Context Matters: Effects of Assortment Variety On Online Purchasing Behaviour and Clicking Patterns Bas Vollenbroek Acacialaan 47A 9741KW, Groningen +31 6 21484036 b.j.vollenbroek@student.rug.nl Student Number: S2690780 16-06-2019 Master Thesis MSc Marketing Intelligence
Faculty of Economics and Business Rijksuniversiteit Groningen
Abstract
Ranking product offerings in an appropriate way is an important consideration for online aggregator firms. Such rankings may influence consumer choice and in turn, the profitability of the firm. Usually products are ranked based on their relevance to the customer but extant research and industry practices do not take the composition of the total presented ranking into account. In the context of online travel agencies we predict and demonstrate that ranking composition, in the form of assortment variety, is related to the purchasing behaviour of
customers. A large dataset from Expedia, a hotel portfolio site, covering 374.780 search sessions, formed the basis for this research. Using logistic regression as well as machine learning methods, the relationship between assortment variety and purchasing probability is shown to be individual and combined attribute dependent. Hidden states relevant to clicking patterns were also
uncovered through the application of a Hidden Markov Model and was found to be related to attribute specific assortment variety levels. This research suggests that optimization of
assortment variety, through ranking algorithms, could lead to improved purchasing probability, clicking patterns and possibly easier decision making processes.
Keywords: Assortment Variety, Purchasing, Ranking, Hotels, Clicking Patterns, Hidden Markov
1. Introduction 5
2. Literature Review 9
Consumer Search and Decision Making 9
Search Costs 10
Choice Conflict 10
Choice Set Heterogeneity 11
Product Attributes 12 3. Method 15 Data Descriptives 15 Data Cleaning 17 Measures 17 GINI Coefficient 18 Coefficient of Variance 18 Skewness 18 Descriptive Statistics 19 Control Variables 20 Model Specification 21 Estimation 22
Logit and Machine Learning 22
Hidden Markov Model 23
4. Results 25
Assortment Variety 25
Hidden Markov Model 26
5. Discussion 29
Overall Homogeneity 29
Attribute Homogeneity 30
Search Costs 32
Choice Behaviour 32
Practical Implications 33
Limitations and Future Research 34
References 37
Appendix 47
Tables 47
1. Introduction
With the advent of the internet people are faced with an ever increasing amount of information. In the context of online purchasing, consumers need to process large amounts of information (Lee & Lee, 2004). Too much information can lead to information overload with consequences such as cognitive fatigue, confusion and lower choice quality (Eppler & Mengis, 2004; Aljukhadar, Senecal, & Daoust, 2012). To deal with large amounts of information
consumers often use heuristics to reduce cognitive load (Malhotra, 1984). Next to these cognitive heuristics, choice overload also has an impact on the use of online decision aids. Information overload augments the conformance to recommendations in online purchasing situation
(Aljukhadar, Senecal, & Daoust, 2012). Further, prior to a purchase decision, online consumers appreciate not having to look at a variety of products (websites) individually making the market for aggregator sites large and relevant. E-tailers can play a role in reducing information overload and improving the purchasing process by using decision support systems such as rankings.
Decision support systems are decision aids that can help free up the decision maker processing capacity whilst maintaining decision accuracy (Häubl & Trifts, 2000). Customized product rankings can be seen as a Recommendation Agent (RA), a good example of a decision support system. RAs try to give customers recommendations that fit them the best from a large amount of option (Xiao & Benbasat, 2007). Online product comparison sites that implement recommendation systems are becoming increasingly popular among online shoppers (Burke, 2002). For consumers, efficient RAs can ease the online decision making process, improve choice quality, reduce information overload and increase choice confidence (Wang & Benbasat, 2007). For firms, efficient RAs can reduce the difficulties associated with the customer
purchasing experience and can be seen as a competitive advantage. Easy to use online stores are also related to higher loyalty rates (Wolfinbarger & Gilly, 2003). Improving RAs can be
beneficial to both the consumer and to the firm. In this paper we will focus on RAs in the form of product rankings.
role. When we assume the standard theory of choice where a consumer tries to maximize value, the introduction of irrelevant alternatives would not influence the preference of the consumer for the option with the highest value (Tversky & Simonson, 1993). However, the context in which alternatives are displayed do have an effect on the preferences of people as shown in effects such as the compromise effect (Simonson, 1989), attraction effect (Huber, Payne & Puto, 1982) and the similarity effect (Tversky, 1972), all of which are relevant in this context. These three effects, also called context effects in aggregate, challenge the assumption of the completely rational consumer. Context effect was defined by Prelec, Wernerfelt and Zettelmeyer (1997) as the effect of set composition on the choice of a subject inconsistent with their stable preference. For Internet Search Intermediaries these context effects can impact the product evaluation and possible consequent choice of consumers when faced with the product ranking.
When a consumer is confronted with a choice set where alternatives are similar on one attribute Tversky and Russo (1969) found that differences on other attributes were enhanced. Mellers and Biagini (1994) later transformed this into the contrast-weighting theory. The basic premise of this theory is that the utility of an attribute is weighed in comparison to the similarity on the other attributes. This, again, shows that the context in which an alternative is presented influences judgment of a consumer. How context influences the intention to make a choice or defer from making a choice is less well established. Dhar (1997) and Meyer (1997) found contrasting results of choice set homogeneity on choice deferral. Dhar (1997) found that choice set homogeneity increases choice deferral whereas Meyer (1997) found no effect of choice set homogeneity on choice deferral. Currently context effects are often not taken into account when constructing a ranking. This research will further investigate the relationship between choice set homogeneity and choice deferral or, in other words, assortment variety and purchase probability. Thereby providing evidence for one of the two opposing findings. Subsequently this evidence can be used for further optimization of rankings.
Usually product ranking algorithms try to rank the most relevant alternative at the top followed by decreasingly less relevant alternatives (Ursu, 2018). Characteristics of the
alternatives, historical clicking and purchasing behaviour and consumer characteristics usually play a role in determining relevance. Relevance scores are calculated at the level of the
are independent offerings. This is a rather strong assumption challenged by context effects and choice heuristics, like previously mentioned. Models as used by Ursu (2018) that try to model consumer choice without including for example these context effects, often do not form an accurate representation of real customer behaviour. Including context effects in choice models improves choice prediction (Berkowitsch, Scheibehenne & Rieskamp, 2014). Including context level effects for one could possibly improve product rankings.
Websites have the advantage of being able to customize and tailor their store for each individual visitor through for example personalised rankings. Hauser, Urban, Liberali and Braun (2009) showed the possibility of morphing the whole website based on the cognitive style of the consumer, inferred from their clicking behaviour. Leading to an estimated increase in purchase intentions of about 20%. Jerath, Ma and Park (2014) discerned customers groups with different levels of search involvement based on their clicking behaviour. Using the search behaviour of customers within an online ranking can serve as information to tailor the user interaction further. Research has been done on patterns within consumer searching behaviour, showing evidence for the existence of sequential searching behaviour online (Chen & Yao, 2017; Kim, Albuquerque & Bronnenberg, 2017). Overall, previous findings regarding clicking and searching behaviour suggest that investigation of clicking behaviour, and possible sequences therein, might be valuable. Evidence for sequences and the context in which they occur could be used to further optimize rankings.
Understanding the effects of rankings on consumers can also be important for policy makers. When ranking characteristics influence consumers in a direct way, influencing their purchase behaviour, it would be possible for search intermediaries to not display the most relevant but only the most profitable product. Increasing the relevancy of rankings can be
customers, intermediaries should ideally provide an assortment of search results which are most helpful for aiding a purchase. For this research our main question will be: how does the
particular assortment of search results influence a customer’s purchase probability? Next to this question there will also be a secondary research question with a more exploratory focus: how do choice sequences within a search result page look like and relate to each other?
This research question will be examined in the context of online travel agents (OTA). The main service that OTAs compete on is ranking products from other sellers in a relevant way. A number of internet search intermediaries also have bidding systems in place. These bidding systems allow companies to pay for a place in the ranking of the intermediary. During the data collection period, Expedia, where the dataset is from, did not have a bidding system in place. As summarized by Ursu (2018), a large portion of the hotel bookings go through OTAs with
Expedia being the largest OTA within the United States. Xiang and Law (2013) noted in their review that intermediaries have become a dominant force in the online hotel booking industry, even causing financial problems for hotel chains. In short, the OTA industry is a large and dominant force in the travel industry. Changes therein can have a widespread impact.
By analysing a large number of search sessions conducted on Expedia, an OTA, this research provides a number of major contributions to literature. First, evidence for a relationship between assortment variety and purchase probability was found. Thereby providing further evidence for a relationship between choice set homogeneity and choice deferral. Second, extending on existing research, this link between assortment variety and purchase probability was found to be attribute specific. Third, initial evidence was found for the existence of hidden states related to clicking patterns within a search session.
The rest of the paper will be structured as follows. In chapter 2 we will discuss the relevant literature on consumer choice and introduce our hypothesis. Thereafter we will describe the specifics of our dataset and the statistical methods we used in chapter 3. Chapter 4 will be used to display the results from the statistical methods. Comparing the results with our
2. Literature Review Consumer Search and Decision Making
Choosing the right product from a large assortment is often difficult. Consumers try to reduce this difficulty by using heuristics (Hauser, 2014). Modeling search and decision making has resulted in multiple theories. One major theory is based on consideration sets. In this
approach consumers first select a subset of products, called a consideration-set, and then choose one item from this set (Hauser, 2014). This decision process is also called a consider-then-choose consider-then-choose approach (Payne, 1976). The same two stage decision making process was also found to be used in online browsing and purchasing (De Los Santos, Hortaçsu & Wildenbeest, 2012).
Hauser (2014) found that a two-stage consider-then-choose approach results in decisions approximating the optimal sequential search models. Optimal sequential search models propose that optimal search happens in a sequential way (Weitzman, 1979). Research done by Weitzman (1979) showed that with any searching task the consumer has the option to either continue search or stop search after an initial search. Which option is the best depends on the utility realization, search costs and valuation of the not-searched products.
Showcasing that optimal search happens in a sequential manner. Morgan and Manning (1985) also found evidence that in most, but not in all, cases optimal search happens
sequentially. In an online retailing settings Weitzman’s (1979) sequential search model has also been successfully applied (Chen & Yao, 2017; Kim, Albuquerque & Bronnenberg, 2017). Looking at research regarding eye movement and information search (Simola, Salojärvi & Kojo, 2008) provides evidence for the sequential movement of the eye during information search tasks. Applying this to the context of a search page this could mean that the visual search task starts at the top of the page and the eye moves down in sequences. Even though there are multiple results shown on the page at once a sequential searching pattern could be assumed. As there is evidence that search happens in a sequential manner we can use a Hidden Markov Model (HMM) to identify hidden states within a search session. Therefore we postulate the following:
Search Costs
An additional construct to explain the searching and purchasing behaviour of a consumer is search costs, which will be discussed in this section. The first product a consumer sees is often not the product the consumer ends up buying. Comparing multiple products and looking up information to make a better decision are often part of the buying process. This process takes some effort and can be seen as a cost to the consumer. In literature the term search costs is often used to describe, or quantify, this process. Search costs encompasses the cost of searching for more information to resolve utility uncertainty (Wang & Sahin, 2018). Wang and Sahin (2018) also found that search costs have an effect on consumer choice behaviour, making inclusion of search costs in choice models necessary. Koulayev (2014) used hotel clickstream data to estimate search costs and found that only a third of the searchers go further than the first page. Estimated search costs had a maximum of $30 per page. In other industries, such as the car-insurance industry, search costs were found to range from $35 up to $170 (Honka, 2014). A reduction of these search costs could be an important force in creating more consumer welfare.
Internet and information services can play an important role in reducing the search costs for consumers. Bakos (1997) mentioned that electronic marketplaces can lower the costs
associated with acquiring information, consequently reducing search costs. Making it possible for firms to compete on search costs. Search platforms with less search costs experience increased consumer retention and consumer satisfaction with their product (De Los Santos & Forcum, 2018). Improving product rankings for example can be one way to decrease search costs. Improved rankings can lead to a higher probability of a match between a buyer and a seller thereby increasing consumer welfare (Ursu, 2018). Decreasing search costs may not only be beneficial for consumers by increasing welfare but it can also be a viable competitive strategy for online firms.
Choice Conflict
Next to search cost there are a number of other theories that try to explain if someone will make decision or defer from making a decision. One line of research is related to the
homogeneity of the options someone has to choose from. Within this line of research the notion of choice conflict is proposed as a possible mechanism explaining the link between the
within a set of alternatives can lead to more choice conflict because consumers have to make larger trade offs (Festinger, 1957). Choice conflict happens when a person has difficulty trading off costs and benefits, risk against value, and immediate satisfaction against future discomfort (Tversky & Shafir, 1992). Tverksy and Shafir (1992) also noted that large amount of choice conflict can lead to choice aversion or choice delay. When the difference between alternatives is very small and the choice has relatively little consequences it doesn’t really matter which option someone chooses. When the overlap between alternatives is complete choice can become trivial and people can become indifferent (Dhar, 1997). If larger differences between alternatives leads to more choice conflict, this could mean that consumers are more likely to avoid choosing when differences between alternatives are large.
Extant choice overload literature can also be relevant to this notion of choice conflict. Choice overload hypothesis states an increase in the number of options to choose from can decrease motivation to choose and satisfaction with the choice (Scheibehenne, Greifeneder & Todd, 2010). One could argue that the amount of options increases as the differences between options increases. As fully overlapping options can be seen as one and the same choice. When overlap decreases the amount of choices increases possibly leading to reduced motivation to choose as found in choice overload literature (Scheibehenne, Greifeneder & Todd, 2010).
Choice Set Heterogeneity
Research by Meyer (1997) found results not consistent with the research done by Dhar (1997). Meyer (1997) found that choice-set difficulty, e.g. availability of a dominant option, and utility similarity had no effect on choice deferral. Their explanation for this finding was the setup of their experiment. During the experiment participants looked at a number of choice sets one after another. As a result of this, judgment of alternatives within a choice set was mainly done in comparison with the alternatives in the last choice set. Therefore the characteristics of the set become less salient drivers. A comparison between the methodological setup of the research done by Dhar (1997), Meyer (1997) and our research can be found in Table 1.
(Insert Table 1)
In conclusion previous research gives no consistent answer on the relationship between choice-set homogeneity and choice deferral. Research by Dhar (1997) suggest that a
homogeneous choice set will be related to more choice deferral. Whereas research on choice conflict (Tversky & Shafir, 1992; Festinger, 1957) suggests the opposite. Meyer’s (1997) findings on utility similarity and choice deferral were also not in line with Dhar (1997). Therefore we hypothesize the following:
H1: Choice set homogeneity is positively related to purchase probability.
and classification standards can differ across countries. Review score is the average review score given by previous customers.
On the search result page the price of each offering takes up a substantial amount of space. Research undertaken by Chiang and Jang (2007) shows that price plays an important role in the online hotel choice. This research also shows that price influences quality and value of an offering directly. Confirming a finding that was previously found by Oh (2000). Price has also been identified as one of the most important motivators to buy online (O’Connor, 2002). There is sufficient reason to assume that price is an important attribute related to purchase probability.
Next to price location has also been identified as an important attribute in the hotel choice. Bull (1994) found location to be the most important attribute of a hotel because of its static nature and complexity. Further confirmation to the importance of location was provided by Lee and Jang (2012). They showed strong agglomeration of and competition between hotels around central areas. Not only for hotel owners this attribute holds importance. Also for consumers location has been identified as an important factor in selection and satisfaction (Shoval, McKercher, Ng & Birenboim, 2011; Lee & Jang, 2011).
Location often also plays a role in determining the amount of stars a hotel will receive. According to Fang, Ye, Kucukusta and Law (2016) the quality of a hotel can often be inferred from the star rating it received. Star ratings have also been identified as being a quality signal which users can use to determine the quality of the hotel (Abrate, Capriello & Fraquelli, 2011). Star ratings could offer useful quality information and turn out to be often used by consumers during their hotel choice (Núñez-Serrano, Turrión, & Velázquez, 2014). But with the rise of user generated content suchs as online reviews the star rating might become obsolete (Torres, Adler, & Behnke, 2014).
Next to the overall homogeneity we will also check for the effect of homogeneity per attribute on purchase probability. Following the same theory as for the overall homogeneity results in the following four hypotheses:
H1a: Price homogeneity is positively related to purchase probability.
3. Method Data Descriptives
Over a period ranging from November 1, 2012 up until June 30, 2013, 9.331.096 offerings were displayed over 374.780 search sessions containing 133.214 unique hotels in 170 countries. The searches were initiated from 34 different Expedia domains with visitors coming from 214 different countries. The dataset contains search level information as well as
information about the specific hotels displayed during the search session. Descriptive statistics and correlations for search level variables are shown in Table 4 and 5. For offering level variables Table 2 and 3 are used. Search level information mainly consists of the following search characteristics: search data and time, destination (city, county or neighbourhood), length of stay, booking window (number of days between the search and the start of the stay), number of adults, number of children, number of rooms and an indicator showing if the stay involves a Saturday night. For a small number of searches (5,3%) the following historical customer data is available: average star rating and average paid price per night. Information on the country and domain from which the search was made was available for every customer.
(Insert Table 2 and 3)
Offering level information mainly covers attributes of each offering visible on the first page of the search results. Data about the following attributes is included: country, star rating, average overall review score, indicator showing if the hotel is part of a chain, location score (desirability of a location), price, indicator showing if the offering has a price promotion tag and position within the search result page. Next to the visible attributes data is also collected for the following: hotel ID, price in the last trading period, distance between customer and hotel (6,8% of the offerings), the probability that a hotel gets clicked on (6,4% of the offerings), if the person clicked, if the person booked.
(Insert Table 4 and 5)
the search sessions 66,8% ended with a booking of one of the hotels and 33,2% with only a click. As a click is necessary to complete a booking all sessions with a booking contained at least one click. On average each customer was shown 24 offerings. During the observation period all bookings made up a total of $78.515.001.
One feature that makes this dataset unique is an experiment that was run during the data collection period. For this experiment a number of customers received a randomly ordered search result page instead of the standard relevancy based ordering. One advantage of this experiment is that the diversity of search result pages that were shown during the data collection period is larger. For our research question this increased diversity is beneficial to finding context effect, as well as finding the effect of overall relevance of the offerings.
No customer identification number was available therefore linking session was not a possibility. The few sessions containing historical data about the customers did not provide enough data to link searches to customers. Because of that each search session will be approached as if it was made by a unique customers. Ursu (2018) showed that in a similar dataset, 40% of the customers only search once, making the lack of customer identification a somewhat smaller downside. Linking hotel ID’s and country ID’s to specific hotels and countries respectively is also not possible as information about the meaning of the ID’s is not provided. However, one country ID stands out by the amount of offerings and visits from this ID.
5.381.216 offerings were displayed located in a country with ID 219 and 201.625 searches were made from within this country. The second country had 532.877 offerings and 34.327 searches. Only few countries have are large enough to have this many offerings. Expedia’s main .com domain is visited mainly from the United States (71,5%) according to Alexa (2019). Similar to Ursu (2018) we can assume that country ID 219 is the United States.
or moved to the second page. Ursu (2018) found in a similar dataset that 67% of the customers only considers the first page. Alleviating the downside of having only one search result a bit.
Data Cleaning
As with most dataset there were some abnormalities within this dataset that had to be filtered out before estimation. One large data cleaning step was related to the price variable. Descriptive statistics showed that there were hotel offerings with prices going up to $19.000.000. This could be the result of either the search criteria set by the customer (e.g. a stay of 4 years) or some error during the creation of the dataset. At the other end there were also offerings with a price of $0. To filter out search sessions that included these extremities, two values were chosen as an upper end and a lower end for the price variable. All search session that included prices lower than $10 or higher than $1000, consistent with Ursu (2018), were removed. Only very few offerings approached a value of $1000. By choosing this value we can be sure that abnormalities are filtered out but a wide range of differently priced offerings are still included. This step removed a total of 317.607 observations (3% of the total) across 13.714 search sessions (3% of the total).
Next to these price anomalies at the offerings level there were also search sessions that included hotel offerings with missing attribute information. 13844 (0,1% of the total)
observations across 10850 (2% of the total) session had missing review score information. This does not include observations with zero reviews. No homogeneity measures were calculated for these sessions. Adding all the cleaning steps up leads to a total of 331.451 removed observations across 24.564 sessions. Because of the size of the total dataset the amount of removed session is not likely to impact our subsequent estimations.
Measures
GINI Coefficient. The GINI coefficient is well known for its use in measuring income inequality (GINI, 1921). To calculate the GINI coefficient a Lorentz Curve is used as well as an equality line. The equality line presents the situation where the distribution is totally equal, e.g. everyone has the same income. Reality is presented by the Lorenz Curve by splitting the population up into groups and plotting their percentage of total income earned in a cumulative manner (Gastwirth, 1971). GINI uses the ratio of the area between the Lorentz curve and the equality line to calculate a coefficient. A GINI coefficient of 1 means that there is complete inequality, one person earns all income, a coefficient of 0 means that there is complete equality, everyone earns the same. Use of the GINI coefficient is not limited to measuring income
inequality. In research about car purchasing behaviour the GINI coefficient was used to measure product variety (Holweg, Miemczyk & STablein, 2011). Other research in the automotive industry used the GINI coefficient as a way to measure the spread of distribution point (Bucklin, Siddarth, & Silva-Risso, 2008). Overall, there is evidence that GINI is not only a good measure of income inequality, but also one of statistical dispersions in general and relevant to this research: assortment variety.
Coefficient of Variance. Coefficient of Variance (CoV), sometimes called the relative standard deviation, is a dimensionless measure of variance making it possible to compare variances across situations with varying range (e.g. Martin & Gray, 1971). CoV is calculated by dividing the standard deviation of a set by its mean. In organisational research CoV finds widespread use as a measure of diversity (Williams & O’Reilly, 1998; Bedeian & Mossholder, 2000). Also in sociological research concerning inequality CoV is often used as a measurement (Allison, 1978). Whilst being similar to the GINI coefficient, Champernowne (1974) showed that the sensitivity of the two measures differed. CoV was more sensitive to extreme values whereas the GINI coefficient was more sensitive to middle-range values. Because of the ease of
computation and flat sensitivity CoV can be a preferred choice over GINI (Allison, 1978). Making CoV a relevant second measurement of assortment variety within this research.
normal distribution has a moments skewness coefficient of 0. As the distribution is skewed more to the left the moments skewness coefficient gets increasingly more negative. When the
distribution is skewed more to the right, the moments skewness coefficient gets increasingly more positive. In research done by Bendel, Higgins, Teberg and Pyke (1989) skewness was also presented as a way to judge inequality, next to the GINI coefficient and CoV.
Descriptive Statistics. All three measures are assumed to measure a similar construct in slightly different ways. In this section we will look further into the similarity of the different similarity measurement. All measures have been calculated for the four product attributes: price, review score, location score and star rating. These four scores are then combined into one overall measurement by summing the logarithms of all four scores. On average the similarity of the search page results was similar across all four attributes. As seen in Table 6, dissimilarity was highest on the price attribute, followed by the location score, star rating and review score.
Skewness did not follow the same pattern, all attributes had a skewed distribution on average but the direction did differ across attributes. Location score, review score and star rating were all skewed to the left with review score being most skewed on average. Price, however, was the only attribute that was skewed to the right. These differences would’ve been lost when an overall similarity measure was used.
(Insert Table 6)
0,19 with the CoV for review score. For star rating this correlation is 0,08. Suggesting a stronger relation between review score and price than for star rating and price.
(Insert Table 7)
One way to decide which of the three measures is preferred is to create three different models, each with one measure, and compare the model fit of the three. Using a R2 measure is one way of displaying the amount of explained variance which can be used as a way to judge model fit. One popular R2 method is the McFadden R2 (McFadden, 1974). Two other measures that are often used in model selection are Akaike Information Criteria (AIC) and Bayesian Information Criteria (BIC). An in depth comparison between the two was done by Burnham and Anderson (2004), here we will use both. As shown in Table 8 the model where CoV is used as a similarity measure has the highest explained variance and the lowest AIC and BIC score for the overall model as well as the attribute model. Closely followed by the models with the GINI coefficient. Differences are small, but are also significant. Based on these three measures the preferred option is CoV. The GINI coefficient and skewness measure will be used to estimate the robustness of the results.
(Insert Table 8)
Control Variables
behaviour between different customer segments can be, for the most part, controlled for by including their search characteristics. Assuming that the segments are partially reflected in their search characteristics.
Next to these search characteristics the domain from which customer is visiting is available. Research by Fan, Shen, Wu, Mattila and Bilgihan (2018) showed that cultural
differences exist when people are offered product recommendation in the hotel industry. Where Chinese customers from a more collectivistic culture are more sensitive to in-group information than to out-group information. This effect was not observed for American customers,
characterised by their more individualistic culture. A number of different studies showed that cultural differences exist in hotel attribute preference and evaluation (Mccleary, Choi & Weaver, 1998; Bodet, Anaba & Bouchet, 2017; Royo Vela, Wang & Tyler, 2008). One way to try to control for cultural differences in preference and possibly purchase behaviour is including the website domain which the person is using to access Expedia as a control variable.
Like explained previously, one feature specific to this dataset is the random rankings that were displayed to customers during the data collection period. It is possible that the random ranking condition influenced purchase behaviour, because the relevancy of the offerings differs across these conditions. We controlled for this experiment during estimation.
A last control variable that was added to the model is the number of items that were viewed during the search sessions. One reason for including this variable is that a small sample influences our previously discussed measure for assortment variety. Deltas (2003) showed that the GINI coefficient is prone to small sample bias and recommended controlling for sample size. The same bias also applies to CoV as was shown in research done by Bedeian and Mossholder (2000). Controlling for the sample size allows for more accurate effects of both the GINI coefficient and CoV.
Model Specification
Overall Homogeneity Model
Attribute Homogeneity Model 𝑆𝑒𝑠𝑠𝑖𝑜𝑛𝐵𝑜𝑜𝑘𝑖𝑛𝑔 = 𝛼 + 𝛽1𝑆𝑖𝑡𝑒𝐼𝐷 + 𝛽2𝑅𝑎𝑛𝑑𝑜𝑚𝐵𝑜𝑜𝑙 + 𝛽3𝐿𝑒𝑛𝑔𝑡ℎ𝑂𝑓𝑆𝑡𝑎𝑦 + 𝛽4𝐵𝑜𝑜𝑘𝑖𝑛𝑔𝑊𝑖𝑛𝑑𝑜𝑤 + 𝛽5𝑁𝑢𝑚𝑏𝑒𝑟𝑂𝑓𝐴𝑑𝑢𝑙𝑡𝑠 + 𝛽6𝑁𝑢𝑚𝑏𝑒𝑟𝑂𝑓𝐶ℎ𝑖𝑙𝑑𝑟𝑒𝑛 + 𝛽7𝑁𝑢𝑚𝑏𝑒𝑟𝑂𝑓𝑅𝑜𝑜𝑚𝑠 + 𝛽8𝑁𝑢𝑚𝑏𝑒𝑟𝑂𝑓𝑂𝑓𝑓𝑒𝑟𝑖𝑛𝑔𝑠 + 𝛽9𝑃𝑟𝑖𝑐𝑒𝐶𝑜𝑉 + 𝛽10𝑅𝑒𝑣𝑖𝑒𝑤𝐶𝑜𝑉 + 𝛽11𝑆𝑡𝑎𝑟𝐶𝑜𝑉 + 𝛽12𝐿𝑜𝑐𝑎𝑡𝑖𝑜𝑛𝐶𝑜𝑉 Estimation
Logit and Machine Learning. To estimate the models, as specified before, multiple Machine Learning Methods (MLM) were used. One often used method to estimate a model with a binomial dependent variable is logistic regression, often called Logit. Next to the logit method the following MLM were also estimated and compared: elastic net regression, lasso and ridge (Zou & Hastie, 2005). These methods apply regularization and variable selection and have an advantage over logit in reducing overfitting, subsequently increasing predictive performance, and easing interpretation. In marketing literature these methods are usually applied to models
containing a large number of variables (e.g. Gelper, Wilms & Croux, 2016; Rutz, Sonnier & Trusov, 2017). Comparing predictive performance of the methods will be done using hit-rate, the GINI coefficient and Top-Decile Lift (TDL). These are often used measures to validate models in for example churn literature (e.g. Lemmens & Croux, 2006; Neslin, Gupta, Kamakura, Lu, & Mason, 2006) where the decision to leave or stay with a company is also a binomial decision.
Hit-rate is a measure that expresses the amount of correctly categorized predictions as a percentage of the total number of predictions. Usually a probability of 0.5 is used in binomial data to categorize people in the yes or no condition. For predictive performance the GINI
coefficient can also be used. Instead of using a Lorentz Curve a Cumulative Lift Curve is used to calculate the GINI coefficient. A lift curve plots the performance of the model against the
Out of all four methods Ridge offers the best predictive performance in terms of hit-rate. As shown in Table 9, Ridge outperforms Elastic Net Regression, Lasso and Logit on this metric for both models. On the GINI metric all estimation methods score similarly, with logit as the highest scoring estimation method for both models. Elastic Net Regression has the highest TDL score for both models. Based on these three metrics no estimation method stands out. Therefore we also took factors into account outside of predictive performance. Out of the four methods Logit is the most widely used method. Next to that estimation is also very efficient, time and computer resource wise. Even though the predictive performance was marginally lower than the other methods we find that the positives sides outweigh this negative side.
(Insert Table 9)
For the previous estimations the relation between the independent variables and
dependent variable was assumed to be linear. Research on assortment variety suggests that there might be an optimal amount of variety (Ryzin & Mahajan, 1999; Kim, Allenby & Rossi, 2002). To test if this relation is also found in our dataset a new logit model is estimated with quadratic variables of the different CoV measures. Comparing the quadratic model with the linear model shows that there is support for a quadratic relationship between the CoV measures and purchase probability. This is reflected in a lower AIC and BIC and a higher amount of explained variance, as shown in Table 10. For the model with overall CoV this difference is visible but not
statistically significant (p = 0,47) as tested with a Likelihood Ratio Test. The model with the CoV for each attribute does show a significant difference between the quadratic and linear form (p < 0,01).
(Insert Table 10)
modelling a series of observations through a number of a priori specified hidden states. After estimation, possible patterns or sequences can be identified through these hidden states. For a more lengthy introduction to HMM we refer to Rabiner and Juang (1986).
Applying this to the context of this research would mean the following. After a customer is presented an offering on the search result page there is a possibility to interact with it, i.e. click on it, or not. We assume that a customer goes through each offering one by one until he or she reaches the end of the page. At every offering on the page this decision to click or not to click has to be made. This series of decisions result in a time series of click / no click observation for every customer. Running a fitting HMM would, on average, result in a similar series of choice as observed in the data. Through application of a HMM the series of choices is attempted to be explained through a number of so called states. Each state has its own probability ratio of
clicking or not clicking, often called emission probabilities. Transitioning between the states also comes with different probabilities, often called transition probabilities. As a HMM is seen as a Markov Process which satisfies the Markov Property it means that the emission and transition probabilities of the next observation only depend on the present active state. These states are not observed directly but inferred from the sequence of choices, making the states ‘hidden’. To make these ‘hidden’ states interpretable different data is needed to describe them. Due to the relatively high computational demands that come with estimation of a HMM a randomly selected subset of the total dataset will be used. A total of 1011 search sessions were included in the subset. A total of 25185 offerings were shown across these 1011 search sessions.
To find an optimal number of hidden states to describe the choice sequences a data-driven approach was used. Metric such as AIC, BIC and log-likelihood scores can be used in selecting the optimal amount of hidden states (Celeux & Durand, 2006). All metrics for models with two, three, four, five and six hidden are displayed in Table 11. Comparing the models based on AIC and BIC metrics indicate that the model with three hidden states is the best fit for the data, as this model has the lowest score on both metrics. Therefore we will continue with a 3 state HMM.
4. Results Assortment Variety
Based on the previous discussion of the predictive performance of the different estimation methods we’ll continue with the interpretation of the model for the linear logit
method. Next to the predictive performance, the descriptive capabilities of our model are vital to our research question and hypothesis. For the interpretation, beta coefficients will be used. The direction of these effects will then be compared to the hypothesized direction. Before interpreting the results multicollinearity of the variables will be checked. No multicollinearity is one of the assumptions of a logit model. To test this assumption VIF scores have been calculated. No variable in our model exceeded a VIF-score of 1.15, therefore we will assume that there is no multicollinearity within our model.
Our first hypothesis was related to the overall homogeneity of the choice set. We hypothesized that as homogeneity increased booking probability would increase. The beta
coefficient effect of the variable for overall homogeneity was found to have a significant positive effect on booking chance. At a significance level of p < 0,001 the beta coefficient was 0.015. Our hypothesized direction differs from the observed direction, lending no support to our hypothesis. CoV for the price attribute had a beta coefficient of -0,193 (p < 0,001), which is in line with our hypothesized direction. For the review attribute CoV the beta coefficient was also negative at -0,116 (p < 0,001) which was also in line with our hypothesis. Star rating and location score CoV both had positive beta coefficient, 0,194 (p < 0,001) and 0.1 (p < 0,001) respectively, which are both not in line with our hypotheses. All coefficient estimates, odds ratios and comparisons to the hypothesized effects are summarized in Table 12. Estimations done with the GINI coefficient and skewness measure showed similar directions, effect sizes and significance levels to the ones estimated using the CoV measure. Indicating that the results found by our model are likely to be robust.
(Insert Table 12)
condition. As shown in Table 13, beta coefficients and significance levels changed compared to the previously discussed results shown in Table 12. The relation between overall CoV and purchase probability decreased, review CoV and location CoV are no longer significantly related to purchase probability and the effect of star CoV on purchase probability also decreased. One interesting difference is the beta coefficient of price CoV, in the random condition this variable has the largest, in an absolute sense, effect out of all CoV attributes. Compared to the previous estimation the effect of price CoV also decreased with 40,4%.
(Insert Table 13)
Hidden Markov Model
As previously explained, estimated hidden states need to be described through other types of data to make them interpretable. To describe the hidden states the following offering level variables were selected: price (standardized), review score, star rating, location score, promotion flag, position, click and booking. As explained during the literature review, a number of these variables might be related to purchasing decisions. Position and promotion flag were not specifically mentioned. Research done by, for example, Ursu (2018) shows the effect that position can have on purchasing probability. Promotion flag is a binomial variable indicating if the offering is visibly on promotion or not. Close and Kukar (2010) show that sales promotion can increase the possibility that customers interact with the product, through placing it in a cart. Promotion and position next to the other variables can therefore possibly explain the
probabilities of belonging to a given state. Thereby providing a description of the states.
Next to the description of the states, transition and emission probabilities will be used to further describe the estimated HMM. Lastly search level variety measures will be used to describe how often the states were observed across different searches. Offering level variables are regressed on the previously estimated probabilities of belonging to state one, two or three. By using a standard linear regression we can make inferences about the type of offerings and
(Insert Table 14)
All observations that were classified as state 1 contained no clicks whatsoever. State 2 stands out by a large margin with 94,2% of the observations containing a click. For state 3 3,1% of the observations that were classified as state 3 contained a click. These observations make clear the differences in emission probabilities of each state. State 1 has a probability of less than 0,01 for a click and 0,99 for no click. For state 2 the probability to observe a click is 0,86 with no click having a probability of 0,14. Lastly, state 3 has a probability to click of 0,04 and the
probability of no click is 0,96. As mentioned earlier, clicks are necessary for a booking, making the both related to each other. This relation is also visible through observing booking behaviour related to the different states. States with a high probability of clicking also have a high
probability of booking. Of the offerings related to state 1 no booking was detected. For state 2 61,4% of the observations were booked. Out of all observations related to state 3 0,02% of the offerings were booked.
As shown in the transition matrix (Figure 1) observing state 2 is almost always followed by state 1. Once state 1 is observed no other transitions to different states, or clicks for that matter, are likely to happen. As seen by the coefficient estimates state 1 is related to offerings that score low on review score, star rating, location score and have no promotion flag. Usually these offerings are positioned at position three or lower, with a position further down the page increasing the chance of being related to state 1. Offerings related to state 2 are not seen in a specific position based on the non-significant coefficient for position. High star rating and location score are positively related to the probability of belonging to state 2. A promotion flag negatively affects the probability of belonging to state 2. In contrast to the estimates for state 1 and state 3, review score is not related to belonging to state 2. Offerings related to state 3 are most likely to be highly positioned, score high on review score, star rating, location score and are likely to have a promotion flag.
(Insert Figure 1)
across the different search sessions. This was done by calculating the amount of observations with hidden state X divided by the amount of observations in the within the search session. Using this ratio as a DV and continuing with the same IV’s as we did for assortment variety resulted in six different models. Two for each state, one with Overall CoV and one with attribute level CoV. The results for the variety measures are displayed in Table 15. Session with a high star rating or location score CoV, i.e. high variety, are likely to have a relatively large number of state 1 observations. Similarly, albeit smaller, a high star rating CoV is also related to a relatively large number of state 2 observations. An opposite effect can be found for review CoV, having a negative relationship to the number of state 2 observations. In contrast to the models of state 1 and state 2, state 3 has one significant effect. Location CoV is negatively related to the amount of state 3 observations. The overall CoV measure is also significantly related to the ratios for each state. Overall CoV is positively related to state 1 ratio and negatively related to state 2 and state 3 ratio.
5. Discussion Interpretation and Theoretical Implications
In this section the results presented in the previous section will be discussed and
compared with the theory that was mentioned during the introduction. For the options where the theory and results do not align possible alternative explanations are proposed. During this research the main research question was how does the particular assortment of search results influence a customer’s purchase probability?. This question resulted in five different hypothesis, all related to homogeneity of the assortment but differing in the aggregation level and attribute type. The first hypothesis was related to the overall, a combination of displayed attributes,
homogeneity on purchase probability. Disaggregating this overall homogeneity into homogeneity of four different attributes: price, review score, star rating and location score was the basis for the other four hypothesis. For all hypothesis a positive relationship was predicted, as homogeneity increases purchase probability increases. All five hypothesis will be discussed in the same order, starting with overall homogeneity. After that, the results related to our second research question, how do choice sequences within a search result page look like and relate to each other?, will be interpreted and discussed.
Overall Homogeneity. Our results indicated that the overall choice set homogeneity was negatively related to purchase probability. In other words, as options within the search results page become more dissimilar the chance that a customer will book one of those offerings increases. This effect was not in line with our hypothesized effect. A possible explanation could be the online context in which the data was collected differs from the context in which previous research was conducted. As mentioned by Johnson, Moe, Fader, Bellman and Lohse (2004) search costs are a lot lower shopping online compared to offline. Exiting a search or making a large number of searches comes with relatively little consequences, as starting a new search is only a few seconds or clicks away. In other words, because of relatively low search costs making the barrier to a new search lower, choice deferral is likely to manifest itself differently in an online context. Previous research on choice deferral and choice set homogeneity (Dhar, 1997; Meyer, 1997; Tversky & Shaffir, 1992; Festinger, 1957) was mostly done in an offline
As discussed in detail in the introduction, previous literature was also not consistent on the direction of our hypothesized effect. Dhar (1997), suggested that a decrease in choice set homogeneity could make it easier to form a consideration set, making the choice process easier. Decreased homogeneity did not impact choice deferral according to Meyer (1997). Research on choice conflict (Tversky & Shaffir, 1992; Festinger, 1957), a possible consequence of choice set heterogeneity, was also inconsistent with the findings of Dhar (1997). Our research contributes to the literature concerning choice set homogeneity by providing evidence for the theory
provided by Dhar (1997). Thereby getting one step closer to an answer on how overall choice set homogeneity influence subsequent choice behaviour. One comment that could be made is that even though our effect was significant the beta coefficient was relatively small (0,016). How strong the support of our results are for the theory of Dhar (1997) is therefore debatable.
Attribute Homogeneity. This research extended on previous research by disaggregating the overall homogeneity into homogeneity of the separate attributes. It is possible that splitting the overall homogeneity up into different attributes does not capture the whole choice set homogeneity. Principles of Gestalt psychology (e.g. Köhler, 1967; Spelke, 1990) suggest, in simple terms, that the whole is more than the sum of its parts. A theory that is often applied to human perception in cognitive psychology (Johan et al., 2012). By splitting overall homogeneity up, we assume that customers judge each attributes’ homogeneity as an independent construct. This assumption is not in line with the notion of Gestalt psychology and therefore possibly also not in line with reality. Interpretation of these separate homogeneity effect should therefore be interpreted and applied with caution.
One important attribute that was investigated was price. A significant positive and large, relative to the overall homogeneity measure, effects was found for the price attribute
In contrast to the price and review score; the homogeneity measures of location score and star rating turned out to have a negative effect on purchase probability. As location score
similarity increases, the chance of booking one of the offering decreases. When offerings
become more similar on star rating the chance of booking one decreases. Both effects were in the opposite direction as hypothesized. A possible explanation for the negative effect of star
homogeneity on price is that stars staring is depicted using symbols and not presented as a number like review score. Research by Kahn and Townsend (2013) has shown that customers tend to prefer a visual depiction of information. This preference can influence the perceived complexity and variety of an assortment. Kahn (2017) also showed that visual elements have a higher perceptual fluency than compared to textual representations. Less homogeneity in star rating might have smaller influence on perceived complexity and subsequent choice. In other words, the capacity to process dissimilarity might be larger for star rating due to its visual depiction. When a large variety is easily processed customers might prefer this variety as information for their subsequent purchasing decision.
Out of all four attribute homogeneity measures location homogeneity had to smallest negative effect on purchase probability. A possible explanation for the negative effect for
location is that the location score that was used as a measure for location is not an accurate proxy for the location information that is displayed on the search results page. As mentioned in the method section, on the search result page customers the location of the offering is usually
described in the form of a neighbourhood. Names such as center or a location specific translation might be recognized by most people. More city specific names such as The Grachtengordel in Amsterdam might not be immediately recognized by every customer. Therefore the homogeneity of location scores might not fully translate to a perceived homogeneity of hotel locations by the customer. This might explain the smallest effect size out of all four attributes.
on the variety effects in an online hotel booking setting and provide evidence for attribute dependent variety effects on customer purchasing behaviour.
Search Costs. As discussed in the previous section, context effects tend to affect customer choice. A view that could arise from this conclusion is that customers can be
manipulated into buying by changing context, to their detriment. However, one could also argue that customers could benefit from a changed context, making it a manipulation to their benefit. One way to understand the effect of context on customer choice is by using the search cost construct. Even though our research has no direct measure of search costs, it could be inferred from the booking behaviour of customers. Thereby assuming that the decision making difficulty is accurately reflected in the booking behaviour of customers.
When this assumption is made increasing the purchase probability would mean that a customer is not only more likely to purchase, it would also make arriving at this purchase decision easier. Making the purchase decision easier would mean a decrease in search costs, benefitting the customer and increasing customer welfare as explained by Ursu (2018). Optimizing the variety of the search result page would than not only benefit the firms through increased sales, it could also increase customer welfare by decreasing search costs.
Choice Behaviour. We postulated that the clicking behaviour of customers could be explained through application of a HMM and its hidden states. A number of different clicks and sequences of clicks were modelled based on this postulation. Three different states were
investigated further. State 1 was related to a large number of no clicks in the data. Next to that state 2 was related to a large number of clicks in the data. Lastly state 3 was related to a
most found in search sessions with large star rating and location score variety. State 2 was mainly related to search session with little review score variety and or more star rating variety. Search sessions with little location score variety were mostly related to State 3. Overall variety was positively related state 1 and negatively related to state 2 and 3.
One could say that state 2 and 3 are most beneficial to the firms as these contain the most chance of clicking on an offering. A click could increase the chance of future clicking and consequently increase the chance of future purchase (Xu, Duan & Whinston, 2014). If increasing the chance of observing clicking patterns with a large number of clicks firms should, consistent with our previous variety findings, keep the attribute variety in mind. With little review score, location score variety and overall variety, but high star rating variety, the chance of observing state 2 or 3 increases. Because of the nature of our dataset, a click is necessary to make a booking, these chances of purchase after a click are likely to be higher. These results show that application of a HMM to clicking data can help to uncover clicking sequences and identify context in which these sequences are likely to occur. Providing evidence for sequential search and decision making theories as theorized by Payne (1976), Hauser (2014) and Weitzman (1979). Showing not only how assortment variety and offering characteristics are related to one single click observation but also how they are related to, potentially beneficial, clicking
sequences.
Practical Implications
Next to theoretical contributions our results can also be applied in practice, by managers for example. As mentioned during the introduction, Expedia mainly used the relevancy of offerings, to construct their rankings. Relevancy was estimated through customer characteristics and historical information about the offering. When purchase probability is of concern, the inclusion of similarity effects in their ranking algorithms could be a valuable addition to their already intricate algorithm. More specifically, a set of displayed offerings should be similar on price and review score and dissimilar on star rating and location score. Overall the set should be dissimilar for the combination of these four attributes. This implication does hold not only within the hotel industry but possibly also in other industries where ranking algorithms are used.
variety, but high star rating variety, are likely to be related to clicking sequences including a large chance of observing clicks. As increasing clicking behaviour of consumers can increase purchase probability it is beneficial to the company to investigate clicking patterns. Optimizing assortment variety can increase the chance of observing beneficial click patterns over no click patterns.
Optimizing the offering variety is possibly not only advantageous to the company, it can also be advantageous to the customer. For a customer making a choice could become easier when the variety is at an optimal level. An easier decision making process could be reflected in an increased amount of bookings. This would provide support for alignment of the goals of the customer and the company. A company could increase the amount of purchases by making the decision making process of the customer easier through variety optimization. Making it a win-win situation on both sides. However, changing context is limited by the relevancy of the offerings. In search result pages where offerings were ordered randomly similarity effects changed or even disappeared. A challenge for practice would be to combine the most relevant offerings in such a way that the variety is optimal.
Limitations and Future Research
The results displayed in this research also come with a number of limitations, whilst also implying possibilities for future research. Both limitations and avenues for future research will be discussed within this section. One limitation of the dataset is that only the search session that used the default ranking were included. During search a number of filtering and sorting options are available to the customer. As argued in research by Hauser, Urban, Liberali and Braun (2009) website visitors usually have different preferences for the way information is displayed on a website. One explaining construct was one’s cognitive style. This relatively stable construct was defined as a person’s preference in terms of information gathering, processing and analyzing (Hayes and Allinson 1998, p. 85). More analytical cognitive styles usually have a preference for more detailed information, whereas the more holistic styles would prefer a simple
probably prefer filtering and sorting themselves over a computer generated ranking in most cases.
Another limitation of this dataset, and thus our results, is that only search sessions that included an interaction were included. One could argue that these customers do not form a total representation of the whole population. In terms of customer journey as defined by Lemon and Verhoef (2016) these customers would most likely be in the purchase stage. It is possible that the choice set homogeneity effects differ across the different purchase stages. With the current dataset, this proposition could not be tested. However, this limitation can also be seen as a strength of our research. As our results would imply that tracking a customer is not necessary to obtain applicable insights or optimize relevance algorithms. Companies currently have the ability to track customers over time with, for example, the use of cookies. Ways of tracking customers and collecting data are faced with increased regulatory pressure from governments. With the GDPR in the European Union being one major example. It is possible that in the future companies will have to comply with stricter laws regarding the collection and storage of data. Companies who ignore growing privacy concerns among customers are faced with lower customer trust (Bart, Shankar, Sultan and Urban, 2005) and lower stock returns (Acquisti, Friedman & Telang, 2006). Therefore companies may need to ask themselves the question; are the benefits of and tracking and collecting customer data along the full customer journey outweighed by the risks?
Application of the HMM came with a large but known assumption. By treating the clicking behaviour of the customer as a time series we assumed that there was a clear beginning, the start of the page, and end, the last offering on the page, to the search behaviour of a customer. When a customer is specifically searching for one offering, or has a clear purchasing goal in mind (Goal-directed search; Janiszewski, 1998) this assumption might hold. For more exploratory search (Janiszewski, 1998) usually undirected searching patterns are visible, e.g. moving up and down the page, the assumption will probably not hold. Additional information about the order in which the offerings were viewed could help relax this assumption. A possible avenue for firms or future research to consider.
preference can however also influence the perceived complexity and variety of an assortment. Quantifying the complexity and similarity of unstructured data such as images and video has a lot of potential for marketing (Wedel & Kannan, 2016). Research done by Pieters, Wedel and Batra (2010), for example, quantified visual product advertisement complexity and found that customers preferred designs that were creatively complex but not in terms of visual features (colors, contrast & luminance etc.). Which images should be displayed next to an offering on a search result page, or which combination of images should be displayed on search result page? These are interesting question for future research to answer.
Product images also tie in with the overall design of the search result page. Research about the differences in perceptual fluency for text or visual presentations suggest that the design of online webshops has a large influence on the perception of the assortment (Kahn, 2017) and decision quality (Mosteller, Donthu & Eroglu, 2014). With increasing channel switching (e.g. from mobile to desktop) during shopping (Verhoef, Kannan & Inman, 2015) customers are exposed to a variety of designs. Testing for differences of homogeneity effects across designs might be an interesting and very relevant area of research.
Product attributes such as star rating and location score are relatively specific to the hotel industry. In other industries there usually are commercial companies that rank a selection of products and hand out star ratings. But this might not be comparable to the standardized way of the hotel industry. Generalizing the attribute homogeneity effects to products in other industries might be problematic. Our research does give some support to the notion that similarity
References
Abrate, G., Capriello, A., & Fraquelli, G. (2011). When quality signals talk: Evidence from the Turin hotel industry. Tourism Management, 32(4), 912-921.
doi:10.1016/j.tourman.2010.08.006
Acquisti, A., Friedman, A., & Telang, R. (2006). Is There a Cost to Privacy Breaches? An Event Study. Paper presented at the Proceedings of the Twenty Seventh International Conference on Information Systems, 94.
Alexa, S. (2019). expedia.com Traffic Statistics. Retrieved from https://www.alexa.com/siteinfo/expedia.com
Aljukhadar, M., Senecal, S., & Daoust, C. (2012). Using Recommendation Agents to Cope with Information Overload. International Journal of Electronic Commerce, 17(2), 41-70.
doi:10.2753/JEC1086-4415170202
Allison, P. D. (1978). Measures of Inequality. American Sociological Review, 43(6), 865-880. doi:10.2307/2094626
Bakos, J. Y. (1997). Reducing Buyer Search Costs: Implications for Electronic Marketplaces. Management Science, 43(12), 1676-1692. doi:10.1287/mnsc.43.12.1676
Bart, Y., Shankar, V., Sultan, F., & Urban, G. L. (2005). Are the Drivers and Role of Online Trust the Same for All Web Sites and Consumers? A Large-Scale Exploratory Empirical Study. Journal of Marketing, 69(4), 133-152. doi:10.1509/jmkg.2005.69.4.133
Bedeian, A. G., & Mossholder, K. W. (2000). On the Use of the Coefficient of Variation as a Measure of Diversity. Organizational Research Methods, 3(3), 285-297.
doi:10.1177/109442810033005
Bendel, R. B., Higgins, S. S., Teberg, J. E., & Pyke, D. A. (1989). Comparison of skewness coefficient, coefficient of variation, and Gini coefficient as inequality measures within populations. Oecologia, 78(3), 394-400. doi:10.1007/BF00379115
Bodet, G., Anaba, V., & Bouchet, P. (2017). Hotel Attributes and Consumer Satisfaction: A Cross-Country and Cross-Hotel Study. Journal of Travel & Tourism Marketing, 34(1), 52-69. doi:10.1080/10548408.2015.1130109
Bucklin, R. E., Siddarth, S., & Silva-Risso, J. (2008). Distribution Intensity and New Car Choice. Journal of Marketing Research, 45(4), 473-486. doi:10.1509/jmkr.45.4.473 Bull, A. O. (1994). Pricing a Motel’s Location. Int J Contemp Hospitality Mngt, 6(6), 10-15.
doi:10.1108/09596119410070422
Burke, R. R. (2002). Technology and the customer interface: What consumers want in the physical and virtual store. Journal of the Academy of Marketing Science, 30(4), 411. doi:10.1177/009207002236914
Burnham, K. P., & Anderson, D. R. (2004). Multimodel Inference: Understanding AIC and BIC in Model Selection. Sociological Methods & Research, 33(2), 261-304.
doi:10.1177/0049124104268644
Celeux, G., & Durand, J. (2008). Selecting hidden Markov model state number with cross-validated likelihood. Computational Statistics, 23(4), 541-564. doi:10.1007/s00180-007-0097-1
Champernowne, D. G. (1974). A Comparison of Measures of Inequality of Income Distribution. The Economic Journal, 84(336), 787-816. doi:10.2307/2230566
Chen, Y., & Yao, S. (2017). Sequential Search with Refinement: Model and Application with Click-Stream Data. Management Science, 63(12), 4345-4365. doi:10.1287/mnsc.2016.2557 Chiang, C., & Jang, S. S. (2007). The Effects of Perceived Price and Brand Image on Value and
Purchase Intention: Leisure Travelers' Attitudes Toward Online Hotel Booking. Journal of Hospitality & Leisure Marketing, 15(3), 49-69. doi:10.1300/J150v15n03_04
Close, A. G., & Kukar-Kinney, M. (2010). Beyond buying: Motivations behind consumers' online shopping cart use. Journal of Business Research, 63(9), 986-992.
doi:10.1016/j.jbusres.2009.01.022
De Los Santos, B., Hortaçsu, A., & Wildenbeest, M. R. (2012). Testing Models of Consumer Search Using Data on Web Browsing and Purchasing Behavior. American Economic Review, 102(6), 2955-80. doi:10.1257/aer.102.6.2955
