by
Boy Zelle
January 12, 2015
2 by
Boy Zelle
University of Groningen Faculty of Economics and Business
Master Thesis Marketing Intelligence
January 12, 2015 Supervisor: Dr. H. Risselada Second supervisor: Dr. Ir. M.J. Gijsenberg Billitonstraat 20a 9715 ES Groningen (06) 14444360 b.zelle@student.rug.nl Student number: 1919970
3
“Don't ask me who has influenced me. A lion is made up of all the lambs he's digested, and I have been reading all my life.”
― Giorgos Seferis
“Example is not the main thing in influencing others. It is the only thing.” ― Albert Schweitzer
ABSTRACT
This research entails the effectiveness of the depth of price discounts on sales volumes under the condition of online consumer review characteristics. The results are based on tablet computers from Amazon.com and are discussed in the context of four different price tiers. The results clarify that the combined effects of discount depth and review valence negatively impact sales in tier 2 (average value tablets) and tier 3 (high value tablets). New curious customers are attracted to these products by this combination of a negative review and a discount, which results in an increase of sales. No other moderating effects of reviews on the relation between discounts and sales are found, so it is concluded that businesses that operate online do not have to consider any other simultaneous combination of effects between reviews and discounts for their online platform. However, the individual powers of both reviews and discounts occur in a staged way and should still be considered with care as additional findings of this research confirm and also contradict conventional findings. By this, the research contributes to the fields of pricing and consumer reviews. Managers can use the results to their advantage by critically considering their pricing strategies and managing their reviews.
4 TABLE OF CONTENTS
MANAGEMENT SUMMARY ... 5
I. INTRODUCTION ... 6
1.1. Research objectives ... 8
II. CONCEPTUAL MODEL, LITERATURE REVIEW AND HYPOTHESES ... 9
2.1. Conceptual model ... 9
2.2. Literature review ... 9
2.3. The effect of price discounts on sales ... 10
2.3.1. The effect of price discounts on sales for high involvement products ... 11
2.4. The effects of online consumer reviews on sales ... 12
2.4.1. The effects of online consumer reviews on sales for high involvement products ... 13
2.5. The moderating effect of reviews on the relation between discounts and sales ... 14
2.6. Different effects of a price discount among product tiers ... 15
2.7. Different effects of reviews among product tiers ... 16
III. RESEARCH METHODOLOGY ... 17
3.1. Data collection and preparation ... 17
3.2. Variable adjustments and recalculations ... 19
3.3. Variable creation, estimation and validation: sales volume ... 22
3.4. Including product tiers in the model ... 22
3.5. Research method and plan of analysis ... 22
3.6. Model equation ... 24
IV. RESULTS ... 25
4.1. Descriptive statistics and normality assumption ... 25
4.2. Approach of modelling and information criteria ... 27
4.3. Results ... 28
4.4. Validation of the model ... 32
4.5. Results in relation to the hypotheses ... 34
V. DISCUSSION ... 35
5.1. Conclusions ... 35
5.2. Managerial implications ... 37
5.3. Research limitations ... 38
5.4. Future research directions... 39
REFERENCES ... 41
APPENDIX ... 48
Appendix 1. Calculation, estimation and validation of the sales volume variable ... 48
5 MANAGEMENT SUMMARY
This research investigates the effectiveness of the depth of a price discounting strategy on sales for products with different levels of online consumer review characteristics over time. Quantitative review characteristics are utilized, which are: review volume (amount of reviews), review valence (average rating) and review variance (variance of the rating). The results are analyzed over four distinct price tiers and based on data of the tablet computer market from Amazon.com. The analysis that was performed is based on a hierarchical multiple linear regression model that differs per price tier.
The results indicate that the combination of discount depth and review valence negatively affects sales in tier 2 and tier 3 (respectively average value tablets and high value tablets). This implies that negative reviews combined with a price discount attract as much new curious customers that sales are stimulated for these product groups. Furthermore, no other interaction effects between reviews and discounts are existent at any tier, which indicates that the simultaneously combined effect of both these marketing instruments is not relevant to consider for companies who operate online besides the negative interaction of review valence and discount depth.
However, previous research already pointed out that during the purchasing process, reviews are more influential in the initial consideration phase and price is more influential in the decision stage. This staged relationship is proven to be even more relevant to consider since this research clarified that simultaneous combined effects are predominantly nonexistent.
This research also approves findings of other research in the context of the tablet market, which are that discounting strategies are not always effective in the long run - except for high value products, post-promotion sales dips are found to exist widespread - especially for high priced product tiers and that a high variance in review ratings decreases sales. The findings imply that products that are priced high should not apply a discounting strategy and that products of high value should. Lastly, during a new product introduction and during a new product category introduction (which were the circumstances for tablet computers in 2012), higher prices generate higher sales as a result of the high involvement and the low price sensitivity of pioneering consumers.
6 I. INTRODUCTION
The volume of online purchases is forecasted to continue increasing rapidly (Seghal, 2014; Gill, 2013; Wigder, Miglani, Kashiwagi & Katz, 2013; Centre of Retail Research, 2012) as well as the amount of consumers that consult online sources during (online) purchasing (PricewaterhouseCoopers, 2012). These trends are mainly driven by the benefits that the consumer has during online shopping (PricewaterhouseCoopers, 2012). One major benefit of online shopping are the possibilities of comparison that it offers, of which consumer reviews are an important constitute (PricewaterhouseCoopers, 2012). In fact, online reviews are the most influential source of information to consult before purchasing (Fretwell, Stine, Sethi & Noronha, 2013) and reviews are perceived as a reliable source of information as 70% of the consumers trusts them (Nielsen, 2012).
Since consulted consumer reviews are part of the first stage of the consumer purchase decision process - the consideration stage -, the choice of consumers is significantly influenced by online consumer reviews (Jang, Prasad & Ratchford, 2012). More than three quarters of all consumers who consult online reviews indicate that the reviews have a significant impact on their purchase decision (Lipsman, 2007). The impact of consumer reviews on sales (i.e. the final choices of consumers) is investigated in a vast amount of research. Most of the findings are limited to their research context and the research field as a whole is inconclusive about the general impact of review characteristics on sales. A recent meta-analysis accomplished to generalize most of the findings from this field so far and found that reviews influence sales significantly (Floyd, Freling, Alhoqail, Cho & Freling, 2014). Specifically, ratings of reviews (review valence) exert more influence on sales (both positive and negative) compared to the amount of reviews (review volume) (Floyd et al., 2014).
7
instance, the perceived product quality, which can change dramatically through consulting reviews, strongly influences the considerations of products by consumers while price has a stronger effect on the final decision of product choice (Floyd et al., 2014; Jang et al., 2012). Therefore it is recommended to clearly present consumer reviews in an early stage to consumers who visit retailing websites (Jang et al., 2012). Another interesting finding in the field of pricing and online reviews is presented by Li & Hitt (2010), who found that price in relation to online consumer reviews is a very relevant factor to consider when developing an online marketing and sales strategy. It can be essential to set a new product at the optimal price when introducing it. If the initial price is lower than expected for most consumers (i.e. the value for money is good), this boosts customer satisfaction and results in more positive reviews, which leads to a positive virtuous circle and a market diffusion that is substantially greater (Li & Hitt, 2010). The importance of gaining a more thorough understanding about the effects of pricing strategies on sales and the altering effects of reviews on this linkage is stressed by these findings since online sales are proven to be heavily dependent on this interplay of effects (Jang et al., 2012; Li & Hitt, 2010).
A relevant topic in extension of price strategies and reviews that has no large knowledge base yet is the effect of discounting strategies (temporary price changes) on sales for products in different conditions of consumer review characteristics. This is interesting to investigate further because price discounting is a very commonly applied and popular technique to increase (online) sales (Chen, 2013). Yet, comprehensive research on its exact effectiveness is limited to the context of supermarket products (Nijs, Dekimpe, Steenkamp & Hanssens, 2001), not to mention in combination with main characteristics of consumer reviews such as review volume, review valence and review variance (Floyd et al., 2014; Jang et al., 2012). For supermarket products it is known that discounting is highly influential in the short run and of minor impact in the long run for sales volumes (Nijs et al., 2001). The latter is due to the post promotion sales dip that is caused by accelerated demand for stockpiling (Nijs et al., 2001).
Logically, it is also not known yet how price discounting strategies work out exactly for products with higher levels of involvement compared to the supermarket assortment. This provides a relevant addition to the context of this research since the effects are likely to differ and generate new insights on discounting effectiveness for high involvement products. The focal products for performing the analyses of this research are tablet computers, which can be classified as high involvement goods (Richins & Bloch, 1986; Senecal & Nantel, 2004).
8
research can specifically be allocated to different product tiers. As prices differ over product tiers, so does the price elasticity (Chuang & Tsaih, 2013; Gordon, Goldfarb & Yang, 2013; Hamilton, East & Kalafatis, 1997; Nijs et al., 2003; Kalra & Goodstein, 1998). Therefore sales results per tier may be expected to differ by the effectiveness of discounts and their interplay with reviews.
With this research direction, it can be elucidated whether several consumer review characteristics reinforce or debilitate the positive relationship between a price discount and sales (Nijs et al., 2001) and thereby whether these two marketing instruments should be actively combined or operationalized separately. By the direction of this research, the findings can contribute relevant extensions to the academic field of the dynamics between pricing and reviews and can also increase the understanding about effective price discounting strategies for professionals in organizations that operate in the online domain.
1.1.Research objectives
The main objectives of this research are as follows: (1) to identify how online consumer reviews impact the relationship between price discounts and online sales over time based on commonly used quantitative review characteristics: review volume, review valence and review variance (Floyd et al., 2014; Jang et al., 2012) and (2) to identify how these effects differ for different product tiers that are categorized by price levels. These objectives induce the following research question:
How do online consumer reviews in terms of volume, valence and variance, impact the effect of price discounts on online sales and how does this effect differ for products from different price categories?
9 II. CONCEPTUAL MODEL, LITERATURE REVIEW AND HYPOTHESES
The conceptual model is presented initially in order to ensure a that a proper overview of the material is provided before discussing the literature review so that the material that is covered is comprehended more conveniently.
2.1. Conceptual model
Based on the literature review, the conceptual model that is displayed below in figure 1 is developed. The dotted arrows and their signs are the main effects that are described by the existing literature and the conventional arrows and their signs are the relations as expected from the hypotheses. The expected relations are labelled by their hypothesis number.
Figure 1. Conceptual model: The main effect of a price discount on sales, the main effects of reviews on sales and the moderating effects of review variables and product tiers on this relationship
The contents of the material where the conceptual model is based on are discussed thoroughly in the next section.
2.2. Literature review
10
discussed based on theory regarding product categories and price elasticity. The overview of the literature identifies the current state of the research field and clarifies the key areas and expected relations of this research.
Before discussing the present state of the literature with regard to online consumer reviews and discounting, the most striking findings of the literature review in terms of variables and their effect on sales are summarized in table 1:
Article(s) Main variable(s) Relation to sales
Akaichi, Nayga & Gill, 2014; Hanssens, Parsons & Schultz, 2001;
Nijs, Dekimpe, Steenkamp & Hanssens, 2001
Price discount Highly positive in the short term
Slightly positive in the long term
Hoch, Dreze & Purk, 1994 Price discount Positive
Chen, 2013 Price discount Positive for price sensitive products
Negative for non-price sensitive products Jang, Prasad & Ratchford, 2012;
Floyd, Freling, Alhoqail, Cho & Freling, 2014
Reviews Price
Positive Negative
Lipsman, 2007 Reviews Positive
Zhu & Zhang, 2010 Review valence Positive
Zhang, 2010;
Chen, Wang & Xie, 2011
Review volume Positive
Sun, 2012 Review variance Negative
Diehl & Poynor, 2010 Customer expectations
Customer satisfaction
Negative Positive
Kalra & Goodstein, 1998 Price sensitivity over price tiers Positive
Laurent & Kapferer, 1985; Richins & Bloch, 1986 Senecal & Nantel, 2004
Involvement goods Positive
Table 1. Summary of the main contributions of the literature review
Based on the pivotal findings of the effects that are summarized in table 1, the theoretical consequences of combining them are substantiated to construe the expected relations as presented in the theoretical framework (figure 1).
2.3. The effect of price discounts on sales
11
price promotions are only slightly effective in the long run and thereby in general (Kahn & McAlister, 1997; Abraham & Lodish, 1990), the effect of price promotions in the short run is high (Hoek & Roelants, 1991; Hanssens, Parsons & Schultz, 2001; Akaichi, Nayga & Gill, 2014), more specifically twenty times as high as in the long run (Ataman et al., 2010). Nijs et al. (2001) found similar results: out of 500 supermarket product categories, price promotions result in a boost of category sales in 58% of the cases in the short run, but only in 2% of the cases this effect lasts in the long run. With an average short term price elasticity of -4, this implies that the short term effects of discounts are very effective, even though this effect evens out in the long run by the post promotional sales dip (Nijs et al., 2001). Still, a price discounting strategy increases profits by 15% while an everyday low pricing (EDLP) strategy results in a reduction of profits of 18% (Hoch, Dreze & Purk, 1994). This is because retailers trigger consumers to switch stores by discount offers, which is an advantage that the EDLP strategy can’t benefit from (Hoch et al., 1994).
Note that even though the long term effects of price discounting on sales volumes are rather low, it must be emphasized that these findings hold for a product category of low involvement goods (supermarket products) that cost relatively low amounts of money. It is assumed that for high involvement goods, the effects of price discounts are more intense in the short run and endure better in the long run, since the benefits of purchasing products with price discounts are larger for consumers of high involvement products, due to their higher prices. Also, high involvement goods should be better able to acquire market share and retain this, because consumers are more attached to these goods (Richins & Bloch, 1986).
Van Heerde, Gupta & Wittink (2003) accomplished to gain understanding of how the additional short term sales are exactly generated. The three sources of a short term sales increase during a price discount are: 1. brand switching, 2. forward buying (stockpiling) and 3. increased consumption for the category as a whole. All three sources contribute in equal proportions to the short term sales increase (Van Heerde et al., 2003).
2.3.1. The effect of price discounts on sales for high involvement products
12
before purchasing (Richins & Bloch, 1986). It is also assumed that stockpiling of tablets and increased purchasing in the product category are not very realistic, which automatically also pleads for a more dominant role for brand switching as a contributor to market share during price discounts for these high involvement goods.
With regard to the main effect of this research (the effect of discounting on sales), it is found from the literature of this section that the effectiveness of discounting on sales can be summarized as follows: Price discounts boost sales dramatically in the short run (Hoek & Roelants, 1991; Ataman et al., 2010) and to a minor extent in the long run. (Nijs et al., 2001; Akaichi et al., 2014). This effect is expected to be more positive for high involvement goods (Richins & Bloch, 1986).
2.4. The effects of online consumer reviews on sales
In the majority of the research concerning online consumer reviews and online sales, one or more of the variables review valence, review volume and review variance are focal in relation to sales. Jang et al. (2012) underline the importance of these three main quantitative review elements by stating that the perceived quality and value of a product are affected by these online consumer reviews characteristics. The relative value of these review characteristics is evident: review ratings (valence) are over seven and respectively fifty times more valuable for online retailers in terms of sales impact than the rating variance and the amount of reviews (volume) (Jang et al., 2012). Other research also points out that review valence is more impactful to consumer behavior and ultimately sales in comparison to review volume (Floyd et al., 2014; Lipsman, 2007). For products with a low popularity (‘the long tail’) and thus a relatively low review volume, review valence is even more influential on online sales (Zhu & Zhang, 2010).
13
The effects of the variance of reviews on sales should always be considered relative to average ratings and can be explained as follows: a product that has a high variance in its ratings (which is the characteristic of a niche product) and a low average rating implies that only a certain target segment could still appreciate the product (Sun, 2012). In this case, demand is boosted through the reviews, whereas products with a high average rating and a high variance of review ratings show a reduced demand, since they discourage marginal consumers who can see this as a sign of uncertainty (Sun, 2012). However the impact of review variance on sales volumes differs based on the combination with the review valence of the focal product (Sun, 2012; Jang et al., 2012), for regular, non-niche products with average review valence it can be concluded by the extrapolation of these research results that review variance negatively impacts sales volumes, since this triggers uncertainty about the product characteristics and value.
Next to the quantitative review characteristics, research shows that qualitative review characteristics impact sales significantly as well by their in-depth information, by the quality inferences that are made by consumers (Chevalier & Mayzlin, 2006; Zhang, 2010; Yoo & Gretzel, 2009) and by the trust that consumers have in the written content of the reviews (Yoo & Gretzel, 2009; Lee & Ma, 2012). In fact, written data has a stronger impact on sales than review ratings (Chevalier & Mayzlin, 2006; Zhang, 2010).
The current research regarding the impact of reviews on sales is not completely congenial due to their research limitations that inhibit generalization (Floyd et al., 2014; De Maeyer, 2012). However many findings show significantly positive results of the influence of reviews characteristic on sales, other findings show that for instance negative reviews can boost sales as well in some contexts (Berger, Sorensen & Rasmussen, 2010). In this instance the recent meta-analysis of the field by Floyd et al. (2014) helps to generalize the main findings with regard to review volume and review valence: both review elements have a significant relation with sales. This means that the following holds regarding the effect of review volume and review valence: a large amount of positive reviews increases sales while a large amount of negative ones make sales decrease; and a high rating makes sales increase while are low rating results in a sales decrease. The following holds regarding the effect of review variance: the general effect of review variance on sales is negative; the greater the variance of a rating, the greater the negative impact on sales (extrapolated from Sun, 2012).
2.4.1. The effects of online consumer reviews on sales for high involvement products
14
that typically receive above-average effort from consumers who are willing to purchase them (Richins & Bloch, 1986). This is caused by high prices and product specifications that are hard to assess (Richins & Bloch, 1986). Due to the high involvement and the high susceptibility of consumers who buy high involvement goods, the impact of online consumer review characteristics (valence, volume and variance) on consumer behavior and thereby sales (Jang et al., 2012) is assumed to be larger compared to low-involvement goods. This line of reasoning is approved by other research: consumers who are considering high-involvement goods are more susceptible to consumer reviews (Senecal & Nantel, 2004). Therefore, the impact of online consumer review volume, valence and variance on sales is relatively strong in case of high involvement goods (Senecal & Nantel, 2004).
2.5. The moderating effect of reviews on the relation between discounts and sales
The main effect of price discounts on sales (Van Heerde et al., 2003; Bijmolt et al., 2005; Ataman et al., 2010; Hoek & Roelants, 1991; Akaichi et al., 2014; Nijs et al., 2001) and the main effects of volume, valence and variance of consumer reviews on sales (Jang et al., 2012; Floyd et al., 2014; Lipsman, 2007; Zhu & Zhang, 2010; Zhang, 2010; Sun, 2012; Chen et al., 2011) have theoretical implications for the consideration and choice stages that consumers experience through the purchasing process which are identified by Jang et al. (2012). Both reviews and price information significantly influence the considerations and choices of consumers, only this effect is stronger for reviews in the consideration phase and stronger for price in the choice stage (Jang et al., 2012). This means that consumers combine the knowledge they acquire about the experiences of other consumers and the product prices in their decision making. More specifically, since reviews are more influential in the consideration phase, they should form a condition for the effectiveness that the price or price discount can have in the choice stage. This is because the price discount is only effective for a selection of considered products. Next to this, the effectiveness of the discount is also determined by the review characteristics of this selection of products since price and review information are scrutinized together by the consumers (Jang et al., 2012). In conclusion: reviews do not only directly impact sales numbers (Jang et al., 2012; Floyd et al., 2014; Lipsman, 2007; Zhu & Zhang, 2010; Zhang, 2010; Sun, 2012; Chen et al., 2011) but also are also expected to impact sales numbers via the relation between price discounts and sales during the consideration-choice process of purchasing (Jang et al., 2012).
15
review valence and review variance are expected to show the same directions as their direct (main) effects on sales. This is expected because the positive impact of a price discount on consumer choice (Floyd et al., 2014; Jang et al., 2012) for a focal product should logically be reinforced in case this product was selected on the basis of review characteristics that have a positive impact on consumer considerations (which are a high review volume, a high review valence and a low review variance). Indeed, the expectations of the consumer in terms of value for money will be exceeded in case a product of the preferred selection is discounted (Diehl & Poynor, 2010). Because of this it is expected that discount effectiveness is enhanced by beneficial review characteristics. From this theoretical foundation the following relations are expected to arise in this research:
Hypothesis 1: The positive relationship between price discount and sales is positively impacted by review volume
Hypothesis 2: The positive relationship between price discount and sales is positively impacted by review valence
Hypothesis 3: The positive relationship between price discount and sales is negatively impacted by review variance
2.6. Different effects of a price discount among product tiers
Not only the power of discounts and reviews are important to consider when implementing price discounting strategies online. The different levels of price sensitivity (i.e. the price elasticity) of different price categories of products are important to contemplate as well in order to decide whether applying a price discount is prudent for a product from a specific price tier. Price promotions may not be an effective strategy for every brand and their associated price class (Chuang & Tsaih, 2013) and price elasticities can differ dramatically across and within product categories and across brands (Gordon et al., 2013; Hamilton et al., 1997; Nijs et al., 2003).
16
In conclusion, because products from a higher price tier have a higher price sensitivity, a greater importance to consumers and a corresponding higher involvement, a discount should be more effective in higher tiers. This linear relationship of price sensitivity based on the price and the value of a product might change under certain circumstances, for instance by including the amount of advertisements to the analysis (Van Heerde, Gijsenberg, Dekimpe & Steenkamp, 2013) or by including consumers price perceptions (Chandrashekaran, 2012). Since these variables are outside the aim of this research, a linear relationship of price sensitivity is adopted. A proper guideline in any case for discount strategies in relation with price sensitivity of products originates from a common wisdom in retailing: for products that are known to be price sensitive, a frequent discounting strategy to increase sales should be operationalized (until a threshold frequency) and for products with a low price sensitivity, the opposite holds (Chen, 2013).
Since products with a high value and a high price are most price sensitive in general (Karla & Goodstein, 1998; Lamb et al., 2009; Jang et al., 2012), a discounting strategy is expected to be most effective for products that are classified as high price tiers. This expectation is reflected in the next hypothesis:
Hypothesis 4: The positive relationship between price discount and sales is positively impacted by the price tier of products
2.7. Different effects of reviews among product tiers
Differences across price tiers are expected as well for the effects of reviews on the relationship between price discount and sales. The interfering effects of review volume, review valence and review variance are supposed to be stronger for higher price tiers because consumers are more involved with products from a high tier (Laurent & Kapferer, 1985; Richins & Bloch, 1986) and are therefore more sensitive for review information as involved consumers perform a more active search, more active information processing and a more extensive consideration of decisions (Laurent & Kapferer, 1985). Indeed, in case of high involvement goods like goods from a high price tier, consumers are also more susceptible for consumer reviews (Senecal & Nantel, 2004).
17
the relatively large amount of involved consumers (Laurent & Kapferer, 1985). A reinforcement of the effects of reviews is predicted for higher price tiers:
Hypothesis 5: The impact of review volume, review valence and review variance on the positive relationship between price discount and sales is positively impacted by the price tier of products
All relevant theories and the ensued hypotheses have been discussed now. The methods of testing these hypotheses in order to answer the research question are discussed in the subsequent chapter.
III. RESEARCH METHODOLOGY
First, the collection and preparation of the data are discussed. Next, adjustments to the dataset and the calculation and validation of the dependent variable sales volume are covered, as well as the modelling options with regard to the product tier variable (e.g. a unit-by-unit model per tier or a pooled model for all tiers together). Finally, the exact research method is discussed and the hypothesized model is constructed.
3.1. Data collection and preparation
As discussed in the previous chapter, the main effect and interaction effects of this research subject are respectively the effect of a price discount on sales volumes and the interaction of either the volume, valence or variance of online consumer reviews with a price discount on sales volumes. These relations are analyzed with respect to different product tiers.
18
Based on a preliminary analysis, a selection of the dataset of 129 products is suitable for further analysis since this selection contains a complete set of information over time in units of weeks. As opposed to the rest of the products in the dataset, this selection is not missing any values with regard to price and review variables. However the complete dataset is larger than this selection of 129 products and for that reason the outcomes of the research may differ, including all products with many missing values of two variables that are critical for this research (price and reviews) in the analysis may yield even more deceptive findings. Because of this, continuing with this selection of products is considered as the best alternative.
To give some initial insight in the dataset, table 2 represents the selection of relevant variables that are present in the original dataset:
Variable group Time series variable(s) per product
Product Title, Brand, Model
Price Amazon price
Review Review volume, Review valence, Verified purchase
Sales Sales rank
Table 2. Summary of variables in the dataset of Wang et al. (2014)
19 3.2. Variable adjustments and recalculations
New variables are created by customizing the original ones in order to perform an analysis that is more suitable for this research. The information in table 3 summarizes the variables that are created by recalculation of the original variables:
Variable group Time series variable(s)
Product Tier (tier 1 = $750 or higher, tier 2 = $749 - $500, tier 3 = $499 - $250, tier 4 = $249 or lower)
Price Discount depth, Lagged discount depth
Review Review variance, Cumulative review volume, Cumulative review valence, Cumulative review
variance
Sales Lagged sales volume
Table 3. Summary of customized variables based on the dataset of Wang et al. (2014)
The customized product variable ‘tier’ that is presented in the synopsis in table 3 is of relevance for this research since it allows to identify differences among product tiers by the price classifications of product groups. Based on the price levels of the data, this variable is created on the basis of four arbitrary cut-off prices that are approximately equally distributed across the price range of all products, which are: tier 1 = $750 or higher, tier 2 = $749 - $500, tier 3 = $499 - $250, tier 4 = $249 or lower. One should be aware of the fact that this generates tier groups that are unequal in size due to the Pareto distribution of most of the included variables (group sizes are presented with the descriptive statistics in table 5).
20
The discount depth and the two lagged discount depth variables (t-1 and t-2) identify temporary price discounts (so no long-term price reductions). The lagged equivalents of the discount depth variable are included in the model in order to account for longer term effects of pricing on sales, for instance the effects of a post promotion dip could be detected. The discount depth variables are value zero in case there is no price discount and are the absolute value of the price difference of the focal week compared to the previous week in case a price discount is present. For further clarification, when the discount depth variable identifies a price increase after a price decrease, this is indicated with the value zero, which marks the end of the temporary price discount period.
The inclusion of review variables in the model of analysis requires adjustments to the original data as well: the variables of the total amount of reviews, the current average rating and the current average variance of the rating are recalculated in order to create the cumulative equivalents of them. By this approach lagged, versions of these variables are not necessary to include in the model anymore. This is because the review variables already include their past effects next to their present effects, since the cumulative variables are adjusted by the latest value for each week that is recorded.
The fact that all historic information is present in one variable does not only simplify and improve our analysis, but also prevents issues concerning multicollinearity as is often the case with lagged variables. Multicollinearity issues are also expected with regard to the lagged variables that are included in the model (lagged sales volume over one week, lagged discount depth over one and two weeks), but due to their theoretical relevance and expected contribution to the model they are still included. Also, since the most suitable model proves to be a multiplicative one in later analysis (see section 3.6), potential multicollinearity issues are of less concern (for further elaboration see section 4.3: Friedrich, 1982; Brambor, Clark & Golder, 2005). Another option is to exclude or multiply some variables that cause multicollinearity issues in case these problems are too severe (Miller & Yang, 2007)
21
cause direct effects to sales because of their theoretical relevance, which is the potential occurrence of post-promotion dips (Nijs et al., 2001).
In addition, note that the sales rank variable of table 2 is not included in this table, since it is not edited but replaced by the newly calculated variable sales volume (discussed in section 3.3).
The pretreatment of data can be essential for the findings of the research, since the transformation(s) should be suitable to the data in order to improve it and there should be a theoretically sound reason for transforming the data (Van den Berg, Hoefsloot, Westerhuis, Smilde & Van der Werf, 2006). These recommendations are incorporated in the discussion that follows now.
22 3.3. Variable creation, estimation and validation: sales volume
Since sales volume is an interesting and more relevant variable to include in this research compared to sales rank, this variable was created as well. Sales volume is a better alternative as a dependent variable compared to sales rank since it contains explicit dynamic and detailed numerical sales information instead of solely the rank order of the amount of sales per product. The exact methodology and theoretical fundaments for the construction of the sales volume variable can be found in appendix 1.1. The general principle is to create the sales volume variable based on existing data and theoretical grounds and afterwards estimate its alpha - and beta coefficient values in a model. The coefficient values are validated on the basis of results of other research that works with sales volumes that are obtained from sales ranks as well. Since the model is approved to be valid (appendix 1.1), it is used to create the sales volume data by estimation.
3.4. Including product tiers in the model
In the previous sections of the research methodology chapter, all variables (and the required adjustments to them) that are included in the model of analysis are covered. Before discussing the actual research methodology and the full model of analysis (e.g. which variables to include exactly), it is essential to determine whether the tier variable should be an independent variable which is an element of a pooled model or that the tier variable demands separate models of analysis in the form of a unit-by-unit model approach. This can be determined by comparing the fit of the models by the Chow-test. The test results indicate that a pooled model is not allowed for this data and therefore the unit-by-unit approach per product tier is applied, as is further discussed in section 3.6.
3.5. Research method and plan of analysis
With regard to the research questions, the tablet computer dataset from Amazon.com of Wang et al. (2014) and the modified and added variables allow us to perform the analyses that can ultimately answer the research questions. The analyses are based on the relations concerning all relevant variables as discussed in the hypotheses of this research (also see figure 1, the conceptual model). The Amazon price variable is included as well because it is so closely related to discounts and therefore increases the realism and explanatory power of the models.
23
evaluation of several models as the method of analysis of the hypothesized relationships it is concluded that these make suboptimal use of the tablet dataset variables compared to the approach of stepwise hierarchical multiple linear regression and are therefore inappropriate to serve as the basis of analysis.
Hierarchical multiple linear models allow for including specific interaction effects and the comparison of model fit to detect a potential (significantly) improved model fit of models that include interaction effects. This hierarchical modelling approach matches the purposes of this research since it has the benefit of assessing the exact impact of the interaction effect between temporary price discounts and review variables on sales volume and enables to identify the optimal model. Therefore, hierarchical multiple linear modelling is selected as the method of analysis. More specifically, the procedure that is deployed during this analysis is summarized in table 4:
Model number Main effects Core interaction effects
Model 1: Full model Lagged sales volume
Price Amazon Discount depth (DD)
Lagged discount depth (DDt-1)
Lagged discount depth (DDt-1)
Cumulative review volume (CRVol) Cumulative review valence (CRVal) Cumulative review variance (CRVar)
DD * CRVol DD * CRVal DD * CRVar
Model 2: Basic model Same as model 1 None
Table 4. Overview of the hierarchical multiple linear regression model approach
The initial investigation comprises discussing the descriptive statistics of the model. Then, the full model is estimated, which includes the main effects the interaction effects as specified in table 4. Afterwards, the basic model is estimated, which consists of the main effects only. The results are discussed together. Both models involve automated step-wise modelling in order to identify the extent to which each variable is adding explanatory power to the models. The variables are assessed based on the outcomes of the models by their direction, strength, explanatory power and significance of the relations that are expected between them. After clarifying this, conclusions based on these outcomes can be drawn.
24
heteroscedasticity and a normal distribution. This validation and re-estimation is done in order to assure the correct interpretation of the analyses since if the just-described conditions are not met, results may give wrong estimates of parameters and significance levels, which may result in erroneous conclusions.
3.6. Model equation
The full form of the multiplicative logarithmic model is presented in equation (1). Based on the conditions of this data, a unit-by-unit approach per tier (as indicated with the indices) is applied (as suggested by Leeflang, Bijmolt, Pauwels & Wieringa, 2015). Logically, the alpha- and beta-values differ per estimated model across tiers and the beta-values of the variables differ across products. This is notified with the index variable i: if the indices are positioned next to the α or a β, this indicates different values for the tiers, while if the indices are placed next to a variable parameter this relates to different values for a specific product. The full model equation represents the following elements (12 independent variables in total including the intercept):
ln(Sit) = αi + lnβ1i(Sit-1) + lnβ2i(Pit) + lnβ3i(DDit) + lnβ4i(DDit-1) + lnβ5i(DDit-2)
lnβ6i(CRVolit) + lnβ7i(CRValit) + lnβ8i(CRVarit) + lnβ9i(DDit*CRVolit) +
lnβ10i(DDit*CRValit) + lnβ11i(DDit*CRVarit)
(1)
Where the variables represent the following:
t = indicator of time; represents the specific week that pertains the value of the
variable
i = index for product tier (1 = $750 or higher, 2 = $749 - $500, 3 = $499 - $250,
4 = $249 or lower) or individual product
αi = constant for tier i
βi = coefficient of the variable
Sit = sales volume
Sit-1 = lagged sales volume (one week lag)
Pit = Amazon price
DDit = depth of price discount
DDt-1 = lagged depth of price discount (one week lag)
DDt-2 = lagged depth of price discount (two weeks lag)
25
CRValit = cumulative review valence
CRVarit = cumulative review variance
PDit*CRVolit = interaction between price discount and cumulative review volume
PDit*CRValit = interaction between price discount and cumulative review valence
PDit*CRVarit = interaction between price discount and cumulative review variance
This full model serves as the basis of the analyses that are discussed in section IV.
IV. RESULTS
The findings of the data analysis based on hierarchical (stepwise) multiple linear regression modelling can be seen as the result of an iterative process of optimizing the model and finding the effects of each variable in relation to each other and the dependent variable. For each model, four different variants based on the different tiers are discussed (since a unit-by-unit model is applied). Based on the developments and significance of changes of the models per tier, conclusions regarding the expected relationships among the variables can be drawn. Initially, the descriptive statistics of the dataset are discussed briefly. In the section that follows (section 4.3), the validity of the models is discussed in detail. Prior to section 4.2, it must be noted that the models discussed all abide the validation criteria and thus it can be concluded the findings from these models are valid.
4.1. Descriptive statistics and normality assumption
26
Variable All tiers Tier 1 Tier 2 Tier 3 Tier 4
N 3225 300 500 750 1675
Mean sales rank 3683.02 3944.60 6071.80 1186.86 4040.79
Mean sales volume 19.96 19.51 19.19 18.51 20.92
Mean Amazon price 335.07 1134.70 542.58 306.76 142.58
Mean amount of price discounts 0.30 0,39 0.28 0.34 0.26
Mean discount depth 4,76 11.80 7.91 5.09 2.41
Mean cumulative review volume 193.45 13.02 48.69 137.67 293.95
Mean cumulative review valence 3.27 3.64 3.77 3.63 2.90
Mean cumulative review variance 1.09 0.58 1.03 1.36 1.08
Segment classification All tablets Premium
tablets Average value tablets High value tablets Budget tablets
Table 5. Mean values of variables and product tier typologies and characteristics
The exact differences between the tiers of the price categories of products are discussed earlier in this chapter (section 3.2): it was indicated that all tiers differ significantly from each other in terms of sales volumes. Most tiers differ significantly from all other tiers on all variables. Since a full description of these significant differences is inefficient and transcends the objectives of the analysis of the tiers, describing the core differences among the tiers is applied as an approach. Table 5 summarizes the differences across tiers by classifying them into product segments. The classification is based on the descriptive statistics of the tiers and the Post Hoc LSD results, which also include the exact differences of effects between tiers.
27 4.2. Approach of modelling and information criteria
In order to test the hypotheses of this research, interaction effects are investigated since the relations of interest are characterized as moderators. In order to optimally test the influence of interaction effects, two models have to be compared by the application of nested models. First, the full model which includes all interaction effects is estimated and afterwards the basic model is estimated. By this, the differences between the models in terms of the effects of the variables and the model fit can rule whether the expected interaction effects of reviews are significant and if this is the case what their exact impact is. Obviously, this approach of stepwise hierarchical multiple linear regression modelling is applied four times, separately for all tiers (since a unit-by-unit model is found to be optimal). The variables are included in the models stepwise, e.g. the variable with the highest explanatory power is included first, the variable with the second highest explanatory power is included secondly, and so on. By performing the regression stepwise, the least significant variables that are added to the model can be discovered. This is of importance because of the parsimoniousness criteria for good models, or as Keough & Quinn (2001) state the model should not become over-fitted by adding just one variable and not increasing the explanatory power of the model.
There are multiple methods for analyzing the model fit of regression models by information criteria. In practice, the values of the most common methods are compared across the nested models in order to assess whether all information criteria identify a change in model fit (Dziak, Coffman, Lanza & Li, 2012). The model with a higher value of the information criterion has a better model fit (Dziak et al., 2012).
Maximum Likelihood Estimation and Log-likelihood estimation are common methodologies of comparing model fit, where a comparison is made whether the model performs better than the null model (Engle, 1983). In case these information criteria are significant, the model of analysis outperforms the null model (Engle, 1983).
Several well-known variants on the Log-likelihood functions are AIC (Akaike’s Information Criterion), AICC (finite sample corrected AIC), BIC (Bayesian Information Criterion) and CAIC (Bozdogan’s consistent AIC) (Akaike, 1973; Schwarz, 1978; Bozdogan, 1987) which main differences with Log-likelihood are their penalties by their emphasis on parsimonious models or the predictive value of models (Dziak et al., 2012; Claeskens & Hjort, 2008; Lin & Dayton, 1997). As suggested by Dziak et al. (2012), the fit of the models of this research are compared by all information criteria that are discussed here.
28 4.3. Results
The results of the stepwise hierarchical multiple linear regression modelling of the full model and the basic model are presented for each tier in table 6:
Tier Model Sign. R^2 Adj. R^2
AIC AICC BIC CAIC Likelihoo d-ratio Log-likelihood 1 Full 0.000 0.827 0.825 564.908 566.181 613.057 626.057 529.702 -269.454 Basic 0.000 0.827 0.825 561.194 561.955 598.231 608.231 527.416 -270.597 2 Full 0.000 0.807 0.803 1140.179 1140.928 1194.968 1207.968 822.171 -557.089 Basic 0.000 0.805 0.803 1137.485 1137.935 1179.631 1189.631 818.865 -558.743 3 Full 0.000 0.836 0.834 1366.009 1366.504 1426.070 1439.070 1370.232 -670.005 Basic 0.000 0.836 0.834 1373.121 1373.419 1419.322 1429.322 1357.121 -676.560 4 Full 0.000 0.796 0.796 3832.930 3833.149 3903.436 3916.436 2668.113 -1903.465 Basic 0.000 0.796 0.796 3828.602 3828.734 3882.838 3892.838 2666.441 -1904.301
Table 6. Overview of model significance and model fit of all tiers and all models
All models are significant and have a sufficiently high explanatory power (R-squared and adjusted R-squared values). The explanatory power between the full models and the basic models does not differ within the tiers except for tier 2, where the full model has a slightly higher R-square value. For each tier, the full model has a better fit according to the information criteria, which makes sense since the full model includes additional variables, which are the interaction effects between discounts and reviews. The Likelihood-ratio tests all show significant results so all models perform significantly better compared to the null model.
However the information criteria show that the full model is a better model compared to the basic model across all tiers, this is not reflected in by the values of the R-squared statistics for the full model and the basic models of tier 1 and 4 since there are no differences between them. This implies that the interaction effects that are added to the basic models of tier 1 and 4 have no additional explanatory power to the basic model.
29
presented in this section (table 7 and table 8). The other tables with an oversight of the impact of the variables of the full models of tier 1 and 4 are displayed in Appendix 2 (respectively table A2 and A3) since these results don’t contribute to answering the research question any further. The effects of the variables of all tiers are still discussed and they are also summarized in comparison with the expectations from the hypotheses in table 9.
Variable Sign. Rank of
explanatory power Unstandardized β Standard error Standardized β Intercept 0.000 1 -0.713 0.041 n.a.
Lagged sales volume 0.000 2 0.511 0.024 0.510
Amazon price 0.000 3 0.810 0.027 0.602
Discount depth 0.333 n.a. 0.003 0.095 0.003
Lagged discount depth t-1 0.000 6 -0.85 0.023 -0.74
Lagged discount depth t-2 0.000 8 -0.050 0.023 -0.044
Cumulative review volume 0.000 4 -0.260 0.029 -0.221
Cumulative review valence 0.572 n.a. -0.213 0.070 -0.130
Cumulative review variance 0.000 7 -0.196 0.067 -0.063
Interaction Discount depth and Cumulative review volume
0.519 n.a. -0.049 0.036 -0.147
Interaction Discount depth and Cumulative review valence
0.000 5 -0.054 0.015 -0.075
Interaction Discount depth and Cumulative review variance
0.362 n.a. 0.052 0.051 0.039
Table 7. Significance, strength and direction of effects and standard error of variables of tier 2 (bold = sign. variable)
Variable Sign. Rank of strength of effect Unstandardized β Standard error Standardized β Intercept 0.000 1 -0.244 0.025 n.a.
Lagged sales volume 0.000 3 0.373 0.018 0.376
Amazon price 0.000 2 0.824 0.019 0.640
Discount depth 0.006 n.a. 0.502 0.181 0.453
Lagged discount depth t-1 0.000 6 -0.059 0.017 -0.052
Lagged discount depth t-2 0.169 n.a. -0.023 0.017 -0.021
Cumulative review volume 0.000 4 -0.269 0.016 -0.319
Cumulative review valence 0.000 5 -0.598 0.122 -0.090
30
Interaction Discount depth and Cumulative review volume
0.118 n.a. -0.020 0.013 -0.079
Interaction Discount depth and Cumulative review valence
0.011 n.a. -0.311 0.122 -0.436
Interaction Discount depth and Cumulative review variance
0.178 n.a. 0.055 0.041 0.052
Table 8. Significance, strength and direction of effects and standard error of variables of tier 3 (bold = sign. variable)
Many main effects that were expected according to the literature are also present in the results but just like it is the case with interaction effects, these differ also per tier in some cases (table 9). Also, some effects are found by the inclusion of additional variables that are not present in the conceptual model but these are still findings that are worth mentioning. An overview of all present (significant) effects and also the absence of effects is given in table 9. The findings are shortly pointed out in this section. The main effects and the interaction effects are discussed in more detail in the conclusion section (section 5.1.)
Variable Tier 1 Tier 2 Tier 3 Tier 4
Impact on sales Impact
sales Comparison with hypothesis Impact sales Comparison with hypothesis Impact sales Comparison with hypothesis Impact sales Comparison with hypothesis
Lagged sales volume + a.f. + a.f. + a.f. + a.f.
Amazon price + a.f. + a.f. + a.f. + a.f.
Discount depth n.e. x n.e. x + n.e. x
Lagged discount depth t-1 - - - n.e. x Lagged discount depth t-2 - - n.e. x n.e. x Cumulative review volume - ≈ - ≈ - ≈ - ≈ Cumulative review valence - ≈ n.e. x - ≈ - ≈ Cumulative review variance - - n.e. x - Interaction Discount depth and Cumulative review volume
31 Interaction Discount depth and Cumulative review valence n.e. x - ≈ - ≈ n.e. x Interaction Discount depth and Cumulative review variance
n.e. x n.e. x n.e. x n.e. x
Table 9: Overview of effects of variables across tiers in relation to their hypothesis
Interaction effects are only present in tier 2 and tier 3, specifically review valence moderates discount depth so that the combination of the effects of these variables becomes negative for both tiers. Note from table 7 and A3 that during the stepwise modelling this interaction effect increased the explanatory power of the model for tier 2 while this was not the case for tier 3. This implies for tier 3 that even though the effect is present, the effect is not very large in terms of explanatory power in comparison with other significant variables. However the main effect of discount depth is insignificant for tier 2, the significantly negative review valence variable moderates this effect so that the combination of a discount and the valence of a review result in a negative impact on sales (table 7 and table 9). The significant positive impact of discount depth on sales in tier 3 becomes negative when it is moderated by the negative review valence variable (table 8). This negative interaction effect for both tiers cannot indisputably be explained from the literature review, however an attempt is done in section 5.1 (conclusions). With regard to the tiers that are involved with this significant interaction effect, it was expected that higher tiers would see stronger effects, which is to a small degree the case since this effect is not found in tier 4 but not in tier 1 as well. The hypotheses of this research are compared to these findings in section 4.5.
32
results unreliable. In conclusion, the negative two-way interaction of review valence and discount depth for tier 2 and tier 3 are the only justifiable interaction effects.
Now, the individual variables and their main effects as presented in table 7-9, A2 and A3 are discussed briefly. The variables lagged sales volume and Amazon price relate significantly positive to sales across all tiers. Discounts only have a significant positive impact for tablets from tier 3, while the lagged versions of discount show significant negative effects for tablets from tier 1 to 3 (for a lag of t -1) and tablets from tier 1 and 2 (for a lag of t-2). Cumulative review volumes and cumulative review valences are proven to be significantly negative for sales outcomes across all tiers and cumulative review variance is significantly negative for all tiers except tier 3. These findings are discussed in the context of theoretical backgrounds in section 5.1.
4.4. Validation of the model
In order to validate the optimal models of each tier, their statistical validity, face validity, multicollinearity issues and predictive validity are discussed and the assumptions regarding the residuals of the dependent variable (sales volume) are tested. These assumptions are respectively: the disturbance terms are uncorrelated over time, e.g. no autocorrelation, no heteroscedasticity and a normal distribution of the disturbance term. The conditions for a valid model are discussed for the full models of each tier.
The face validity, which includes the assessment of the directions of the significant variables, is impossible to judge accurately for this model because of adjustments to the data (log linearization specifically) of the original dataset. However, the face validity is assessed to be not completely sufficient, since some unexpected, unrealistic effects are observed in several tiers: negative effects of review volume and valence and negative effects of the interactions of review valence and discount. This is striking but not solvable because the research complies to all preliminary requirements and all potentially relevant or explanatory variables are included in the analysis.
33
The condition of the absence of autocorrelation in the disturbance terms are met since the Durbin Watson statistic values of all models are close within the benchmark boundaries given the circumstances of this research. More specifically, the values of all models (so both basic and full) vary between 1.200 and 1.467, while the boundaries for the basic models are 0.75048 – 2.17427 and for the full models these are 0.43119 and 2.76111 (Savin & White, 1977). Nevertheless, an attempt has been done to improve the Durbin Watson statistic value by making it closer to the optimal value 2 with means of applying GLS (Generalized Least Squares) to the data. Unfortunately, this attempt did not result in any changes in the Durbin Watson value that are worth mentioning. It is assumed that this is a result of the Pareto distribution of the data since this persistent characteristic also endured after other adjustments to the data in attempts to change its distribution.
The condition of the absence of heteroscedasticity in the disturbance terms of the dependent variable (sales volume) is met since the statistic of the Goldfeld-Quandt test for heteroscedasticity is insignificant for all focal models of the tiers (which are the full models) as displayed in table 10:
Tier SSR # Observations # Parameters Degrees of freedom F-value P-value
1 203,1392 vs. 257,8331 150 vs 150 11 vs 11 139 vs 139 0,787871 0.65093474 2 165,1315 vs. 278,1153 250 vs 250 11 vs 11 239 vs. 239 0,593752 0.80249741 3 173,6919 vs. 158,698 375 vs. 375 11 vs 11 364 vs. 364 1,094481 0.44445584 4 831,3895 vs. 815,8394 837 vs. 838 11 vs 11 826 vs. 827 1,020294 0.48719545
Table 10. Overview of input and output for detecting heteroscedasticity by the Goldfeld-Quandt test
Since all tests are found to be insignificant, the null hypothesis is accepted, which implies that there is no heteroscedasticity among the residuals of the dependent variable of the focal model.
34
outliers and further in the process even the exclusion of values that are at the borders of the distribution but definitely no outliers, the normality condition is still unmet. This is also supported by the Shapiro-Wilk and Kolmogorov-Smirnov tests that prove that the residuals are not completely normally distributed since these tests have significant outcomes. Since no other solutions are possible to overcome this restriction, it is emphasized that the plots show that the distribution of error terms is very close to normal in case the outliers are neglected. This is theoretically allowed since the power law of the Pareto distribution is applicable to this data, which explains why the distribution is not perfectly normal. Any further improvements regarding the normality of the data are impossible as the dataset is already log linearized and mean centered as well.
The predictive validity of the full models across the tiers cannot be assessed as a result of that the dependent variable that is focal in this research is already based on variables that are predicted by the model.
Based on this analysis of the distribution of the data and the distribution of the residuals of the full models of all tiers, it is concluded that all optimal models are valid to use despite some peculiarities in the data due to its Pareto distribution. Therefore, it is concluded that it is safe to draw conclusions from the research outcomes. The section that follows therefore describes the results in relation to the hypotheses.
4.5. Results in relation to the hypotheses
The hypotheses are tested on the basis of the results in this section (as is summarized in table 9 as well). The next chapter bases its conclusions on these inferences.
Hypothesis 1: rejected: the positive relationship between price discount and sales is not significantly influenced by review volume.
Hypothesis 2: rejected: the positive relationship between price discount and sales is significantly negatively influenced by review valence in tier 2 and tier 3.
Hypothesis 3: rejected: the positive relationship between price discount and sales is not significantly influenced by review variance.
35
Hypothesis 5: rejected: the expected moderating impact of the review variables where hypothesis 5 is based on do not exist and next to this do the main effects of reviews not differ per tier.
Despite the fact that none of the expected relations is actually present in the data, there are still sufficient interesting findings to discuss in the next chapter.
V. DISCUSSION
In the final section, the conclusions are discussed thoroughly, managerial implications are presented, research limitations are discussed and future research directions are provided based on the results and conclusions of this research.
5.1. Conclusions
36
From this it is concluded that regarding the tablet computer market, no positive moderating effects of reviews are present on the impact of discounts on sales. This contradicts what was expected on the basis of the consideration-choice process during purchasing (Jang et al., 2012). From this, it is inferred that discounts, review volume and review valence definitely impact sales positively (Hoek & Roelants, 1991; Hanssens, et al., 2001; Akaichi et al., 2014; Ataman et al., 2010; Nijs et al., 2001; Hoch et al., 1994), but that they are regarded as separate elements of influence which don’t enhance their combined power positively in a simultaneous way. However, the staged relationship between reviews and price which is characterized as reviews being more influential in the initial consideration stage and price being more important in the stage of final choice (decision stage) (Jang et al., 2012) is now considered to be of even greater relevance to consider. Indeed, when both reviews and price are so influential in separate stages, one should critically take this into account when designing or optimizing an online sales structure.
However the effects that were expected for this research are not found, some interesting main effects of variables are found and confirmed within the context of this research, which helps to generalize the findings of related research.
First of all, the main effect of price discount depth is to a absent to a large extent across tiers while it was expected to be clearly present throughout all tiers (Hoek & Roelants, 1991; Hanssens et al., 2001; Akaichi, et al., 2014; Nijs et al. 2001; Hoch et al., 1994). Also, this would not have been expected due to the fact that during almost one third of all the weeks a price discount is given (table 5). As previous research already indicated, discounting policies are in many cases not very effective in the long run (Hoek & Roelants, 1991; Nijs et al., 2001; Ataman et al., 2010; Akaichi et al., 2014; Kahn & McAlister, 1997; Abraham & Lodish, 1990). The dataset covers data of 24 weeks, so the short term positive impact of a discount on sales may still diminish over a timeframe of this size. Tier 3, which includes high value tablets, is the only tier where a price discount has a significant positive effect. This can be explained by the fact that high involvement products of a high value (i.e. good price-quality ratio) are more price sensitive (Kalra & Goodstein, 1998; Lamb et al. 2009; Jang et al., 2012) and therefore the gain of consumers who purchase is large for tier 3 in comparison with other tiers.
37
be recalled that higher tiers have a higher price elasticity (Kalra & Goodstein, 1998) and promotions are foremost highly effective in the short run (Nijs et al., 2001).
The Amazon price variable was included in the models of this research to increase the realism of the models as consumers do not only consider discounts and reviews but also judge the regular price in relation to this during online purchasing (Jang et al., 2012). This additional finding shows an unorthodox direction that is present in all tiers: a higher price implies a positive impact on sales. This finding can possibly be explained by the fact that involved consumers consider products based on price and inferred quality, amongst others (Kalra & Goodstein, 1998; Lamb et al. 2009; Jang et al., 2012). However, this concept especially holds when one realizes that the research was conducted in the unique circumstances of the introduction of an entire new product category with many new product launches. Tablet computers were just introduced, which attracts pioneering consumers who are most likely not very price sensitive and are willing to pay for new products of high quality.
For the cumulative review volume variable, a significant negative impact on sales is found across all tiers while a positive impact was expected. This unexpected finding cannot be underpinned by theoretical grounds and is regarded as an unexplainable peculiarity caused by the context of this research. The same holds for cumulative review valence: for all tiers except tier 2, a negative significant effect is found instead of a positive one, which is a remarkable result. However, as was suggested earlier for the significantly negative moderating effect of review valence on sales through price discounts, there is a counterexample of a unique case as discussed by the theory of Berger et al. (2010). Here, negative reviews of new products trigger the attention of curious consumers who would otherwise never had searched for the product, which ultimately attracts more consumers who actually make an online purchase.
Furthermore, cumulative review variance also has a significant negative impact on sales except for tier 2. In case of this variable, this direction of effect was expected on theoretical basis. Indeed, in case a product has a review with an average or high rating and its variance is high, this causes a negative influence on sales (Sun, 2012).
5.2. Managerial implications
