The effect of eWOM volume and type on information
overload and the moderating role of summary ratings
30 – 01 - 2017
The effect of eWOM volume and type on information overload and the
moderating role of summary ratings
Master Thesis University of Groningen Faculty of Economics and Business
MSc Marketing Management and Marketing Intelligence 1st Supervisor: Dr. H. Risselada
2nd Supervisor: Dr. F. Eggers
Author: Marin Denise Teinsma Date: 30 of January, 2017
Adress: Tuinbouwstraat 116B, 9717 JP Groningen Telephone: 0646710925
Abstract
With the growing internet and the corresponding increasing amount of reviews, a risk for information overload becomes present. Current research investigated mostly the effect of volume count on purchase intention and sales, but not the effect of volume content. Next to that, this research focused on the behavior of consumers, instead of the amount of sales by looking into the information overload and subjective states. Based on previous research, there are two different types of reviews, each having different effects. This research investigates the attribute-value and simple-recommendation reviews.
200 Participants had to indicate their purchase decision when exposed to one of the eight different conditions and their corresponding subjective states. Hereby summary ratings expressed in star values were included in the conditions to investigate if this moderates the expected relationship between type and volume on information overload and subjective states.
The research found a significant influence of volume content on information overload, but no significant evidence type of review and the subjective states. However, the volume count appears also to have different effects on the information overload and subjective states.
Preface
Table of content
Introduction ... 7
Moderator ... 9
Theoretical Framework ... 11
Word Of Mouth ... 11
Electronic Word of Mouth ... 11
Valence and Volume ... 12
Information Overload ... 14
eWOM overload ... 14
Star summary ratings ... 16
Covariates ... 17 Research Design ... 19 Research Method ... 19 Study design ... 19 Product ... 19 Reviews ... 19 Pre-test ... 20 Experimental procedure ... 21 Measurements ... 21 Independent variables ... 22 Dependent variables ... 22 Interaction ... 22 Covariates ... 22
Manipulation and control variables ... 23
Plan of Analysis ... 24 Results ... 26 Descriptive ... 26 Scale validation ... 26 Correlation analysis ... 26 Factor analysis ... 27 Reliability analysis ... 27
Control and manipulation checks ... 28
Regression ... 29
Model ... 32
Information overload ... 33
Satisfaction ... 34
Confidence ... 34
Confusion ... 36
Latent Group Regression ... 37
Discussion ... 41
Theoretical Implications ... 41
Managerial Implications ... 43
Limitations and further research ... 43
References ... 45
Introduction
With the growing electronic environment due to technological developments, eWOM gains more popularity and becomes more important in the market (Rosario et al. 2016; Godes et al., 2005). eWOM - electronic word of mouth - is information on the internet created by consumers about goods, services, brands or companies to other consumers (Rosario et al. 2016). This kind of information are mostly opinions and recommendations about consumers experiences (Chen, Wang and Chie, 2011). Due to the Internet, consumers can quickly access information from eWOM sources about products that influences intention of the purchase (Chen et al. 2016). It is becoming so essential that 'eWOM-related efforts is considered one of the greatest challenges interactive marketers face today' (Rosario et al. 2016, p. 297). Hereby eWOM serves as a leading sign for success of the product (Godes and Mayzlin, 2005). According to IBM research, 53% of the customers connect online to get information and see reviews and rankings (Spencer and Freeman, 2012). Although 83% says they trust completely in recommendations from friends or family, 66% also claims that they trust online generated opinions. Of those 66% people, 69% takes action upon that opinion (Nielsen, 2015). Next to that, Bickhart and Schindler (2001) found that reviews produces more product interest than market-generated sources of information in an online environment.
In the literature eWOM has been researched heavily and it can be concluded that it has a big impact on behavior (Reichelt, Sievert and Jacob, 2014; Chevalier and Mayzlin, 2006; Senecal and Nantel, 2004). It reduces uncertainties of the customers and therefore leads to higher sales (Rosario et al. 2016; Duan, Gu and Whinston, 2008a; Liu, 2006). The metric volume of eWOM especially is proved to be an effective metric to indicate a positive impact on choice decision and sales (Rosario et al. 2016; Chevalier and Mayzlin, 2006, Senecal and Nantel, 2004). However, this metric for volume has always been a count metric.
overload is very much present nowadays. Information overload can appear when consumers decide that the costs for searching no longer compensates for the information given. When there is too much information to process, they stop the searching (Branco, Sun and Villas-Boas, 2015). Toffler already mentioned in 1970 the term information overload: '' When the individual is plunged into a fast and irregularly changing situation, or a novelty-loaded context, his predictive accuracy plummets. He can no longer make the reasonably correct assessments on which rational behavior is dependent.” (p. 350 - 351).
The phenomena of information overload is investigated intensively over the decennia (Iyengar and Lepper, 2000; Keller and Steallin, 1987; Malhotra, 1982; Jacoby, Speller and Berning, 1974a, b) and has been connected to causing frustration, stress (Krishen, Rascke and Kachroo 2011), less satisfaction, confidence and more confusion (Lee and Lee, 2004; Malhotra, 1982; Jacoby, Speller and Berning, 1974a, b).
Overload caused specifically by a volume of eWOM has not been researched a lot. A reason for this is that volume of eWOM always has been linked to a count metric. This count metric of volume doesn’t cause information overload, since this is not hard to process. Park and Lee (2007) and Kwon et al., (2015) did look at the other aspect of a high volume of information, namely at the content volume of information. However, these researches had contradicting results. Park and Lee (2007) found in their research that information overload does exists when content volume increases, but Kwon et al., (2015) didn’t. Forman, Ghose and Wiesenfield (2008, p. 291) phrased that ‘’research in this area is fragmented, we have yet to understand why, how, and what aspects of online consumer-generated product reviews influence sales”. This shows that the underlying process of eWOM influencing sales is not clear yet and needs to be researched more. Due to contradicting results of eWOM content volume causing information overload and due to the limited research into the content aspect of eWOM, this research will investigate this relationship more extensively. Based on the original information overload literature (Iyengar and Lepper, 2000; Keller and Steallin, 1987; Malhotra, 1982; Jacoby, Speller and Berning, 1974, a,b) it is expected that an increase of eWOM content will cause information overload. This leads us to the following research question;
Q: To what extent does eWOM content volume lead to information overload?
simple-recommendation only recommends the product and shows feelings of the writer. Both structures creates different effects on the receiver and therefore type of review, existing out of attribute-value and simple-recommendation, will be included in the research. Moreover, Park and Lee (2007) found a significant relationship between the type of the review and perceived information and indicated an effect on information overload.
Next to that, since count volume is researched so heavily and linked many times to purchase decision (Rosario et al., 2016), it is most probably influential in the underlying decision process. Therefore the count volume of information is also included in this research.
Moderator
Following the traditional Elaboration Likelihood Model, people take two paths in order to form a decision; the central and peripheral route. The central route focuses on text content and highly analyzes all the information given, motivated by a high involvement. The peripheral route is mostly seen as a shortcut. There is less cognitive effort to process the information since this route is mainly focusing on cues like visuals or colors (Petty and Cacioppo, 1986). Therefore it is expected when people experience too much information and therefore overload, they will take the peripheral route and look for cues to release the cognitive load and make the decision process easier. One of those cues often represented with eWOM are numeric cues in the form of star summary ratings. Hence, numeric cues may reduce the effect of content volume on information overload.
Most of the websites are using these numeric cues, but it hasn’t been researched a lot if this is actually helpful. Helpful in order that it diminishes the information overload and doesn’t stop people from processing the information for their purchase decision.
This research will investigate the research question by conducting a questionnaire whereby consumers make a purchase decision based on the information given. This information exists out of different sources; namely count and content of eWOM volume, type, product description and summary ratings. The results shows that information overload does exist when increasing in content volume, but is not influencing the subjective states significantly. Hereby summary ratings are not moderating the relationship between volume of information on information overload and subjective states.
measured by the subjective states; satisfaction, confidence and confusion and by the construct information overload. Next to that, it will be researched if the expected positive relationship between volume and information overload is moderated by summary ratings. To our knowledge, only a few studies have researched the influence of numeric cues on evaluations of reviews, but never in combination with information overload.
This research has implications from a managerial perspective in the sense that managers can understand the effects eWOM has on consumer behavior. With the results of this study, managers can develop guidelines for creating more efficient online reviews, understanding the underlying structure of eWOM influencing sales and be aware of the risk of information overload. This risk may be reduced by giving a clear overview of reviews using numeric cues such as star summarizing statistics.
Theoretical Framework
In the following sections, a framework of definitions from the concepts of WOM, eWOM and in particular the count and content volume and type of eWOM will be given. Followed by the expected influence of these concepts on information overload and the interaction effect of summary ratings. An overview of the theoretical framework is given in figure 1.
Word Of Mouth
Word of Mouth is defined as ‘‘all informal communications directed at other consumers about the ownership, usage or characteristics of particular goods and services of their sellers’’ (Westbrook, 1987 p. 261). This informal communication is highly influential due to the underlying herd behavior (Banerjee, 1992). Herd behavior means that people will follow others actions resulting in people following all the same behavior. It shows that consumers are able to ignore their own knowledge in favor of information from actions of others (Banerjee, 1992). WOM is also influential in the sense that it helps reducing uncertainty since people can share their experiences with each other (Roselius, 1971). Electronic Word of Mouth
Electronic Word of Mouth differs from WOM since it occurs in an electronical environment. Therefore eWOM is mostly written, while WOM is a spoken word and occurs in a face-to-face situation. One advantage of this face-to-face situation is WOM being more influential and having immediate impact on the receiver (Herr, Kardes and Kim, 1991). Advantages of eWOM compared to WOM are that one can read the messages in his own speed and more information is available for the consumer about the product or service (Bickart and Schindler, 2001). It is however questionable until which threshold the amount of information stays an advantage. Besides that, since eWOM has an online surrounding, the personal ties are weak due to a missing personal relationship between the reviewer and the person reading the review and because eWOM is mostly anonymous (Chatterjee, 2001).
Although the ties are weak, eWOM is widely being used. Ong (2011) discovered in his research that two third of his participants are first reading the reviews before purchasing. It helps consumers to investigate the expectations when buying the product or service which will result in customers being
more satisfied with the purchase (Ong, 2011). Next to that, eWOM has empirically been found to influence the behavior, intentions and attitudes of consumers in the online and offline decision to buy (Reichelt et al., 2014; Liu, 2006). This can be explained by the fact that consumers follow others and are influenced by them through social interactions while making a purchase decision (Godes et al. 2005).
Valence and Volume
eWOM has been researched intensively with a focus on the metrics valence and volume (Rosario et al., 2016). Valence and volume create two different effects; the persuasive and the awareness effect (Liu, 2006).
Persuasive effect of valence
The persuasive effects occurs when the valence of reviews is showed. Hereby valence indicates the nature of the messages (positive or negative)(Liu, 2006). A positive review indicates a direct or indirect recommendation and a negative review shows mostly deprecation and private complaining. Hence, positive reviews demonstrate expected quality and negative reviews reduces it (Liu, 2006). According to the amount of negative or positive messages, consumers create a positive or negative attitude and evaluation towards the product or brand.
Awareness effect of volume
The volume of reviews indicates an awareness effect. When the volume is high, more people are aware of the product existence and make it part of their consideration set (Liu, 2006; Duan, Guan and Whinston, 2008b). This literature discusses volume and focuses on the count metric of volume. When the count volume increases, people will be more aware of the product.
the reviews are mostly already positive and when it is negative, consumers expect that it is created by competitors from the brand or by frustrated consumers and therefore customers are more influenced by positive than negative reviews (Ong, 2011).
Ewom volume
EWOM volume is defined by Liu (2006, p. 75) as the ''total amount of eWOM interaction''. The volume of information and eWOM has a positive influence until a certain threshold, this can indicate information overload (Rosario et al., 2016; Chevalier and Mayzlin, 2006, Senecal and Nantel, 2004). Next to that, a higher volume creates a higher ''stickiness'' - the likelihood to continue with a purchase - when consumers can easily gather the information about the product (Spenner and Freeman, 2012). These literature investigated the effect of real reviews, but measured with the count metric of volume. Therefore their research includes both content as count metrics of volume. For this reason, it is hard to distinguish the separate effects of count and content information on decision behavior.
Recommender and informative role of volume
As discussed before, the count volume creates an awareness effect. Next to that, volume fulfills two different roles according to Park and Lee (2007). They found that volume fulfills an informative and a recommender role. When looking at the recommender role this can be linked at the awareness effect of count aspect of volume. Volume of reviews provide a positive or negative signal for the popularity of the product. When volume increases, more people will talk about the product, the popularity increases, which results in more customers who are informed by the product (Liu, 2006; Godes and Mayzlin, 2004). This higher awareness will indirectly instruct customers that a lot of people already used this product and shows the popularity (Park and Lee, 2007), hence fulfilling the recommendation role. The underlying constructs are the herd behavior (Banerjee, 1992), the bandwagon effect, whereby others follow the one whom starts with an action (Rosario et al., 2016) and the awareness effect as mentioned earlier. People will act according to the information that is reflecting in the high volume (Godes and Mayzlin, 2004), this will reduce the uncertainty when choosing a product (Chen, Wang and Xie, 2011) and results eventually in higher sales (Rosario et al., 2016).
Simple-recommendation and attribute-value
Simple-recommendation shows emotions, are subjective and abstract and are based on the feelings of a customer. An attribute-value review is objective, rational and gives facts about the product (Park and Lee, 2007). When the volume increases of attribute-value reviews, the information set is more informational since different arguments are given for supporting the recommendations. When the volume increases of simple-recommendation, they only deliver more feelings, but not more reasons. Park and Lee (2007) could indicate the perceived informativeness with an increase of simple-recommendation and attribute-value. They found that the informativeness is stronger and gives more impact with attribute-value reviews compared to simple-recommendation.
It can be concluded that eWOM volume has a lot of positive consequences, especially on influencing sales. However, with so much content information, the perceived informativeness can be reached early and decline after (Park and Lee, 2007) and risk of overload can exist. According to Spencer and Freeman (2012), the rising volume of messages is overwhelming. It creates a stickiness, but only when information can be gathered in an easy way. This is not the case when there is too much information to process and information overload is created.
Information Overload
When there is too much information to process, cognitive overload is created and can cause less likelihood to purchase (Iyeangar and Lepper, 2000), a decrease in decision quality (Keller and Steallin, 1987; Malhotra, 1982; Jacoby, Speller and Berning, 1974a), frustration and stress (Krishen, Rascke and Kachroo, 2011) and a negative impact on the subjective states. The subjective states in the literature of information overload are satisfaction, confidence and confusion. When more choice sets are introduced, satisfaction decreases, consumers are less confident and confusion increases (Malhotra, 1982b; Scammon, 1977). This is due to the pressure of processing too much information in a limited time and having a limited information processing capacity (Malhotra, 1982). When information expands, the effort to convert this information also increases. But, since it is limited and when it crosses the threshold, consumers will be overloaded by the information and the response rate will decrease (Schroder et al., 1967; Grise and Gallupe, 2000). This can even cause delay or avoidance (Iyengar and Lepper, 2000) and can be paralyzing (Spencer and Freeman, 2012).
eWOM overload
consumer experiences a choice as too difficult, s/he can create the feeling to stop comparing the alternatives and make the choice on impulse (Hendrick et al. 1968). Therefore a risk for choosing the wrong product will exist.
Although information overload is a common term in the marketing world, research is limited when it comes to knowledge about overload of electronic Word Of Mouth. It has been discovered that too much information can decrease the perceived informativeness (Zu and Zhang, 2009; Park and Lee, 2007). Hereby, Park and Lee (2007) investigated mainly content volume of eWOM causing overload. They divided volume in three levels; low, moderate and high amount of reviews, and into the two types of reviews; review-attribute and simple-recommendation. It appears that attribute-value is causing sooner perceived informativeness than recommendation. Reason for this, is that simple-recommendations can be understood easily and will take less effort than an attribute-value review. To process that information, it takes more cognitive effort (Park and Lee, 2007). Therefore more reviews, means more to process. Since this is very time consuming, most of the consumers will scan the reviews when volume is increasing, but with risk for missing important information. This can result into less satisfaction, confident and more confusion (Park and Lee, 2007). Therefore, the first hypothesis will be:
H1: The change of review type from simple-recommendation to attribute-value review leads earlier to information overload.
H2: The change of review type from simple-recommendation to attribute-value reviews leads earlier to a) less satisfaction, b) less confidence and c) more confusion.
Park and Lee (2007) didn’t focus especially on the subjective states of the customers. Kwon et al., (2015) continued with researching the eWOM overload with investigating the subjective states. However, no clear information overload was existing in their research. They did discover that when using a F-distribution with mainly positive reviews, participants felt more information load than when more negative reviews were included. A reason for this could be that the readers read more reviews when mainly positive reviews were included (Kwon et al., 2015). A F-distribution is the distribution of reviews, whereby most of the reviews are really positive and just a few really negative which shows a F figure as distribution.
content metric and need different levels of cognitive processing. Therefore information overload can be straightforward. Next to that, Kwon et al., (2015) had a relative easy product whereby consumers were not highly dependent on reviews and could therefore ignore most of them. The purpose of this paper is however to investigate when information given by the producer is not clear and customers are dependent on reviews, how this affects the consumer behavior.
When people have to read a lot of reviews, it is time consuming and therefore the customers would perceive less helpfulness of the reviews (Ong, 2011). Hence, looking at the informant role, even though more information is given, the overall perceived quality will lower. This gives us the following hypothesis:
H3: The increase of volume content from moderate to high leads to an increase in information overload. H4: The increase of volume content from moderate to high leads to a) less satisfaction, b) less confidence and c) more confusion.
Next to that, Park and Lee didn’t focus especially on subjective states, but Kwon et al., (2015) did. Based on hypothesis 3 and 4, it is assumed that when volume increases, information overload exists and will influence negative subjective states. Based on hypothesis 1 and 2, attribute-value reviews lead earlier to information overload and subjective states. Hence, when looking at the increased volume, it is expected that when volume of reviews increases of both attribute-value reviews and simple-recommendation, an increase of information overload will exist. This will be reflected in a decrease of satisfaction and confidence and an increase in confusion.
Star summary ratings
In order to decrease the information overload, managers can make the eWOM more easy to understand (Rosario et al., 2016). Websites like amazon.com often use numeric cues for this. As explained in the introduction, when consumers experience a high amount of information, they can look for efficient methods to process information. Hereby they can rely on numeric cues like star summary ratings (Mudambi and Schuff, 2010). This will save their cognitive processing (Mudambi and Schuff, 2010) and decreases the risk for overload. The researchers found that star summary ratings can positively influence the perceived quality of the reviews and the helpfulness.
Another reason for including summary ratings, is because this research contains the investigation of a search product. Search products often include a high financial risk, which means that it can cost consumers a lot of money and therefore they need to choose the product carefully (Roselius, 1971). When there is a high financial risk, consumers look more for eWOM (Lin & Fang, 2006) and need to decide in a clear way which information is available. Schneider (1987) shows that overload happens when the situation is 'uncertain, ambiguous, novel, complex or intense'. As Malhotra (1982b) also describes, in a situation of overload, consumers will look for ways to simplify the information process. Therefore summary ratings can become a helpful device and causes less information overload. A critical note on the research of Park and Lee (2007) was that they didn’t include star summary ratings although this is very standard for online reviews. Therefore, together with the reasons mentioned above, this paper is including this numeric cue in the research. Kwon et al., (2015) included the star summary ratings to reflect reality as close as possible. They didn’t find a significant relationship between volume and information overload. This raises the question if the star summary ratings were maybe causing this absence of relationship. The star summary ratings makes it easier for consumers to process information and therefore it can be a reason that the participants in the Kwon et al., (2015) research didn’t experience information overload. Based on this, the following hypothesis is created: H5: Including summary ratings will negatively moderate the relationship between eWOM volume content on information overload and a) less satisfaction, b) less confidence and c) more confusion.
H6: including summary ratings will negatively moderate the relationship between eWOM type on information overload and a) less satisfaction, b) less confidence and c) more confusion.
Covariates
In order to see if information overload and subjective states are maybe caused by other factors, extra covariates are added to the model. First of all, age is added, since people from 21 to 34 years old have the highest level of trust on consumer opinions posted online (70%) (Nielsen, 2015). This can indicate that this age group is experiencing less information overload, since they trust the reviews more. People above the 65 years old have the least trust in online opinions (47%).
information overload (Chen et al., 2016), since they are more depending on reviews and see it as something valuable in order to reduce uncertainty.
Involvement is added, due to the Elaboration Likelihood Model. As described in the introduction, when involvement is high, people are more motivated to process the available information and will follow the central route. Therefore it is expected that people with higher involvement experience less information overload.
Attitude towards advertisements is measured. When attitude is positive, it is expected that people experience less information overload (Park and Lee, 2007).
Experience in online buying is measured by asking how often people buy online and how much experience they have with buying the product used in the survey. How higher the experience in online buying and with buying the specific product, how lower the expected information overload is. People who are often online, are also experienced in reading and scanning reviews online. Therefore they will experience less cognitive load.
Lastly, the information source is asked. Consumers base their decision on different parts of information. This covariate existed out of; product title, price, product description and reviews.
Conceptual model
Based on the former theory and hypothesis, the following conceptual model can be created:
Research Design
In order to research the relationship between the volume of eWOM and respectively information overload and subjective states, a survey has been conducted. Hereby a between subject design was performed. This section will describe the method, followed by the study design used for the questionnaire, which instruments are used for the measurements and concludes with the plan of analysis.
Research Method
The research has a factorial design of 2 x 2 x 2. Whereby the independent variables are; review type (attribute-value and simple-recommendation), review quantities (moderate and high) and two conditions of the moderator (summary ratings available and not available). The survey was distributed among social media and email. All the participants filled in the questionnaire voluntarily and any incomplete surveys were discarded. In order to make certain every condition got equally respondents, the survey was randomized.
Study design Product
A laptop was chosen as product to consider for purchasing in the survey. Reasons for this decision were; the reviews of the survey could be based on real reviews, a lot of people are familiar with the product, people are more buying their electronical products online (CBS) and lastly, since a laptop is a
complex and experience product, people look more at the evaluation process
(Gupta and Harris, 2010). The product was shown in the summary with a brand name and some corresponding information. An example is given in figure 3.
Reviews
Based on research of (Kwon et al., 2015) the reviews used in the survey were derived from real reviews of the website Amazon.com. This reflects reality as close as possible. A best-seller laptop on Amazon.com was chosen to make sure enough reviews and a F-distribution were available. As mentioned in the introduction, this F-distribution was used since most reviews are positive on the internet (Chevalier and Mayzlin, 2006). Kwon et al., (2015) also included this distribution in their
research. The F-distribution of this laptop exists out of 64% five star reviews; 17% four star; 5% three star; 4% two star and 10% one star in both conditions of volume. Hereby a five star represents an extremely positive review and a one star an extremely negative one. In total 54 reviews were created. This amount was based on the research of Park and Lee (2007) who discovered that the amount of 27 was experienced as a high amount of content to read. The amount of reviews needed to be doubled, since there were two conditions whereby a high content volume of 27 attribute-value or 27 simple-recommendation reviews were needed. Each review contained a title, name of the reviewer, date, star rating (if this condition was active) and text. The attribute-value and simple-recommendation reviews were created following the examples and explanation of Park and Lee (2007). Hereby an attribute-value review only gave information about the product and its corresponding attributes. A simple-recommendation review recommended the product and only showed feelings of the reviewer. Examples are given below:
Attribute value:
This laptop is very good and the price is good as well. The lighted keyboard is a nice feature and the battery seems to last forever. The SSD as the primary drive is a nice feature because it allows the computer to start quickly and the programs run quickly.
Simple-recommendation:
Highly recommend you can’t go wrong with this. This laptop is incredible. No person could say this was not what they were looking for. I looked around. I’m buyer that reads tons of reviews before buying so I hope this helps someone in their future decision to purchase.
Next to that, since the moderator needs to give every review star ratings, the reviews are divided in five to one stars according to the real F-distribution from the laptop on Amazon.com. For example, a one star review on amazon is transformed to a one star review in the survey. Hereby the different sort (attribute-value or simple-recommendation) of reviews were acknowledged.
Pre-test
compared to a high amount of reviews indicated that they wanted to see more reviews and experienced less information overload than the participants who saw a high amount of reviews. This pre-test was also designed to indicate the duration of the timer. The timer was a countdown and visible in the top of the page were the product information and reviews were displayed. This stimuli was included based on Biehal and Chakravarti (1983) in order to make sure people would continue with the survey and to experience an information overload. Next to that, the timer was necessary for this research in order for people really reading the reviews. However, the 5 minutes duration of the timer was experienced as too long in the pre-test. In order not to lose the attention of people, but still encourage people to read the reviews, we heavily decreased the duration to 35, 25 or 20 seconds, depending on the volume and type condition.
Experimental procedure
People first read a scenario they were looking for an affordable laptop for everyday use. In order to give the respondents a task while reading the information, they had to decide if they would purchase this laptop after reading the description and the reviews. After reading this, they could click on the next page were the timer, product information (description, price, name of the laptop), distribution of stars, stars per review (depending on the condition given) and the reviews were given. The people received beforehand information that even though the timer was finished, people could spend as much time as they wished on the page with description and reviews. In this way, people did read the reviews due to the timer, but could decide for themselves how much time they wanted to spent on the page. After reading this page, people were given the task to indicate to what extent they agreed with the following statements regarding the purchase decision; ‘I would buy this product’ and ‘I would recommend this product to my friends’. They could indicate this based on the Likert scale ranging from ‘strongly disagree’ till ‘strongly disagree. This measurements were based on the literature of Park, Lee and Han (2006). These statements were thereafter called the ‘purchase decision’. After these statements, the respondents were asked to fill in the survey whereby the dependent variables, the covariables and control variables were asked.
Measurements
Independent variables
The independent variables in this research consist out of volume and type of the review. Hereby volume is measured in count with the quantity variable (discussed at the control variables) and with content volume having two different levels depending on the condition. These two conditions are ‘moderate’ with 9 reviews and ‘high’ with 27 reviews displayed. As mentioned before, these amounts are based on the literature of Park and Lee (2007). Next to that, also based on Park and Lee (2007), two different types of review are shown depending on the condition given. This are simple-recommendation and attribute-value. Examples are given in the study design.
Dependent variables
In this research we try to investigate if information overload is caused by the volume and type of reviews. Therefor our dependent variables are based on the original overload literature. This exist out of 4 questions regarding the subjective states and the information overload itself. Based on Jacoby et al., (1974), Malhotra (1982) and more current literature of Park, Lee and Han (2006) and Kwon et al., (2015), we created four items: ‘I am satisfied about my purchase decision’, ‘I am confident about my purchase decision’, ‘I felt confused while reading the reviews’ and ‘I experienced information overload while reading the reviews’. All questions in the survey were asked with the same Likert scale to avoid confusion, ranging from 1 (strongly disagree) until 7 (strongly agree).
Interaction
The interaction term in this research, namely the summary ratings and star rating per review, are given in some of the survey conditions. In total there are eight conditions, whereby interaction term is given at both different levels of volume and type of the reviews. The stars were given per review and a F-distribution was shown below the product description. An example of this F-distribution is given in figure 4.
Covariates
In order to see if there are some latent groups, covariates were added to the survey. Susceptibility to informational influence (.82 alpha). The instrument is measured with four different items based on Bearden et al., (1989) and Chen et al., (2016); ‘To make sure I buy the right product, I often observe what others are buying and using’, ‘If I have little experience with a product, I often ask my friends about the products’, I often consult other people to help choose the best alternative available’ and ‘I frequently gather information from friends or family about a product before I buy’.
Product involvement (Yang and Yun, 2002). The participants had to indicate the level of interest according to 1.’importance’, 2. ‘relevance’, 3. ‘meaning a lot’, 4. ‘excitement’ and 5. ‘appealing’. Online Experience: (Park and Lee, 2007). The participants had to indicate to what level they agreed with the following statements; ‘I often buy a laptop online and ‘I often buy online’.
Attitude towards reviews: (Park and Lee, 2007). The respondents had to indicate their level of agreeing regarding the following statements; ‘When I buy a product online, 1. I always read reviews that are presented on the website, 2, the reviews on the website are helpful for my decision making, 3, the reviews on the website make me confident in my purchase decision, 4. If I don’t read the reviews, I worry about my decision’.
Sources of information: the respondents were asked to indicate on which source of information they based their purchase decision on, namely; product title, description, price, reviews, average rating and distribution of ratings. This had to add up to 100% (Kwon et al., 2015).
Manipulation and control variables
In order to check if the manipulation of positiveness, quality, quantity and type were perceived correctly, participants had to answer several questions in the survey.
For the perceived quality, the instrument based on Lee (2009) was used; ‘The reviews that I read were 1. credible, 2. presented good arguments and 3. provided facts to support their position’.
The perceived quantity also based on Lee (2009); ‘I felt there were too many reviews available’ and ‘I would have preferred to have more reviews to read’.
For the perception or review positiveness based on Park and Lee (2007): 'The reviewers positively evaluate the product’, ‘In general, the reviewers recommend the product’ and ‘The information about the product is objective’.
And the type of review with the instrument of Park and Lee (2007): ‘Each review has sufficient reasons supporting the opinions’, ‘Each review is objective’, ‘Each review is product-relevant’, ‘Each review is rational’ and ‘Each review has information about the product itself’.
according to the condition, it is expected that people perceive type as attribute-value versus simple-recommendation in the intended way.
Since the survey was randomly assigned to different conditions, it was controlled for individual differences. Brand familiarity was controlled due to changing the brand name to brand E. The letter E was chosen since most of the laptop brands don’t start with an E and for example the letter A can be linked to popular computer brands or seen as good quality. Lastly, the demographics will be used as control variables. The variables asked were; age, nationality, gender and education level.
Plan of Analysis
In order to test if the independent variables (volume and type), the moderator (summary ratings) are influencing the dependent variables (information overload and subjective states), a moderated multinomial regression will be performed. In order to perform a regression, several conditions needs to be met; outliers, linear, multicollinearity, homoscedasticity, normal distribution and autocorrelation.
There shouldn’t be significant outliers, this is inspected when cleaning the data and will be controlled for when performing the regression by looking at the case wise diagnostics to see which residuals are more than 3 standard deviations away from the mean.
The relationship between the independent and dependent variables needs to be linear.
There shouldn’t be a case of multicollinearity, otherwise it is hard to interpret the results, since it will not be clear which independent variable causes the variance explained in the dependent (Leeflang, 2014). This can be tested by looking at the tolerance (> .2) and VIF values (< 5) (Leeflang et al., 2015). The variables should be randomly divided to show homoscedasticity. Hereby the error variances are equal for all combinations of the independent variable and the dependent, otherwise the efficiency of the parameters estimates will be lowered (Leeflang et al., 2015). This can be done by looking at the scatterplots of the residuals and by looking at the Levene’s test = homogeneity of variance. When this is not significant, it indicates there is no heteroscedasticity.
The data should have independence of error terms/disturbance term. Otherwise the error terms follows an AutoRegressive (AR) process, also called the autoregression (Leeflang et al., 2015). Hereby the Durbin-Watson for autocorrelation is used. These values range from 0 to 4, whereby 2 means perfectly non-autocorrelation. When there is a small number close to 0, there is a positive autocorrelation and when there is a high number close to 4, there is a negative correlation (Leeflang et al., 2015). Normally, the Durbin-Watson needs to be between 1.5 and 2.5.
Goodness of fit
To look at the overall goodness of fit of the model, we will look at the adjusted R², which is the proportion of total variance in the dependent variable explained by the model (Leeflang et al., 2015) and at the significance of the whole model in the ANOVA test.
Latent Class Regression
Results
The results of the research are described below. First the descriptive of the dataset will be given, followed by the validation and reliability of the measurements. The results of the multinomial moderated regression will be given, concluded by a latent group regression.
After analyzing the control question, eleven people did not fill in the correct answer. Those respondents were excluded from the dataset, since it is not sure how truly they filled in the survey. One question was removed from the dataset, namely the question if participants could report what source of information affected their purchase decision. This question was removed due to incomplete answers.
Descriptive
In total 268 participants filled in the survey, whereby 200 people completed the whole questionnaire and were included in the dataset. Of those people, 99 filled in the survey with recommendation as condition and 90 with attribute. 92 people saw the moderate amount and 97 the high amount of reviews. 99 saw star and 90 without.
The respondents are 53,4% female and 46,6% male with an average age of 34 (SD = 14,74). The education level scaling from no degree (0% of the respondents) to university master (50,8%, mean = university master, SD = 1,192). Their nationalities were mostly Dutch 91,5 %, the other nationalities exist out of; Turkish (3,2%), German (1,6%), French (1,1%), Chinese (1,1%), Swiss (,5%), Bulgarian (,5%) and American (,5%). There were no outliers which needed to be excluded from the dataset.
Scale validation
In order to see if we could factor the measurements, a correlation analysis was performed. After this a reliability analysis is conducted to be certain factor and sum variables could be formed. The manipulation and control checks show if the survey was perceived as intended. We started with the scale validation by recoding three questions, namely; ‘I experienced information overload’, ‘I felt confused while reading the reviews’ and ‘I felt there were too many reviews available’.
Correlation analysis
Factor analysis
Since all items correlate with each other according to the design, a factor analysis is performed including the DV, IV and the covariates in order to increase parsimonious. In order not to clutter the analysis too much, we excluded the manipulation and control variables from the analysis. Performing a factor analysis, the KMO measure of sample adequacy of .722 shows that the variables are probably factoring well, since it is greater than 0,5 which is desirable (Malhotra, 2004). Next to that, the Bartlett’s Test of Sphericity is significant p < 0.00 (1659,120, df 231) which means that the variables are correlated. We used the principal component method with rotation varimax and put the maximum iteration for convergence at 25 and we suppressed the small coefficients below the absolute value of .40.
The eigenvalue of 7 factors is higher than 1, the total explained cumulative variance of 6 factors is higher than 60%, which should be the minimum of total explained cumulative variance, whereby 6 factors each explain more than 5% of the total variance and the scree plot shows 6 factors is most appropriate (appendix IV). Based on these criteria, we will focus on 6 factors. The rotated component matrix is shown in appendix III.
Reliability analysis
The factor subjective states, existing out of satisfaction, confusion and confidence, was not created, since the Cronbach’s alpha is only .626. When deleting the item for confusion; ‘I felt confused while reading the reviews’, the Cronbach’s alpha increases to .772. However, the subjective states will not be measured appropriately anymore. Next to that, it will be more interesting for this research, if subjective states will not be combined to see the different effects. Also ‘I often buy online’ and ‘I have a lot of experience with buying a laptop online’ is not combined to one variable, since it has a low Cronbach’s alpha of .591. Based on Malhotra (2004), this alpha is too low, since it needs to be over .6 to indicate internal reliability. Next to that, all items for the manipulation and control variables can be combined into four variables; quantity, quality, positiveness and type.
Assumptions parametric test
In order to see if the control and manipulation variables are perceived as planned, several statistical tests needs to be conducted. In order to apply an ANOVA, the data needs to be independent, there needs to be a homogeneity of variances and the data should have a normal distribution. Since the surveys are randomized and voluntarily, we accounted for independent data. The normality test of Kolmogorov-Smirnov and Shapiro-Wilk (appendix XX) show that Quantity, Quality and Positiveness don’t have a normal distribution of the data (p <.05), while Type has (p > .05). These results are congruent with the Q-Q plots. Therefore an ANOVA will be conducted for type and a Kruskall Wallis for quantity, quality and positiveness. To see if the data has a homogeneity of variances, we conducted a Levene’s test of normality for the type variable. It shows that the test is not significant (table 3), which means variance is equal between groups.
Control and manipulation checks
In order to see if the dataset is controlled and manipulated in the right way, the mean values need to be compared. Since the data is not normally distributed for the variables quantity, quality and positiveness, a nonparametric test needs to be applied. Conducting the Kruskall Wallis test for the manipulation checks, the results are described in table 2. For the degree of positiveness, it was clear that people experienced the same amount of positiveness among the conditions of summary rating
Variable: Eigenvalue for Factor Cronbach’s Alpha
Factor Involvement 4,626 .830
Factor Purchase decision 1,940 .888
Attitude towards reviews ,806
Susceptibility towards Interpersonal Influence .834
Quantity .746
Quality .796
Positiveness .699
Type .776
and type, but not for volume. This could be reasoned by people perceiving the type ‘simple-recommendation’ as more positive than ‘attribute-value’, since it is easier to progress. Next to that, although the distribution of positive reviews was the same among all conditions, it could be that in the summary rating condition, the perceived positiveness became more salient due to the amount of stars. Quality is not significant different among conditions of summary ratings and volume, but it is among the condition of type. This was expected and intended, since attribute-value is perceived as higher quality reviews than simple-recommendation. Quantity only differs among the condition of volume, which is logical, since an increase in reviews stands for an increase in volume. Lastly, type is manipulated correctly (table 3), since this differs significantly among the conditions.
Quantity will be included in further analyses to control for the volume count. Quantity is not reflected in the independent variable volume, since volume is reflecting information content instead of count.
Type t -6,484 df 187 sig ,000 Levene F ,011 signifance ,918
Table 2:Kruskall Wallis test. Table 3: T-test
Next to that, education and nationality will not be included in further analysis, since most of the respondents were Dutch and got a master degree at the university. The other variables were mean centered for interpretation purposes for the moderation regression. From here on the variables ‘I often buy online’ and ‘I have a lot of experience with buying a laptop online’ will be called ‘Buy online’ and ‘Experience laptop’. Lastly the interaction terms (volume * summary and type * summary) were created.
Regression
To test the hypothesis, a multiple regression is performed. Four different models will be conducted, each with a different dependent variable. It is tested if volume, type and the interaction significantly influence the information overload, satisfaction, confusion and confidence. The results of the regression for the hypothesis will be explained in the conclusion section.
Conditions for multiple regression
Although the risk for multicollinearity is already a bit reduced due to looking at the correlation table, performing a factor analysis and creating new variables, we performed a multicollinearity analysis by looking at the Variance Inflation Factors and the Tolerance level. Since most of the values of VIF were below 2 and therefore congruent with the < 5 level and all tolerance levels at least .313 and therefore congruent with the > .2 level (Leeflang et al., 2015), it can be concluded there is no multicollinearity. Only the interaction terms were having a VIF of 3 and a tolerance level close to .3.
Model Information Overload
Satisfaction Confident Confusion Old Confusion New Breusch-Pagan LM 16,133 21,515 23,237 39,615 23,237 Sig ,305 ,089 ,057 ,000 ,057 Koenker LM 17,217 19,196 20,326 36,370 20,326 Sig ,245 ,158 ,120 ,001 ,120
Table 4: homoscedasticity test
Linearity is shown in the figures above, whereby all r2 are 0 and therefore linear.
In order to see if the standardized residuals are normally distributed, a histogram and a P-Plot are conducted. All plots show a non-normal distributed pattern, except the first histogram and P-plot of information overload. Therefore we bootstrapped all the models, except the model for information overload.
In order to be completely sure the standardized residuals are normally distributed, a Kolmogorov-Smirnov and a Shapiro-Wilk test of Normality are conducted for every model. As can be seen in table 5, the values for information overload are non-significant, which means the standardized residuals are normally distributed, but the other three models are significant and therefore non normally distributed.
Model Information
Overload
Satisfaction Confident Confusion
Kolmogorov-Smirnov
Statistic .051 ,092 ,091 ,106
Significance ,200 ,001 ,001 ,000
Shapiro-Wilk Statistic ,989 ,960 ,966 ,964
Df ,189 189 189 189
Significance ,144 ,000 ,000 ,000
Table 5: normality test
In order not to have wrong estimation of the variance of effects, we conducted a Durbin-Watson test to look for autocorrelation. In the table 6 the values for the Durbin-Watson test per model are given. Looking at the table, it shows that there is no autocorrelation since all values are between 1.5 and 2.5.
Model Information Overload Satisfaction Confident Confusion
Durbin-Watson 1,741 1,666 1,941 2,089
Table 6: Durbin-Watson
The Cook’s distance shows there is not an overly influential case interfering with the regression. When greater than 1, it needs to be excluded, but the maximum value that appeared in all four models is at maximum for information overload; ,05114, satisfaction: ,05316, confident: ,05434 and for confusion ,13330.
Model
After adjusting the models as a result for the multicollinearity, normality, autocorrelation and heteroscedasticity, four different models were regressed. In order to test the hypothesis, four different dependent variables were analyzed. This are the subjective states existing out of confidence, satisfaction and confusion and the information overload. Since we also performed a latent class regression in our further analysis, a lot of additional instruments were measured in the questionnaire. Therefore these variables will be controlled in the hierarchical regression. Next to those variables, the variable volume count will also be included in the model as a control variable. Reason for this is the expected awareness effect influencing information overload and the subjective states. The regressions are hierarchical models in the sense that they existed of three different models. The first model included the control variables age and gender, susceptibility, purchase decision, attitude, involvement, buy online, experience laptop and quantity (volume count). The second model included the main effect volume (content), type and star summary and the last model included the interaction effects. We will include the main effect in the last model as well, otherwise it is possible that the main effect is omitted in the product term. This gave us the following formula:
Y, = α + β1x1 + β2x2 + β3x3 + β4(x1*x3) + β5(x2*x3) + ε
Controlled for: age, gender, involvement, buy online, experience laptop, attitude, purchase decision, susceptibility and quantity count.
Information overload
The ANOVA table tells us that the model in total is significant (F= 14,937, p <,000 ), model 2 F = 11,778, p and model 3 F = 10,278.
*p <.05, **p<.01
Results: Looking at the regression model with dependent variable information overload, it shows that both quantity (B = ,600, p < .001) and volume (B = -,196 p, < ,05) have a significant influence on information overload. Hereby we can accept our H4 which states that if volume increases from moderate to a high level, it will lead to information overload. So when people see more reviews to progress, they will be overloaded with information. Type however doesn’t have a significant influence on information overload, rejecting our H1, stating that when changing from simple-recommendation to attribute-value leads earlier to information overload. Lastly, our moderator summary ratings star is not significantly influencing the feeling of information overload. So when people see a lot of reviews
or attribute value reviews, including summary ratings will not lower the feeling of information overload. This leads us to rejecting our H5.
Satisfaction
Looking at the coefficients, none of the variables is influencing satisfaction and none of the models is significant. None of the models is significant, model 1: F = 1,427, sig ,180, model 2: F = 1,149, sig ,324, model 3: F = ,983, sig ,473.
*p <.05, **p<.01
Results: The model of satisfaction shows no significantly relationships influencing the subjective state of satisfaction. This means that our H2a, H4a and H6a will be rejected, showing that our independent variables are not influencing the satisfaction level of the consumers.
Confidence
Looking at the coefficients, only quantity volume is significantly influencing the confusion.
Model 1 and 2 are significant, model 1: F 2,357, p = 015 and model 2: F 1,864, p = 042. Model 3 is not: F 1,725, and p= ,054.
Results: Looking at the model of the subjective state confidence, we see Age only being significant in the models without the interaction effects (B -,162, p <,05). Hereby it shows that if people become older, this has a negative effect on the confidence level of consumers. However, this significant effect disappears once the interaction effects are included. Gender, SII and quantity are significant influencing confidence among all models. Interpreting gender (B -,168, p <,05), it shows that when gender goes from female to male, this has a negative effect on confidence. When SII (B ,196, p <,05) goes up, this has a positive influence on confidence. Concluding, quantity (B -,199, p <,05) has a negative effect on confidence. So when people see a higher amount of reviews, they will feel less confident. The model shows that our control variables are more influencing confidence than our independent variables are. This means that our H2b, H4b and H6b are rejected. Volume doesn’t have significant influence on confidence, next to that, the type will not differ feeling of confidence and lastly, including summary ratings will also not differ the feeling of confidence.
Confusion
All models are significant, model 1: F 3,929, p = ,000, model 2: F 3,780, p=, 000 and model 3, F 3,264, and p =,000.
*p <.05, **p<.01
Results: Controlling for several variables, it shows that in our first model, age is significant influencing confusion (B ,150, p <,05). How older people get, how more confused they feel when making a purchase decision. However, this effect diminish when including the main and interaction variables. Purchase decision is significantly influencing (B, 283, p <,000) the feeling of confusion. Which is interesting, since a higher purchase decision means that people will more confused about their decision. Next to that, volume is only significant in the model without the interaction effects (B,173 p <,05). This means that when this interaction effect is not there, so with other words, when there is not a high volume and stars included, volume has a negative effect on confusion. So when volume increases, this will make consumers feel less confused. This makes us accept H2c, but reject H4c and H6c, since type and summary are not influencing confusion significantly.
Model 2 Model 3 Model 4 Step and variables
B Bias SE B B Bias SE Stan.
Latent Group Regression
In order to identity latent segments based on the attributes volume, type, star, volume*star and type*star influencing respectively; information overload, satisfaction, confusion and confidence, we performed a latent group regression. The dependent variables were treated as ordinal and the independent as nominal.
All the covariates: susceptibility, attitude, involvement, experience and purchase decision and control variables: quantity, age and gender were added in the second model. In order to decide upon the amount of classes, the information criteria and number of parameters will be used.
The model with the lowest criteria will be selected. We requested 2 - 8 different class models. After estimating, the following summary is displayed. Looking at the information criteria showed in figure 9, the first model with 2 classes is most appropriate.
Information overload
Considering the information criteria (BIC, IAC, AIC3, CAIC) whereby BIC and CAIC are most appropriate for large sample sizes, the first model is chosen with 2 classes with a R2 of ,5360. The LL increases with every class added. As the table below shows, volume has a significant effect (p<,05) on information overload, whereby both classes respond negatively (class 1 = -,5021 and class 2 =,5547). This effect is not significantly different among classes (wald(=) ,0097, p = ,92). Next to that, gender has a significant influence (wald = 6,1348, p <,05), when going from female to male, information overload lowers for class 1, but rises for class 2. Involvement (3,9313, p <,05) and experience online (6,3453, p <,05) have the same effect as gender, but quantity is higher for information overload (6,3453, p <,05).
Class 1 Class 2 Wald p-value Wald(=) p-value mean
Segment Size ,5878 ,4122
R2 ,1389 ,2307
Intercept Strongly disagree -2,5972 4,40 59,41 ,000 14,29 ,027 ,2891
Disagree -,6115 4,7385 1,5939
Somewhat disagree
0,0852 3,6794 1,5669
Neither agree nor disagree
-0,0262 1,0667 0,4243
Agree 1,7785 -2,2713 0,1091 Strongly agree 0,3596 -7,6846 -2,9565 Type -0,0196 ,5412 1,3327 ,51 1,2601 ,26 ,2116 Volume -,5021 -,5547 7,8599 ,020 ,0097 ,92 -,5238 Star -,2177 1,1352 3,9869 ,14 3,7927 ,051 ,3400 Type*star -,2940 -,1798 1,3659 ,51 ,0308 ,86 -,2469 Volume*Star ,2949 -,1628 1,3636 ,51 ,4282 ,51 ,1063 Covariates Intercept 3,3228 -3,3228 ,7627 ,38 Age ,0489 -,0489 3,2757 ,070 Gender -2,7924 2,7924 6,1348 ,013 Purchase -,0727 ,0727 ,1221 ,73 Involvement -1,0746 1,0746 3,9313 ,047 SII -,8296 ,8296 4,0979 ,043 Attitude ,6739 -,6739 3,1001 ,078 Oftenonline ,1204 -,1204 ,2561 ,61 Experience online -1,1908 1,1908 6,3453 ,012 Quantity 3,0803 -3,0803 6,3453 ,0042
Looking at the profile of the classes, it appears that class one is the biggest (58,78%) and they also experience the highest information overload (mean = 4,5312) whereby class 2 experience a low information overload (2,1137). Both classes are equally of age (class 1 = 35 and class 2 = 31 years old). Class one has an equally divided gender, but class 2 has more females (62%) than males (38%). Both classes purchase almost equally (4,57 vs 4,52), have the same amount of involvement (4,76vs 5,10), susceptibility (4,75 vs 4,91), attitude (5,23 vs 5,04), experience online (5,26 vs 5,63), experience laptop (3,01 vs 3,86). But they differ highly on quantity, whereby class one is 4,52 and class 2 2,12. So looking at the profiles, it can be concluded that class one experience high information overload and therefore also a high quantity. Class 2 experience a very low information overload and therefore a low quantity, next to that, they have more females than males in their group.
Satisfaction
In order to avoid clutter of information, the other tables of the latent regressions will be given in the appendixes and the main points will be described here.
Looking at the information criteria, there is again chosen for 2-class regression. Class 1 has a R-square of ,0173 and size 58,96% and Class 2, R-square of ,2125 and size 41,04%.
Looking at the parameters, it shows that only the volume has a significant influence on satisfaction (wald = 10,2632, p <,01), class 1 a negative influence (-,2319) and class 2 a positive (2,0002). This differs significantly among classes (wald (=) 10,1266 p<,01). Of the covariates, purchase, SII, attitude and quantity are significantly influencing information overload.
Purchase SII Attitude Quantity Class 1 -2,2743 -1,0052 2,2663 -1,2007 Class 2 2,2743 1,0052 -2,2663 1,2007 Wald 5,2436 4,4623 5,4437 6,1681
Class 1 experienced satisfaction with a mean of 4,9631, has an average age of 33 years, equally divided among male and female, mean purchase of 3,9554, SII of 4,5608, attitude of 5,2874 and quantity of 3,0194.
Class 2 experiencing information overload a slightly higher satisfaction of 5,4274, has an average age of 35 years, more female than male (56%), purchase intention of 5,4064SII of 5,1804, attitude of 4,9478 and a quantity of 4,2627.
Confusion
Looking at the information criteria, there is chosen for a 2-class regression. Class 1 has a R-square of ,0734, size 70,37% and Class a R-square of 2,4474 and size 29,63%.
Looking at the parameters, it shows that type has a significant influence (wald 9,6533, p <,01) whereby class 1 has a negative effect (-,2366) and class 2 a positive effect (2,2170). This also differs significantly (wald(+) 8,6799, p <,05). Of the covariates, almost all are significantly influencing confusion.
Age Purchase SII Experience Laptop Class 1 -,0609 -1,4878 -,4571 -,5295 Class 2 ,0609 1,4878 ,4571 ,5295 Wald 8,5366 9,3943 5,4729 6,5842 p-value ,0035 ,0022 ,019 ,010
Class 1 experienced confusion with a mean of 3,0082, has an average age of 31 years, slightly more female (53%), mean purchase of 4,1115, SII of 4,6078 experience laptop 3,5220.
Class 2 experiencing information overload a slightly higher confusion of 4,2503, has an average age of 40 years, more female than male (56%), purchase intention of 5,5979, SII of 5,30894, experience laptop 3,9847.
Confidence
Looking at the information criteria, there is chosen for a 2-class regression. Class 1 has a R-square of ,0426 and size 52,87%. Class 2 a R-square of ,3051 and size 47,13%.
Looking at the parameters, it shows that none of the independent variables significantly influence confidence. Of the covariates, age, gender, purchase decision, experience online and quantity significantly influence.
Age Gender Purchase Experience online Quantity Class 1 -,0580 2,0179 -1,1381 -1,0917 ,7056 Class 2 ,0580 -2,0179 1,1381 1,0917 -,7056 Wald 5,6777 5,8553 6,2355 6,0901 4,8991 p-value ,017 ,016 ,013 ,014 ,027
Discussion
In this section, the implications of the analyses will be discussed as well as the managerial implications. This study tried to investigate the research question; does review volume and type lead to information overload? using the variables of review type and volume content influencing on information overload and its subjective states. The research is based on the original information overload literature as well as specific review volume content causing information overload (Park and Lee, 2007 ; Kwon et al., 2015).
Theoretical Implications
Main and interaction effects regression
The study investigated if changes of type and volume content influences information overload. Reason for this is that the type simple-recommendation is perceived easier to progress than attribute-value reviews (Park and Lee, 2007). It was also expected that changing the level of volume content from moderate to high, would lead to information overload. This hypothesis was based on the original literature of information overload. From the analysis, it appears that type of review has not a significant influence on information overload. However, volume content does. When increasing from 9 reviews to 27, consumers feel more overloaded with information. Next to that, the control variable volume count, reflected in the variable quantity, also shows a significant influence on information overload. This has even a bigger effect (.600) than volume does (.313). This is not in line with the expectations, since it expected that volume count would lead to less information overload, due to the awareness effect.
