Dynamic effects of positive and negative online reviews on product sales

(1)

- Analyses on tablet product reviews from Amazon.com

By Lan Huang

University of Groningen Faculty of Economics and Business

Master thesis Supervisors: Dr. Hans Risselada (h.risselada@rug.nl) Co-Accessor: Maarten Gijsenberg (m.j.gijsenberg@rug.nl) Student E-mail: l.huang.8@student.rug.nl Student phone number:

+31 (0)621180492 Student number:

S2622459 Address:

Jacobastraat 270, 2512JG, The Hague

(2)

ABSTRACT

Product online reviews have played an important role when consumers are searching for what to buy. Although many empirical studies have found that online reviews have a significant influence on sales, there is a gap that limited studies have investigated the dynamic effects of online reviews have on sales. Focusing on the polarity of online review ratings, the author examined what drive the dynamic effects of positive reviews and negative reviews on sales. The analyses were performed by using secondary datasets of tablet computers collected from Amazon.com from 1st February, 2012 to 11th July, 2012. The first step was to obtain the effects of reviews from 23 regressions and the second step was to analyse the factors of the dynamic effects generated by the first step. The results of the first step showed that positive reviews and negative reviews were not significant on sales throughout the entire period but only several weeks. Positive reviews and negative reviews from recent weeks had positive influence on sales while cumulative positive reviews from previous weeks had negative influence on sales. Time was found to have a positive relationship with the effects of cumulative negative reviews. The effects of cumulative positive reviews were positively driven by its lag effects. The effects of positive reviews of the corresponding weeks were negatively related to the average number of positive reviews per week. However, there were no sufficient evidences to find out the factor of the effects of negative reviews of the corresponding weeks.

(3)

Acknowledgements

(4)

1

1 Introduction

With the development of Internet, product online reviews have played an important role when consumers are searching for what to buy, especially in the e-commerce environment. This phenomenon changes consumers’ buying behaviour because of the growth of the online communication. People tend to rely on the opinions of other consumers in deciding what products to buy. Since the access to online communication is increasingly easier than before, consumers make their decisions based on online recommendations even in offline circumstances (Lee, Park, and Han, 2008). On most of the ecommerce websites, people are allowed to leave reviews and rate every product. Reviews are able to provide consumers with information and experience, eliminating their uncertainty of purchasing online (Dellarocas, 2003). The number of online reviews in the large online-based retailers has grown significantly in recent decades. For instance, from June, 1995 to March, 2013, the largest online-based retailer Amazon.com had received 34,686,770 reviews for all product categories1. The average number of monthly reviews had increased by 10,000 from January, 2000 to January, 20122. Electronic word of mouth (eWOM) marketing is a crucial part of the online marketing strategies. eWOM has not only become a way for online marketers to use to influence consumers’ purchasing behaviour, but also become the access for companies to know about the opinions of customers (Trusov, Bucklin, and Pauwels, 2009). In marketing, eWOM is defined as the information communicated on internet about products, services, brands and companies (Babic, Sotgiu, de Valck, and Bijmolt, 2015). Particularly, online reviews play an important role in eWOM, through which consumers receive information about product usage, quality, functions, etc.

In the past 15 years, over 100 studies have investigated the effectiveness of online reviews. Those studies have proved that online reviews have a significant impact on sales performance (Chaveliar, and Mayzlin, 2006; Cui, Lui, and Guo, 2012; Yin, Bond, and Zhang, 2014). Although the recent decrease of relevant studies might suggest the saturation of this topic, a meta-analysis showed that there was still room for further research (Babic, et al 2015). In order to study the effects of online reviews, many researchers were focused on two key metrics of online reviews, which were the volume and the valence. Previous research has examined and compared the different effects of the volume and the valence on sales performance. The volume of reviews can create awareness among consumers. Valence involves the emotional side of the reviews, containing the information about quality and satisfaction (Dellarocas, Zhang, and Awad, 2007; Godes, and Mayzlin, 2004), which can be positive, negative or neutral (Liu, 2006). According to Babic, et al (2015), it has revealed the inconsistency in which particular metrics of online review actually have the predicting impacts on sales. For example, some of the studies proved that the volume of online reviews could predict sales (Duan, Gu, and Whinston, 2008; Gu, Park, and Konana, 2012; Ho-Dac, Carson, and Moore, 2014; Liu, 2006; Xiong and Bharadwaj, 2014) whereas other studies argued that valence was more effective in predicting sales (Chintagunta, Gopinath, and Venkataraman, 2010; Dellarocas, et al., 2007). Another

1

https://snap.stanford.edu/data/web-Amazon.html

2

(6)

3

meta-analysis on online reviews showed that volume elasticity was higher than valence elasticity (You, Vadakkepatt, and Joshi, 2015).

Regarding the valence, the negative and positive sides of valence have been the focus of the studies in online reviews. Some studies showed, the impact that negative reviews had on decreasing sales was stronger than the impact that positive reviews had on increasing sales (Chevalier and Mayzlin, 2006; Sun, 2012). In addition, others found that the presence of negative eWOM was not harmful to sales because it helped to boost product attention (Doh and Hwang, 2009; Hiura et al., 2010; Kikumori and Ono, 2013). The meta-analysis from Babic, et al (2015) also showed that negative eWOM did not harm sales in general. The inconsistent results on negative and positive reviews have given the current research space to further investigate their impacts on sales. Besides, the current research was mainly focused on the extremity sides of the valence, which is defined as polarity of review ratings. In the five-star rating system, it refers to one-star rating and five-star rating, which are the extreme points of the scale. Several studies showed that this polarity of review ratings had a significant influence on sales (Chevalier and Mayzlin 2006; Ghose and Ipeirotis 2006; Zhang, Li, and Chen 2012).

However, it was found that most of the previous studies including the meta-analysis from Babic et al (2015) did not treat the effects of the reviews dynamically. It is unclear if all those impacts of online reviews changed overtime, e.g. if negative reviews of this week had the same effect as the ones of last week did on sales. By reviewing previous literature, the dynamics effects of positive and negative reviews on sales were limited but not zero. A few studies provided some discussions about the effects of the reviews overtime. For instance, the dynamic figures of volume and ratings have been briefly discussed in a penal data analysis from Duan, Gu and Whiston (2008). They showed that eWOM ratings were changing over time and volume only soared in the beginning but dropped after one week. It is also confirmed that the effects of reviews decrease over the product lifetime in general (Chen, Wang & Xie, 2011) and the effects of WOM can last around 21 days (Trusov, Bucklin, and Pauwels, 2009). Different from the study of Trusov et al (2009) that how long the effects lasted, I examined how the effects of online reviews evolved during a given period of time. Sonnier, McAlister, and Rutz (2011) argued that ignoring the dynamic effects would understate the effects of online reviews on sales. In a study of the movie industry, Liu (2006) showed that effects of reviews were changing during 8 weeks. This study suggests that the effects of reviews are dynamic for at least an 8-week period, which provided the foundation in dynamic effect analysis. However Liu (2006)’s study was limited to 40 movie reviews. As it is stated by Cui, Lui, and Guo (2012), the volume of reviews matters in experience goods like movies, but valence is more important in search goods. Consumers are able to search for all the features online prior to buying search goods like tablets. Therefore, further investigations into dynamic effects over a longer period than 8 weeks and over other product categories with more observations are needed.

(7)

4

Amazon.com. The data contains reviews information, sales raking, price, and brands of tablet products over 24 weeks from 1st February, 2012 till 11th July, 2012. In this paper, two steps were used for analysing the dynamic effects of the online reviews. First, standard regression models were performed on 24 weeks basis. Secondly, the dynamic effects of positive reviews and negative reviews associated with sales were evaluated by analyzing the changes of coefficients over this 24-week period.

The results of the current research show that positive reviews of the corresponding week were significant and were positively related to sales inthree regressions. The cumulative positive reviews of the previous week were also significant but negatively related to sales in four regressions. Negative reviews of the corresponding week were significant and positively related to sales in two regressions and the cumulative negative reviews of the previous weeks were not significant in any regressions. Basically, the positive reviews and negative reviews did not have significant relationships with sales throughout the entire 24-week period. Meanwhile, the results of the current study showed that negative reviews had relatively larger effects than positive reviews had on increasing sales. Surprisingly, the negative reviews of the recent weeks showed positive effects on sales while cumulative positive reviews of previous weeks in the regression of week 2 showed negative effects on sales. Those findings are contradictory to most of the previous studies. It indicates that highly positive reviews of five-star ratings are not always positive for sales and negative reviews of one-star ratings are not always bad for sales.

The results of the analyses on dynamic effects showed that, during this observed period, the impacts of cumulative positive and negative reviews of previous weeks presented linear trends while the impacts of positive and negative reviews of the corresponding weeks changed dramatically each week. Time showed a positive effect on the coefficients of cumulative negative reviews of the previous weeks. As time passed by, the effects of cumulative negative reviews of previous weeks became stronger. However, the effects of cumulative positive reviews are positively driven by its lag effects but not by time. As for the dramatic changes of the effects of positive and negative reviews of the recent week, the results showed that the average ratio of positive reviews per week had a negative influence on the coefficients of the ratio of the positive reviews received per week.

(8)

5

help to increase awareness for the products, while, if the highly positive reviews are not convincing enough, they cannot assist in boosting sales. The management team of the eWOM should be careful with this.

This paper is organized as follows. Chapter 2 provides theoretical backgrounds including literature review, conceptual model and hypotheses. The author explains the methodology in Chapter 3 and illustrates the analysis and results in Chapter 4. Lastly, Chapter 5 is the discussion of results. Meanwhile, in Chapter 5, the author offers implications for management as well as the limitations of the current research and implications of future research.

2 Literature review, conceptual model and hypotheses

2.1 Dynamic effects of online reviews

The studies about eWOM’s effect on sales are not brand new. eWOM refers to all the word of mouth communication on the internet, which includes the communication in online forums, electronic bulletin board, blogs, review websites and social network sites (Blackshaw and Nazzaro, 2006; Goldsmith, 2006). The current research is only focused on a particular mechanism of eWOM, which is the online product review from Amazon.com. Despite the fact that there are already over 100 studies in eWOM, there still has a lot of space for further research in online product reviews. For example, the dynamics of the effects that online reviews have on sales still need to be investigated (Babic, et al, 2015). Therefore the current study aims to discuss dynamic effects of reviews on sales in an ecommerce environment. In order to demonstrate the theoretical background and contribution of current study, based on the recent meta-analyses of Floyd, et al (2014), You, et al (2015) and Babic, et al (2015), table 2.1 summarizes the key findings of 20 articles according to the similarities of research objectives, which is the study of online reviews’ impacts on sales and the dynamics of the impacts.

Articles Data Metrics Effect of online reviews on

sales

Discussion about

dynamic effects Liu (2006) Movies (Yahoo.com) Volume &

Valence

Volume is highly significant but valence is not significant

Eight models are presented based on eight weekly data; Coefficients changed during the eight weeks

Duan, Gu, and Whinston (2008)

Movies (Yahoo.com) Volume & Valence

Volume is highly significant but valence is not significant

WOM volume had significant dynamic relationship with positive valence

Sonnier, McAlister, and Rutz (2011)

A technology firm Valence All valence including positive, negative and neutral comments are significant; Positive and

(9)

6

negative reviews have stronger impact than neutral reviews in the long run

reviews Chevalier and Mayzlin (2006) Books (Amazon.com, BN.com) Volume & Valence

Negative reviews have stronger impact

Not discussed Li and Hitt (2008) Books (Amazon.com) Valence &

Variance

Positive reviews increase sales; the variance matters

Timing of the review affected sales

Chen, Dhanasobhon, and Smith (2008)

Books (Amazon.com) Volume & Valence

Higher ratings generate higher sales; Stronger effect found in niche products

Not discussed

Forman, Ghose, and Wiesenfeld (2008)

Books (Amazon.com) Review features

Reviewers identities and locations influence sales

Not discussed

Dewan and

Ramprasad (2009)

Music albums (Nielsen SoundScan,

Amazon.com)

Volume & Valence

Both volume and valence are significant; volume has stronger effect

Not discussed

Pathak, Garfinkel, Gopal, Venkatesan, & Yin. (2010)

Books (Amazon.com) Volume & Valence

Ratings and volume are positively related to sales

Not discussed

Archak, Ghose, and Ipeirotis (2011)

Digital cameras and camcorders

Volume & Valence

Rating and volume are positively related to sales; neutral reviews do not help to increase sales

Not discussed

Amblee and Bui (2011)

Amazon short stories (Amazon.com)

Volume & Valence

Both volume and valence are significant

Not discussed Chen, Wang, and Xie

(2011)

Digital cameras (Amazon.com)

Volume & Valence

Negative reviews have stronger impact than positive ones; higher volume benefits sales

The impact of reviews decreases over the product lifetime

Ghose and Ipeirotis (2011)

Audio and video players, digital cameras, DVDs (Amazon.com)

Volume & Valence

Objective reviews are considered more helpful

Not discussed

Cui, Lui, and Guo (2012)

Video games, consumer electronics

(Amzon.com)

Volume & Valence

Valence is more important for search goods while volume for experience goods

Volume had impact after the launch of product and such effect decreased overtime

Gu, Park, and Konana (2012) Digital cameras (Amazon.com, Cnet, DPreview, Epinion) Volume & Valence Internal review platform(Amazon) has limited impact on high involvement products

Not discussed

Sun (2012) Books (Amazon.com, BN.com)

Volume & Valence & Variance

Negative reviews have stronger impact on sales; high variance matters only when the rating is low

(10)

7 Zhang, Li, and Chen

(2012)

Books (Amazon.com, BN.com)

Valence Sentiment divergence of reviews have impact on sales

Increase of sentiment divergence of reviews results in higher sales overtime

Floyd et al. (2014) Meta-analysis with 26 studies

Volume & Valence

Valence has higher average elasticity than volume on sales

Not discussed You, Vadakkepatt, and Joshi (2015) Meta-analysis with 51studies Volume & Valence

Valence has higher elasticity than volume; Negative reviews have more significant effect on valence elasticity

Not discussed

Babic et al. (2015) Meta-analysis with 96 studies

Volume, Valence & Variance

Cumulative volume is the most important; valence and variance can also affect sales

Not discussed

Table 2.1 Comparison of previous studies on the effects of online reviews on sales (Floyd, et al 2014; You, et al 2015; Babic, et al 2015)

Among the above 20 articles, the studies that focused on dynamic effects of online reviews on sales are from Liu (2006), Duan et al (2008), and Sonier et al (2011). The study from Liu (2006) can be considered most similar to the current study based on its perspectives, in which an eight-week movie dataset was used for the analysis. As a result, valence of reviews was not significantly related to revenue, but the results revealed that the effects of both volume and valence were dynamic. Later on, another study on movie industry proved again that the effects of online reviews are dynamics. Duan et al (2008) found that although valence did not influence sales, positive reviews could dynamically drive the volume. By the analysis of dynamic effects that online reviews have on sales, we can obtain more useful insights about online reviews and how the influence changes overtime (Duan et al 2008). Besides, Cui et al (2012) argue the influence of reviews differs between experience goods and search goods. Their study states that volume only mattered in the beginning after the product was launched, the effect decreased afterwards. Thus, applying dynamic analysis on different product may have different results from the studies of Liu (2006) and Duan et al (2008). This study extends their dynamic analysis by using data from search goods. Sonier et al (2011) used a more complicated approach (dynamic model) to examine the carryover effects of positive, neutral, and negative comments on generating revenue. However, the key findings from Sonier et al (2011) are the effects of positive, negative, and neutral comments lasted around one week and the positive comments brought the higher elasticity for sales than negative and neutral comments. Their study did not explicitly discuss about the change of effects over time and what influenced the dynamic effects. Moreover, the study was limited by using data from a single durable goods company.

(11)

8

dynamic features of the online reviews but not the dynamic effects. These discussions can provide the current study insights and background about dynamic effects. The timing can be one of the possible predictors for the dynamics. Li and Hitt (2008) mentioned that the time when the reviews were posted has impact on sales. The reviews from the early stage influence the opinions of the consumers in later period. In general, the previous studies reveal that the effects of reviews decrease overtime (Chen, Wang, and Xie 2011; Cui, Lui, and Guo 2012). Zhang, Li, and Chen (2012) stated that increase of sentiment divergence of reviews results in higher sales overtime. Sentiment divergence refers to the split of opinions, which can be seen as the polarity of ratings.

2.2 Polarity of online review ratings

According to Zhang, Li, and Chen (2012), sentiment divergence has impact on sales overtime. Increasing the sentiment divergence of reviews can contribute in better sales. In online review rating system, the sentiment divergence means the polarity of ratings. It is considered one-star rating as ―poor‖ and five-star rating as ―excellent‖ in the five-star rating system (Turney, 2002). It is more interesting to study about the polarity of the reviews rather than the middle part. Because the lowest rating (one-star) represents an extremely negative point of view and the highest rating tells an extremely positive opinion about the products (Mudambi and Schuff 2010). From Table 2.1, not all the previous studies pay attention to the polarity of online review ratings. However, according to Ghose and Ipeirotis (2006), the polarity of reviews has significant influence on sale. In the studies of Chaverlier and Mayzlin (2006) and Babic et al (2015), they defined five-star rating as positive valence and one-star rating as negative valence. Therefore, in order to differentiate from the previous studies on valence, the current study focuses on the polarity of online review ratings, which are represented by one-star rating reviews and five-star rating reviews.

2.3 Conceptual model and hypotheses

(12)

9

Figure 2.1 Conceptual model

(13)

10

Step 1:

The three meta-analyses from Floyd, et al (2014), You, et al (2015) and Babic, et al (2015) showed a significant positive relation between review volume and product sales. Accordingly, the current study combines the volume with valence. Babic et al (2015) introduced the variable of positive volume and negative volume which accounts for both valence and volume at the one variable. On the study of valence of reviews, many studies pointed out the importance of negative reviews. They found that negative reviews had a stronger effect on decreasing sales than positive reviews did on increasing sales (Chaveliar, and Mayzlin, 2006; Chen, et al, 2011; Sun, 2012). Consumers associate negative reviews with low quality while positive reviews with better quality. A better-perceived quality in the end leads to better demand and creates better sales (Etzion, and Awad, 2007). In a study of hotels, Blal, and Sturman (2014) mentioned that ratings were crucial for sales. Their study showed that valence was positively related to the sales performance. Based on the previous literature reviews, the first hypothesis is proposed as:

H1a: Valence of reviews at weekt has a significant relationship with sales at weekt..

H1b: The number of positive reviews at weekt has positive effect on sales at weekt., whilst the

number of negative reviews at weekt has negative influence on sales.

In order to include the potentially important variables in the model, another meta-analysis from Babic, et al (2015) showed that cumulative volume had the most predicting power in sales. So cumulative reviews variables were added in the conceptual model. Previous studies have proved the fact that the reviews from the previous week still influenced the sales regardless the effects decrease after one week (Trusov, et al, 2009; Ghose and Ipeirotis, 2011). The positive reviews and negative reviews from previous weeks are also expected to have a significant influence on sales. The impacts of the valence from cumulative review variables should be the same as the ones used in the last hypothesis. Based on that, the second hypothesis is proposed and it can be extended as follows:

H2a: Valence of reviews before weekt has a significant relationship with sales at weekt.

H2b: The cumulative number of positive reviews before weekt has positive effect on sales at

weekt, whilst cumulative number of negative reviews before weekt has negative influence on

sales at weekt.

Step 2:

(14)

11

be proposed as:

H3: The effects of positive reviews and negative reviews in the observing period decrease with the age of week increases.

Another factor for the dynamic effects can be found from the study of Zhang, Li, and Chen (2012). They state that the increase of sentiment divergence of reviews resulted in higher sales overtime. Although the current study was unable to examine the development of sentiment divergence directly, it can be regarded as the extremity of review ratings. Therefore, the current study used the number of one-star rating and the number of five-star rating as the drivers for the change of the effects. Because the polarity of ratings had a stronger significant influence on sales than the natural ratings did on sales (Ghose and Ipeirotis, 2006), it is expected that the effect is stronger when the number of one-star rating and five-star ratings increases. The fourth hypothesis is formed as:

H4: The strength of the effects from positive reviews and negative reviews on sales at weekt

positively related to the numbers of one-star rating and the numbers of five-star rating at weekt. In the studies of movie industry, increasing volume of reviews could create high awareness in the pre-release period and was used as a prediction for sales (Liu, 2006). High volume reviews also created a buzz which had a strong influence on sales in return (Duan et al, 2008). For beer products, volume of reviews is an essential indicator for sales (Clemons, Gao, and Hitt, 2006). In the meta-analyses of Floyed et al (2014), You et al (2015) and Babic et al (2015), they all stated that the volume of the reviews was an important factor for sales. In this conceptual model, recent volume and cumulative volume are considered as control variables together with price, and price promotion. Apart from that, the products’ features should also were accounted as control variables. There is no relevant analysis about tablet products. But tablets are also portable digital devices. Based on the research done for mobile phones, screen size was significantly related to consumer buying behavior (Chae and Kim, 2004). Competitive brands are also involved as the indicator of competition. Thus, brands and screen size are added into the model.

After the theoretical background is discussed in this section, the author of this paper is able to explain the methodology of the research in order to examine the proposed hypotheses.

3 Methodology

(15)

12

3.1 Data description

With the availability of time series data from Wang, et al. (2013), the current research is able to account for the effects of market dynamics in the study of how online reviews impact on sales. The data set used in this research consists of two parts of the original data. First part is the market dynamics of 2,163 tablet products, which contains 24-week information of sales rank, price, promotion, and cumulative number of reviews. Second part contains 40,741 product reviews including information on product ID, item ID and reviewers’ ID, reviews’ date, and rating. The current study combined these two parts of data. Table 3.1 lists the description of the useful data based on Wang, et al. (2013).

Variable Description

Item ID Amazon standard identification number (ASIN) of the product Sales rank Sales rank of the item in the category of tablets and table PCs List price Retail price suggested by the manufacturer

Amazon price Current selling price of the product

Number of reviews Total number of reviews posted up to the collection day Review date Date of the review was posted

Rating Brands Screen size

Numerical rating of the review (one-five stars: one means really bad and five means really good)

Two most competitive brands namely Samsung and Apple The screen size of the product measured by inch

Table 3.1 description of the data is used for this research (Wang, et al. 2013)

3.2 Measurement of variables

Dependent variable: sale rank. Although Amazon.com does not publish the sales volume for each product, the information on sales rank is available and therefore could be used as a proxy variable. Chevalier and Mayzlin (2006) stated that the sales volume was in a liner relation with the sales rank. Therefore the sales rank was used directly as a dependent variable in the studies of Chevalier and Mayzlin (2006) and Cui, et al (2012). However, it is important to mention that the lower values of the sales rank actually mean higher volume sales. The interpretation of the coefficients of the independent variable is different. Specifically, a negative coefficient actually indicates a positive relationship with the dependent variable. On the other hand, positive parameters mean a negative relationship with the dependent variable.

(16)

13

of reviews by calculating the number of negative reviews and positive reviews. The number of negative reviews and positive reviews from the corresponding week was calculated individually, which was then used to calculate the number of cumulative positive and negative reviews. Reviews with all the ratings: the volume of total reviews. Although the current study is focused on polarity of online reviews, the volume of total reviews was included into the analysis as a control variable. Because the volume of total reviews played a significant role in the previous studies because they found that the volume of total reviews had a positive influence on sales (Floyd et al 2014; You et al 2015; Babic et al 2015). Since the dataset crawled by Wang et al (2013) only had the cumulative number of reviews that each product received weekly, the number of reviews that each product received in the corresponding week needed to be calculated. According to the conceptual model, the individual number of reviews per week and the cumulative number of reviews per week were included.

The control variables included price, competitive brands dummy variables, and price promotion dummy variable. Besides the aforementioned independent variables, the analysis also included other factors as control variables, namely competitive brands, screen size, price, and price promotion. The most competitive brands in this dataset are Samsung and Apple Wang, et al (2013). The Samsung dummy variable was computed as 1 if this product was from Samsung and as 0 if not. The Apple dummy variable was also calculated in the same way. However, there was more than one selling price from the data because the sales rank from Amazon.com includes the sales from other distribution channels (Landahl, 2011). Thus the price variable in the analysis was calculated by taking the average of different selling prices from all channels including the new product price, the refurbished product price and used product price. Price promotion is also a dummy variable, for which 1 and 0 represented the presence and absence of discount respectively. To see whether or not there was a price promotion for the product in Amazon, I checked if the price listed by manufacturers was higher than the price listed by Amazon. The purpose of introducing this dummy variable was to test if there was a distinction between two groups (with discount and without discount) in the dataset.

3.3 Aggregation of data

The data obtained from Wang, et al (2013) was separated into market dynamics and consumer generated reviews. The sales rank, price and the cumulative number of total reviews per product were already calculated based on 24 weeks. The two data sets were combined, in which the 40,741 reviews were aggregated into each product based on 24 weeks. In addition the (individual/cumulative) numbers of negative reviews and positive reviews were categorized for each product for 24 weeks.

(17)

14

week. In this case, numbers of total reviews were not applicable for the first week. The analysis was only conducted from the second week. Secondly, among those calculated individual numbers, I extracted the individual numbers of negative reviews and positive review for each product per week. In the end, the cumulative numbers of negative reviews and positive reviews were calculated by summing up all the individual numbers from previous weeks.

The new cumulative dataset was combined with the market dynamics dataset based on the common column product ID (ASIN). However, there were missing cases in the merged dataset i.e. many products did not have sales rank information or did not have review information. Although some of the products did not have reviews information directly in the dataset, by using the ASIN on Amazon.com I found out that the missing values for reviews actually meant that this product did not receive any reviews yet and hence the missing values of reviews information were replaced with zero. The missing cases without information on sales rank and reviews were dropped out of the dataset. As a result, only 404 products, which have the complete information, were kept for the data analysis. After that, this aggregated dataset was used for the data analysis in the next stage.

3.4 Transformation of reviews variables

After computing and combining all the selected variables into one dataset, the first linear regression pre-analysis was applied on the data of the 23 weeks. From the result of the pre-analysis, there was a high level of multicollinearity issue because the Variance inflation factor (VIF) values for six variables (individual/cumulative numbers of negative reviews, individual/cumulative numbers of positive reviews and individual/cumulative numbers of total reviews) were way larger than 10. Meanwhile, the Pearson’s Correlation between those six variables was higher than 0.5. This phenomenon has been discussed by Liu (2006), where using absolute (cumulative) numbers of positive and negative reviews together with total review volumes could have a potential high level of multicollinearity. It is proposed to use the percentages of positive and negative reviews based on the (cumulative) numbers. Therefore, the four variables (individual/cumulative numbers of negative reviews and individual/cumulative numbers of positive reviews) were all transformed to percentages. The ratio of positive reviews at weekt was calculated was defined as the number of positive reviews divided by the number of

total reviews at weekt. And the ratio of cumulative positive reviews was defined as the number

of cumulative positive reviews divided by the number of cumulative total reviews at weekt-1.

The same method was applied to calculate the ratio of negative reviews at weekt and ratio of

cumulative negative reviews weekt-1.

Meanwhile, an outlier with more than 7000 reviews per week was deleted from the dataset because the rest of the products only have lower than 500 reviews each week. This extreme case could raise the non-normality issue. Finally, all the variables used in the analysis are summarized in Table 3.2.

Variables

(18)

15

Independent variables RPt (Ratio of positive reviews at weekt)

RNt (Ratio of negative reviews at weekt) NRt (Number of total reviews at weekt)

RPCt-1 (Ratio of cumulative positive reviews at weekt-1) RNCt-1 (Ratio of cumulative negative reviews at weekt-1) CRt-1 (Cumulative number of reviews at weekt-1)

Pt (Price at weekt)

SCR(Screen size for each product) PPt (Price promotion dummy at weekt) Samsung (Samsung brand dummy variable) Apple (Apple brand dummy variable)

Table 3.2 Variables for analysis after transformation

The lag sales ranks was another potential independent variable but was excluded from the analysis from the very beginning. Based on an experiment in which I included the lag sales ranks in the model, the result showed that the lag sale ranks had a high correlation with the dependent variable and its coefficients were dominated in the regression. That’s because the sales rank was only updated once a month if the ranking was higher than 10,000 (Chaveliar, and Mayzlin, 2006), and there were many products that had very high sales ranks during some weeks in the data. For those high-ranking products, the sales ranks in the previous week were actually the same as the sales ranks in the current week. Alternatively, RPCt-1, RNCt-1, and CRt-1,

which were the cumulative numbers of corresponding reviews from the last week, were the representatives of lag variables for RP, RN, and NR. Using RPCt-1, RNCt-1, and CRt-1 as lag

variables here have three advantages: 1) RPCt-1, RNCt-1, and CRt-1 included not only the reviews

of one week before but also the cumulative reviews from a long time ago, so that I could also evaluate how the reviews received before can influence the current sales; 2) instead of using direct lag variables of RP, RN, and NR, RPCt-1, RNCt-1, and CRt-1 avoided multicollinearity

issues; 3) according Leeflang et al (2015), a model should be simple and include only the necessary variables, so RPCt-1, RNCt-1, and CRt-1 solved the problem of including too many

parameters.

3.5 Descriptive Statistics

In order to get more insights into the data, I checked the descriptive statistics. Among those variables, except for SCR (Screen size for each product), Samsung (Samsung brand dummy variable) and Apple (Apple brand dummy variable), the rest of the variables were changing throughout the 24-week period.

First of all, the descriptive statistics of the non-constant numerical variables are presented in

Appendix 1. For SRt, CR (Cumulative number of review at weekt-1), RPCt-1 (Ratio of

cumulative positive reviews at weekt-1) and RNCt-1 (Ratio of cumulative negative review at

weekt-1), there are 9672 (403 product items times 24 weeks) observations in total. However,,

(19)

16

reviews at weekt), RPt (Ratio of positive reviews at weekt) and NRt (Number of total reviews at

weekt). It can be seen that there are missing values in price variable, for which it is not feasible

to obtain anymore. But the overall observations of each week are still enough to continue with the analysis, since the week with the smallest sample size had 388 observations.

For the other dummy variables such as promotion, Samsung, and Apple, the frequencies were generated.. There are only 4.7% and 5.5% of the total products belonging to Samsung and Apple respectively. The average percentage of the products, for which there was a price promotion, is 57.5%. As for the screen size, the minimum screen size is 2.8 inches and the maximum is 14 inches. Average screen size is 8.6 inches.

As it is shown in Appendix 1, the differences in the range of numerical data values are large the range differences of the values in each numerical variable are large. For example, the values of SR ranges from1 to 229445; the values of RP, RN, RPC and RNC range from 0 to1 (0% to 100%); the price also varied from US$36 to US$2964. As it is shown that the scales of the variables are quite different, therefore the log transformation is needed and further explanations are provided in the modeling approach.

3.6 Modeling approach

Before analyzing the dynamic effects of reviews, a multiplicative linear model was applied to the data of each week. The reasons and purposes of choosing multiplicative linear model are: 1) the ranges of different variables are quite large. Multiplicative model is known as a log-log model, which means that it requires natural logarithm transformation for all numerical variables. As a result, non-linear independent variables can be linearizable and coefficients can be considered as elasticities. It is easier to interpret coefficients as elasticity terms because they are able to describe the percentage difference in the dependent variable in response of one degree of change of the independent variable; 2) it is more convenient to use the elasticity terms generated from the first step when analyzing the dynamic effects in the second step; 3) a multiplicative model was chosen because this model allows interactions between the various independent variables directly (Leeflang, Wieringa, Bijmolt and Pauwels, 2015). The model of the first step is demonstrated as follows:

Equation for H1 and H2 ①: 𝑆𝑅_𝑡 = 𝛼_𝑡 RP_𝑡𝛽1𝑡_RN 𝑡 𝛽2𝑡_𝑁𝑅 𝑡 𝛽3𝑡_𝑅𝑃𝐶 𝑡−1 𝛽4𝑡_𝑅𝑁𝐶 𝑡−1 𝛽5𝑡_𝐶𝑅 𝑡−1 𝛽6𝑡_𝑃 𝑡 𝛽7𝑡_𝛽 8𝑡 𝑃𝑃𝑡 𝛽_9𝑡𝑆𝑎𝑚𝑠𝑢𝑛𝑔𝛽_10𝑡𝐴𝑝𝑝𝑙𝑒𝑆𝐶𝑅𝛽11𝑡_𝜀 𝑡 Where,

𝛼_𝑡, 𝛽_𝑡 = model coefficients in week t (t=2, 3, 4,…24).

(20)

17

all the zero values of numerical variables into an extremely small value of 0.000001 before taking the natural logarithm. All the numerical variables of this model were transformed using natural logarithm including RPt, RNt, NRt, RPCt-1, RNCt-1, CRt-1, Pt and SCR.

H1 and H2 were the basis of analyzing the dynamic effects in the second step. After obtaining the coefficients of the variables generated for 23 weeks, I was able to use those coefficients to analyze the dynamic effects of the reviews, for which there were 23 observations. Since the independent variables from the multiplicative model (equation ①) were already linearized, a linear additive model was used to test the drivers of the dynamics. Besides, the effects of reviews were expected to have a linear downward trend, because the effects decreased after a long period (Cui et al, 2012). Thus, For H3, the factor that drove the effects of cumulative positive and negative reviews was time. Therefore time was an independent variable for the effects of cumulative positive and negative reviews, which was defined as week t (t=2, 3, 4,…24). In equation ①, the coefficients from RPCt-1 andRNCt-1 were used as the dependent

variables. Week number was added as the independent variable to analyze the dynamic effects. Finally, the dynamic effects of cumulative positive and negative reviews were analyzed using the equation as follows:

For H3 ②:

𝛽𝑅𝑃𝐶 = 𝜇𝑅𝑃𝐶 + 𝜃𝑅𝑃𝐶𝑊𝑒𝑒𝑘𝑁𝑢𝑚 + 𝜀𝑅𝑃𝐶 and 𝛽𝑅𝑁𝐶 = 𝜇𝑅𝑁𝐶 + 𝜃𝑅𝑁𝐶𝑊𝑒𝑒𝑘𝑁𝑢𝑚 + 𝜀𝑅𝑁𝐶

For H4, the purpose was to investigate the strengths of the effects of positive reviews and negative reviews received from the corresponding weeks. The dependent variables were the 23 pairs of coefficients of RPt and RNt , which were taken from the equation ① generated in the

23 regressions. H4 proposed that the strengths of the effects were influenced by the average number of positive reviews per week and the average number of negative reviews per week. And the numbers of negative reviews and positive reviews were transformed to the ratios of negative reviews and positive reviews. Hereby, the independent variables were measured by the average of RPt andRNt per week. Finally, the dynamic effects of positive and negative reviews

received from corresponding weeks were analyzed as follow: For H4 ③:

𝛽_𝑅𝑃 = 𝜇_𝑅𝑃+ 𝜃_𝑅𝑃RP∗_{+ 𝜀}

𝑅𝑃 and 𝛽𝑅𝑁 = 𝜇𝑅𝑁+ 𝜃𝑅𝑁RN∗+ 𝜀𝑅𝑁

Where, RP∗ and RN∗ refer to the average of RPt and RNt for each week.

(21)

18

4 Analyses and results

4.1 Analysis of OLS

The key results of the 23 regressions by using the multiplicative model (ordinary least squares) are reported in Appendix 2. Every the regression achieved significant outcomes in F test (p < .05) which means that each model fits the data. The adjusted R square is low in all the results. This indicates that the proportion of variances that can be explained by the predictors is not very large. Based the results in Appendix 2, there are only a few independent variables showing a significant relationship with the dependent variable in 23 regressions. Except that price promotion dummy variable and Apple brand dummy variable were significant in every regression, the other independent variables were not significant as expected in the hypotheses. Besides, the individual number of total reviews was significant with sales in week 4. Table 4.1 summarizes the results of the regressions with significant results of reviews variables, which are the key variables of the current study.

The values of standard error of the estimates are quite small in each regression (<0.500), meaning that the coefficients of the parameters are relatively accurate. Overall, the VIF values in every regression were lower than 5. In each regression, correlations among all predictor variables are not significant and lower than 0.500. The VIF values and Pearson’s correlation tables are not demonstrated here as there are too many regressions. Therefore, no multicollinearity is detected in OLS regressions.

OLS W2 N:393 W3 N:395 W4 N:395 W9 N:395

B Sig. Std. E B Sig. Std. E B Sig. Std. E B Sig. Std. E

(22)

19

Table 4.1 Results of the regressions with significant results of reviews variables

The values of standard error of the estimates are quite small in each regression (<0.500), meaning that the coefficients of the parameters are relatively accurate. Overall, the VIF values in every regression were lower than 5. In each regression, correlations among all predictor variables are not significant and lower than 0.500. The VIF values and Pearson’s correlation tables are not demonstrated here as there are too many regressions. Therefore, no multicollinearity is detected in OLS regressions.

To check the robustness of the model, unstandardized residuals of each regression were saved for further investigation of heteroscedasticity issue according to Leeflang, et al (2015). The scatterplot are used to check if there is heteroscedasticity issue. By plotting unstandardized residuals against the predicted values of dependent variables, the variances of the regression are supposed to distribute in the scatterplot up and down without patterns. However, the residuals of the current study in the scatterplot were relatively dispersed in the plot abnormally (an example of the scatterplot from the regression of week 4 is shown as below in Figure 4.1). The abnormal sign from Figure 4.1 shows that the variances were not constant. As it can be seen from Figure 4.1, some variances are smaller and some variances are larger. It indicated potential heteroscedasticity issue in the regression.

OLS W14 N:397 W17 N:397 W18 N:397 W24 N:397

B Sig. Std. E B Sig. Std. E B Sig. Std. E B Sig. Std. E

(23)

20

Figure 4.1 Scatterplot of regression of week4

Levene’s tests were used in every regression to confirm and to diagnose the source of heteroscedasticity. The opposite side of heteroscedasticity is homoscedasticity. Levene’s homoscedasticity test of variance was employed to detect if there was presence of homoscedasticity. If p-value of Levene’s test is lower than 0.050, it means there is heteroscedasticity issue. Before applying Levene’s test, the cause of the fluctuation of the variances in the scatterplot had to be identified. Dummy variables are used to find out the cause. Compared to brands dummy variables, I selected the dummy variable of price promotion as the source of heteroscedasticity. Because 1) promotion variables were significant in every regression, it means that the observations of the group with discounts significantly differed from the observations of the group without discounts. Thus, the variances between the two groups (with discounts and without discounts) could also be significantly different; 2) there were 57.5% of the products had discounts. Compare to the frequency of 5.5% in Apple brand, the variance of the two groups of discount dummy variables were more sufficient than the two groups of Apple brand.

(24)

21

a result, by using price promotion dummy variable as the source of heteroscedasticity, it was detected that the regressions of week 13, week 17, week 18, week 19, and week 20 had heteroscedasticity issue. The regression of week 4 was very close to having the heteroscedasticity issue. As for the rest of the regressions, there was no sufficient evident in identifying the heteroscedasticity issue in these weeks.

4.2 GLS correction

One of the best-known methods to remedy heteroscedasticity is to apply weighted least squares (WLS), which is a special case of generalized least square (GLS) (Leeflang et al, 2015),. This method was performed to the regressions which are found significant in Levene’s test, namely week 13, week 17, week 18, week 19, and week 20. Week 4 reported a very close significant level in Levene’s test (p=0.051). Thus week 4 was also included in the operations of GLS.

The residuals of the OLS regressions were saved for the operations of GLS. Since the source of heteroscedasticity came from price promotion dummy variable, the residuals were split into two groups(with discounts and without discounts). The standard deviations of the residuals of both groups were computed. Next, all the predictor variables and the dependent variable were split into two groups. Then each variable was divided by the standard deviation of the residuals based on two groups(with discounts and without discounts) respectively. Accordingly, all the variables of the problematic weeks (week 4, week 13, week 17, week 18, week 19, and week 20) were transformed by taking out the standard deviations of the residuals. Afterwards, these regression models were estimated again after operations of GLS.

The results of GLS were similar with the results of OLS. Table 4.2 summarizes the results from the regressions of week 4, week 13, week 17, week 18, week 19, and week 20 after applying GLS. It is obvious that these remedied regressions have better model fit with lower p-values of F-test and higher values of adjusted R square. Overall standard errors were reduced by taking out the standard deviations of the residuals. The residuals from these GLS regressions were also saved to test if the heteroscedasticity issue was solved or not. From Table 4.2, it can be seen that the results of Levene’s test are no longer significant (p > 0.050). The heteroscedasticity issue was solved by GLS.

GLS W4 N:395 W13 N: 397 W17 N: 397

B Sig. Std. E B Sig. Std. E B Sig. Std. E

(25)

22

Table 4.2 GLS results

As it is shown in Table 4.2, price promotion dummy variable and Apple dummy variable are still highly significant (p < 0.050). Same as the results of OLS, RPCt-1 is significant in week 4

after using GLS; RPt is significant in week 18 after using GLS; RNt is significant in week 17

after using GLS. After the GLS, the author continues with the second step which is mention in the methodology section. The coefficients of the OLS regressions without heteroscedasticity issue and the coefficients of the GLS regressions are used to analyze the dynamic effects.

4.3 The impacts of positive reviews and negative reviews on sales

The results of the first step, which was to obtain the impacts of positive and negative reviews. So I combined the results from Table 4.1 and Table 4.2 and reported the significant results of the variables of positive reviews and negative reviews. The results of the positive reviews and negative reviews are shown in Table 4.3. As for other control variables, same as the results of OLS, price promotion dummy variable and Apple brand dummy variable were significant in every regression. In addition, the individual number of total reviews was still significant with sales in week 4.

Samsung -0.11 0.658 0.247 0.053 0.833 0.253 -0.025 0.920 0.249 Apple -0.673 0.004 0.232 -0.880 0.000 0.232 -0.902 0.000 0.234 PPt -0.75 0.000 0.116 -0.633 0.000 0.120 -0.833 0.000 0.120 F-test 8.134 0.000 6.221 0.000 11.745 0.000 Adjusted R 0.166 0.127 0.23 Levene’s test 0.827 0.363 2.242 0.135 2.006 0.158 GLS W18 N: 397 W19 N: 397 W20 N: 397

B Sig. Std. E B Sig. Std. E B Sig. Std. E

(26)

23

Table 4.3 Results combined from OLS and GLS

Previously in the literature review part, H1a and H1b proposed that ―Valence of reviews at weekt

has a significant relationship with sales at weekt.” and “The number of positive reviews at weekt

has positive effect on sales at weekt., whilst the number of negative reviews at weekt has negative

influence on sales”. The key variables of H1a and H1b are RPt and RNt . As it shown in Table

4.3, RPt are significant in three regressions(week 2: β = -0.071, p-value < 0.050, Std.Error =

0.033; week 9: β = -0.105, p-value < 0.050, Std.Error = 0.036; week18: β = -0.092, p-value < 0.050, Std.Error = 0.036) and RNt are significant in two regressions (week 2: β = -0.082,

p-value < 0.050, Std.Error = 0.037; week 17: β = -0.119, p-value < 0.050, Std.Error = 0.044). In addition, RPt is marginally significant in week 4 after GLS operation (β = -0.068, p-value =

(27)

24

results. The coefficients of the significant RPt and RNt variables are all negative. But a negative

values represent a positive relationship with sales. It can be concluded that the variables of ratios of positive reviews and negative reviews at weekt are positively related to sales at weekt.

The statement of the number of positive reviews at weekt has positive effect on sales at weekt is

supported. However, the statement of the number of negative reviews at weekt has negative

influence on sales at weekt is not supported. Thus, H1a was not supported and H1b were

partially supported by the results.

Previously in the literature review part, H2a and H2b proposed that ―Valence of reviews before

weekt has a significant relationship with sales at weekt” and “The cumulative number of

positive reviews before weekt has positive effect on sales at weekt, whilst cumulative number of

negative reviews before weekt has negative influence on sales at weekt”. RPCt-1 and RNCt-1 are

the key variables of H2a and H2b. As it shown in Table 4.3, RPCt-1 are significant in four

regressions (week 2: β = 0.047, p-value < 0.050, Std.Error = 0.022; week3: β = 0.052, p-value < 0.050, Std.Error = 0.022; week4: β = 0.047, p-value < 0.050, Std.Error = 0.027; week24: β = 0.043, p-value < 0.050, Std.Error = 0.022). But RNCt-1 does not show any significant results in

the 23 weeks. Thus, variables of ratios of the cumulative positive and negative reviews before weekt did not have significant relationship with sales ranks in every week. The coefficients of

RPCt-1 are all positive, which means they have negative relationships with sales. Therefore, H2a

and H2b were not supported by the results.

4.4 Analysis of the dynamic effects

Since the goal of this research is to address the dynamic effects of positive and negative reviews on sales, the coefficients of RPCt-1, RNCt-1, RPt and RNt were taken into the analyses

based on equation ②, and equation ③. For those weeks which have heteroscedasticity issue, the coefficients of GLS regressions were used. The rest of the weeks which had no heteroscedasticity issue, I continued with the coefficients of OLS regressions. The coefficients of RPCt-1, RNCt-1, RPt and RNt are plot and illustrated in Figure 4.2.

Figure 4.2 Dynamic effects of RPCt-1, RNCt-1, RPt, and RNt,

(28)

25

4.4.1 Dynamic effects of RPC

t-1

and RNC

t-1

As expected, the coefficients of RPCt-1 during the 23 weeks manifested a downward trend.

However, the coefficients of RNCt-1 shows an upward trend. According to Cui et al (2012), the

effects of reviews decreased with time passed by. As it is mentioned before, time was expected to be the driver of the effects of RPCt-1 and RNCt-1, so week number was created as the

independent variable for the analysis of the dynamic effects of RPCt-1 and RNCt-1. Equation ②

was used to test the driving forces of the effects of RPCt-1 and RNCt-1. The regression results are

reported in Table 4.3.

Model 1 (N=23) Model 2 (N=23) Model 3 (N=23) Model 4 (N=23)

DV IV Coefficients RPCt-1 Coefficients RNCt-1 Coefficients RPCt-1 Coefficients RNCt-1

B Sig. Std.E B Sig. Std.E B Sig. Std.E B Sig. Std.E Week number -.000362 .273 .000 .001 .012 .000 .000 0.590 0.000 .001 .005 0.000 Lag RPCt-1 - - .636 0.003 0.187 - Lag RNCt-1 - - - .374 .064 .190 F-test 1.266(0.273) 7.642(0.012) 6.050(0.009) 11.460(0.001) Adjusted R2 _0.012 _0.232 _0.325 _0.499 Durbin-wartson 0.740 1.020 2.054 1.989

Table 4.4 regression results of equation ②

From the results, the original proposed model for RPCt-1 was not significant in the F-test

(p=0.273), which means that the model does not fit the data of RPCt-1. The proposed model for

RNCt-1 was significant with 7.642 in F-test (p<0.050), which represented that the model fitted

the data of RNCt-1. According to Liu (2006), the reviews from last week still had significant

effect on the sales of the current week. It could be possible that the effects of the reviews from last week could have influenced the effects of the reviews of the current week. So in order to have more insights of the driving forces of the dynamic effects, I added predictors of Lag effects of RPCt-1 and Lag effects of RNCt-1 in the analyses. The results of the regressions after using lag

variables are shown in Table 4.4 as Model 3 and Model 4. Both of Model 3 and Model 4 reported significant levels in F-test (p<0.050) and had relative higher adjusted R square values after adding extra predict variables. Moreover, Durbin-wartson was used to check if there is autocorrelation issue because the models are analyzed in time dependent basis (Leeflang, 2015). The Durbin-wartson values of Model 1 and Model 2 were way lower than 2, which suggested there was autocorrelation issue. But after adding lag variables, it is shown from Table 4.3 that Durbin-wartson values of Model 3 and Model 4 are close to 2, which indicated that there was no serial correlation issue. Therefore, Model 3 and Model 4 fitted better with the data. Surprisingly, in Model 3, the week number was still not significant with the coefficients of RPCt-1(p>0.050)

but lag effects of RPCt-1 showed a significant positive relationship with RPCt-1 (p<0.050). In

Model 4, the lag RNCt-1 did not have a significant relationship with RNCt-1 (p<0.050) whilst the

week number was positively significant with RNCt-1 (p>0.050). Based on the analyses till now,

(29)

26

4.4.2 Dynamic effects of RP

t

and RN

t

As it proposed previously in H4 that the strength of the effects was driven by the numbers of positive reviews and negative reviews from the corresponding week. For equation ③, the means of the ratios of one-star rating (RNt) for each week and the means of the ratios of

five-star rating (RPt) for each week were created. Mean RPt per week and mean RNt per week

are the independent variables. During the 23 weeks, the means of RPt per weekvaried from

0.0135 to 0.0601 and the means of RNt per week varied from 0.008 to0.0331. The means of RPt

had an average standard deviation of 0.1739 and the means of RNt had an average standard

deviation of 0.1283. The variances of mean RPt and mean RNt were sufficient to continue the

analyses with equation ③. The results of the regressions from equation ③ are reported in Table 4.5. DV IV Coefficients RPt Coefficients RNt

B Sig. Std.E B Sig. Std.E

Mean RPt -1.209 .025 .501

Mean RNt -.302 .813 1.259

F-test 5.827(0.025) 0.58(0.813)

Adjusted R2 0.180 -.045

Table 4.5 regression results of equation ③

From Table 4.5, the values of the F-test show that the proposed model for coefficients of RPt

was significant (p<0.050) while the proposed model for coefficients of RNt was not significant

(P>0.050). So the proposed model for coefficients of RPt fitted the data but the proposed model

for coefficients of RNt did not. The results show that Mean RPt per week was significantly

related to the effects of RPt.

As it can be seen from Figure 4.2, the 23 coefficients of RNt scattered dramatically in the graph

and they did not present a linear trend as the coefficients of RPt. So it could be the reason that

why the proposed model was not significant. As some of the coefficients are negative, it is not possible to analyze in non-linear model like exponential regression model. In order to figure out what drove the changes of the coefficients of RNt, I made an exploration with dummy variables.

I created two dummy variables which could have the possibility of driving the fluctuation of the 23 coefficients of RNt. One was dummy variable of large mean RNt. It was defined as 1 when

the corresponding week had more than average level (0.022) of mean RNt and otherwise it was

0. Another one was promotion dummy. It was made based the frequencies of the products with discounts per week. It was defined as 1 when the corresponding week had a larger than 57.5% of the products with discounts and otherwise it was 0. I performed one-way ANOVA with the 23 coefficients of RNt with these two dummy variables. The results were not significant to show

(30)

27

4.5 Results of the analyses of dynamic effects

From Figure 4.2, it can be seen that all the effects of reviews are dynamic and changing throughout the 23 weeks. After the analyses of the coefficients of RPCt-1, RNCt-1, RPt and RNt,

the driving force of the dynamic effects were explored in the current study. The factors of the dynamic effects of RPCt-1, RNCt-1, and RPt were discovered by the regression analyses. But I

was not able to find out the factor of the dynamic effects of RNt with the analyses and given

data.

According to Table 4.3, there is no significant relationship between the numbers of the weeks and coefficients of RPCt-1. However, the coefficients of RNCt-1 show a positive relationship

with the ages of week (β = 0.001, p-value < 0.050, Std.Error = 0.000). The effects of cumulative negative reviews in the observing period increased with time passed by. Thus, the third hypothesis which states that ―the effects of positive reviews and negative reviews in the observing period decrease with the age of week increases‖ is not supported. At the same time, the coefficients of RPCt-1 were positively related to its lag effects (β = 0.636, p-value < 0.050,

Std.Error = 0.187), which indicated that a larger effect of cumulative positive reviews from last week increased the impact of cumulative positive reviews this week.

According to Table 4.4, the coefficients of RPt were significantly affected by the means of RPt

but this relationship is negative (β = -1.209, p-value < 0.050, Std. Error = 0.501). Meanwhile, the coefficients of RNt were not significantly influenced by the means of RNt. Therefore, the

fourth hypothesis which stated ―the strength of the effects from positive reviews and negative reviews on sales at weekt positively related to the numbers of one-star rating and the numbers of

five-star rating at weekt‖ are not supported. The negative relationship between the coefficients

of RPt and means of RPt suggested that less number of positive reviews received on that week

resulted a larger impact.

Since the results of all the analyses are reported, the results are discussed in the next section.

Dynamic effects of positive and negative online reviews on product sales - Analyses on tablet product reviews from Amazon.com