and retirement saving plans. To address this gap in knowledge, the next two sections will provide more information about the characteristics of these annuities, and their differences with regular durable goods.
Annuities and their characteristics
Annuities are contractual financial products that are sold by financial institutions. These institutions are designed to collect and grow funds from the customer. Within the annuity contract many details are included, like for example the age that payout will start and the payment intervals. A lot of different annuities exist and the decision for an annuity depends on the level of income someone is willing to accept in retirement and the amount of risk someone is willing to take.
The factor risk that is related to annuities stems from the fact that these annuities can give some security for coming unforeseen risky situations, and therefore have a huge influence on someone’s future. For example, insurances give protection from financial loss, healthcare
providers provide preventive, promotional, curative or rehabilitative health care services when this is needed (Birkman, 1996) and retirement saving plans are plans for setting aside money to create some security for the period after the retirement.
expectancy (Shu et al., 2016). Most of the time customers do not have this important knowledge to make the optimal choice.
Comparison with general durable products
It is important to make the distinction between the durable products we see on a regular base and products like annuities since they differ in nature, which also result in a different effect of social interaction on consumer’s decision making processes.
The first difference between these two types of products is the number of payments that has to be made by the consumer. For a digital camera a consumer only pays once. For an annuity this may differ. Normally an annuity is funded over time within the so called accumulation phase (Carson, Duran and Dumm, 2011). This means that consumers commit themselves for a longer period of time. This already brings us to a second difference, the long- versus short term orientation that is needed. A general product like a digital camera can be seen as a durable good and should work for a couple of years. This seems long, but compared to an annuity this period is rather short. Within the accumulation phase the funds must be long and frequent enough to provide a nice amount of money for the annuitization phase (i.e. the payout moment).
The moment of delivery of the product can be seen as difference number four. The transfer of a durable good generally takes place at the same time as the moment of purchase. In contrast, the payout of an annuity takes place in the annuitization phase after a long time period. Compared to a camera it seems relevant for a consumer to make a more deliberate and well considerate decision according to annuities, based on their importance for the future.
Table 1 gives a visual representation of the differences that are mentioned. Regular durable good
(Digital camera)
Annuity (Health insurance)
Payment is made once Payments are made over time within the
accumulation phase Relatively short term oriented
(< 5 years)
Long term oriented
Price is determined, and value is only changing due to deprecation
Price is fluctuating due to inflation and value depends on the accumulation phase The moment of delivery of the product
is the same as the moment of buying
The moment of delivery (i.e.
annuitization phase) of the product is delayed to a later moment in time
Table 1: The differences between regular durable products and an annuity
Decision making process
Based on the characteristics of durable goods and annuities it is easy to state that both products are placed on the high involvement end, where the motivation to perform cognitive effort in information processing is large. Figure 2, shows this high involvement end of the FCB grid. The option that is placed on the left side is related to thinking. Here, products associated with a high amount of (financial) risk are placed. Some examples within this option are houses and loans. This option also fits perfectly to products like annuities. The decision making of these kind of products involves the “think”-“feel”-“do” sequence, where consumers first learn about the quality and attributes of the product through carefully processing the available information sources to make a well considerate decisions and then develop an attitude (or evaluation) towards the product followed by the action in line with that attitude.
Figure 2: the high involvement end of the Foote, Cone-Belding grid
The biggest difference in the decision making process between the two products can be found in the moment of processing the available information. In general, consumers of annuities are more open and susceptible to available information about the attributes and quality, compared to consumers of a camera who are more susceptible to the right feeling. Based on this finding we can conclude that the decision behavior for annuities and other durable goods are both triggered differently by available information. Therefore it seems that the attitude towards the different kind of products will also be influenced differently by persuasive communications.
The Elaboration likelihood model
Based on the Elaboration Likelihood Model (ELM), a dual process theory of persuasion, persuasive communications can cause changes in attitude through two different routes of processing; peripheral and central (Fennis & Stroebe, 2016). These two are the endpoints of a continuum of processing intensity. Which route is used depends on the processing ability and processing motivation.
The central route processing implies that consumers will take all available information into account. So both, arguments and cues. Here, persuasion is determined by the quality of the argument. The central route requires high effort. From the analysis on the FCB-grid we already know that both product categories are linked to high involvement and require high effort. Therefore we can conclude that to a high extent persuasion is determined by arguments for both product types. However, since the ELM-model is based on a continuum the different product types can still have a different position at the central route-end.
account to make a well considerate decisions. Therefore, annuities are placed at the central route end of the elaboration continuum (see figure 3).
Figure 2: The elaboration continuum
In contrast to the central route, in the peripheral route persuasion is determined by rules of thumb and heuristics (Fennis & Stroebe, 2016). Peripheral route processing is focused on simple acceptance/rejection cues, and requires low effort.
Based on the FCB-grid we already concluded that the decision making process of regular durable goods and annuities are triggered differently by available information. Regular durable goods were assumed to be less triggered by all available information, and therefore can be placed slightly more towards the peripheral route end of the spectrum. So, consumers of durable goods are triggered more by the peripheral route of the Elaboration Likelihood Model, without realizing. Therefore, these consumers are more susceptible to heuristics like social proof (i.e. if other people like the product, it must be a good product). In the next sections the concept of social interaction will be explained further. Here I also make the link between heuristic cues and OL, and between arguments and eWOM.
Social interaction
combination with new tools and application (e.g., consumer review sites, social networks, weblogs and online discussion forums, etc.) offered consumers a simple way to exchange product information (Sneha, Mishrab and Shrivastav, 2014). The content is the information that is given through this channel, and consists of two important dimensions; objectivity and valence. This research is focusing on different kind of valences; negative, neutral or positive valence. The impact of social interactions is the effect of others’ actions.
Social interaction is a broad concept and can range from the passive observation of what strangers has chosen before you to personal recommendation. According to Chen, Wang and Xie (2011) consumers can exchange information by the use of two different social interactions: observational learning and word of mouth. The next sections will give a further explanation of these concepts.
Observational Learning
The first form of social interaction is called observational learning (OL). The concept of observational learning is less explored in marketing literature, but can be traced back to studies in the psychology science about social learning (Bandura, 1977). OL occurs when someone observes the actions of other people and make the same choices that they have made (Bandura, 1997).
Therefore, an observation learning cue can be seen as a heuristic cue that contains issue irrelevant information that is rather easy to process and allows consumers to arrive at a quick attitude towards a product. Since “actions speak louder than words” the OL-cues may still be perceived with a high credibility (Chen, Wang and Xie, 2011).
It is important to note that the content of observational learning can have different valences, an OL signal can either be positive, negative or mixed. As a result of the way the OL signals will be presented within this research I have decided to exclude mixed OL. Observational learning valence is determined by the percentages of adoptions. In other words, the valence depends on the share of choices among all actions that are made previously (Bikhchandani, Sushil, Hirschleifer an Welch, 1992). For example, an OL signal is getting more positive when the percentage of cumulative purchases among the choices made (by all previous informed customers) is getting larger. The opposite is also true, an OL signal is getting more negative when the percentage of cumulative purchases among the choices made (by all previous informed customers) is getting smaller. The cascade theory suggests that the positive and negative cues have an opposite effect on adoption behavior. Since positive cues will motivate the adoption behavior of consumers, the following hypothesis is formulated:
H1a: “Positive OL cues will positively influence the decision to choose for an
annuity.”
According to the cascade theory, the effect of negative OL cues have an opposite effect, which leads to the following hypothesis:
H1b: “Negative OL cues will negatively influence the decision to choose for an
annuity.”
From their findings we can conclude that negative OL has a smaller impact than positive OL. There is a plausible reasons for this phenomena. Products often differ in two dimensions, in quality (vertically) and in taste (horizontally). Thus, a negative OL cue can also mean two things. The product can either be chosen less due to a perceived low quality, or just due to a different taste of the previous customers. So, when there is a negative OL signal (i.e. only a few customers made a purchase action), this not automatically means that this product is perceived unfavorably by different consumers. It might also be a niche product with high quality. In contrast, the message of positive OL is rather clear. Positive OL is an indicator that a product scores high on both dimensions. A product can only achieve a high amount of purchases when it has got a high quality, in combination with a preferred taste. Therefore, a positive OL signal is perceived as more valid for consumers, than a negative OL signal. Based on the results of the quasi-experimental studies of Chen and his colleagues in 2011, I expect the same asymmetric effects of OL on the decision behavior on more simple durable goods like camera’s.
H1c: “Positive OL cues will positively influence the decision to choose for a regular
durable good (i.e. camera).”
H1d: “Negative OL cues will negatively influence the decision to choose for a regular
durable good (i.e. camera).”
H1e: “The effects of OL cues will have a stronger impact on the decision making
process for a regular durable good (i.e. camera) in comparison with an annuity.” Word of Mouth
Next to OL there is also a second form of social interaction, called word of mouth (WOM) (Chen et al., 2011). WOM occurs when consumers provide information (e.g. opinions or recommendations) about goods, services, companies or brands to other consumers (Rosario et al., 2016). When this information is communicated through the internet it is called electronic word of mouth (eWOM). As mentioned in the section about social interaction, this research focuses on internet as the channel for social interactions. Therefore the term eWOM is used instead of WOM. In contrast with OL, eWOM contains more information. That is because information that is derived from OL-cues only includes the actions of previous customers and not the underlying reasons (Bikhchandani, Sushil, Hirschleifer an Welch, 1992). eWOM, however, often contains both recommendations or opinions and the reasons for them. So, OL and eWOM differ in the amount of information.
Another difference between the two social interactions can be found in the credibility. OL is focused solely on previous actions of other customers. Therefore, the real preference of previous customers is displayed. Since eWOM can include a lot of distracting or unrealistic reasons behind a choice instead of their real preference, OL most often is seen as the more credible one. A reason for eWOM to consist of unrealistic information can be found in the limited capacity of people’s brain, the wish to give moral answers or the intention to influence other people.
From the research of Chen, Wang and Xie (2011) we can conclude that positive eWOM is less influential than negative eWOM.
The first reason can be found in differences of credibility (Chen et al., 2011). The provider of a health insurance, or another interested party, may hype his or her own insurance by writing very positive reviews on websites. Although the provider can post a lot of positive reviews, it is hard to prevent others from posting a negative review. In addition, the study of Chevalier and Mayzlin (2006) also have shown that the presence of negative eWOM is harmful and even more powerful in decreasing sales than positive eWOM is in increasing it.
A second reason can be explained by the idea of losses loom larger than gains. People who score high on risk aversion are more likely to pay more attention to negative signals, instead of positive signals (Mizerski, 1982).
Based on the literature the following hypotheses are formulated:
H2a: “Positive eWOM will positively influence the decision to choose for an annuity.” H2b: “Negative eWOM will negatively influence the decision to choose for an
annuity.”
Based on the results of the research conducted by Chen and his collegues (2011) neutral eWOM has a negative effect on the decision to choose for a product. Therefore I expect neutral eWOM, to also have a negative effect on the decision behavior for an annuity.
H2c: “Neutral eWOM will negatively influence the decision to choose for an
annuity.”
Based on the results of the quasi-experimental studies (Chen et al., 2011), I expect the same asymmetric effects of eWOM on the decision behavior for annuities.
H2d: “Positive eWOM will positively influence the decision to choose for a regular
H2e: “Negative eWOM will negatively influence the decision to choose for a regular
durable good (i.e. camera).”
H2f: “Neutral eWOM will negatively influence the effect of OL on the decision to
choose for a regular durable good (i.e. camera).”
However, I expect the relationship of eWOM to be weaker for cameras compared to annuities since the persuasion of cameras is based on less information according to the FCB-grid (Fennis & Stroebe, 2016). The elaboration continuum in figure 3 also shows that the persuasion towards annuities is stronger influenced by the argument quality of information. In addition, Rosario and his colleagues (2016) mention that consumers consult the opinions of others to reduce uncertainty about perceived risk. Consumers of annuities will therefore be stronger influenced by the effect of eWOM, instead of the other consumers.
H2g: “The effects of eWOM cues will have a stronger impact on the decision making
process for an annuity in comparison with a regular durable good (i.e. camera).”
The moderation effect of eWOM on OL
One of the goals of this research is to find out what will happen when a consumer is facing both kind of social interactions (OL and eWOM) at the same time. Within this paragraph we will take a closer look at this.
Kardes and Kim, 1991: 457). In other words, accessible information is not used as an input for judgment when more diagnostic information is available.
Based on this model there could be expected that when adding eWOM information this will replace the effect of OL for both product types, since eWOM has a higher diagnosticity than OL. To get a deeper understanding of this moderating effect, I separated the social interactions based on their valence. This leads to the following six hypotheses;
H3a: “Positive OL has a weaker effect on product preference when displayed with
positive eWOM.”
H3b: “Positive OL has a weaker effect on product preference when displayed with
negative eWOM.”
H3c: “Positive OL has a weaker effect on product preference when displayed with
neutral eWOM.”
H3d: “Negative OL has a weaker effect on product preference when displayed with
positive eWOM.”
H3e: “Negative OL has a weaker effect on product preference when displayed with
negative eWOM.”
H3f: “Positive OL has a weaker effect on product preference when displayed with
neutral eWOM.”
The study of Chen et al. (2011) reveals that there is a positive complementary interaction between OL and eWOM volume. Though, this study does not find any clear evidence for an interaction between OL and eWOM valence. However, based on the accessibility-diagnosticity model I do expect the presence of eWOM to have an effect on the relationship between observational learning and the decision making process.
will have a smaller effect for this kind of products. This reasoning is based on the findings of the FBG grid (Fennis, & Stroebe, 2016) and the ELM (Petty, Cacioppo and Schumann, 1983), where I concluded that eWOM will have a smaller impact on the decision behavior of regular durable goods. This leads to the last hypothesis;
H3g: “The moderation effects of eWOM on the relationship between OL and the
decision making process that are depicted above will have bigger impact for annuities, compared to regular durable goods.”
In the next section an overview of the different hypotheses will be shown. Hypotheses
Within this section the different hypotheses are ordered, the hypotheses for annuities are shown in table 2. The hypotheses for regular durable goods can be found in table 3, and the hypothesis according to the comparison between the two models are placed in table 4. Table 5 shows the hypotheses according to the moderation effect.
Hypotheses for the annuities
H1a “Positive OL cues will positively influence the decision to choose for an annuity.”
H1b “Negative OL cues will negatively influence the decision to choose for an annuity.”
H2a “Positive eWOM will positively influence the decision to choose for an annuity.”
H2b “Negative eWOM will negatively influence the decision to choose for an annuity.”
H2c “Neutral eWOM will negatively influence the effect of OL on the decision to choose for
an annuity.”
Table 2: Hypotheses for the annuities :
Hypotheses for the regular durable goods
H1c “Positive OL cues will positively influence the decision to choose for a regular durable
good (i.e. camera).”
H1d “Negative OL cues will negatively influence the decision to choose for a regular durable
good (i.e. camera).”
H2d “Positive eWOM will positively influence the decision to choose for a regular durable
good (i.e. camera).”
good (i.e. camera).”
H2f “Neutral eWOM will negatively influence the effect of OL on the decision to choose for a
regular durable good (i.e. camera).”
Table 3: Hypotheses for the regular durable goods
Hypotheses for the comparison between annuities and regular durable goods
H1e “The effects of OL cues will have a stronger impact on a regular durable good (i.e.
camera) in comparison with an annuity.”
H2g “The effects of eWOM cues will have a stronger impact on an annuity in comparison with
a regular durable good (i.e. camera).”
Table 4: Hypotheses for the comparison between annuities and regular durable goods
Hypotheses for the moderation effect
H3a “Positive OL has a weaker effect on product preference when displayed with positive
eWOM.”
H3b “Positive OL has a weaker effect on product preference when displayed with negative
eWOM.”
H3c “Positive OL has a weaker effect on product preference when displayed with neutral
eWOM.”
H3d “Negative OL has a weaker effect on product preference when displayed with positive
eWOM.”
H3e “Negative OL has a weaker effect on product preference when displayed with negative
eWOM.”
H3f “Positive OL has a weaker effect on product preference when displayed with neutral
eWOM.”
H3g “The moderation effects of eWOM on OL that are depicted above will have bigger impact
for annuities, compared to regular durable goods.”
Methodology
Research settingIn the first place we will take a closer look at on of the specific setting of this research, namely the health insurance market in the Netherlands.
Within this paper I focus on health insurances as a representation of the product category called annuities. Since health insurances differ around the globe, and the health care insurance market in the Netherlands is such a specific market I assume it is important to elaborate some more on this specific topic. This paragraph will give some knowledge about the health insurances within the Netherlands and provide the reader with some background information. The health insurance system in the Netherlands changed dramatically in 2006, now placing greater emphasis on consumer choice and competition among insurers (van den Berg, van Dommelen, Stam, Laske-Aldershof, Buchmueller and Schut, 2008). The Netherlands became one of the countries that combine universal coverage with private insurance and regulated market competition (Leu, Rutten, Brouwer, Matter and Rutschi, 2009). The changes in 2006 have resulted in an enormous premium competition in the Netherlands, where insurance carriers accept initial losses to build a substantial market share (Leu et al., 2009). With regard to the higher level of competition, for companies like insurance carriers, it is harder to retain their customers.
2016). It seems that consumers are more willing to change their insurance provider, and therefore these consumers will be more likely influenced by online social interaction with other customers.
Research design
To test the proposed model and its related hypotheses, two Choice Based Conjoint (CBC) surveys were administered through the internet to a wide group of respondents. The subject of the two studies only differ in their product group. The attributes and their attribute levels are held constant among both surveys, to create a setting where the outcomes of the two studies could be compared in a scientific way.
Both surveys were combined in one entry link, so the respondents were randomly assigned to one of both. This between subjects design has some different advantages. An important advantage of this combined entry link, where respondents only fill in one of the two survey, is the fact that respondents are less likely to get to know some thoughts about the researchers’ intention of this survey. Since people will not be able to develop a strategy to do better or worse in the second experiment, which confounds the results, this will strengthen the internal validity of this research. The internal validity of this research is also strengthened by the fact that people won’t have any learning effect after filling in a survey twice, i.e. when people become familiar with this testing environment of the first survey this can also confound the results.
As shown in table 6, the designs of the CBC’s contained four attributes with three to four levels. Subjects were presented with 3 alternative options within each choice task (see figure B1 and B2 in Appendix B for exemplary choice tasks of both surveys).
of people who have viewed that product. The OL cues contains the purchase percentage of that specific product (digital camera/health insurance).
Based on a pre-test (see Appendix A), I consider an OL signal of a product to be negative if the purchase percentage is 15 percent. In other words, most customers who viewed this product have rejected this product and chosen other types of annuities. In contrast, a positive OL signal can be seen as a purchase percentage of 85 percent, which means that most of the customers who viewed this product have chosen for this specific one. The respondents can also face a product alternative with no indication for the OL cue (i.e. “Unknown”). This attribute level is added to see whether its utility differs with the overall mean.
To include e-WOM information, the 12 choice sets each contain one attribute that shows the respondent a review quote of a previous customer. I have made a difference between positive comments, negative comments, neutral comments and no WOM information at all.
The attributes Price and brand age are included as fillers, to ensure that respondents will not get any suspicion about the actual purpose of this research.
Attribute nr. Attribute Levels Specification
1 Price (1) €104,95 (per month*) Linear/
(Filler) (2) €120,95 (per month*) Part-worth**
(3) €124,95 (per month*)
2 Brand age (1) This brand is ten years on the market Part-worth
(Filler) (2) This brand is three years on the market (3) This brand is one year on the market 3 Percentage of
bought items
(1) 85% of the customers who viewed this product, actually bought it
Part-worth
(OL) (2) 15% of the customers who viewed this product, actually bought it
(3) Unknown 4 Customer
Review
(e-WOM)
(1) Perhaps one of the best digital cameras I have ever had. And yes, after 1 year I am still satisfied with my choice!
Part-worth
(2) This is one of the worst digital cameras I have ever had. I am absolutely disappointed about my choice!
(3) This is a standard product (4) No review available
**The preferred specification for price turned out to be linear for digital cameras. However, to make a valid comparison, both are included as Part-worth.
Table 6: Attributes and Levels
Experimental design
For both surveys the number of full factorial alternatives is 108 (i.e. 3 ∙ 3 ∙ 3 ∙ 4) amounting to
e.g., (108−3)!∙3!108! = 204,156 potential choice sets, when kept in mind there are 3 stimuli per
choice set. To reduce this number of possible combination, a random sequence choice design type shows respondents 12 choice sets containing three camera/health insurance alternatives each. Based on a dual-response no-choice question, respondents first had to choose which of the alternatives is most and least preferred, and are also asked to indicate if they would really buy the preferred choice if it was available, yielding 36 decisions in total. In times of study design this type of CBC is chosen to gather additional information. However, in the eventual analyses I did not make use of the additional data according the dual-response or no-choice question since it did not make any contributions in finding the answers on the formulated research questions.
Measurement items
Except for their socio-demographic characteristics (i.e. gender, age, nationality and level of education), respondents also indicated their experience with the product that was related to their questionnaire (see table 7). Here, people filled in if they already had some experience with buying similar products. To prevent biases and use of prior knowledge and established associations I did not include brand names for the specific choice options.
this question really carefully. If respondents did not pass the intention check, they were excluded from the data set.
Construct Variables Answer possibilities
Socio- Demographic Characteristics Gender Age Nationality Level of education Male / female In years Country name o No degree
o Secondary school (Middelbare school) o Community college (MBO)
o University of professional education (HBO) o University - Bachelor's degree (BSc) o University - Master's degree (MSc) o Other
Product experience Experience o Yes, I have bought a digital camera/ health insurance before.
o No, I have never bought a digital camera/ health insurance before.
Respondent Engagement
Behavioral Engagement
Response time to CBC tasks (in minutes)
Intention check This is an intention check, please fill in number three (while default option is number four)
Table 7: Summary of measurement items
Data collection
Data collection did take place in an online setting in a period from November till December, in week 46 till week 50, of 2016. Within this survey participants did get in touch with either digital cameras or health insurances. As mentioned before, this allocation is executed randomly by the use of only one entry link that was connected to both surveys.
Modeling approach
“aggregate model”. In order to test this model in ‘Latent Gold’ a predefined segmentation is
used. In other words, digital cameras and health insurances are assigned to a product category dummy (1 = digital camera, 2 = health insurance). In the analysis the two product categories serve as fixed classes. The combined model will be tested multiple times with inclusion of different covariates (see table 7) to find the best model, by means of model fit. After checking the validity, the model will be used to analyze the main effects and compare them to the hypotheses.
Model Diagnostics
Goodness of fit measurement & predictive validity
Model fit is used to compare different models to each other. By the use of this model fit one can analyse whether a difference between the models lead to a significant better model. Important information criteria are the AIC, the AIC3, the BIC and the CAIC. A lower value represents a better model fit (Vermunt & Magidson, 2015). Hence, these criteria are not the same and should not be compared across each other. On the one hand, AIC and AIC3 tend to favor more complex models, and suit well to small sample sizes. While considering the number of parameters, these important information criteria takes the parsimoniousness (i.e. the number of parameters) of the models into account. On the other hand, BIC and CAIC are almost identical and preferred, especially for large sample sizes. Since this research includes models with a lot of respondents and even more observed choices, the BIC and the CAIC criteria are better suited to test the model fits within this paper.
To test if the estimated model parameters are significantly different from zero, a likelihood ratio tests is used. The likelihood ratio test is a widely used approach to testing the significance of a number of explanatory variables (Crichton, 2001). In order to work with the likelihood ratio test, the Chi-squared test statistic is calculated and the p-value is examined.
𝐿𝐿(𝛽)), where LL(0) represents the log likelihood of the null model, and LL(β) is the log
likelihood of the estimated model.
In order to test how well the models fit, the R² and the Mc Fadden’s adjusted R² are calculated in the following ways: 𝑅2 = 1 −𝐿𝐿(𝛽)
𝐿𝐿(0) , and 𝑅2𝑎𝑑𝑗 = 1 −
𝐿𝐿(𝛽)−𝑛𝑝𝑎𝑟
𝐿𝐿(0) , where LL(0) and LL(β) are defined in the same way as above and npar represents the number of parameters. According to Mac Fadden (1979) the R² represents excellent fit for values of 0.2 to 0.4. The adjusted R² allows one to compare models that have different numbers of parameters.
Finally, the hit rate (i.e. the percentage of times that the alternative with the highest predicted choice in a given task is the alternative actually chosen) is also reported and used to evaluate how well the model predicts a respondents actual choice.
It is important to note that the Log likelihood, the hit rate and the R² always improve with more parameters. Therefore, the most adequate measures to test the validation of the models are the adjusted R² and the information criteria (BIC and CAIC).
Relative attribute importance
The relative attribute importance will be used to test the hypotheses that include a comparison between the two product categories. To asses this relative importance, first the range should be calculated. This is the difference between the lowest and highest utilities of the attributes levels of one attribute. It shows an attribute’s maximum effect (Vermunt and Magidson, 2013), and measures how much influence each attribute has on people’s choices. The relative importance is then calculated by dividing the range of that attribute by the total range of all attributes. Based on the different relative importances a ranking could be made, which represents the order of importance of the attributes.
The interaction effects
will be investigated. This interaction effect occurs in a situation where the effect of two combined attributes differs from the sum of their utilities. To create such effects, new variables are computed. For example, the interaction between OL and eWOM valence is recoded into 6 different variables. positive OL x positive eWOM, positive OL x negative eWOM, and so on. These new variables are added to the model which result in different utilities and different relative attribute importances.
Results
Sample statistics
Since the combined model exists of two data of two different surveys, the statistics of both samples are summarized in table 8. The most important ones are described below.
Digital camera sample- The sample for this survey consists of 246 respondents, of which 35%
completed the entire questionnaire. From all respondents 96.59% did pass the intention check. Among the respondents there were more males (53.5%) than females (46.50%), and the sample average age was about 24 years. Most respondents are from the Netherlands (88.37%).
Health insurance sample- This sample consists of 222 respondents, with a completion rate of
31.5%. This completion rate is slightly lower compared to the digital camera-survey. The reason for this can be the found in the fact that digital cameras are more appealing to them. All respondents (100%) this pass the intention check.
Respondents in the digital camera setting (mean = 10 minutes, S.D. = 7.0) showed longer and more variable response times than respondents in the health insurances setting (mean = 8.7 minutes, S.D. = 5.2). The difference between the two conditions was around 2.0 minutes on average.
Variables Answer possibilities Digital camera Health Insurance
Mean (range) Sample (%) Mean (range) Sample (%) Age Gender Nationality Level of education in years male female the Netherlands Other No degree Secondary school Community college
University of professional education University - Bachelor's degree (BSc) University - Master's degree (MSc) Other 24 (17-61) 53.49 46.51 88.37 11.63 1.16 22.09 18.60 26.74 13.95 17.44 0.00 23 (16-60) 57.14 42.86 91.43 8.57 2.86 28.57 10.00 21.43 21.43 14.29 4.28
Experience Yes, I have bought a camera/ health insurance before.
No, I have never bought a camera/ health insurance before.
43.02
56.98
67.14
32.86
Intention check Passed 96,59 100
Table 8: Sample statistics
CBC Analysis - the main effects
This section discusses the different models based on the Choice-Based Conjoint analysis. Here I provide insights in the validity of the models, and an interpretation of the parameters and the relative importance of the attributes. In addition, the main effects of OL and eWOM are analysed and compared to the hypotheses.
Model Selection
compared to the initial models that included price as part-worth. The results are shown in table 10.
Product Digital camera Health insurance Both
Model Initial Model Pricing linear Initial model Pricing linear Combined Model Npar 9 8 9 8 19 LL -588.01 -588.02 - -524.94 -527.13 -1206.59 BIC 1215.23 1210.90 1086.58 1086.87 2506.66 CAIC 1224.23 1218.90 1095.58 1094.87 1199.06 Hit rate 75.00% 75.00% - 69.49% 69.35% 72.63% R² 0.4280 0.4280 0.3251 0.3223 0,3319 R²adj 0.4193 0.4202 0.3135 0.3120 0.3214
p-value 3.4e-187 ** 1.9e-187 ** 5.4e-201 ** 2.2e-200 ** 0.00 **
= improved model fit,
= no improvement in fit, - = no difference †p < .10, *p < .05, **p < .01 Table 9: Model comparison - price as part-worth/linear modelIn the digital camera setting the difference in log likelihood is rather small. The LLparth-worth is -588.01 and, LLlinear is -588.02 which leads to a chi-square of 0.02, and one degree of freedom (df = (9-8) = 1). This results in p(chisq = 0.02, df = 1) > 0.10, which means that the models are not significantly different in fit (i.e. LL). Moreover, the hit rates are identical. To choose the most parsimonious model price should be considered as linear (numeric), because this model uses less parameters. In addition, inspection of the adjusted R² leads to the same conclusion, since the adjusted R² of the initial model scores higher and therefore shows a better fit (R²adj: 0.4202 > 0.4193).
1086.58, hit rate: 69.49% > 69.35%). To get a definite answer, the adjusted rho-squared should be consulted. The adjusted R² in the part-worth model scores a higher compared to the linear model (R²adj: 0.3135 > 0.3120). Therefore it can be concluded that including price as a linear model will not lead to the same or an improved model fit.
To make a valid comparison between the two, both models should include price under the same circumstances. So, to create a model where the two product categories (regular durable goods and annuities) are included together, pricing should be considered as part-worth for both. The results for this combined model are also included in table 10, and can be found in the last column. The combined model turns out to be highly significant (p = 1.0e-385), and scores perfectly on its rho-squared (R² = 0.3319), since values of 0.2 to 0.4 for rho-squared represent excellent fit (Mc Fadden, 1979)
Covariates
In order to test the proposed combined model and to ensure reliable estimates, the combined model is estimated six times. The difference between these estimations is the inclusion of six different covariates. To check which covariates should be included in the model, different criteria are analyzed. The results are shown in table 11.
Model (1) (2) (3) (4) (5) (6) (7)
Included Covariates
None Gender Age Nationality Education level Experience Education level & Experience Npar 19 20 20 26 24 20 25 LL -1206.59 -1205.94 -1206.28 -1202.92 -1192.55 -1203.50 -1188.98 BIC 2506.66 2510.28 2510.95 2533.75 2503.00 2505.41 2500.78 CAIC 2525.66 2530.28 2530.95 2559.75 2527.01 2525.41 2525.78 R² 0.3319 0.3323 0.3321 0.3340 0.3397 0.3337 0.3417 R²adj 0.3214 0.3212 0.3104 0.3196 0.3264 0.3226 0.3279 Hit rate 72.63% “ “ “ 72.73% “ 72.73%
p-value 1.0e-385** 2.3e-386** 2.3e-386** 3.0e-390 ** 2.7e-385** 2.3e-386** 6.1e-386**
†p < .10, *p < .05, **p < .01
The results show that model 5 (with education level) and 6 (with experience) score best on the BIC and CAIC. Therefore, a seventh model (which includes both education level and
experience) is estimated. Compared to the model without covariates, this model turns out to
score best on the BIC (BIC: 2501 < 2507), and to score almost similar on the CAIC (CAIC ≈ 2526). Moreover, the Log likelihood for the model without covariatesis -1206.59 and the Log likelihood for the model with education level and experience as covariates is -1188.98, which leads to a chi-square of 17.61, and six degrees of freedom (df = (25-19) = 6). This results in
p(chisq = 17.61, df = 6) < 0.01, which means that the model with the two covariates fits
significantly better than the model without covariates, and that the additional parameters increase fit. Besides the information criteria, the most adequate measure to test whether the price attribute should be linear or part-worth, the adjusted R², leads to the same conclusion, since the adjusted R² has a higher score for the model that includes the two covariates (R²: 0.3279 > 0.3214). Therefore this model has been chosen to proceed further analyses with.
Validity
After examining the results for the eventual model (which includes education level and
experience as covariates) we can come to several important findings about the predictive
Parameter interpretation of the main effects
The results for the final model are shown in table 12 on the next page. All attributes, are highly significant. Slightly different results can be found for the difference between the two classes; only the difference between the classes for brand age is not significant at al (p = 0.48). Fortunately, this attribute is only used as a filler. According to the subject of this paper we will focus here on the attributes that are related to the OL cue (percentage of bought items) and e-WOM information (customer review).
OL cues – The main effect of OL cues is significant, such that more positive cues increase
consumers utility. The parameters show that customers who are exposed to a positive OL cue are more likely to choose a product (β = 0.8807 for digital cameras, 0.7421 for health insurances) for health insurances) than the average customers. These finding are in line with
H1a (“Positive OL cues will positively influence the decision to choose for an annuity.”) and H1c (“Positive OL cues will positively influence the decision to choose for a regular durable good (i.e. camera).”). The effect of negative OL cues are also rather clear. Consumers are less likely to choose for a specific product when they are exposed to a negative OL-cue (β = -0.1683 for digital cameras, -0.2829 for health insurances), which leads to confirmation of H1b (“Negative OL cues will negatively influence the decision to choose for an annuity.”) and H1d (“Negative OL cues will negatively influence the decision to choose for a regular durable good (i.e. camera).”).
Attributes
camera
Insurance
Wald p-value Wald(=) p-value Mean Std.Dev. Price (Filler) €104,95 per month 0,1466 0,5730 85,7218 1,1e-17** 22,6660 1,2e-5** 0,3286 0,2109€120,95 per month -0,0114 -0,1485 -0,0699 0,0678
€124,95 per month -0,1352 -0,4245 -0,2587 0,1431
Brand age (Filler) This brand is ten years on the market 0,1438 0,0688 13,9091 0,0076** 1,4678 0,48 0,1118 0,0371
This brand is three years on the market 0,0022 0,1087 0,0477 0,0527
This brand is one year on the market -0,1460 -0,1775 -0,1594 0,0156
Percentage of bought items (OL cue)
85% of the customers who viewed this product, actually bought it 0,8807 0,7421 294,3779 1,8e-62** 5,9648 0,051† 0,8215 0,0685
15% of the customers who viewed this product, actually bought it -0,1683 -0,2829 -0,2172 0,0567
Unknown -0,7124 -0,4592 -0,6043 0,1252
Customer Review (e-WOM information)
Perhaps one of the best digital cameras I have ever had. And yes, after 1 year I am still satisfied with my choice!
1,8527 1,3137 667,55466,2e-141** 21,6507 7,7e-5** 1,6226 0,2666
This is one of the worst digital cameras I have ever had. I am absolutely disappointed about my choice!
-1,5832 -1,4297 -1,5177 0,0759
This is a standard product 0,0509 0,2542 0,1377 0,1006
No review available -0,3204 -0,1382 -0,2427 0,0901
evaluations increase consumers utility. The results show that customers who are confronted with positive e-WOM information about a product are more likely to actually choose that product, than the average customer (β = 1.8527 for digital cameras, 1.3137 for health insurances). In contrast, the likelihood that a consumer will choose for a specific product when they are exposed to negative e-WOM information gets smaller compared with an average customer (β = -1.5832 for digital cameras, -1.4297 for health insurances). Provision of neutral e-WOM has the same effect as positive e-WOM, but its impact is significantly smaller (β = 0.0509 for digital cameras, 0.2542 for health insurances). In other words, customers that are exposed to neutral e-WOM about a product are more likely to choose that product compared to the average customer. The customers that were not exposed to e-WOM turned out to be less likely to choose for a product compared to the average customer (β = -0.3204 for digital cameras, -0.1382 for health insurances).
Therefore, the second hypotheses (H2a, H2b, H2c, H2d, H2e and H2f) are all supported, except for H2c and H2f. The latter two hypotheses stated that neutral e-WOM would have a negative influence on the decision behavior to choose for a product. The results however, show that this influence is positive although its effect are rather small (i.e. the β’s are close to zero).
To assess how preferences differ across the two classes and how OL influences choices, relative attribute importance was calculated. This is the percentaged largest difference between attribute parameters, the results are shown in table 13.
Relative Attribute importance
†p < .10, *p < .05, **p < .01
Table 12: Attribute importance and ranking
Overall, for both product types the relative importance for price seems very unimportant compared to the relative importance of social interactions. One of the most significant differences between the two product groups can be found for price. Price importance is relatively higher for health insurances (ranked third) compared to digital cameras (ranked fourth).
The OL cues are relatively more important in the digital camera setting (28.44%, ranked second), compared to the health insurance setting (22.98%, ranked second). This finding is in line with H1e that stated; “The effects of OL cues will have a stronger impact on a regular
durable good (i.e. camera) in comparison with an annuity.” Therefore H1e is confirmed.
e-WOM is the most important attribute for consumers of both product categories, and is ranked first in both settings. However, the relative importance of e-WOM for a digital camera (61.73%) is even higher than for a health insurance (52.47%). This finding is not in line with H2g;“The effects of eWOM cues will have a stronger impact on an annuity in comparison
with a regular durable good (i.e. camera), and therefore this hypothesis should be rejected.
CBC Analysis - The moderating effect of eWOM valence
In this section, the analysis of the moderating effect of eWOM valence on OL will be discussed. The hypotheses with regard to this moderation effect will be tested and further analysed based on the parameter interpretation and a comparison of the relative attribute importance of the different product categories.
Model Selection
table 14. The last column gives a conclusion if an improvement was observed compared to the model without any interactions.
Model Model
without interactions
Model 2
(OL x eWOM valence)
Improved fit? Npar 25 39 LL -1188.98 -1178.58 2548.75 2587.75 No BIC 2500.78 No CAIC 2525.78 No R² 0.3417 0.3475 R²adj 0.3279 0.3259 No
Hit rate 72.73% 72.86% Yes
p-value 6.1e-386** 6.1e-391**
†p < .10, *p < .05, **p < .01
Table 13: Model comparison interaction effect
To asses the quality of the new model that includes an interaction between percentage of
bought item (OL cue) and customer review (eWOM information) relative to the model
without any interaction different information criteria are calculated. The new model turns out to score worse on the information criteria, the BIC (2549 > 2501) and CAIC (2588 > 2526) are larger for the interaction model compared to the old model. The interaction model only scores better on the hit rate, but this difference is rather small. ). Moreover, the Log likelihood for the model without interactions is -1188.98 and, the Log likelohood for the model with interactions is -1178.58, which leads to a chi-square of 20.80, and six degrees of freedom (df = (39-25) = 6). This results in p(chisq = 20.80, df = 6) < 0.005, which means that the models are significantly different in fit (i.e. LL). Based on this findings the model without interaction effect will make better prediction of the preferences of the respondents than a model with an interaction between OL and eWOM valence. Therefore should not be included in the model.
However, to check the hypotheses according to this interaction it was necessary to include the interaction effects. The results for this model are shown in table 15, on the next page.
