Why do consumers purchase products with low average ratings in the online market?

(1)

Why do consumers purchase products with low

average ratings in the online market?

(2)

Why do consumers purchase products with low average

ratings in the online market?

Saskia van Dijk s1960393

Faculty of Economics and Business Master Thesis Marketing Intelligence June 22, 2015

Zonnelaan 30-76 9742BM Groningen 0625111251

s.i.van.dijk@student.rug.nl Supervisor: Dr. Hans Risselada

(3)

Management Summary

(4)

(5)

Preface

I am interested in this topic on the eect of consumer reviews on sales since I do a lot of online shopping myself. I have read many consumer reviews and I always take the average rating of a product into account. However, I noticed that products with a low average rating are apparently still purchased, as can be seen at the recent posted negative reviews on the product. I wondered why this was the case, therefore, I have chosen this research question: "Why do consumers purchase products with low average ratings in the online market?" This study was a good experience in doing research in the eld of Marketing Intelligence and I would like to thank my supervisor, Hans Risselada for his support and suggestions.

Abstract

(6)

1 Introduction

The sales of products on the internet is becoming more and more important and nowadays, almost every company has an online shop. Furthermore, the number of companies that sell their products only online is also increasing (e.g. Wehkamp). With the increasing use of the internet for purchasing consumer goods, also a huge increase in user generated content (UGC) is seen (Akehurst, 2009), this increase in the usage of UGC is partly because it is free to obtain for consumers in most cases (Krumm, Davies and Narayanaswami, 2008). UGC is generated by consumers who voluntary share data, information and media that become accessible for other consumers, usually on the web (Krumm, Davies and Narayanaswami, 2008). Due to this rapid increase of UGC, consumers do not only use traditional mouth (WoM), they also use electronic word-of-mouth. So, it is important for consumers that UGC can also be found at online retailer websites (e.g. Amazon.com, Bol.com). An example of electronic WoM are online customer reviews and they can be dened as peer-generated product evaluations posted on company or third party websites (Mudambi and Schu, 2010, p186). Online customer reviews are an important resource for consumers to compare the quality of products. Since online consumer reviews are believed to signicantly aect consumers' purchasing decisions (Zhu et al, 2010), rms proactively stimulate their consumers to use online word-of-mouth (Zhu, Feng and Xiaoquan Zhang, 2010). Therefore, when evaluating a product, consumers have to process not only the rm-generated content and information about the product itself (e.g. price), but, consumers have to process user generated content about the product too.

(9)

Mengis, 2004). Synonyms of information overload are cognitive overload and knowledge overload (Eppler and Mengis, 2004).

The eect of reviews on sales of a product has been studied many times, for example by Chevalier and Mayzlin (2006) and Zhu and Zhang (2010). According to Mudambi and Schu (2010) the eect of reviews on sales depends on the perceived helpfulness of a review. The perceived helpfulness of a review is inuenced by the review complexity (Koratis, García-Bariocanal and Sánchez-Alonso, 2012) and review complexity is inuenced by the length of a review. Furthermore, also according to Blal and Sturman (2014) and Wang, Mai and Chiang (2014) the sales of a product are inuenced by the volume (review length) and valence of a review. Besides this, the eect of information overload due to the number of choices of a product and due to the amount of rm-generated information about a product has been studied (Eppler and Mengis, 2004). Thus, the eect of reviews on sales and the eect of information overload due to the many alternatives for one product and due to all rm-generated information and how this inuences their decision making process, has been examined in the literature. However, little research has been done on information overload due to the volume of consumer reviews and the review complexity. When making online purchase decisions consumers should be able to process large amounts of information (Aljukhadar, Senecal and Daoust, 2012). When consumers are con-fronted with high levels of information load, this can result in cognitive fatigue and confusion, due to their limited capacity to process information (Aljukhadar et al., 2012). Information overload is mainly caused by the quantity, frequency, intensity and quality of the information (Eppler and Mengis, 2004). Hence, if there are many reviews and/or reviews are very long, consumers might suer from information overload.

(10)

average rating. Hence, whether there is a positive interaction eect between the number of reviews and the average rating of a product. Also, whether there is a positive interaction eect between the average number of characters of a review with the average rating of a product. Thus, there is examined whether more information given to the consumer is good, it positively inuences the decision made by the consumer and increases the impact of the average rating of a product.

Furthermore, in this study there is going to be examined whether too many and too complex consumer reviews causes information overload and therefore inu-ences the ability of the consumers to make the optimal decision, since consumers alter their decision making process when they are overloaded with information (Aljukhadar et al., 2012). Also, according to Park, Lee and Han (2006) many consumer reviews can cause information overload, which results in adverse judg-mental decision making by the consumer. According to Chevalier and Mayzlin (2006) consumers do not only rely on the summary statistics, consumers do read the reviews. Hence, consumers do want to check on what arguments the average rating is based on. Thus, consumers do not rely on a simple cue, also, according to Park et al. (2006) the product attitude of a consumers does not change with a simple cue. There is going to be examined whether an average rating based on for example 500 reviews is trusted less than a rating based on for example 50 reviews, since now it is harder for a consumer to read all the reviews and examine on which factors the rating of consumers is based, because the consumer is overloaded with information. Furthermore, there is going to be examined whether an average rating based on very long reviews is trusted less than an average rating based on reviews of moderate length. Thus, whether the interaction eect between the average rating and the number of reviews and the interaction eect between the average rating and the average length of the reviews becomes negative when there is an overload of information. Hence, whether after a certain amount of information provided to the consumer it has a negative eect on the decision made by the consumer, hence, it has a negative eect on the impact of the average rating of a product.

(11)

question why do consumers purchase tablets which nevertheless have low average ratings, regression analysis is used. As dependent variable the sales of the tablets is used and as independent variables the number of reviews, the average number of characters per review, the average rating of the product given by consumers, the price of the tablet and the features of the tablet are used.

The results partly support the hypothesis that the average rating of a products positively aects the sales of this product. However, they do not support the hypothesis of information overload due to the number of reviews and the average number of characters per review and they do not support the hypothesis that due to information overload the impact of the average rating decreases. The main eects and the interaction eects are not signicant. Therefore, based on this study the eect of information overload on sales cannot be shown.

A rm might wonder why a comparable product, with a lower average rating, can be preferred over their product. In this study this question is not answered for tablets since the hypotheses in this study are not supported by the data. Further research can examine whether these eects might exist for other prod-ucts or product categories, such as for experience goods, goods that need to be experienced before the quality of the product can be assessed (e.g. a hotel room). A tablet is a search good, a good for which the consumer had the ability to obtain information about the quality of the product before purchasing the product. The results of this study does show that for a search good having a high average rating has a positive eect on the sales of this product. Only, it does not matter on how many reviews this average rating is based. For rms selling search goods this implies that they should use resources to contact con-sumers to give an average rating instead of using resources to contact concon-sumers to write an entire reviews, this will increase the response rate of the contacted consumers.

2 Theoretical Framework

(12)

down to nd the consumer reviews, next to a lot of other information (e.g. product information, supplier information). According to Chevalier and Mayzlin (2006) consumers do read reviews, hence, they do not only rely on the summary statistics. However, the study of Chevalier and Mayzlin (2006) was based on an average of 60 reviews per product and nowadays, products have on average more reviews. However, for a consumer to read 60 reviews is already a lot of information to process. Hence, there is no dierence expected between a study based on an average of more than 60 reviews. Hence, when purchasing a product a lot of information should be processed by the consumer.

2.1 Variables

In this analysis the sales, product specications and review information of tablets (e.g. an Ipad) on Amazon.com are used. The sales will be used as a dependent variable. First, the average rating is used as an independent vari-able. The average rating of a product is the average number of stars given to the product by all consumers who have written a review. Secondly, the number of customer reviews and the review length are used as independent variables. Thirdly, the features and the price of the product are used as independent vari-ables. The features of a tablet used in this analysis are the brand of the tablet and the storage size of the tablet (e.g. 16GB, 32GB).

2.1.1 Average rating

(13)

that the average rating of a product given by consumers in reviews signicantly inuences sales (Chevalier and Mayzlin, 2006; Zhang and Dellarocas, 2006). When a product has a high average rating this has a positive eect on the sales of this product. It is expected that the average rating of tablets has a signicant eect on sales. First, it is examined whether there is a main eect of the average rating of a tablet on the sales of tablets. If there is no main eect of the average rating on sales, it is less likely that there are interaction eects with the average rating that inuences the sales. Since, if there is no main impact of the average rating on sales then the impact of the average rating also cannot be moderated by the average length of a review and the number of reviews of a product. Therefore, the following is hypothesized:

Hypothesis 1:

The average rating of a product given by consumers has a positive eect on the sales of this product.

2.1.2 Review length and number of reviews

There are several studies that have shown a positive relationship between sales performance and volume of electronic WOM for several product categories (Blal and Sturman, 2014). "The volume dimension of eWOM is often quantied as the number of reviews available or the length of the posted reviews" (Blal and Sturman, 2014, p366). Hence, the number of reviews of a product and the average length of the reviews of a product are used in the analysis to capture the eect of the volume dimension of eWOM. The review length is also chosen to capture the review complexity. Review complexity is how easy or hard it is for the consumer to read and understand the reviews (Consti-Ramsden, Durkin & Walker, 2012). The length of a review inuences the complexity of a review, since a longer review contains more information and most likely more arguments. Therefore, it is harder to read and understand a longer review. A review of just one sentence, like "the product is bad" is simpler to understand than a longer story about the product and why it is good or bad.

(14)

ratings to nd the useful ones. When there is no information overload consumers are able to process all information given and are more likely to take the average rating into account, since the trustfulness of this average rating could be veried by the consumer. When consumers are overloaded with information they alter their decision-making process (Aljukhadar et al., 2012), for example to change the sorting tool used by the consumer and to use other sorting tools than to sort on the average rating. This is not logical, however due to information overload consumers don't act rational, information overload due to the number of reviews can result in confusion, cognitive strain and other dysfunctional consequences (Park et al,. 2006) Which results in purchasing a product that is ranked as the most purchased product on the website or purchasing the most recommended product by the website itself for a reason that is not known by the consumer or a consumer can sort the products on price and purchase the cheapest product. Due to the use of these alternatives, the impact of the average rating of the product decreases.

(15)

that this is also due to a positive eect on the reliability of the average rating. This is since a review that is more extensive is more likely to contain arguments that are helpful for the consumer. In contrast, when a consumer review is very short, the reason for the rating given to the product is most likely not explained and argumented well, so, the review is less useful for other consumers. In that case it is harder for a consumer to decide for himself whether he agrees with the rating or not, based on arguments, hence the average rating is less taken into account. Thus, the interaction eect between the average length of a review and the average rating is positive. However, after a certain average length of a review the interaction eect is expected to become negative. Due to the information overload the impact of the average rating is decreased. So the following is hypothesized:

Hypothesis 2:

First, an increase in the average number of characters of a review increases the eect of the average rating, after a certain number of characters of a review, this eect decreases.

(16)

decisions.

First the number of reviews is expected to increase the impact of the average rating. When there are a 100 consumers who have reviewed a product the av-erage rating becomes more reliable than when there are two consumers who have reviewed the product. Hence, the interaction eect between the number of reviews and the average rating is positive. After a certain number of reviews the interaction eect is expected to become negative. Due to the information overload the average rating is less used by consumers. So the following is hy-pothesized:

Hypothesis 3:

First, an increase in the number of reviews increases the eect of the average rating, after a certain number of reviews, this eect decreases.

2.1.3 Features and price

Important factors that inuence the sales of tablets are the features and the price of the tablet. Therefore the eects of these variables should be controlled for in the regression.

2.2 Conceptual Model

In gure 1 the conceptual model can be seen.

(17)

3 Methodology

3.1 Data

For the analysis in this study the data given by Wang, Mai and Chiang (2014) is used. This data contains purchase information of tablets from Amazon.com for a period of 24 weeks, starting from February 1, 2012. The data set contains information about 2163 dierent tablets. Wang, Mai and Chiang (2014) exten-sively discuss the variables of the data. For the analysis the following variables of the data set are used: the sales rank, the price, the number of reviews of a tablet, the brand of the tablet and the memory size of the tablet. In the data all the reviews and their ratings of the dierent products are given. However, the average rating and the average number of characters per review are also required for the analysis. Therefore, using this data the average rating per week per product and the average number of characters of the reviews per week per product are computed.

(18)

Table 1: Descriptive Statistics week 1

Variable N Mean SD

Sales rank 742 8.2 2.1

Number of reviews 2163 13.1 285.9

Average number of characters 2163 245 589

Average rating 503 3.3 1.1

3.2 Analysis Method

(19)

are included in Model 1, the interaction eect between the average number of characters per review and the average rating and the interaction eect between the squared average number of characters per review and the average rating are included in Model 1. Lastly, the price and the features of the tablet are included in Model 1. Such that they can be controlled for.

There is most likely an endogeneity issue with the number of reviews. If the sales of a product increases, it is likely that the number of reviews also increases. If a tablet is hardly sold, the probability of consumers writing a review is also very low. To solve this issue, the copula approach is used (Park and Gupta, 2012). This means that an extra regressor is included in the model for every variable that is expected to be endogenous. This extra regressor is given by the following formula:

Φ−1(H(X)),

where Φ(.) is the normal density with mean zero and variance σ2

and H(X)

is the empirical distribution of the endogenous regressor X. For Model 1 this implies including one extra regressor for the number of reviews and the number of reviews square and one extra regressor for the interaction term of the number of reviews and the average rating and the interaction term of the number of reviews square and the average rating, which results in Model 2. Only two extra regressors have to be included since a main and a squared term have the same empirical distribution. When a copula regressor is signicant, the regressor wherefore it was included is indeed endogenous. However, sometimes when a copula regressor is not signicant, it does lter the eect of the supposed endogenous regressor. This would also suggest that the regressor is indeed endogenous.

(20)

to run a separate regression for all 24 weeks. Between week 1 and week 12 and between week 12 and week 24 there is a signicant increase in the total number of reviews. Therefore, a regression has been done for week 1, week 12 and week 24. Furthermore, since there is non-normality, bootstrap is used. Taking all this into account the following model is formulated to test the hypotheses:

Sales−Ranki= αi+ βi1P ricei+βi2Storagei+ βi3Brand1i+

βi4Brand2i+ βi5Brand3i+ βi6Avg−Ratingi+ βi7N um−Chari+

βi8N um−Chari2+ βi9Avg−Rating ∗ N um−Chari+

βi10Avg−Rating ∗ N um−Chari2+ βi11N um−Reviewsi+

βi12N um−Reviews2i+ βi13Avg−Rating ∗ N um−Reviewsi+

βi14Avg−Rating ∗ N um−Reviews2i+ ui, (1)

where Sales is the log of the sales rank of the tablet, P rice is the list price of the tablet the retail price suggested by the manufacturer in dollars, Storage is the storage size of the tablet, Brand is the brand of the tablet, where 1 is Samsung, 2 is Apple, 3 is HP and the reference category is all other brands, Avg−Rating

is the average rating of the tablet, Num−Reviews is the number of reviews of

the tablet, Num−Char is the average number of characters per review of the

tablet, u is the disturbance term and i is the week number where 1 is week 1, 2 is week 12 and 3 is week 24. To account for endogeneity the copula approach is used, therefore, extra regressors are included in the model. So the following model will be used:

βi4Brand2i+ βi5Brand3i+ βi6Avg−Ratingi+ βi7N um−Chari+

βi8N um−Chari2+ βi9Avg−Rating ∗ N um−Chari+

βi10Avg−Rating ∗ N um−Char2i + βi11N um−Reviewsi+ βi12N um−Reviews2i+

βi13Avg−Rating ∗ N um−Reviewsi+ βi14Avg−Rating ∗ N um−Reviews2i+

βi15Copula1i+ βi16Copula2i+ ui, (2)

where Copula1is the copula term included for Num−Reviewsand Num−Reviews2

and Copula2is the term included for Avg−Rating∗N um−Reviewsand Avg−Rating∗

(21)

4 Results

Firstly, a regression of Model 1 and Model 2 has been done. Secondly, due to multicollinearity issues with the squared and interaction eects, a third model is regressed. This is Model 3, which only takes the number of reviews, the average rating and the control variables into account. Hence, Model 3 is nested in Model 1. Another model is regressed, Model 4, which controls for endogeneity with the copula approach in Model 3. Hence, Model 4 is nested in Model 2. Since the results of this models are still inconclusive for the average rating, another model to assess the signicance of the average rating is performed, Model 5. In Model 5, the average rating is included in the model as a categorical variable. This approach was chosen such that in this model the missing values of the average rating and thus, also the zero's for the number of reviews and average number of characters per review, are taken into account. In Table 2 an overview of the models used can be seen.

Table 2: Overview of the models

Model 1 Model with all variables not controlling for endogeneity Model 2 Model with all variables controlling for endogeneity

Model 3 Model with the main eect of the number of reviews not controlling for endogeneity Model 4 Model with the main eect of the number of reviews controlling for endogeneity Model 5 Model with the average rating as a categorical variable

4.1 Full model

(22)

correlation with sales and therefore a negative correlation with the sales rank, the higher a product is ranked the better, since a tablet is sold more if it is ranked rst than when it is ranked second. Being ranked rst can be seen as a higher sales rank than being ranked second, however, the number is lower, so for example the average rating of a tablet is expected to negatively aect the number of the sales rank. The same negative relationship was expected for the number of reviews of a product and for the average number of characters per review. The brand of the tablet has a signicant correlation with the sales of the tablets. For Apple and Samsung the eect on the sales rank is negative, for HP and other brands the eect on the sales rank is positive. So, tablets of Apple and Samsung are expected to be preferred. The storage size of the tablet has no signicant correlation with the sales of the tablets. The price of the tablet has a positive correlation with sales rank, the higher the price of the tablet the higher the sales rank, hence, the lower the sales.

Table 3: Correlation matrix (p-value)

Variable Sales Rank

Number of Reviews -.202 (.000) Number of Characters -.383 (.000) Average Rating -.290 (.000) Brand Apple -.264 (.000) Brand Samsung -.085 (.021) Brand HP .077 (.037) Brand other .166 (.000) Storage size .000 (.999) Price .408 (.000)

In Table 4 one can see the output of the regression of Model 1 and Model 2. To examine whether Hypothesis 1 holds, the average rating in the models is examined. It can be seen that Avg_Rating is not signicant in Model 1 and Model 2.

(23)

Hence, Hypothesis 2 is not supported in the data, the average number of char-acters per review does not moderate the eect of the average rating.

Table 4: Regression output for Model 1 and Model 2

Model 1 Model 2

Variable beta std.error p-value beta std.error p-value

Constant 8.701*** 2.058 .001 11.866*** 2.375 .001 Brand Samsung .408 1.367 .705 .627 1.396 .665 Brand Apple -2.580*** .294 .001 -2.180*** .338 .001 Brand HP -.749 .861 .259 -.479 .725 .433 Storage -.002 .001 .187 -.001 .001 .352 Price .001*** .000 .001 .001*** .000 .003 Avg Rating -.286 .550 .544 -.277 .536 .552 Num Reviews .005 .030 .832 .042 .033 .115

Num Reviews square .000 .000 .209 1.765E-05 .000 .814

Avg Rating*Num Reviews -.007 .008 .294 -.011* .008 .092

Avg Rating*Num Reviews2 _-1.876E-05 _{3.691E-5 .346} _-3.864E-06 _{3.275E-5 .838}

Num_ Char .000 .005 .945 .001 .004 .753

Num Char square -5.593E-07 2.563E-6 .723 -7.746E-07 2.216E-6 .583

Avg Rating*Num Char 6.609E-05 .001 .934 -1.939E-05 .001 .979

Avg Rating*Num Char2 _1.726E-07 _7.069E-7 ₆₇₂ _1.737E-07 _{6.200E-7 .638}

Copula 1 -2.454 4.423 .598

Copula 2 .889 4.135 .836

Adjusted R2 _.629 _.658

* Signicance at the 10% level. ** Signicance at the 5% level. *** Signicance at the 1% level.

To test Hypothesis 3, the average number of reviews, the average number of reviews squared and both interaction terms with the average rating are exam-ined. It can be seen that without controlling for endogeneity Num_Reviews, Num_Reviews_sq, Avg_Rating*Num_Reviews and

Ag_Rating*Num_Reviews_sq are not signicant (p-value > 0.1). Also, when controlling for endogeneity Num_Reviews, Num_Reviews_sq,

(24)

3 is not supported in the data, the number of reviews does not moderate the eect of the average rating.

Furthermore, the control variables are examined. Only the brand Apple has a signicant eect on the sales of the tablet, as was expected since there was also a signicant correlation between the brand Apple and the sales of the tablet. If a tablet is from Apple this negatively aects the sales rank and positively aects the sales of the tablet. The lower the number of the sales rank the better (i.e. 1 is better than 10). The other brand have no signicant eect on sales, as was not expected due to the signicant correlations with sales. The storage size has no signicant eect on sales, as was expected. The price has a signicant eect on sales, it has a positive eect on the sales rank and therefore, a negative eect on sales, this was also expected.

4.2 Model with the main eect of the number of reviews

In Model 1 the eect of the number of reviews is already insignicant before including the copula regressor. However, due to the inclusion of the squared and interaction term there was an issue with multicollinearity. To examine whether the number of reviews is endogenous, as was originally expected, a separate regression with the independent variables, Avg_Rating and Num_Reviews and the control variables has been done. Therefore, the following model will be used:

Sales−Ranki= αi+ βi1P ricei+βi2Storagei+ βi3Brand1i+ βi4Brand2i+

βi5Brand3i+ βi6Avg−Ratingi+ βi7N um−Reviewsi+ ui (3)

When controlling for endogeneity, again by using the copula approach, the fol-lowing model will be used:

βi4Brand2i+ βi5Brand3i+ βi6Avg−Ratingi+

βi7N um−Reviewsi+ βi8Copula1i+ ui (4)

(25)

controlling for endogeneity Num_Reviews becomes insignicant, also the copula term is signicant (p=0.001) which conrms that Num_Reviews is endogenous. Furthermore, the same model was also regressed with Num_Reviews*Avg_Rating, however, in this model this term was not signicant before and after con-trolling for endogeneity. Hence, it is concluded that the Num_Reviews and Num_Reviews*Avg_Rating has no signicant eect on sales. Furthermore, in Model 3 and Model 4 the Avg_Rating is also not signicant.

Table 5: Regression output for Model 3 and Model 4

Model 3 Model 4

Variable beta standard error p-value beta standard error p-value

Constant 8.389*** .518 .001 11.362*** .680 .001 Brand Samsung .312 1.245 .822 .563 1.311 .640 Brand Apple -2.849*** .323 .001 -2.606*** .291 .001 Brand HP -.446 .844 .560 -.211 .675 .736 Storage -.002 .002 .144 -.002 .002 .236 Price .001*** .000 .001 .001*** .000 .006 Avg Rating -.269 .177 .133 -.226 .151 .200 Num Reviews -.008*** .002 .001 .001 .003 .539 Copula 1 -1.006*** .233 .001 Adjusted R2 _.577 _.623

4.3 Average rating as a categorical variable

(26)

dummy term of the average rating an interaction term should be included. Hence, for the interaction term of the average rating and the number of reviews there are ve terms included in the model instead of one term.

The results of the eect of the average number of characters is the same as in Model 1 and Model 2, where the average rating was included as a contin-uous variable. All interaction terms between the average number of charac-ters per review and the average rating are not signicant and Num_Reviews, Num_Reviews_sq,

Avg_Rating 2-3*Num_Reviews, Avg_Rating 3-4*Num_Reviews, Avg_Rating 4-5*Num_Reviews, Avg_Rating 2-3*Num_Reviews_sq,

Avg_Rating 3-4*Num_Reviews_sq and Avg_Rating 4-5*Num_Reviews_sq are signicant. However, when including the proper copula regressors this eect disappears again. To conclude, including the average rating as a categorical variable such that the missing values are taken into account does not change the signicance of the average number of characters per review and the number of review terms, therefore, the results of this model are not shown.

(27)

the sales rank than an average rating between 3-4 stars (Mudambi and Schu 2010). Only having an average rating below 3 has no signicant dierent eect than having no average rating. This is possible due to that a rating below 3 implies that the products is not good, a missing average rating is also a bad sign for a product, mostly this is interpreted by consumers as that the products is not good or not sold.

Table 6: Regression output for Model 5

Variable beta standard error p-value

Constant 8.213*** .332 .001 Brand Samsung 1.264 1.151 .194 Brand Apple -2.888*** .509 .001 Brand HP -.061 .573b .915 Storage -.001 .001 .274 Price .001*** .000 .001 Avg Rating 0-1 .700 .527 .149 Avg Rating 1-2 .197 .333 .540 Avg Rating 2-3 -.513 .365 .167 Avg Rating 3-4 -1.397*** .343 .001 Avg Rating 4-5 -1.086* .570 .061 Adjusted R2 _.522

(28)

regressed as a continuous variable, in that case the average rating was not signif-icant. This is probably since the data taken into account diers in the models, when the average rating is included as a continuous variable the products with a missing average rating are not taken into account. Based on these results it is concluded that having a high average rating positively inuences sales compared to having no average rating. So, Hypothesis 1 is not completely supported.

4.4 Taking a subsection of the tablets in the data

The impact of the average rating could be inuenced by multicollinearity, there-fore, a regression has been run on a subset of the data. One model has been regressed on the tablets with few reviews and one model has been regressed on tablets with many reviews and the impact of average rating is compared. A regression has been run on the products which have less than 50 reviews and a regression has been run on the products which have more than 50 reviews. In both models the average rating of a product had no signicant eect on sales. Further research has been done to examine whether the eects described in Hy-pothesis 2 and HyHy-pothesis 3 are there for a part of the data, since also here it was expected that multicollinearity inuenced the results of the models. Price plays an important role in the purchasing decision of consumers, it is assumed that products purchased with a low average rating are mainly cheap products. Therefore, it is examined whether the eect of information overload on the im-pact of the average rating does exist for the low priced products. Therefore, a regression is run on subsets of the tablet data. A regression has been performed on the premium tablets, i.e. tablets that have a price higher than 500 dollar. Also on the lower price tablets a regression is performed, i.e. tablets that have a price lower than 500 dollar. However, for the regressions the same results were found as when including the full sample, all eects hypothesized are not signicant.

4.5 Robustness checks

(29)

Hence, the dierence between the Sales Rank in the two weeks is taken and dependent variable and the dierence between the values of the independent variables in the two weeks are taken as independent variables. For example the dierence in the number of reviews, so the independent variable is the increase in the number of reviews between the two weeks. However, these models gave similar results as Model 1 and Model 2. Therefore, the output is not included in this study. This result does also suggest that the model is robust. However, the results of the signicance of the average rating was not robust between the models were the average rating was included as a continuous variable and the models where the average rating was included as an categorical variable. However, this is most likely since in Model 1, 2, 3, 4 the products that had no average rating were not taken into account an these cases were taken into account in Model 5, hence the models were based on dierent data sets. The results were robust for the four models were the average rating was included as a continuous variable. Furthermore, the assumption of normality is checked for the model and there was non-normality, the Shapiro-Wilk test was signicant, hence the sample is signicantly dierent from normal (p=0.000). So the models in the study are bootstrapped. Furthermore, there was no heteroscedasticity in the model. Also, multicollinearity was only an issue with the squared and interaction terms, what would be expected. Therefore, a model without the squared and interaction terms has also been done. With all the other variables there was no multicollinearity issue.

5 Discussion

In this study, there is contributed to the literature on consumer reviews by examining whether, after a certain number of customer reviews for a product, there is a negative eect of customer reviews on the impact of the average rating of that product. Furthermore, it is examined whether the same negative eect exists for the average number of characters per review, that after some average number of characters per review for a product, there is a negative eect on the impact of the average rating of that product. The results of multiple regression models are evaluated.

(30)

reviews was found by Park et al. (2006), however, they only examined whether consumers perceived information overload and not how it impacted their deci-sion making. Also, the positive eects expected are not found in this study, that in the beginning, if a product has more consumer reviews, this positively inuences the impact of the average rating, thus, when there is no overload of information. Hence, an average rating based on 100 reviews is not taken more into account by the consumer than an average rating based on two reviews. Also, the expected positive eect for a product with a longer review instead of a product with a very short review, is not found in the data set. Hence, the results of this study show that the number of reviews and the average number of characters per review of tablets do not inuence the impact of the average rating on the sales of the product.

In this study, the eect of the average rating on the sales rank is also exam-ined. It was hypothesized that the average rating positively inuences sales, and therefore, negatively eects the sales rank. The results show that this is not always the case, the results do show a positive eect on sales if the average rating of a product is high instead of having no average rating. Hence, when a product receives a high average rating, this positively aects the sales of this product. The positive eect of a high average rating on sales was also proved by Chevalier and Mayzlin (2006).

Furthermore, the number of consumer reviews has also no signicant eect on sales. This was a surprising nding, due to the previous literature on consumer reviews (Zhu and Zhang, 2010; Chevalier and Mayzlin, 2006). However, it could be that the eect of the number of reviews found in previous literature was not controlled for endogeneity. In this study there is controlled for endogeneity, by including a copula regressor for the endogenous variable. Including this extra regressor in the model has as a result that the positive eect of the number of reviews on the sales rank disappears. Hence, there is a correlation between the number of reviews and the sales rank, but, no prove is found that the number of reviews inuences the sales rank and not the other way around. The results of this study suggest that a decrease in the sales rank and therefore, an increase in sales, eects the number of reviews.

(31)

(32)

beforehand which aspects of the product are good and bad, therefore, there is no need to read the reviews, and probably therefore, the number of reviews and the average length of the reviews is not noticed. Thus, for an experience good, it is assumed that the consumer reads more consumer reviews, to examine on which factors they based their rating.

Hence, a suggestion for further research is to examine whether the negative eect of too many reviews and too long reviews is present for experience goods. It is suggested to examine the hypotheses in this study on a data set of an experience good. For example data on consumer reviews and bookings (sales) of hotels. Whether in this branch, too many consumer reviews and too long consumer reviews can actually inuence the impact of the average rating by the consumer. In this product group the average rating of a product is also expected to inuence the sales of that product.

6 Managerial implications

(33)

for search goods, since the actual number of reviews does not inuence the sales of this type of products, according to the results of this study.

However, consumer reviews also contain a rating given by the consumer and the average rating given by the consumer does inuence sales, although, only high average ratings positively inuences sales compared to having no average rating. When a rating is on a scale from 1-5 stars, having a low average rating does not improve sales over having no average rating. Therefore, it is important for rms to actively contact consumer to give a rating of the product they have purchased. However, the number of reviews the average rating is based on does not inuence sales, hence, it is assumed that the consumers does not use the reviews. Thus, it is suggested that for search goods, rms only use resources to contact consumers to give a rating and not complete review, since this only takes a minute or two and if consumers don't feel the obligation to write a whole review, the response rate is likely to increase. Furthermore, there is also no harm done if many consumer write reviews, because there is no overload of information due to the number and length of the reviews.

7 Limitations and further research

(34)

(35)

8 References

Akehurst, Gary (2009). "User generated content: the use of blogs for tourism organizations and tourism consumers", Service Business: An International Jour-nal, 3 (1), 51-61.

Aljukhadar, Muhammed, Senecal, Sylvain and Daoust, Charles-Etienne (2012). "Using recommendation agents to cope with information overload", Interna-tional Journal of Electronic Commerce, 17 (2), 41-70.

Archak, N., Ghose, A., Ipeirotis, P.G. (2011). "Deriving the pricing power of product features by mining consumer reviews", Management Sci, 57 (8), 14851509.

Blal, Inès and Sturman, Micheal C. (2014). "The dierential eects of the qual-ity and quantqual-ity of online reviews on hotel room sales", Cornell Hospitalqual-ity Quarterly, 55 (4), 365-375.

Brown, Jo, Broderick, Amanda, J. and Lee, Nick (2007). "Word of Mouth communication within online communities: conceptualizing the online social network", Journal of Interactive Marketing, 21 (3), 1-20.

Chen, Yubo, Wang, Qi and Xie, Jinhong (2011). Online social interactions: a natural experiment on Word of Mouth versus observational learning, Journal of Marketing Research, 48 (2), 238-254.

Chevalier, Judith, A. and Mayzlin, Dina (2006). The eect of Word of Mouth on sales: online book reviews, Journal of Marketing Research, 43 (3), 345-354. Consti-Ramsden, G., Durkin, K., and Walker, A.J. (2012). The message they send: e-mail use by adolescents with and without history of specic language impairment (SLI), International Journal of Language and Communication Dis-orders, 47 (2), 217-228.

Eppler, Martin J., Mengis, Jeanne (2004). "The concept of information over-load: a review of literature from organization science, accounting, marketing, MIS, and related disciplines", The information Society, 20, 325-344.

Hoyer, Wayne, D., Deborah, J. MacInnis and Pieters, Rik (2013). "Consumer behavior", 6th edition.

Krumm, J., Davies N., Narayanaswami C (2008). "User-generated content", Pervasive Computing, 7 (4), 10-11.

(36)

Liu, Y. (2006). "Word of Mouth for movies: its dynamics and impact on box oce revenue", Journal of Marketing, 70(4):7489.

Mudambi, Susan M. and Schu, David (2010). "What makes a helpful online review? A study of customers reviews on Amazon.com", MIS Quarterly, 34 (1), 185-200.

Pan, Yue, Zhang, Jason, Q. (2011). "Born unequal: a study of the helpfulness of user-generated product reviews", Journal of Retailing, 87 (4), 598-612. Park, Do-Hyung, Lee, Jumin and Han, Ingoo (2006). "Information overload and its consequences in the context of online consumer reviews", The tenth pacic asia conference on information systems.

Park, S. and Gupta S. (2012). 'Handling endogenous regressors by joint esti-mation using copulas", Marketing Science, 31 (4), 567-586.

Schindler, Robert M. and Bickart, Barbara (2012). "Perceived helpfulness of online consumer reviews: the role of message content and style", Journal of Consumer Behaviour, 11, 234-243.

Sun, Monic (2012). "How does the variance of product ratings matter?", Man-agement Science, 58 (4), 696-707.

Wang, Xin, Mai, Feng and Chiang, Roger H. L. (2014). "Database submission: market dynamics and user-generated content about tablet computers", Market-ing Science, 33 (3), 449-458.

Zhang, Xiaoquan M. and Dellarocas, Chris (2006), The Lord of the Ratings: how a movie's fate is inuenced by reviews, in Proceedings of the 27th Interna-tional Conference on Information Systems (ICIS), Milwaukee: Association for Information Systems.