Moderating effects for the success
of online personalization in the
hospitality environment
by
PREFACE
This thesis is the final piece of my master Marketing Intelligence at the University of Groningen.
During my job as a conversion specialist, I became interested in personalization and its effect on the behavior of website visitors, as well as the website’s conversion rate. Therefore, I’m
really thankful for the possibility to write my master thesis on this topic. First, I would like to thank my thesis supervisor prof. dr. Peter Verhoef:
Thank you for your feedback and the discussions we’ve had. It has been very insightful. Finally, I would like to thank my parents, my sister, and my girlfriend
for supporting and encouraging me throughout the entire process: Not only during my thesis, but for my entire study.
Thank you for all your support in the past years. I wish you a lot of joy reading this thesis.
Vincent Alkema, Groningen, June 2018
SUMMARY
This study investigates the influence of online personalization on the direct bookings in the hotel industry. The research question central in this study is:
What are the moderating effects of the device, visitor source, and price of the hotel on the relationship between online personalization and bookings in the hospitality environment?
A data set containing cross sectional data on website data from 15 different hotels over a period of 15 months has been used to answer the research question. OLS has been applied to 4 models, of which two are completely pooled and two are partially pooled.
No statistical significant effect has been found for the use of online personalization. The moderation effects of price, source, and the device a visitor uses also showed no significant effects.
TABLE OF CONTENTS 1. INTRODUCTION ... 6 2. THEORETICAL FRAMEWORK ... 9 2.1 Conceptual Framework ... 9 2.2 Personalization ... 10 2.3 Device type ... 13 2.4 Visitor Source ... 15 2.5 Price Segment ... 16 3. RESEARCH DESIGN ... 17 3.1 Data ... 17 3.2 Descriptives ... 18
3.3 Missings and outliers ... 20
3.4 Method... 21 4. RESULTS ... 23 4.1 Nonzero Expectation ... 24 4.2 Heteroscedasticity ... 25 4.3 Correlated Disturbances ... 26 4.4 Nonnormal Errors ... 27 4.5 Multicollinearity ... 30 4.6 Pooling... 31 4.7 Summary ... 32
4.8 Final model specification and estimation ... 33
4.8.1 Estimation Model 1 ... 34
4.8.2 Estimation Model 2 ... 35
4.8.3 Estimation Model 3 ... 36
4.9 Validation ... 39 4.9.1 Face Validity ... 39 4.9.2 Statistical Validity ... 40 5. CONCLUSION ... 41 5.1 Discussion ... 41 5.2 Managerial implications ... 44
5.3 Limitations and future research directions ... 45
7. REFERENCES ... 46
APPENDIX A – HISTOGRAM OF THE BOOKINGS ... 51
APPENDIX B – BAR CHARTS... 52
APPENDIX C – HETEROSCEDASTICITY PLOTS ... 54
1. INTRODUCTION
The hotel industry has seen great changes since the internet has become widely used by the public. One of these changes is the way hotels distribute their rooms (Gazzoli, Kim & Palakurthi, 2008). In the 1990s, online travel agencies (OTAs) like Booking.com, Expedia, and Hotels.com partnered with hotels to provide consumers an overview of multiple hotels and consequently let them book directly with the OTAs (Carroll and Siguaw, 2003). Since then, the worldwide usage of the internet has grown and therewith the use of the internet for online hotel bookings. But, together with the online hotel bookings, the market shares of these so-called OTAs has also grown compared those of the hotels themselves. Simultaneously the costs of these intermediaries has gone up (Toh, Raven and DeKay, 2011). The OTAs were able to gain large market shares, due to using different business models, smarter business practices, and taking advantage of the poor management and application of the online pricing of hotels (Enz, 2003; O’Connor, 2002,2003, Tso and Law, 2005). The success of OTAs resulted in financial problems for hotels, since OTAs often offered lower prices than the hotels did on their own websites. This was possible due to a clause known as ‘room parity’. Room parity restricts hotels from offering rooms for lower prices than displayed on the websites of OTAs (Gazzoli, Kim & Palakurthi, 2008).
The commissions hotels have to pay OTAs for renting out their rooms are typically between 15 and 30 percent of the total fee, whereas the fee airlines have to pay is only around 5 percent (Toh, Raven and DeKay, 2011). The study of Toh, Raven and DeKay (2011) also showed that larger hotel chains can use their position to negotiate lower commissions (around 15 percent), while the smaller hotels have to pay higher commissions (up to 30 percent). In conclusion, we can state that the commission of OTAs for hotels is extremely high, and these commissions can only be lowered if hotels agree to exclusively work with one particular OTA.
The increasing market share of OTAs has even led to conflicts between these two channels in the hotel industry, since both the hotels and the OTAs aim to sell the same rooms (Myung, Lan, and Bai, 2009). In response to this unhealthy situation where hotels are highly dependent on OTAs, we’re seeing an increasing number in hotels that promote to book directly with them. An example that demonstrates this recent development is the worldwide campaign of Hilton, called ‘Stop Clicking Around’. This campaign was introduced in 2016, and was Hilton’s largest marketing campaign the organization’s 97-year history (Hotel News ME, 2018). The aim of this campaign was convincing their hotel guests to book directly with them. The results of the campaign exceeded the expectations with an increase of 1.6% for the online direct bookings, and an increase of 27% of the online direct revenue compared to 2015 for the EMEA region (Europe, the Middle East and Africa). But, despite the hotels trying to convince their customers to book directly, the market shares of the OTAs is still growing (Business Travel IQ, 2018).
According to Garrow et al. (2006) consumers that shop online will first visit the hotel’s website, after finding them at the OTAs, to check for information, or to make a booking. This is known as the ‘the billboard effect’. If a consumer is however not able to find what he/she is looking for on the hotel’s website, they will go back to the OTAs to proceed the booking. A reason for this phenomenon might be the fact that hotels are not able to persuade the visitor to book directly with them. However, this does mean there still is an opportunity for hotels to ‘win’ the booking from OTAs.
Hotels are facing a high dependency on OTAs and don’t have the right online marketing strategy to compete with these giants. According to Toh, Raven and DeKay (2011), a solution for hotels could be to improve their online marketing strategy. Carrol and Siguaw (2003) argue that personalization may be key in this solution. In the current literature there is a lack of information on the effectiveness of personalization in the hotel industry, and the driving factors of this effect. This study aims to fill this gap in the literature and provide insights on the effectiveness of personalization in an online hospitality environment. More importantly: the role of different website visitor and hotel characteristics for this effect. Therefore the research question, central to this this study, is:
What are the moderating effects of the device, visitor source, and price of the hotel on the relationship between online personalization and bookings in the hospitality environment?
In order to answer this research question a data set containing cross sectional data from 15 different hotels will be used. The data set consists of A/B-testing data, where half of the visitors have seen a ‘normal’ version of the website, and half of them a personalized version. Data on more than 2 million visitors and a total of 48.000 online bookings are included in the data set.
The personalized versions of the websites consist of smart solutions that show personalized and persuasive messages. These messages appear in different ways on the websites and are triggered by various click- and search actions. The personalized items can be seen as an extra layer that has been put over the website. However, these tools could be interpreted as pop-up advertisements, matching the website’s layout. Characteristics for these personalized messages can be based on previous visits (new visitor vs a returning visitor). The pages visited, combined with the customer’s location.
2. THEORETICAL FRAMEWORK
In this chapter you will find the theoretical framework of this research. First the conceptual framework of this study will be discussed, followed the possible effects of personalization, device type, price segment, and the different types of sources through which a customer can visit the hotel website.
2.1 Conceptual Framework
This section will introduce the conceptual framework developed for this study. This conceptual framework has been developed based on the theoretical framework discussed in the next sections of the study. It is a representation of the relationships of the previous discussed attributes and it’s assumed relations.
2.2 Personalization
In literature, there are many terms used to describe personalization. Most commonly used are customization; segmentation; targeting; profiling and tailoring; one-to-one marketing (Wind & Rangaswamy, 2001). Peppers and Rogers (1997) describe personalization as using ones information to deliver targeted solutions to a specific customer. Personalization is about delivering the right message to the right person at the right time, to maximize business opportunities (Tam and Ho, 2006). Dijkstra (2008) states that personalization is about using one or more recognizable characteristics from an individual in a persuasive text. In this study a combination of the definition from Peppers and Rogers (1997) and Dijkstra (2008) is used. Therefore, personalization is described as using ones information to deliver targeted solution in a persuasive context.
Online personalization
The process of online personalization can be described as: Personalization web content for individual users in an automated manner (Treiblmaier et al. 2004). Users will be identified while entering a website, data on the navigation pattern of a user will be collected, the data will be compared to data of similar users, estimations will be made for the preferences of this user compared to the known data and consequently the content on a website will be updated according to this data (Lavie et al. 2010). Smith (2006) defines online personalization as providing content or recommendations that are relevant for a user based on past behavior, similar to other users, explicitly defined preferences or individual characteristics. A great example of online personalization is Amazon. They welcome returning visitors with personalized messages, and offer book recommendations based on their customer’s past behavior.
Effects of (online) personalization
The use of personalization will reduce the perceived complexity, while browsing on a website. This results in lower search efforts and transaction times, which in turn result in an increase in customer satisfaction (Thongpapanl & Ashraf, 2015; Thongpapanl & Ashraf, 2015).
Benefits of personalization for visitors are better products, service, communication and customer experience (Allen et al., 2001; Coner, 2003; Peppers et al., 1999). This in turn increases both the customer satisfaction and the purchase intention (Peppers, Rogers and Dorf, 1999; Thongpapanl & Ashraf, 2015). Personalization has been assumed to positively influence loyalty (Peppers and Rogers, 1993; Rust et al., 2000). Users are more likely to accept offers that are self-referent, rather than offers that are seen as non-self-referent (Tam and Ho, 2006). However, there it is still uncertainty about the amount of information and the level of personalization needed to increase the satisfaction and purchase intention, and consequently an increase in sales (Thongpapanl & Ashraf, 2015).
Possible negative effects of personalization
Although it seems personalization only results in a variety of positive effects, there are also studies claiming the opposite. A possible negative effect from personalization may be discomfort. This can occur when consumers realize their information has been collected without their knowledge (Tucker, 2012; Aguirre, et al, 2015). When companies do not inform their customers about the fact that they collect their data, and show them personalized content that contains distinct personal information, customers may think the company is acting in its own self-interest (Aguirre, et al, 2015). This would lower the customer’s perception on the company’s intentions (Shen and Ball, 2009). Another research claims that a higher degree of personalization increased the feeling of intrusiveness, which in turn results in a negative influence on ones purchase intention (Van Doorn and Hoekstra, 2013).
As mentioned earlier, the use of personalization reduces the perceived complexity while, browsing on a website. This results in a lower search effort and transaction time, which results increased customer satisfaction (Thongpapanl & Ashraf, 2015; Thongpapanl & Ashraf, 2015). As stated before, users are also more likely these self-referent accept offers. Despite the possibility of a perceived higher level of intrusiveness (Van Doorn and Hoekstra, 2013).
In summary, Tam and Ho (2006) found that both a better ad-fit, and the use of personalization, positively influence the choice behavior. The reaction to personalized content depends on both the degree of personalization and the possible benefits of the personalized message (Van Doorn and Hoekstra, 2013). Since the degree of personal information used in the personalization tools for this study is moderate, it can be assumed that personalization won’t result in the perception of intrusiveness, which in turn results in an decrease of the amount of bookings. Also the fact that the online travel agencies currently already widely use personalization tactics, and thus the visitors are used to these strategies, brings us to the following hypothesis:
2.3 Device type
Mobile devices are becoming more and more popular these days and websites see a continuous increase in mobile visitors every year. Are these visitors from mobile devices, like smartphones and tablets, acting differently Are these visitors contributing to more sales or a shift in sales from desktop to mobile? And will these visitors react differently to personalization? This section gives an overview of the current literature on the situation described above.
Differences between desktop and mobile users
Although hotel websites see an increase in mobile users, the amount of mobile bookings is lagging behind. Data shows that conversion rates for desktops in the travel industry are on average 2.4%, and 0.7% on mobile (Smart Insights, 2018). A research from Google expresses the uncertainty of consumers with mobile booking. More than 50 percent of the consumers switches from a mobile device to a desktop to continue the booking, and double check the prices (Think with Google, 2018). The amount of mobile bookings is only sixteen percent of all bookings hotels get. This percentage is much higher for online travel agencies, mostly due to a better user experience, simplified browsing and a better booking process (Travel Tripper, 2018a). A clear difference between desktop users and mobile users is that desktop users are in a fixed place, and take more time to visit a website, compared to mobile users who are likely on the move, and therefore will be more open to distractions (Travel Tripper, 2018b).
The usage of different devices in the customer journey
According to De Haan (2016) the difference between conversion rates on desktop devices versus mobile devices is lower for customers who have more experience with the online retailer, but on the other hand higher rises when the price goes up. This is in line with research from Chin et al. (2012), which states that a consumer is likely to move from a mobile device to a desktop device when the risk of transaction is high on mobile. It can be concluded that mobile devices are more used in the early stages of the customer journey to look up information and quick searches, whereas later on in the customer journey the customer will switch to a desktop device to make the final purchase, because of the lower perceived risk and better usability (small screens on mobile can be a limitation (Shankar et al. 2010)).
In the current literature less is known about the influence of different device types on the effect of personalization. As described above people still rely on their desktop PC or laptop for making a booking. Although the amount of traffic from mobile devices is increasing, these devices are mostly used for looking up information, rather than booking or making a purchase. Since it can be concluded that the desktops are still mostly used for making the bookings or purchases, we assume that the use of a desktop device has a more positive moderating effect on the effect of personalization on sales in regard to mobile devices like a smartphones or tablets. Therefore the hypothesis is as followed:
2.4 Visitor Source
This section of the theoretical framework will give an overview of the existing literature on the differences between the two sources a web visitor can come from (direct and search engine) and how these differences might influence the effect of personalization on bookings.
Difference between direct and search engine
If a visitor goes directly to a website, without being referred by a third party like a search engine or website such as an OTA, it is labeled by Google Analytics as a direct visitor. The visitor types the URL of the website directly in the internet browser and will end up on the hotel’s website or visits the website via bookmarks. The fact that the visitor is able to find the hotel’s website unaided, means it is likely that the visitor is already familiar with the brand and the corresponding website (Kemmis, 2018). However if a visitor finds the website through a search engine, it is possible they did a more undefined search on an engine (e.g. Google, Yahoo, or Bing). In this case it is more likely that the visitor is not familiar with the hotel or the hotel’s website beforehand. Therefore, it can be assumed that most of these visitors are new visitors, or are visitors that are in an early stage of their customer journey.
One could argue that returning visitors are more likely to visit the website via entering the URL directly in their browser or a bookmark. These returning visitors are most likely in a further stage of their customer journey, and will therefore be more likely to convert. Visitors who visit the website through a search engine on the other hand are more likely to be in an earlier stage of the customer journey, and therefore more likely to be just browsing the website for more information. As there is more information available about the visitors who are in a later phase of the customer journey, it can be argued that personalization tactics are better able to make the messages more relevant and therefore more successful. Therefore, we argue that personalized messages will have more impact on visitors that directly visit a hotel’s website than visitors who visit the website through search engines. The hypothesis is as follows:
H3. The effect of personalization is expected to be stronger with visitors that visit the hotel’s
2.5 Price Segment
This paragraph discusses the possible effects of price on the effect of online personalization on direct bookings. This study aims to investigate whether there is a difference between low and high priced hotels for the effect of online personalization. Since literature on the moderating effect of price on personalization is not widely available, this study looks at the elaboration likelihood model to come to a proper hypothesis.
The elaboration likelihood model describes two ways of processing information and how this processing influences ones attitude change. The two ways of processing are via the central route and via the peripheral route, where motivation and ability are variables that determine which of these two routes will be used. When motivation and ability are high, one will process stimuli via the central route, and when both are low one will process stimuli via the peripheral route. One can be persuaded via the central route when the preferences meet the expectations or preferences. Personalization can be used to get a match between the content and the website visitor’s expectations. Based on the elaboration likelihood model (Petty and Cacioppo, 1986) it can be expected that personalization is more effective for hotels in the higher segment since it is assumed that the motivation is higher with more expensive purchases.
The fourth and final hypothesis of this research argues that online personalization will be more effective for hotels in the higher segment.
3. RESEARCH DESIGN
This chapter will describe the data and the methodology that is used in order to achieve the goal of this study. The data that was used for this research, is a data set that covers cross sectional data of 15 hotels over a period of 15 months. The data set contains data on A/B tests that have been performed on the websites of these 15 hotels. During these A/B tests the visitors of those website were randomly assigned to either the original variant of the hotel website, or to a personalized variant. The next section will cover more information on the data set.
3.1 Data
As described in the introduction of this chapter the data set in this study contains cross sectional data on A/B tests that have been performed on websites of 15 different hotels in the period from 01-01-2017 until 31-05-2018. The data can be considered cross-sectional data, since it consist of a sample of different hotels taken at a certain period in time (Leeflang et al. 2015). Furthermore, the data set consists of data from 2.166.673 visitors, with a total of 48.631 transactions. This data has been aggregated on hotel level, which resulted in 180 observations. An overview of all variables that can be found in the data set are described in table 3.1.
Overview of all the variables in the data set Dependent
Bookings = The amount of bookings Ratio
Independent
Pricing Segment = Price segment of the hotel, low (0) or high (1) Binary Personalization = If sessions were personalized, yes (1), or no (2) Binary Device Type = Type of device during session, desktop, mobile or tablet Nominal Source Type = Type of source the session came from, direct (0) or search (1) Nominal Visitors = The amount of visitors in a particular group Ratio
Bookings = The amount of online direct bookings Count
Data set
3.2 Descriptives
The previous sections gave a short introduction on the data in the data set. This section will dive a little deeper into the data, by discussing the most important descriptives. The entire data set contains data from fifteen different hotels. For research purposes these hotels been divided into two price segments: High and low. The distribution of these hotels amongst the two price segments is as follows: Nine of the hotels are in the low price segment (60%) and the other six are in the high price segment (40%).
The 15 hotels in this data set are located in 10 countries: China, Czech Republic, Dubai, France, Germany, Ireland, Italy, Switzerland, The Netherlands, and United Kingdom. Considering the various countries, sizes, price segments and brands of these hotels, it can be assumed that the hotels perform differently. Table 3.1 gives us an overview of the total visitors and bookings that are in the data set per hotel. The conversion rates range from 0,48% to 6,58%, with an average of 2,02%.
Hotel Visitors Bookings Conversion Rate
1 787055 14598 1.85% 2 287216 5048 1.76% 3 95229 1236 1.30% 4 203816 2662 1.31% 5 255405 16799 6.58% 6 24123 217 0.90% 7 142717 1321 0.93% 8 56158 893 1.59% 9 20211 491 2.43% 10 13062 281 2.15% 11 37452 890 2.38% 12 68895 1497 2.17% 13 45400 1404 3.09% 14 68408 998 1.46% 15 61526 296 0.48%
As can be seen in table 3.2, the majority of the visitors come from desktops (46,8%) followed by mobile devices (40,6%). The smallest amount of visitors are tablet users (12,6%). Table 3.2 also shows that desktops has the highest conversion rate amongst the three different devices (3,26%), followed by tablet (2,59%) and lastly mobile devices (0,96%).
Desktop Mobile Tablet
Visitors 1.013.541 (46,8%) 880.061 (40,6%) 273.071 (12,6%)
Bookings 33096 8471 7064
Avg. CR 3,26% 0,96% 2,59%
Table 3.2 Amount of visitors and bookings per device
Table 3.4 shows a more detailed overview of the amount of visitors and bookings per source (direct and search) and the different devices (desktop, mobile, and tablet). Contradicting to the literature, it seems that visitors who enter the website through a search engine, have a higher conversion rate. However, there are no statistically significant differences between the conversion rates of direct and third party sources found, determined by one-way ANOVA (F(0.693) = 3.94, p = .41).
Direct Search
Visitors 470.186 (21.7%) 1.696.487 (78.3%)
Bookings 9855 38776
Avg. CR 2.10% 2.29%
Desktop Mobile Tablet Desktop Mobile Tablet Visitors 208.823 209.241 52.122 804.718 670.820 220.949
Bookings 6.358 2.092 1.405 26.738 6.379 5.659
Avg. CR 3.04% 1.00% 2.70% 3.32% 0.95% 2.56%
3.3 Missings and outliers
The data set has no missing values, hence there is no further analogy or technique required to solve this problem. There are two cases where the amount of bookings was zero. In order to log transform the booking variable, these cases have been replaced with values based on the average conversion rate.
As described in section 3.1 this data set contains data from different hotels, with different sizes, and different marketing strategies. Table 3.2 showed there are differences between the hotels with respect to the amount of bookings and conversion rates. These differences in bookings may result in outliers that are problematic for the analyses that will be performed in this research. The next chapter will discuss if the outliers are problematic. Figure 3.1 shows the boxplots of the amount of bookings across the different hotels.
3.4 Method
This section discusses the method that has been selected to estimate the effects of personalization, source type, device type and price on the bookings of hotels and their moderating role. A linear regression (OLS) was used to analyze the effects of personalization, source type, device type and hotel type on sales (see equation 1 for the first model). The dependent variable (the amount of bookings) is a count variable. Besides continuous dependent variables, regression models can also be used for models that have count data as a dependent variable, because the error of applying a model for continuous data to count data is really small (Leeflang et al. 2015).
In the figures below you can find the several models that have been designed for this research. The right side of the equation explains the variation in the dependent variable. This side is of the equation has two parts: The first part contains the constant (intercept) and the independent variables and the parameters that belong to those variables. The second part is the error term, which is the stochastic part of the model (Leeflang et al. 2015). The constant is the value of the dependent variable of all independent variables at a value of zero.
The parameters, or so-called beta’s, explain how strong the dependent variable (left part of the equation) responds to one-unit change in the independent variable. In order for the model to be as complete as possible, all independent variables combined, should be the most important contributors to a change in the dependent variable. A model that lacks this completeness may result in an omitted variable bias.
LN B= α + β1 * Pers + β2 * D + β3* D2+β4 * S* β5 *P + ε
Equation 1. Model 1 - Fully pooled model
LN Bi = αi + β1 * Pers+ β2 *D + β3 * D2+β4 *S + β5 * P+ ε
Equation 2. Model 2 - Partially pooled model
LN B= α + β1 * Pers+ β2 *D + β3 * D2+β4 *S + β5 * P + β6 (D* Pers) + β7 (S * Pers) +
Β8 (P * Pers) + ε
Equation 3. Model 3 - Pooled model with interaction effects
LN Bi = αi + β1 * Pers+ β2 *D + β3 * D2+β4 *S + β5 * P + β6 (D* Pers) + β7 (S * Pers)
+ Β8 (P * Pers) + ε
Equation 4. Model 4 - Partially pooled with interaction effects Where:
α = Constant
αi = Constant for hotel i
B = Bookings
Bi = Bookings for hotel i
Pers = Dummy variable for personalization (No versus yes)
D = Dummy variable for device being used (Desktop, mobile, and tablet)
S = Dummy variable for the source of the visitor (Direct versus search)
4. RESULTS
This chapter discusses the estimates that are a result of the performed analyses. These estimates will be used to answer the research question that was defined earlier in this study. The specified models will first be compared to each other by looking at the general model statistics. Secondly, the different models will be tested for the different model assumptions, that are known for OLS models. Lastly, the models will be interpreted.
Model Multiple R2 Adj. R2 F-Statistic DF RSE P-Value
Model 1 0.3239 0.3045 16.67 174 1.551 1.931e-13
Model 2 0.83 0.811 43.67 161 0.8088 < 2.2e-16
Model 3 0.3248 0.2891 9.087 170 1.569 3.597e-11
Model 4 0.8309 0.8072 35.06 157 0.8169 < 2.2e-16
Table 4.1 Model statistics
Table 4.1 shows the different model statistics for the four different models. Looking at the p-values, we see that all models are highly significant (p < 0.01). The differences between the multiple R2 and the adjusted R2 are relatively low for all four models. This is due to the adjusted
R2 controls for the amount of variables in a model, which are relatively low in all four models.
When comparing the R2 values across the models, some differences can be noted. The R2 tells
us how much of the dependent variable is explained by the variables in our model. The adjusted R2 for the fully pooled model (regular OLS) is 30.45%, which is relatively low. For the partially
pooled model (OLSDV) the adjusted R2 is 81.1%, which is relatively high, and a lot higher than
the fully pooled model.
The R2 values are 28.91% for the pooled moderation model (model 3) and 80.72% for the
4.1 Nonzero Expectation
The first assumption, which is the nonzero assumption, is the most serious one for the OLS model. Possible reasons for violation of this assumption are incorrect functional form(s), omitted variables(s), and varying parameter(s), which causes biased parameter estimates (Leeflang et al. 2015). Solutions to a violation of the assumption are a modification of the model specification, adding other relevant predictors and allowing the parameters to vary. In order to test this assumption, the Ramsey Regression Equation Specification Error Test (RESET) was performed. The purpose of the RESET-test is to be able to see if the residuals are a function of the independent variables. The outcome of the RESET-test can be found in table 4.2.
Model Reset DF1 DF2 P-value
Model 1 0 6 168 1
Model 2 0 36 123 1
Model 3 0 6 164 1
Model 4 0 36 119 1
Table 4.2 RESET-test
4.2 Heteroscedasticity
This paragraph discusses the heteroscedasticity assumption for OLS models, which is the second assumption. This assumption is based on the error terms to be homoscedastic, meaning the variances are the same for all cases. Please note that violation of this assumption is not as critical as the first assumption. A violation only decreases the efficiency of the parameter estimates (Leeflang et al. 2015). As a result of heteroscedasticity the estimates of the beta variances become less efficient. According to Leeflang et al. (2015) heteroscedasticity occurs especially in data sets with cross-sectional data.
One can detect heteroscedasticity visually or by using statistical tests. To get an idea of a possible violation of the assumption, we will first look at plots and afterwards we will discuss the results of the Breusch Pagan test, that has been performed to give a clear conclusions on the assumption. Remedies for heteroscedasticity are modification of the model specification or using an heteroscedasticity consistent estimation like GLS.
Figure 4.1 Residual plots for model 1 (Pooled model)
The Breusch Pagan test has been performed to confirm the presence of heteroscedasticity. Table 4.2 presents the outcome of the Breusch Pagan test for the four models. The null hypothesis states that the variance of the residuals is constant. Since the p-values for all four models are much lower than .05, the null hypothesis can be rejected, hence it can be concluded that heteroscedasticity is indeed present in all four models.
Model BP DF P-value
Model 1 30.669 5 1.089e-05
Model 2 81.087 18 5.526e-10
Model 3 31.917 9 0.0002058
Model 4 81.414 22 9.47e-09
Table 4.2 Studentized Breusch-Pagan test
A possible remedy is estimating the models with heteroscedasticity consistent estimation like GLS. Since the data set consists of cross-sectional data, only for one period in time it is not possible to transform the variables for GLS estimation. As a result the models will be estimated with the presence of heteroscedasticity. Like mentioned in the introduction of this paragraph, a violation of the assumption will result in less efficient parameter estimates, which has to be taken into account when interpreting the estimates.
4.3 Correlated Disturbances
4.4 Nonnormal Errors
This paragraph discusses the non-normality assumption for OLS. Outliers can influence the normality, hence an inspection of the histogram can give an idea whether the non-normality assumption should be supported. Non-normality can be spotted by forms of skewness or kurtosis in the histogram. Other ways to detect non-normality are the Shapiro-Wilk normality test, Kolmogorov-Smirnov test, and the Jarque-Bera test. For all tests, the null hypotheses assumes that the investigated residuals are normally distributed. This means that insignificant p-values indicate significant differences from normality.
Non-normality tests, like the ones mentioned above, may show deviations from the normal distribution, because of outliers in the residuals. This is especially the case with relatively small data sets, where outliers will get a lot of weight, which in turn results in a violation of the assumption. Considering the small data set used in this study it is likely that outliers will have more impact, therefore it is possible that the null hypothesis will be rejected.
However, if the model specification seems right, following strict rules with regards to such violations would not be appealing (Leeflang et al. 2015). To see the effect of outliers on the estimates, Leeflang et al. (2015) suggest estimating the model without outliers as well. On the other hand, they agree that assessing which outliers have substantial effect on the OLS is a difficult task.
The residuals of all four models were plotted with a QQ plot and a histogram, to get an idea of the distribution of the residuals. When looking at the residual plot and histogram, we can see the residuals of model 1 (Pooled model) follow the red line quite well. There are a few outliers but they do not seem to be problematic. Furthermore, we can see it has a normal shaped (bell-shaped) distribution (see figure 4.1 and 4.2). For model 2 (Partially pooled model) we see that the residuals follow the linear line quite well. In the histogram it can be seen that the distribution follows slightly positive kurtosis and negative skewness (see appendices D1 & D2). For model 3 (Pooled moderation model) we see the residuals follow the linear red line quite well, and the distribution seems normal (see appendices D3 & D4). For the fourth model (Partially pooled moderation model), we can see some positive kurtosis and negative skewness in the distribution of the residuals (see appendices D5 & D6), which was also the case with the other partially pooled model (model 2).
Besides plots and histograms of the residuals, the Shapiro-Wilk test, Kolmogorov-Smirnov test, and the Jarque-Bera test have been performed. As mentioned in the introduction of this paragraph, significant P-values indicate non-normality patterns. The outcomes of the three tests can be found in table 4.3.
According to the Shapiro-Wilk test and the Kolmogorov-Smirnov test, both model 1 and model 3, are not normally distributed. However, when looking at the Jarque-Bera test we can see the P-value is not significant, indicating a normal distribution. Considering the clear normal distribution in the histograms of model 1 and 2, and the insignificant Jarque-Bera test, it is doubtful whether the distribution is really deviating from a normal distribution. For models 2 and 3, the P-values are higher than .05 for all tests, indicating normally distributed residuals.
Model Shapiro-Wilk Kolmogorov-Smirnov Jarque-Bera
W P-value D P-value AJB P-value
Model 1 0.97494 0.00249 0.074531 0.01639 3.5126 0.137
Model 2 0.99268 0.5036 0.039063 0.7212 1.203 0.5195
Model 3 0.97523 0.002701 0.07448 0.01652 3.4413 0.141
Model 4 0.99214 0.4384 0.038411 0.7446 1.141 0.5325
4.5 Multicollinearity
This paragraph will discuss the multicollinearity assumption for OLS. Multicollinearity occurs when independent variables correlate with each other. Other possible causes of multicollinearity are a bad selection of variables or a lack of data. Although multicollinearity is quite common in marketing science research, it can result in unreliable estimates with too large variances (Leeflang et al. 2015). Symptoms of multicollinearity are insignificant parameter estimates for variables that are theoretically important, or different signs than expected from theory. A way of detecting multicollinearity is by looking at the VIF scores. VIF scores for variables equal to, or larger than 5 are problematic. Multicollinearity can be decreased by changing the variable. Possibilities are recoding, deleting or combining variables. Another possibility is adding different data. An example of this are competitive effects. R was unable to show the VIF scores for the partially pooled models (model 2 and 3). The reason of this was aliased coefficients in the model. After performing an Alias-test in R, it became clear the variable for hotel 15 (h15) and the price variable were perfectly multicollinear. In order to get the VIF scores for the other variables in model 2 and 3, both h15 and the price variable were excluded from the model. Table 4.4 shows the VIF scores for the four models. Besides the relatively high VIF scores for personalization in model 3 and 4, no extreme VIF scores were observed. The high VIF scores for personalization in model 3 and 4 make sense, because of the interaction effects with personalization in the model. These scores are not problematic (VIF < 5).
Variable Model 1 Model 2 Model 3 Model 4
4.6 Pooling
This paragraph discusses the Chow test. This test checks if pooling is allowed and whether pooling or partial pooling is the best option for this study. Since the goal of this study is to design a model which is generalizable, a unit-by-unit model is not favorable. Therefore, the Chow test in this paragraph compares the fully pooled (model 1) versus the partially pooled (model 2) model.
𝐹 = (𝑆𝑆𝑅 𝑝𝑜𝑜𝑙𝑒𝑑 − 𝑆𝑆𝑅 𝑝𝑎𝑟𝑡𝑖𝑎𝑙𝑙𝑦 𝑝𝑜𝑜𝑙𝑒𝑑)/ (𝑑𝑓 𝑝𝑜𝑜𝑙𝑒𝑑 − 𝑑𝑓 𝑝𝑎𝑟𝑡𝑖𝑎𝑙𝑙𝑦 𝑝𝑜𝑜𝑙𝑒𝑑) 𝑆𝑆𝑅 𝑝𝑎𝑟𝑡𝑖𝑎𝑙𝑙𝑦 𝑝𝑜𝑜𝑙𝑒𝑑 / 𝑑𝑓 𝑝𝑎𝑟𝑡𝑖𝑎𝑙𝑙𝑦 𝑝𝑜𝑜𝑙𝑒𝑑
Equation 4.1 Chow test
4.7 Summary
The previous paragraphs discussed the different model assumptions. In paragraph 4.1 the non-zero expectation was discussed. The RESET test was insignificant (P>.05) for all four models, meaning the assumption is not violated, hence there is no misspecification all four models. Paragraph 4.2 discussed the heteroscedasticity assumption. This assumption was violated for all four models. A violation of this assumption results in a decrease in efficiency of the parameters. This violation was expected, given the nature of the data set. Paragraph 4.3 discussed the correlated disturbances assumption. No tests have been performed here since the data set consists of cross-sectional data. In paragraph 4.4 the non-normal errors assumption was discussed. When looking at the plots, it was expected that the assumption would be violated for models 2 and 4. However, did statistical tests show different results. The Shapiro-Wilk test and the Kolmogorov-Smirnov test were significant for model 1 and 3, indicating a deviation from normal distribution. On the other hand the Jarque-Bera test was not significant (P>.05), meaning the distribution would be normal. Paragraph 4.5 discussed the multicollinearity assumption. For models 1 and 3 there were no problematic VIF scores, for the other two models however, the variables for hotel 15 and price were perfectly multicollinear. These variables have been excluded for both models. Table 4.5 gives an overview of the statistical checks. Lastly, has the Chow test been performed to check if pooling is allowed. The insignificant P-value from the Chow test indicates that pooling is allowed. The upcoming sections will discuss which model, the pooled or partially pooled, has the best model fit.
Model Nonzero Heteroscedasticity Nonnormal Errors Multicollinearity
Model 1 Passed Not passed Passed Passed
Model 2 Passed Not passed Not passed Not passed
Model 3 Passed Not passed Passed Passed
Model 4 Passed Not passed Not passed Not passed
4.8 Final model specification and estimation
This paragraph discusses the final model specification, which is based on the outcomes of the tests that have been performed in the previous paragraphs. The estimates of the four models will be provided and discussed in the next paragraphs.
Only a few changes have been made with respect to the models specified in paragraph 3.4. For the partially pooled models (model 2 and 4) the variables h15, price, and the interaction effect between price and personalization were excluded due to multicollinearity reasons.
LN B= α + β1 * Pers + β2 * D + β3* D2+β4 * S* β5 *P + ε
Equation 4.2 Model 1 - Fully pooled model
LN Bi = αi + β1 * Pers+ β2 *D + β3 * D2+β4 *S + β5 * P+ ε
Equation 4.3 Model 2 - Partially pooled model
LN B= α + β1 * Pers+ β2 *D + β3 * D2+β4 *S + β5 (D* Pers) + β6 (S * Pers) + ε
Equation 4.4 Model 3 - Pooled model with interaction effects
LN B= αi + β1 * Pers+ β2 *D + β3 * D2+β4 *S + β5 (D* Pers) + β6 (S * Pers) + ε
4.8.1 Estimation Model 1
This section discusses the estimates of model 1. Table 4.6 shows the overall model statistics for model 1. The model is very significant (P<.01). The R2 is 32,4%, which tells us that the model
explains 32,4% of the variance in the dependent variable, which is the amount of direct bookings in this study. The adjusted R2 is 30,4%. The adjusted R2 applies a penalty for the
amount of the variables that are used in a model. Both the R2 and the adjusted R2 are quite
small, indicating the model does not fit the data very well. However, there still are some significant estimates in the model, which can be very valuable and will be discussed below.
Residual Std. Error R2 Adjusted R2 F-statistic P-value
1.551 (df = 174) 0.324 0.304 16.673*** (df = 5; 174) 1.931e-13
Table 4.6 Output model 1
Table 4.7 represents the estimates from model 1. Model 1 shows us the fully pooled model without any moderation effects. This means that the output of table 4.7 only tells us something about the direct effects on the amount of direct bookings. Price significantly has a negative effect on the amount of bookings (β = -0.82533, P < .01). Personalization has no effect on the amount of bookings (β = 0.08618, P > .05). This is result is remarkable and will be discussed later on in this study. Mobile has a significantly negative effect on the amount of bookings (β = -1.56949, P < .01). Compared to the base case, which is desktop, as expected. Tablet has a significantly negative effect on the amount of bookings (β = -2.01265, P < .01), even more negative than mobile. Source has a significantly positive effect on the amount of direct bookings (β = 0.90309, P < .01). Direct has been coded as 0 and search as 1.
Coefficient Estimate Std. Error T-Value Pr(>|t|)
4.8.2 Estimation Model 2
Model 2 is very significant (P < .01). The R2 is much higher than model 1 (83%). The adjusted
R2 is 81,1%. Given the high R2 values it can be concluded that the model fits the data quite
well. Table 4.9 displays the output from model 2. Model 2 represents the partially pooled model without the moderation effects. The price and h15 variables were both excluded from this model due to multicollinearity. Besides the individual intercepts per chain all the other estimates are the same as with model 1. All intercepts are significant (P < .01) except H6 & H10 (P > .05).
Residual Std. Error R2 Adjusted R2 F-statistic P-value
0.809 (df = 161) 0.830 0.811 43.668*** (df = 18; 161) < 2.2e-16
Table 4.8 Output model 1
Coefficient Estimate Std. Error T-Value Pr(>|t|)
4.8.3 Estimation Model 3
This section discusses the output of model 3. The overall model statistics for model 3 are presented in table 4.10. The overall model is, like the other models, very significant (P<.01), as can be seen in table 4.10. The R2 is 32,5%. The adjusted R2 is 28,9%, which is a bit lower, but
still makes sense since model 3 has the moderation variables as well, and the adjusted R2 applies
a penalty for the amount of variables. Both, the R2 and the adjusted R2 are quite small, indicating
the model does not fit the data very well. However, just as with model 1, there still are some significant estimates in the model which can be very valuable. Again, the price variable is influencing the amount of bookings significantly negative manner (β = -0.84475, P < .01). Personalization and the moderation effects of personalization with the variables of interest are significant (P > 0.1).
Residual Std. Error R2 Adjusted R2 F-statistic P-value
1.569 (df = 170) 0.325 0.289 9.087*** (df = 9; 170) 3.597e-11
Table 4.10 Model statistics
Coefficient Estimate Std. Error T-Value Pr(>|t|)
(Intercept) 4.94717 0.35718 13.850 < 2e-16 *** Price -0.84475 0.33751 -2.503 0.013261 * Personalization 0.25100 0.50513 0.497 0.619903 Mobile -1.45296 0.40501 -3.587 0.000436 *** Tablet -1.93405 0.40501 -4.775 3.85e-06 *** Source 0.95336 0.33069 2.883 0.004449 ** Personalization:Price 0.03884 0.47731 0.081 0.935245 Personalization:Mobile -0.23305 0.57277 -0.407 0.684611 Personalization:Tablet -0.15720 0.57277 -0.274 0.784066 Personalization:Source -0.10054 0.46766 -0.215 0.830038 Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
4.8.4 Estimation Model 4
The overall model statistics for model 4 are presented in table 4.12. The model is very significant (P < .01). The R2 is 83% and the adjusted R2 is 81,1%. Given the high R2 values, it
can be concluded that the model fits the data quite well. Table 4.13 shows the output of model 4. Model 4 represents the partially pooled model with the moderation effects included. The price and h15 variables were both excluded from this model, due to multicollinearity.
Residual Std. Error R2 Adjusted R2 F-statistic P-value
0.814 (df = 158) 0.831 0.808 36.959*** (df = 21; 158) < 2.2e-16
Table 4.12 Model specifications
Coefficient Estimate Std. Error T-Value Pr(>|t|)
(Intercept) 2.6899 0.2847 9.448 < 2e-16 *** H1 4.4435 0.332 13.365 < 2e-16 *** H2 3.2242 0.332 9.698 < 2e-16 *** H3 2.0097 0.332 6.045 1.04e-08 *** H4 2.9408 0.332 8.846 1.72e-15 *** H5 4.6743 0.332 14.060 < 2e-16 *** H6 0.3748 0.332 1.127 0.261281 H7 1.8568 0.332 5.585 9.98e-08 *** H8 1.2160 0.332 3.658 0.000346 *** H9 0.7354 0.332 2.212 0.028398 * H10 0.1646 0.332 0.495 0.621313 H11 1.1552 0.332 3.475 0.000660 *** H12 2.1054 0.332 6.333 2.39e-09 *** H13 2.2564 0.332 6.787 2.18e-10 *** H14 1.6330 0.332 4.912 4.912 Personalization 0.2665 0.2428 1.098 0.273972 Mobile -1.4530 0.2103 -6.910 1.12e-10 *** Tablet -1.9341 0.2103 -9.198 < 2e-16 *** Source 0.9534 0.1717 5.553 1.16e-07 *** Personalization:Mobile -0.2330 0.2974 -0.784 0.434384 Personalization:Tablet -0.1572 0.2974 -0.529 0.597790 Personalization:Source -0.1005 0.2428 -0.414 0.679366 Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
4.9 Validation
This paragraph discusses the face validity and the statistical validity of this study. Section 4.9.1 addresses the face validity and section 4.9.2 discusses the statistical validity of the models.
4.9.1 Face Validity
For face validity one checks if the signs and sizes of the estimates in a model are as expected. Table 4.14 shows the hypothesized effects and the final effects from the four models. The effects were only reported if the estimates were significant. Despite the moderation effects being not significant, the direct effects did show some surprising effects. Please note that due to violation of the heteroscedasticity assumption some parameter estimates might be inefficient and biased.
Even though the direct effects of price, device, and source on the amount of sales were not within the scope of this study, it is also interesting to look at those effects as well. As expected has price a negative effect on the amount of bookings. This is in line with the expectations and general common sense. For device it was expected that mobile and tablet would have a negative effect on the amount of bookings compared to a desktop. The outcomes of the direct effects are in line with this expectation. Source on the other hand, has an opposite effect. It was expected that visitors from search engines would have a less positive effect on the amount of bookings compared to visitors that entered the website directly. The outcomes of models 1, 2, 3, and 4 show a positive effect of source.
Coefficient Hypothesized effect Model 1 Model 2 Model 3 Model 4
Personalization Positive Not sig. Not sig. Not sig. Not sig.
Price Negative Negative - Negative -
Mobile Negative Negative Negative Negative Negative
Tablet Negative Negative Negative Negative Negative
Source Negative Positive Positive Positive Positive
4.9.2 Statistical Validity
This section discusses the statistical validity of this study. The statistical validity has largely been discussed in sections 4.1 – 4.7. This section will therefore give a short overview of the outcomes of the performed tests.
After performing a RESET-test it was concluded that none of the four models had violated the misspecification assumption. Heteroscedasticity was found present in all four models. This resulted in less efficient parameters and slightly biased results. The histograms and plots showed some deviance from normality for models 2 and 4. However, the Shapiro Wilk and Kolmogorov Smirnov test indicated non-normality for models 1 and 3. The Jarque Bera test concluded differently, and showed the distribution of all four models to be normal. Multicollinearity was found present at models 2 and 4 for the variables H15 and price. Those variables were excluded, hence it was not possible to find any effects for those variables and the interaction effect of personalization with price.
All models seemed to have a very high significance level for the overall significance (P < .01). The R2 and adjusted R2 are quite low for model 1 (R2 0,324 and adjusted R2 0,304) and 3 (R2
0,325 and adjusted R2 0,289). Unfortunately, no significant effects were found for
5. CONCLUSION
This chapter provides the conclusion, discussion and the limitations of this study. The aim of this study was to gain insights in the effects of online personalization on direct bookings, and the moderating effects of certain characteristics of the visitor (used device and the source), and the product (price). In a time where big online travel agencies are responsible for a large part of the online bookings, and literature about the use of online personalization for the hotel industry is not widely available, there is a great gap in literature to be filled. This study aimed to do so by investigating the effects addressed above. For small and independent hotels that do not have large marketing budgets it can be really helpful to have more knowledge on the effects of online personalization. This knowledge can help these hoteliers to persuade their website visitors to book directly with them, and in doing so becoming less dependent on large online travel agencies. This resulted in the following research question for this research:
What are the moderating effects of the device, visitor source, and price of the hotel on the relationship between online personalization and bookings in the hospitality environment?
5.1 Discussion
This section discusses the estimation and the outcomes of the different models that have been used, by discussing both the main and moderation effects.
Main effects
The main effect this study aimed to investigate is the effect of online personalization on the amount of direct bookings of hotels. The hypothesis that was states was the follows:
H1. The use of personalization will have a positive effect on the amount of direct bookings.
However the effect of personalization seems to be positive (0.08618, 0.08618, 0.25100, 0.2665)
This research also discovered some other direct effects on the amount of direct bookings for hotels. Although these are not in the scope of this research, it is interesting to briefly discuss these findings.
All four models showed that mobile devices (tablets or smartphones) significantly influence the amount of direct bookings negatively compared to desktop devices. This is in line with prior research, that showed average conversion rates are much higher for desktop devices than for mobile devices (2.4% versus 0.7%) (Smart Insights, 2018). Like mentioned by Chin et al. (2012), it is likely that visitors move from a mobile device to a desktop device due to the higher perceived risk of transactions on mobile devices.
Moderation effects
This section discusses the moderation effects that have been investigated in this study. The first moderation effect is the effect of the devices used on the effect of personalization. The following hypothesis was designed for this moderation effect.
H2. The effect of personalization will be stronger with desktop devices than with mobile
devices and tablets.
After performing several analyses it has been concluded that no significant moderation effects have been found for the devices used on personalization. Therefore, it can be concluded that there is no statistical proof for any positive, or negative effects for the devices used on personalization.
The second moderation effect that has been investigated in this study is the moderation effect of the source of the visitor on the effect of personalization. The effect of this moderation was hypothesized as follows.
H3. The effect of personalization is expected to be stronger with visitors that visit the hotel’s
website directly.
None of the two models that investigated the moderation effects found any significant moderating effects for the source of the visitor on the effect of personalization. Although the effect seems to be as hypothesized (β = -0.1005, P > .1), the results were not significant and therefore will not be further interpreted.
The third, and final moderation effect that was investigated, was the effect of price on online personalization. The relationship was hypothesized as follows.
H4. The effect of personalization is expected to be stronger with hotels in the higher price
5.2 Managerial implications
In times where hotels are highly dependent on external parties like online travel agencies, knowledge on increasing direct bookings has becomes more important than ever. Marketing managers of hotels feel the pressure to increase the amount of direct bookings and are therefore looking for possibilities to do so. This study investigated if personalization could be a possible solution to this problem. By using personalization hotels could be more successful in persuading their visitors to book directly with the hotel. No statistical significant evidence was found for the effect of using personalization on the amount of direct bookings, nor for the moderating effects that were investigated in this study.
Although no evidence was found for the usage of personalization and the possible moderating effects, some other results might be very insightful for marketing managers. This study found that the amount of direct bookings is higher for visitors who visit the website on desktop devices, rather than visitors who visit the website on a mobile phone or tablet. Where desktop is the leading device for making direct bookings, it was found that mobile phones are in second place and tablets in the third place. This insight might be very interesting for marketing managers, to know which device is most interesting to focus on.
5.3 Limitations and future research directions
Just as any other study, this study does have some limitations and drawbacks. This paragraph discusses these limitations and provides a few points for future research.
This study did not control for possible endogenous variables. A Hausman-Wu test could have solved this limitation by testing for endogeneity. Performing this test and solving the possible problem could have improved the parameter estimates and the statistical validity of this study.
The presence of heteroscedasticity in all four models has led to less efficient parameter estimates and biased results. Thirdly, the amount of observations in the data set that was used for this study is relatively low (N = 180). A larger amount of observations might have led to more significant parameters. The data set used in this study consisted of cross-sectional data. Due to the nature of cross-sectional data, it is not possible to see the effect of parameters over time. Data sets with data like time series data, would have been able to do so. By using a data set consisting of time series data it would have been possible to control for seasonality within the data set.
7. REFERENCES
Aguirre, E., Mahr, D., Grewal, D., De Ruyter, K., and Wetzels, M. (2015) Unraveling the Personalization Paradox: The Effect of Information Collection and Trust-Building Strategies on Online Advertisement Effectiveness, Journal of Retailing, Volume 91-1, 34-49
Allen, C., Kania, D., & Yaeckel, B. (2001). Internet world guide to one-to-one web marketing. New York: John Wiley & Sons, Inc.
Business Travel IQ. (2018). OTA market share and rate parity. [online] Available at: https://www.businesstravel-iq.com/article/2017/05/18/ota-market-share-and-rate-parity
[Accessed 26 Apr. 2018].
Carroll, B., & Siguaw, J. (2003). The evolution of electronic distribution: Effects of hotels and intermediaries [Electronic version]. Cornell Hotel and Restaurant Administration Quarterly, 44(4), 38-50. [Accessed 26 Apr. 2018], from Cornell University, School of Hospitality
Chin, E., Felt, A.P., Sekar, V., & Wagner, D. (2012). Measuring User Confidence in Smartphone Security and Privacy. Symposium on Usable Privacy and Security (SOUPS) 2012, July 11-13, 2012, Washington, DC, USA
Cho, CH. (2012) How advertising works on the WWW: Modified Elaboration Likelihood Model
Cöner, A. (2003). Personalization and Customization in Financial Portals, Journal of American Academy of Business, 2(2), pp.498-504
Cristian Morosan, John T. Bowen, (2018) "Analytic perspectives on online purchasing in hotels: a review of literature and research directions", International Journal of Contemporary Hospitality Management, Vol. 30 Issue: 1, pp.557-580
Electronic Channels of Distributions: Challenges and Solutions for Hotel Operators Hospitality Review Volume 24/Issue 2
Pearl Brewer, Andrew Hale Feinstein, Billy Bai, January 2006
Enz, C.A. (2003), “Hotel pricing in a networked world”, Cornell Hotel and Restaurant Administration Quarterly, Vol. 44 No. 1, pp. 4-5.
Eunha Myung , Lan Li & Billy Bai (2009) Managing the Distribution Channel Relationship With E-Wholesalers: Hotel Operators' Perspective, Journal of Hospitality Marketing & Management, 18:8, 811-828.
Gabriel Gazzoli, Woo Gon Kim, Radesh Palakurthi, (2008)
"Online distribution strategies and competition: are the global hotel companies getting it right?", International Journal of Contemporary Hospitality Management, Vol. 20 Issue: 4, pp.375-387
Hinz M., Fiala Z., Wehner F. (2004) Personalization-Based Optimization of Web Interfaces for Mobile Devices. In: Brewster S., Dunlop M. (eds) Mobile Human-Computer Interaction - MobileHCI 2004. Mobile HCI 2004. Lecture Notes in Computer Science, vol 3160. Springer, Berlin, Heidelberg
Ho, S. and Bodoff, D. (2014). The Effects of Web Personalization on User Attitude and Behavior: An Integration of the Elaboration Likelihood Model and Consumer Search Theory. MIS Quarterly, 38(2), pp.497-520.
Kobsa, A., Cho, H., Knijnenburg, B. (2016) The Effect of Personalization Provider Characteristics on Privacy Attitudes and Behaviors: An Elaboration Likelihood Model Approach
Lavie, T., Sela, M., Oppenheim, I., Inbar, O., and Meyer, J. 2010. “User Attitudes Towards News Content Personalization,” International Journal of Human-Computer Studies (68:8), pp. 483-495.
O’Connor, P. (2002), “An empirical analysis of hotel chains online pricing practices”, Information Technology and Tourism, Vol. 5 No. 2, pp. 65-72.
O’Connor, P. (2003), “On-line pricing: an analysis of hotel-company practices”, Cornell Hotel and Restaurant Administration Quarterly, Vol. 44 No. 1, pp. 88-96.
O’Connor, P. 2007. Online consumer privacy: An analysis of hotel company behavior. Cornell Hospitality Quarterly 48 (2): 183-200.
Peppers, D. and Rogers, M. (1997), Enterprise One-to-one: Tools for Competing in the Interactive Age, Doubleday, New York, NY
Peppers, Rogers and Dorf, (1999), “Is Your Company Ready for One-to-OneMarketing”Harvard Business Review, 77, p.p 151– 160.
Peppers, Don. Rogers, M. Dorf, R. (1999). The one to one fieldbook: the complete toolkit for implementing a 1 to 1 marketing program, Double Day, New York
Petty, R. E., Cacioppo, J. T., & Goldman, R. (1981). Personal involvement as a determinant of argument-based persuasion. Journal of Personality and Social Psychology, 41(5), 847-855.
Shankar, V., Venkatesh, A., Hofacker, C. and Naik, P. (2010). Mobile Marketing in the Retailing Environment: Current Insights and Future Research Avenues. Journal of Interactive Marketing, 24(2), pp.111-120.
Smart Insights. (2018). Ecommerce conversion rates 2018 - Smart Insights. [online] Available at: https://www.smartinsights.com/ecommerce/ecommerce-analytics/ecommerce-conversion-rates/ [Accessed 4 May 2018]
Smith, A. (2006) Exploring Service Marketing Aspects of E-Personalization and Its Impact on Online Consumer Behavior
Tam and Ho (2006). Understanding the Impact of Web Personalization on User Information Processing and Decision Outcomes. MIS Quarterly, 30(4), p.865.
Thongpapanl, N. & Ashraf, A. R. (2015) Enhancing Online Performance through Website Content and Personalization, Journal of Computer Information Systems, 52:1, 3-13
Think with Google. (2018). Travel Booking Trends Revealed in Let's-Book-It Moments. [online] Available at: https://www.thinkwithgoogle.com/marketing-resources/micro-moments/travel-booking-trends-book-it-moments/ [Accessed 22 Apr. 2018]
Toh, R., Raven, P. and DeKay, F. (2011). Selling Rooms: Hotels vs. Third-Party Websites. Cornell Hospitality Quarterly, 52(2), pp.181-189.
Travel Tripper. (2018). Important mobile booking stats for hotels in 2018 | Travel Tripper. [online] Available at: http://www.traveltripper.com/blog/important-mobile-booking-stats-for-hotels-in-2018/ [Accessed 24 Apr. 2018].
