IN
DEGREE PROJECT
TECHNOLOGY,
FIRST CYCLE, 15 CREDITS
,
STOCKHOLM SWEDEN 2017
Factors affecting the proportion of
smartphone usage at Flygresor.se
GABRIELLA ANDERSSON
LOUISE KARLSSON
Factors affecting the proportion of
smartphone usage at Flygresor.se
GABRIELLA ANDERSSON
LOUISE KARLSSON
Degree Projects in Applied Mathematics and Industrial Economics Degree Programme in Industrial Engineering and Management KTH Royal Institute of Technology year 2017
Supervisor at CEO of Flygresor.se.: Mattias Nyman Supervisors at KTH: Thomas Önskog, Per Thulin
TRITA-MAT-K 2017:01 ISRN-KTH/MAT/K--17/01--SE
Royal Institute of Technology
School of Engineering Sciences
KTH SCI
SE-100 44 Stockholm, Sweden URL: www.kth.se/sci
Sammanfattning
Fler och fler användare går över till att använda internet från desktop till mo-bila enheter. Denna utveckling har lett till att vi idag står inför ett skifte från e-commerce till m-commerce. Ett problem med detta skifte är att många e-handelsaktörer har svårt att hålla konverteringsgraden på samma nivå i mo-bilen som på desktop. Med detta som bakgrund syftade denna studie till att kartlägga de faktorer som påverkar andelen smartphonebesökare på en hem-sida, i den här studien på jämförelsesajten Flygresor.se. Metoden som användes var multipel linjär regressionsanalys. För att identifiera om vikterna på förk-laringsvariablerna skiljde sig åt mellan andelen Sessioner och andelen Transak-tioner genomfördes två regressioner. Sessioner refererade till andelen inklick på hemsida via smartphone och Transaktioner refererade till andelen vidareklick från hemsida till slutgiltig bokningssajt via smartphone. De förklaringsvari-abler som användes delades in i olika kategorier; Marknadsföring, Trafikslag, Säsong och Övriga, där kategorin Övriga bestod av variablerna Totalt antal besök på hemsida och Dataförbrukning per smartphone per dag. Studien visade att alla kategorier innehöll variabler med signifikant påverkan på responsvari-ablerna. Studien visade också att endast en variabel hade olika påverkan på de två analyserna, nämligen Totalt antal besök på hemsida. Detta indikerar att smartphoneanvändare tenderar att, i förhållande till desktopanvändare, i lägre grad klicka sig vidare till den slutgiltiga bokningssidan. Då ingen annan variabel visade sig ha betydande effekt enbart på regressionsanalysen med
responsvari-Abstract
Digitization has changed the way people access the internet. Smartphones is soon to be the preferred internet access device leading us into a new generation of e-commerce, namely mobile commerce or m-commerce. The on-going transition, from desktop to smartphone has led to an uprising problem for companies within the area of e-commerce. Visitors coming from a smartphone device tend to not go through with the purchase. With this transition in mind, the thesis aimed to identify the factors that affect the proportion of smartphone visitors on a website, more specifically at the flight comparison site Flygresor.se. The method used was multiple linear regression analysis. To see whether the chosen factors affected the proportion of smartphone transactions or just the proportion of smartphone sessions two regression were performed. One with response variable Sessions and one with response variable Transactions, where Sessions refer to the number of visitors on the website and Transactions refer to the number of visitors moving on to the final booking website. The explanatory variables used were divided into four categories; Marketing, Channels, Season and Other, where the category Other contained the variables Total number of visitors and Amount of MB used per smartphone subscription. The study showed that all categories contained variables with significant impact on both of the response variables. There was only one variable that had different impact on the models, namely the Total number of visitors. The result indicates that smartphone users tend to, in comparison with desktop users, to a less extent continue to the final booking website. Since there were no other variables that only had an impact on
Preface
This bachelor’s thesis was written by the students Gabriella Andersson and Louise Karlsson as a part of a five year degree program in Industrial Engineering and Management at KTH, The Royal Institute of Technology.
We would like to thank our supervisor in Applied Mathematics, Thomas Önskog and our supervisor in Industrial Engineering and Management, Per Thulin for giving us feedback and guidance. In addition, we would like to express our grat-itude to Mattias Nyman, the CEO of Flygresor.se, and his colleague Kristoffer
Contents
List of Tables List of Figures 1 Introduction 1 1.1 Background . . . 1 1.2 Problem Statement . . . 4 1.3 Aim . . . 4 1.4 Research Questions . . . 5 1.5 Earlier Research . . . 5 2 Mathematical Theory 7 2.1 The Multiple Regression Model . . . 72.2 Ordinary Least-Squares Estimation of the Regression Coefficients 8 2.3 Fundamental Assumptions . . . 8
2.4 Violations of the Assumptions . . . 9
2.4.1 Endogeneity . . . 9 2.4.2 Heteroskedasticity . . . 9 2.4.3 Multicollinearity . . . 10 2.5 Types of Variables . . . 10 2.6 Model Validation . . . 11 2.6.1 R2and Adjusted R2 . . . . 11
2.6.2 AIC and BIC . . . 12
2.6.3 F-test, P-value and Confidence Intervals . . . 12
3 Method 14 3.1 Collection of Data . . . 14
3.2 Model Creation . . . 14
3.2.1 Response Variables . . . 14
3.2.2 Explanatory Variables . . . 15
3.2.3 Types of Explanatory Variables . . . 17
3.3 Initial Model: Sessions . . . 18
3.3.1 Model Validation . . . 19
3.3.2 Reducing the Model . . . 21
3.4 Initial Model: Transactions . . . 22
3.4.1 Model Validation . . . 23
4.1 Final Model: Sessions . . . 26 4.2 Final Model: Transactions . . . 29
5 Discussion 32
5.1 Model Accuracy . . . 32 5.2 Explanatory Variables . . . 33 5.3 Other Variables Affecting Smartphone Usage . . . 35
6 Conclusion 37
List of Tables
1 Types of Explanatory Variables . . . 17
2 Table of content, Initial Model: Sessions . . . 18
3 Goodness of fit, Initial Model: Sessions . . . 19
4 Statistics for reduction, Initial Model: Sessions . . . 21
5 VIF-values, Initial Model: Sessions . . . 21
6 Table of content, Initial Model: Transactions . . . 22
7 Goodness of fit, Initial Model: Transactions . . . 23
8 Statistics for reduction, Initial Model: Transactions . . . 25
9 VIF-values, Initial Model: Transactions . . . 25
10 Goodness of fit, Final Model: Sessions . . . 26
11 Table of content, Final Model: Sessions . . . 26
12 Standardized β -coefficients, Final Model: Sessions . . . 27
13 Goodness of fit, Final Model: Transactions . . . 29
14 Table of content, Final Model: Transactions . . . 29
15 Standardized β -coefficients, Final Model: Transactions . . . 29
List of Figures
1 Residuals vs. Fitted Values, Initial Model: Sessions . . . 192 Normal QQ-plot, Initial Model: Sessions . . . 20
3 Histogram of Residuals, Initial Model: Sessions . . . 20
4 Scale-Location plot, Initial Model: Sessions . . . 20
5 Residuals vs. Fitted Values, Initial Model: Transactions . . . 23
6 Normal QQ-plot, Initial Model: Transactions . . . 24
7 Histogram of Residuals, Initial Model: Transactions . . . 24
8 Scale-Location plot, Initial Model: Transactions . . . 24
9 Residuals vs. Fitted Values, Final Model: Sessions . . . 27
10 Normal QQ-plot, Final Model: Sessions . . . 28
11 Histogram of Residuals, Final Model: Sessions . . . 28
12 Scale-Location plot, Final Model: Sessions . . . 28
13 Residuals vs. Fitted values, Final Model: Transactions . . . 30
1
Introduction
1.1
Background
During all times of human history, the development of new technologies has changed the way of doing business. The Industrial Revolution brought large-scale changes in both the social and the economic structure, leading to a new era where transformed production processes and business models created new industry models. (Rosenberg and Birdzell, 1987) The society is right now going through another technological revolution called the digitization (MacDonald, Couldry, and Dickens, 2015). According to Zehng digitization is the technology revolution that, after the first Industrial Revolution, have had the biggest impact on the society so far (Qin et al., 2014).
The Oxford English Dictionary (OED) defines digitization as “the action or process of digitizing; the conversion of analogue data (esp. in later use images, video, and text) into digital form”. Digitization is in other words the process of converting something to digits, ones and zeros, and therefore make it on a digital form. The technological revolution today is different from previous revolutions in terms of the pace. Technological improvements have carried on faster than ever since the digitization started. John M Jordan instantiate this through telling about an advertisement that Intel launched in 2005 about air travel saying that if air travel in 1978 would have improved in the same pace as the technological improvement of microprocessors performance and price (known as Moore’s law) a flight ticket would cost one penny from New York to Paris (2012).
Digitization and the evolution of internet has affected many aspects in people’s everyday lives. It has changed the way people communicate, collect information, shop, book trips etc. The development has also changed many aspects of how businesses operate, and has created new business opportunities. The digitization is availing services and products online, not forcing customers to physically go to a specific location to purchase what they want.
Before 1990s, the digitization projects were small scaled, but from then the ap-proach was taken on a larger scale (Rikowski, 1993). In 1990 Tim Berner Lee created the world’s first web browser, the “WorldWideWeb”. Shortly thereafter in 1991 the internet became available for commercial use which enabled the first step of the development towards electronic commerce, commonly referred to as e-commerce. (Madan, 2016) In 1994 the Internet became even more ac-cessible when the browser Netscape was released. Netscape had a new type of
security technology, called the Secure Sockets Layer, the SSL (Chaouchi and Laurent-Maknavicius, 2010). It enabled sensitive information such as credit card numbers to be securely transmitted from one device to another leading to the world’s first internet purchase in 1994. Pizza Hut then sold a pizza online and became the first vendor to receive an online payment (Pizza Hut, 2007) . One year later, two of the pioneers in e-commerce, Amazon and Ebay launched their websites, becoming two of the first companies to sell products through the internet, paving the way for e-commerce (Mirchandan, 2012).
Next big break-through for e-commerce came 1998 with the launch of the money transfer service Paypal (initially Confinity). Paypal made it even easier to purchase and sell products and services online. (Williams, 2007) E-commerce continued to evolve quickly and in the year of 2000 Google launched Google Adwords (Google, 2000). Google Adwords did not only had a big impact on search results but it also changed the e-commerce marketing landscape. The ad-marketing strategy was born. Since then e-commerce has fully exploded and revolutionized industry after industry, not least the travel business accounting for 51 % of the total Swedish e-commerce market with a turn over exceeding 51 billion SEK in 2016 (DIBS, 2016).
Businesses are now facing yet another transition. The development of the smart-phone has changed the way we access the internet leading to next generations e-commerce, the mobile commerce or m-commerce. The transition put new de-mands, as well as it presents new opportunities, for companies operating on the internet.
The expression m-commerce was first introduced in 1997 of Kevin Spacey at the launch of the Global Mobile Commerce Forum. He defined m-commerce as “The delivery of electronic commerce capabilities Directly into the customer’s hand, anywhere, via wireless technology”. (Madan, 2016) The same year the term “smartphone” was coined, this when Ericsson released the mobile device “Penelope” (Ericsson, 2017). The device did however not receive any massive adoption, instead Japan was the first country where the smartphone reached out to a broader public after the Japanese firm NTT DoCoMO released its smart-phone device (Madan, 2016). During this time the rest of the world still focused on regular mobile phones and it was not until the early 2000s, when phone manufacturer BlackBerry started to gain popularity, the smartphone started to became more common in Europe and the USA. These early smartphones were however mainly marketed towards business users. (Halevy, 2009)
During the following years phone manufacturers started to search for the best way to capitalize on the smartphone. In 2006 the resistive touch screen found
its way onto the smartphone (Yamamoto, 2013). The touch screen was however notcompletely well reasoned; it was not suitable for large fingers and demanded a lot of finger pressure. It was all about to change.In 2007 Apple released the first Iphone. The interface was finger-friendly with a large display and capac-itive digitizer. The real context switch, however, came through the operating system, iOS, that offered a WebKit browser. The browser enabled fully rendered websites through the smartphone device.(Apple, 2007) It was the real start of the transition towards m-commerce.
Since then the development of smartphones has continued with high pace. The smartphone screens have become larger and larger in order to create an even better mobile user experience and the smartphone’s processor is soon as good as the ones in personal computers.
Now when the smartphone is on the move to be a fully good complement or alternative to the desktop the advantages with m-commerce have become visible. One of the most obvious advantages with m-commerce is the accessibility of the smartphone. According to Madan people do not think it is convenient to carry around a laptop everywhere they go, and it is not easy to access while on the move, for instance while commuting (2016). The smartphone on the other hand is, because of its ubiquitous presence, always ready to use. In today’s society where people value their time, the accessibility is becoming vital, people want to optimize their time. M-commerce embraces this fact. With m-commerce people do not have to wait until they get back to their desk to purchase their desired trip or product.
Another factor favouring the transition to m-commerce is the development of mobile payment methods. Today there are several mobile payments methods such as Swish, Google Wallet and Apple and Android Pay. These payments method has made it easier for companies to make smoother and simpler check-out processes on mobile sites and in applications. In Singapore mobile pay-ment already stands for 70 % of the country’s total paypay-ments (Hammel-Bonten, 2015).
All put together speaks for that we today are facing a new technology shift, taking e-commerce from the desktop into the smartphone.
One company that is confronting the transition from e-commerce to m-commerce is Flygresor.se. Flygresor.se is Sweden’s largest flight comparison website and was founded in 2007. The company offers a tool that compares prices of flight tickets from 30 different online travel agencies, OTA’S, and thousands of airline companies. They mean that traveling should be a positive and social experience
and aims to make the flight booking a part of that experience. Flygresor.se do not have any product of its own, instead their revenue streams are based on the referral of customers to their partners, namely the online travel agencies or OTA’s and airlines, websites. (Flygresor.se, 2017)
1.2
Problem Statement
The on-going transition, from desktop to smartphone has led to an uprising problem for companies within the area of e-commerce. Visitors coming from a smartphone device tend to not go through with the purchase. For instance, in 2016, 67% of the total time Americans spent on online shopping was through a smartphone in comparison to desktop’s 33%. However, even though people spent more time searching and comparing products through their smartphone, the money spent from each platform differed substantially. 80% of the online spending was through a desktop compared to only 20% on smartphones. (Com-Score, 2017) The same issue is occurring in Sweden. (DIBS, 2016)
To avoid the problem described above, Flygresor.se wants to improve its, as well as helping their partners improve their mobile websites and offers, in order to keep up the profitability and not get behind competitors that fully embraces the transition towards m-commerce.
One important input and knowledge for Flygresor.se is to identify what factors that affect the amount of smartphone usage at the website. The knowledge could be used to see if they could to some extent be able to influence the pace of the on-going transition to smartphone devices on their websites as well as identify if there exist any factors that have a high impact on the smartphone conversion rate.
1.3
Aim
The aim of this thesis is to analyze what factors affect the proportion of smart-phone usage, in comparison to desktop, on the flight comparison website Fly-gresor.se. It is done through multiple regression analysis.
The thesis concerns two regression models. One with response variable pro-portion of smartphone sessions, i.e. the propro-portion smartphone visitors on the website, and another one with response variable proportion of smartphone trans-actions, i.e the proportion smartphone visitors moving on to the final booking website. The regression models will include the same explanatory variables,
coming from categories such as Marketing, Channel and Season, where Channel refers to where the smartphone users come from. Moreover additional variables such as the daily total number of visitors and the average amount of mega bytes used per smartphone subscription per day are included in the models.
1.4
Research Questions
In order to reach the aim, the following specific research questions have been formulated:
• What factors affect the proportion of smartphone visitors at the flight com-parison website Flygresor.se?
• Do the factors affect the proportion of smartphone transactions or just the proportion om smartphone sessions?
1.5
Earlier Research
M-commerce is one of the fastest growing technologies of today and the on-going transition towards m-commerce have led to a large number of studies have been analyzed factors leading to m-commerce adoption. Alain Yee-Loong Chong investigated for instance factors that can be used to predict the adoption of m-commerce on the fast growing Chinese market, through neutral networks, in the study “Predicting m-commerce adoption determinants: A neutral network approach" (2013). Chong concludes that Perceived value, Trust and Social influence are the most significant factors of m-commerce adoption and puts great importance on the influences from peers, family, friends and media. Moreover is the study "Assessing User Attitudes Toward Mobile Commerce In The U.S. Vs. Korea: Implications For M-Commerce CRM", exploring costumers attitude and behaviour towards m-commerce through simple linear regression analysis (Cho, 2008). The factors that were considered most important for cus-tomers were Information and Convenience. The study also shows a high user acceptance towards m-commerce. Hence, there are enormous opportunities for m-commerce. The study points out that all activities performed on Computer-Mediated Communication (CMC) will be replicated by Mobile-Computer-Mediated Com-munication (MMC) in the nearest future.
Lee discusses the possible transition to m-commerce in the study “Strategy, Adoption, and Competitive Advantage of Mobile Services in the Global Econ-omy” (2012). Lee concludes that companies active on the digital market need
to adopt a mobile first approach in order to stay competitive. Madan comes to the same conclusion in the book “Securing Transactions and Payment Systems for M-Commerce” (2016). Moreover, Madan discusses strengths, possibilities, weaknesses and threats with the transition to m-commerce. Madan concludes that the largest challenges with the transition for businesses are that they need to adapt their websites so they are suitable for smartphones screens. Further-more, they need to create a mobile website that feels safe and reliable, in terms of user interface and payments processes.
Mutual for these studies are that they are focusing on factors describing user behaviour and usage intention. However there is a small amount of research analyzing variables that businesses actually might be able to influence, disre-garding user interface, such as marketing and incoming channels. The aim for this thesis is to fill this gap to some extent.
2
Mathematical Theory
In following section the mathematical theory that has been used to perform the multiple regression analysis is presented. The main source for the mathematical theory is the book "Introduction to Linear Regression Analysis" written by Douglas C. Montgomery, Elizabeth A. Peck and G. Geoffrey Vining (2012) and is the reference used if nothing else is referred.
2.1
The Multiple Regression Model
Regression analysis is a widely used method for analyzing the relation between a response variable and one or more explanatory variables. In this thesis a regression from a structural perspective has been performed, hence, it follows the assumption that the explanatory variables affect the response variable and not vice versa. It is based on an observational study through a multiple regression analysis.
A multiple regression model is defined as:
yi= n
X
j=0
xijβj+ εi i = 1, ..., n (1)
where the value of the response variable yidepends on the explanatory variables
xij and the stochastic error term ei. The coefficients βj is estimated from the
acquired data with n number of observations. Eq. (1). can also be written in matrix form:
Y = Xβ + ε (2) where the vectors and the matrix are defined as:
Y = y1 .. . yn β = β1 .. . βk ε = ε1 .. . εn X = 1 x11 · · · x1k .. . ... . .. ... 1 xn1 · · · xnk
2.2
Ordinary Least-Squares Estimation of the Regression
Coefficients
Ordinary least squares is a method for estimating the regression coefficients in Eq.(2). so for every observation, the sum of the squared difference between the explanatory variables and the response variable is minimized.
The vector of least-squares estimates, ˆβ, is obtained by minimizing the func-tion: S(β) = n X i=1 ε2i = ε0ε = (y − Xβ)0(y − Xβ) (3) that is ∂S ∂β|βˆ= −2X 0y + 2X0X ˆβ = 0 (4) hence ˆ β = (X0X)−1X0y (5) given that the explanatory variables are linearly independent, and the inverse correlation matrix X0X−1 exists. When five fundamental assumptions, pre-sented in Section 2.3, are satisfied, ordinary least-squares gives the best linear unbiased coefficient estimates.
2.3
Fundamental Assumptions
There are five fundamental assumptions for ordinary least squares 1. Linearity
The relationship between the response variable y and the explanatory variables are linear, at least approximately.
2. Exogeneity
The error term is expected to have zero mean i.e E(ei) = 0
3. Homoscedasticity
The error term has constant variance i.e V ar(ei) = σ2
4. No auto-correlation The errors are uncorrelated. 5. Normality
The errors are normally distributed.
2.4
Violations of the Assumptions
2.4.1 Endogeneity
Endogeneity occurs when the error term correlates with at least one of the explanatory variables. When endogeneity exists the estimated coefficients will not be unbiased, that is E (βi) 6= 0.
There are three common sources to endogeneity: 1. Simultaneity
Simultaneity is said to be present whenever the response variable causes changes in the explanatory variables, hence the flow goes in both direc-tions.
2. Model miss-specification or omitted variables
If an important explanatory variable is excluded from the regression model the effect of that variable will instead be hidden in the models residual. 3. Measurement errors in the explanatory variables
(Lang, 2016)
2.4.2 Heteroskedasticity
Heteroskedasticity occurs when the assumption of equal variance, V ar(ei) = σ2,
is violated. If OLS is assumed when the residuals are heteroskedastic rather than homoskedastic, the estimated standard deviations will be inconsistent and thus, test-statistics using the standard deviations invalid. (Lang, 2016)
Detecting Heteroskedasticity
The presence of heteroskedasticity can be investigated at a visual level by ex-amine the plot of residuals against fitted values. If the spread of the residuals seems to depend on the fitted values, the residuals are heteroskedastic, and the ordinary least-squares estimates inefficient. (Lang, 2016)
2.4.3 Multicollinearity
Multicollinearity emerges when explanatory variables are correlated and implies that there is a near-linear dependence among two or more explanatory variables. Multicollinearity causes problem since it increases the variance of the regression coefficient estimates and thereby making the coefficients less precise and difficult to interpret.
Detecting Multicollinearity
There are several diagnostics for detecting and quantifying the severity of mul-ticollinearity. A simple method is to examine the off-diagonal elements in the correlation matrix, X’X. If two explanatory variables are highly correlated, the corresponding matrix value will be close to unity. The method is, however, only useful in examine pairwise correlations. If there are more than two variables causing near-linear dependencies there are no guarantee that the pairwise, off-diagonal elements will be large, hence, additional diagnostics must be used to detect multicollinearity. For instance the Variance Inflation Factors, commonly referred to as VIF’s, which is the diagnostic used in this thesis.
The variance inflation factor for the j:th regression coefficient estimate is defined as: V IFj = 1 1 − R2 j (6) where R2
j denotes the coefficient of determination obtained from regressing xj
on the other explanatory variables. Further explanation of R2j will be given in Section 2.6.1. VIF indicates how much a regression coefficients variance increases due to correlation between the explanatory variables.
A VIF-value equal to one indicates that there is no correlation at all between the explanatory variables. There exist various recommendations of the maximum level of VIF. One of the more common recommendations are a maximum VIF level of 10, which is the level applied in this thesis.
2.5
Types of Variables
Variables used in regression analysis can be parted into two groups, quantitative variables and qualitative variables. Quantitative variables have a well-defined scale of measurement, i.e proportion, meters, seconds etc. Qualitative variables on the other hand, do generally not have a scale of measurement, instead a set of
levels is assigned to the variable, representing the effect the qualitative variable has on the regression. Hence indicator variables, also called dummy variables, are used. An indicator variable is constructed so it takes the value one when the event the variable represents occurs, and the value of zero whenever the event not occurs.
2.6
Model Validation
2.6.1 R2 and Adjusted R2
The coefficient of determination or R2 measures the goodness of fit, in other words, how well the model fits the observation. R2is defined as:
R2=SSR SST = 1 −SSRes SST (7) where SSRes = n X i=1 ε2i = n X i=1 (yi− ˆyi)2 SSR= n X i=1 (ˆyi− ¯y)2 SST = SSR+ SSRes
SSRis the regression sum of squares, SST is the corrected sum of squares of the
observations and SSRes is the residual sum of squares. Evidently, the fit of the
model will be better if the residual sum of square is minimized. R2 is generally
said to be the proportion of variation in the response variable explained by the explanatory variables. Since 0 ≤ SSRes ≤ SST, it implicates 0 ≤ R2≤ 1, where
a value close 1 indicates that most of the variability in the response variable can be explained by the model.
The R2 will always increase if more explanatory variables are added to the
model, since it presumes that all the explanatory variables explain the variation in the response variable. Therefore, it can be better to use the modified version called adjusted R2, which gives the percentage of the variation explained only
from those explanatory variables that actually influence the response variable. In other words, the adjusted R2 only increases if the new explanatory variable
improves the model. It is always lower than the R2. The adjusted R2 can be calculated from: R2adj = 1 −(n − 1)(1 − R 2 ) n − k − 1 (8) 2.6.2 AIC and BIC
Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) are useful measures while performing model selection. AIC and BIC measure the relative quality of statistical models, for instance between an initial model and a reduced model: The criterion’s are defined as:
AIC = nln|ε|2+ 2k (9)
BIC = nln SSRes n
+ kln(n) (10) where n and k is the number of observations and explanatory variables respec-tively. The model that minimizes the AIC and BIC is the model preferred. (Lang, 2016)
2.6.3 F-test, P-value and Confidence Intervals
To determine the statistical significance of a regression models estimated coef-ficients the F-test could be used. That a regression coefficient is statistically significant means that it contributes significantly to the model and the null hypothesis below is rejected:
H0: β1= · · · = βk= 0 where k, i ∈ R
The test-statistic for the F-test is defined as:
F = SSR/k SSRes/(n − k − 1)
(11)
where n and k is the number of observations and explanatory variables respec-tively. From this statistic the p-value, can be calculated:
p = P r(X ≥ F ) X ∈ F (1, n − k − 1) (12) which is the probability that the random variable X is larger than the F-value. The null hypothesis is rejected if the p-value is smaller than the significance level α, hence p < α.
If the F-test determines that at least one coefficient in the regression model is statistically significant, the next step is to find out which coefficient or coeffi-cients it is. This procedure, of testing an individual regression coefficient, can be performed through the t-test.
t =
ˆ βj
Std.Error( ˆβj)
(13)
The confidence interval is defined as 1-α, defined by Eq. (14) below. In this thesis the significance level α is sat to 0.05, resulting in a 95 % confidence interval. b βj− tn−k;α 2Std.Error( bβj) < β 0 j < bβj+ tn−k;α 2Std.Error( bβj) (14)
Thus, for a result to be statistically significant the p-value should be less than α and the confidence interval should not contain the value zero. If one of the conditions are satisfied, then the null hypothesis could be rejected.
3
Method
3.1
Collection of Data
The regression was run on data from Flygresor.se and Telia. The original data was retrieved for the period 1 January 2016 until 31 December 2016, where one day equals one data point in the regression models. Due to changes on Flygresor.se’s mobile website in the middle of March, the data set was reduced to only include data from the period 17 March until 31 December 2016. The reduction of days will be further discussed in Section 5.2.
The explanatory variables that were chosen to be included in the models were divided into four groups, Marketing, Channels, Season and Others. The data used for the categories, Marketing and Channels was collected from Flygresor.se. To divide the year’s dates into the seasons the Swedish Meteorological and Hydrological Institutes guidelines were used. The data from Telia described the amount of mobile data used per smartphone subscription per day.
3.2
Model Creation
3.2.1 Response Variables
Two regression analyses were performed with following response variables: Sessions: The number of incoming smartphone sessions relative to the total number of sessions, where sessions refer to the number of visitors entering Fly-gresor.se’s website.
Number of smartphone sessions Total number of sessions
Transactions: The number of smartphone transactions relative to the total num-ber of transactions, where transactions refer to the numnum-ber of visitors continuing from Flygresor.se’s website to the final booking website.
Number of smartphone transactions Total number of transactions
3.2.2 Explanatory Variables
The marketing variables are included to investigate which marketing channels that attract smartphone users to a greater extent than desktop users, as well as identify the ones that do not. The results could be used as an indicator to help companies to set up their marketing strategy to influence the proportion of smartphone users on their website in the desired direction. A hypothesis is that TV commercials will have a positive impact, since one can assume that people are likely to take out the smartphone during commercial breaks.
Variables representing the channels the visitors coming from, are included to review which channels that currently do or do not bring in a satisfactory pro-portion of smartphone users. As mentioned in Section 1.1, smartphone internet traffic has substantially increased the last couple of years, and will most likely continue to increase. Consequently, it is important for online vendors to be ready for the transition towards m-commerce, making sure all incoming chan-nels are well adapted for smartphone users. The result could be an indication of how the resources should be allocated.
Seasonal changes might influence the proportion of smartphone users and sea-sonal patterns can occur. It can be hypothesized that during summer and holidays people spend more time outdoors and are more likely to use their smartphones than their desktop devices.
It is now known that there is a increasing trend of smartphone usage. It might well be the most influential factor that effect the proportion of smartphone users on Flygresor.se’s website. The increasing trend is represented by the data re-trieved from Telia.
Marketing variables
• Facebook: daily amount of SEK invested in marketing on Facebook. • Google: daily amount of SEK invested in marketing on Google. • Bing: daily amount of SEK invested in marketing on Bing. • TV: daily amount of SEK invested in TV-marketing. • Radio: daily amount of SEK invested in Radio-marketing.
Channel variables
All channel variables are smartphone proportions of the daily number of total visitors.
• Direct: visitors typing the sites URL direct in browser.
• Organic: visitors incoming from searches, including generic keywords and branded keywords, where branded keywords are specific to the company’s name or brand and generic keywords are words that have a much broader reach, e.g "Flight Tickets".
• Email: visitors incoming from hyperlinks sent out through emails. • Referral: visitors incoming from hyperlinks from sources outside of its
search engine.
• Paid_search: visitors incoming from searches including paid generic key-words and paid branded keykey-words, where branded keykey-words are specific to the company’s name or brand and generic keywords are words that have a broader reach, e.g "Flight Tickets"
• Social: visitors incoming from social medias.
Seasonal variables
• Holiday: refers to if there was a holiday in Sweden at the specific day, where no holiday was set as the benchmark.
• Season: refers to what season there at the specific day. It was divided into four various season according to SMHI, where Summer was set as the benchmark.
Other variables
• Data_day: the average amount of mobiledata used at the specific day. The data from Telia was collected monthly, therefore an average per day was calculated.
Average amount of megabyte mobile data a person uses at the specific month The number of days at the specific month
• Total_visits: the total number of visitors on the website on the specific day.
3.2.3 Types of Explanatory Variables
In Table 1 all the explanatory variables are visualized, describing what type and in which unit each one of them was measured.
Explanatory variables Type Unit Marketing
Facebook Quantitative SEK Google Quantitative SEK Bing Quantitative SEK TV Quantitative SEK Radio Quantitative SEK Channels Direct Proportion % Organic Proportion % Email Proportion % Referral Proportion % Paid_search Proportion % Social Proportion % Season Holiday Indicator 0 or 1 Winter Indicator 0 or 1 Spring Indicator 0 or 1 Autumn Indicator 0 or 1 Other Data_day Quantitative MB Total_visit Quantitative Number
3.3
Initial Model: Sessions
The initial model contained all the explanatory variables described in Section 3.2.2. The variables were inserted in the Eq.(1) from Section 2.1 and is presented below.
Sessions = β0+β1(F acebook)+β2(Google)+β3(Bing)+β4(T V )+β5(Radio)+
...β6(Direct) + β7(Organic) + β8(Email) + β9(Ref erral) +
...β10(P aid_search) + β11(Social) + β12(Holiday) + β13(W inter) +
...β14(Spring)+β15(Autumn)+β16(Data_day)+β17(T otal_visit)
In Table 2 and Table 3 the Initial Model for Sessions coefficient estimates and diagnostics are featured. The p-value for the entire model was less than 2.2 · 10−16, thus the hypothesis that all explanatory variables are equal to zero was
rejected. However, a few of the variables had individual p-values that were greater than the significance level and a confidence intervals that contained zero. Namely, Radio, Paid_search, Autumn, Google, Bing, Email and Facebook. These detections will be further investigated in Section 3.3.2.
Estimate Std.Error p-value lower upper (Intercept) 7.510 · 10−1 5.978 · 10−2 0.000 6.333 · 10−1 8.688 · 10−1 Marketing Facebook 5.569 · 10−7 2.547 · 10−7 0.029 5.533 · 10−8 1.058 · 10−6 Google −2.094 · 10−7 1.235 · 10−7 0.091 −4.527 · 10−7 3.387 · 10−8 Bing −8.436 · 10−7 5.384 · 10−7 0.118 −4.0201 · 10−6 2.166 · 10−7 TV 1.509 · 10−7 2.929 · 10−8 0.000 1.640 · 10−7 3.167 · 10−7 Radio 8.197 · 10−8 2.447 · 10−7 0.738 −3.999 · 10−7 5.638 · 10−7 Channel Direct −1.944 1.175 · 10−1 0.000 −2.175 −1.712 Organic 1.000 1.138 · 10−1 0.000 7.762 · 10−1 1.225 Email 3.381 · 10−1 1.693 · 10−1 0.169 4.576 · 10−3 6.715 · 10−1 Referral 2.363 · 10−1 1.085 0.030 2.261 · 10−1 4.499 Paid_search −6.368 · 10−1 6.275 · 10−2 0.169 −3.698 · 10−2 2.102 · 10−1 Social −6.368 · 10−1 2.881 · 10−1 0.028 −1.204 −6.936 · 10−2 Season Holiday 2.751 · 10−2 3.758 · 10−3 0.000 2.011 · 10−2 3.492 · 10−2 Winter 2.505 · 10−2 6.621 · 10−3 0.000 1.200 · 10−2 3.809 · 10−2 Spring 2.505 · 10−2 6.411 · 10−3 0.000 −3.632 · 10−2 −1.107 · 10−2 Autumn 9.376 · 10−3 5.399 · 10−3 0.084 −1.255 · 10−3 2.000 · 10−2 Other Data_day 2.751 · 10−2 4.662 · 10−4 0.002 5.709 · 10−4 2.407 · 10−3 Total_visit 1.670 · 10−6 2.414 · 10−7 0.000 2.145 · 10−6 1.194 · 10−6
Table 2: Table of content, Initial Model: Sessions
As can be seen in Table 3, the models R2 is 0.848 indicating that the created model explains 84.8 % of the variation in the data. The adjusted R2is 83.8 %,
which is slightly smaller because the calculation reckons the degrees of freedom.
R2 Adjusted R2 Degrees of freedom Residual standard error 0.848 0.838 257 0.019
Table 3: Goodness of fit, Initial Model: Sessions
3.3.1 Model Validation Linearity assumption
To verify the assumption of linear relationship between the response variable and the explanatory variables a scatter plot with the residuals on the y axis and the fitted values on the x axis was used. The plot shows that the residuals are fairly equally spread around the horizontal axis, not showing any distinct pattern or having any extreme outliers. Consequently, it indicates the linearity assumption is met.
Figure 1: Residuals vs. Fitted Values, Initial Model: Sessions Normality assumption
The normality assumption was checked by looking at the plot of the standard-ized residuals plotted against the quantiles, i.e the normal Q-Qplot, and the histogram of the standardized residuals. According to the plots, the residu-als seems fairly normal distributed. The QQ-plot shows tendencies of a light
tail, however, the majority of the observations follows the line. The histogram has a pattern similar to a bell-shaped curve which indicates normal distribu-tion.
Figure 2: Normal QQ-plot, Initial Model: Sessions
Figure 3: Histogram of Residu-als, Initial Model: Sessions
Homoskedasticity assumption
To verify the assumption of homoscedasticity the Scale-Location plot was used. The plot shows randomly spread residuals indicating that the assumption is met.
Figure 4: Scale-Location plot, Initial Model: Sessions
3.3.2 Reducing the Model Significance
After looking at the p-value and the confidence interval for each explanatory variable, the variables with a high p-value and a confidence interval that con-tained zero, were removed. The variables were removed one at the time where the reduced models adjusted R2, AIC and BIC were analysed. After remov-ing all insignificant variables the adjusted R2decreased and the AIC increased
the slightest, however the BIC value decreased and the reduced model did not contain any statistically insignificant variables.
Removed variable Adjusted R2 AIC BIC
Base(none) 0.8384 -1387.012 -1318.293 Radio 0.8389 -1388.891 -1323.790 Paid 0.8383 -1388.819 -1327.333 Autumn 0.8378 -1388.885 -1331.016 Google 0.8375 -1389.223 -1334.972 Bing 0.8363 -1388.134 -1337.499 Email 0.8347 -1386.533 -1339.515 Facebook 0.8333 -1385.161 -1341.759
Table 4: Statistics for reduction, Initial Model: Sessions Variance Inflation Factors
The presence of multicollinearity was examined through VIF’s, variance infla-tion factors. All values were below ten, indicating that there is no severe mul-ticollinearity in the model. Consequently, no more explanatory variables were removed.
Explanatory variable VIF
Data_day 7.179 Organic 5.176 Direct 4.159 Winter 2.732 Spring 2.428 Total_visit 2.037 Social 1.718 TV 1.659 Holiday 1.645 Referral 1.321
3.4
Initial Model: Transactions
A second regression was performed with the same explanatory variables, but this time with the response variable Transactions. The equation is presented below.
Transactions = β0 + β1(F acebook) + β2(Google) + β3(Bing) + β4(T V ) +
...β5(Radio) + β6(Direct) + β7(Organic) + β8(Email) +
...β9(Ref erral) + β10(P aid_search) + β11(Social) +
...β12(Holiday) + β13(W inter) + β14(Spring) + β15(Autumn)
...+ β17(Data_day) + β18(T otal_visit)
In Table 6 and Table 7 the Initial Model for Transactions coefficient estimates and diagnostics are featured. Similarly to the Sessions Model, the hypothesis that all variables were equal to zero was rejected (p-value < 2, 2·10−16). However the hypothesis that none of the variables equaled zero was not rejected. P-values that were bigger than the significance level and confidence intervals containing zero were discovered. In this case, Radio, Paid_search, Autumn, Google, Bing, Email and Facebook. These detections will be further investigated in Section 3.4.2.
Estimate Std.Error p-value lower upper (Intercept) 4.347 · 10−1 6.181 · 10−2 0.000 3.130 · 10−1 5.564 · 10−1 Marketing Facebook 3.133 · 10−7 2.634 · 10−7 0.235 −2.054 · 10−7 8.319 · 10−7 Google −5.670 · 10−8 1.277 · 10−7 0.658 −3.082 · 10−7 1.949 · 10−7 Bing −7.777 · 10−7 5.567 · 10−7 0.164 −1.874 · 10−6 3.186 · 10−7 TV 1.535 · 10−7 3.029 · 10−8 0.000 9.390 · 10−8 2.132 · 10−7 Radio −1.442 · 10−7 2.530 · 10−7 0.569 −6.425 · 10−7 3.540 · 10−7 Channel Direct −1.348 1.216 · 10−1 0.000 −1.588 −1.109 Organic 8.034 · 10−1 1.177 · 10−1 0.000 5.715 · 10−1 1.035 Email 6.444 · 10−1 1.751 · 10−1 0.000 2.995 · 10−1 9.892 · 10−1 Referral 6.400 · 10−1 1.122 0.570 −1.569 2.849 Paid_search 6.134 · 10−2 6.489 · 10−2 0.345 −6.645 · 10−2 1.891 · 10−1 Social −3.644 · 10−1 2.979 · 10−1 0.222 −9.511 · 10−1 2.223 · 10−1 Season Holiday 2.590 · 10−2 3.886 · 10−3 0.000 1.824 · 10−2 3.355 · 10−2 Winter 3.061 · 10−2 6.847 · 10−3 0.000 1.713 · 10−2 4.409 · 10−2 Spring −7.816 · 10−3 6.630 · 10−3 0.240 −2.087 · 10−2 5.240 · 10−3 Autumn 2.610 · 10−2 5.583 · 10−3 0.000 1.511 · 10−2 3.709 · 10−2 Other Data_day 2.295 · 10−3 4.820 · 10−4 0.000 1.346 · 10−3 3.245 · 10−3 Total_visit −1.621 · 10−6 2.497 · 10−7 0.000 −2.113 · 10−6 −1.130 · 10−6
Table 6: Table of content, Initial Model: Transactions
As can be seen in Table 7, the models R2 is 0.805 indicating that the cre-ated model explains 80.5 % of the variation in the data. The adjusted R2 is
79.2
R2 Adjusted R2 Degrees of freedom Residual standard error
0.805 0.792 257 0.019 Table 7: Goodness of fit, Initial Model: Transactions
3.4.1 Model Validation
The Initial Transactions Model was validated using the same tools as the Ses-sions model.
Linearity assumption
As can be seen in Figure 5 the residuals are equally spread along the x-axis not showing any clear pattern. Hence, the assumption of linear relationship between the response variable and the explanatory variables is considered verified.
Figure 5: Residuals vs. Fitted Values, Initial Model: Transactions
Normality assumption
The normal QQ-plot shows a light positive tail. However, the the majority of the observations follows the line and the histogram shows a pattern of a bell-shaped
curve, indicating normal distribution. Hence, the model met the assumption of normal distributed residuals.
Figure 6: Normal QQ-plot, Initial Model: Transactions
Figure 7: Histogram of Residu-als, Initial Model: Transactions
Homoskedasticity assumption
The Scale-location plot shows that the residuals are equally spread along the range of fitted values, hence it is concluded that the model meets the assumption of homoskedasticity.
Figure 8: Scale-Location plot, Initial Model: Transactions
3.4.2 Reducing the Model Significance
In order to identify and remove non-significant variables, the same procedure was used as in the regression with Sessions. Changes in AIC, BIC and adjusted R2 were analyzed after removing one variable at a time, starting with the one
with the largest P-value. The diagnostics suggested different models, although after analyzing the confidence interval for each variable in each model it was concluded to remove all of the below listed variables, resulting in a model of nine variables, all statistically significant.
Removed variable Adjusted R2 AIC BIC Base (none) 0.7916 -1368.581 -1299.863 Google 0.7923 -1370.371 -1305.269 Radio 0.7928 -1372.019 -1310.534 Referral 0.7932 -1373.394 -1315.526 Social 0.7929 -1374.000 -1319.738 Facebook 0.7932 -1575.293 -1324.658 Bing 0.7929 -1375.860 -1328.842 Paid_search 0.7917 -1375.241 -1331.840 Spring 0.7905 -1374.630 -1334.845 Table 8: Statistics for reduction, Initial Model: Transactions
Variance inflation factors
None of the variables had a VIF-value exceeding 10, hence no severe multi-collinearity is present.
Explanatory variables VIF
Organic 5.094 Direct 4.028 Data_day 3.695 Winter 3.082 Total_visit 2.003 Autumn 1.670 TV 1.551 Holiday 1.381 Email 1.061
4
Results
4.1
Final Model: Sessions
Sessions = β0+β1(T V )+β2(Direct)+β3(Organic)+β4(Ref erral)+β5(Social)
...+ β6(Holiday) + β7(W inter) + β8(Spring) + β9(Data_day)
...+ β10(T otal_visit)
The goodness of fit as well as the adjusted R2 have decreased slightly, however as can be seen in Table 11, none of the variables have confidence intervals that contain zero. It implies that all of the explanatory variables contribute significantly to the model. Moreover, the degrees of freedom has increased and the residual standard error has neither increased or decreased.
R2 Adjusted R2 Degrees of freedom Residual standard error
0.839 0.833 264 0.019 Table 10: Goodness of fit, Final Model: Sessions
Estimate Std.Error p-value lower upper (Intercept) 8.444 · 10−1 5.097 · 10−2 0.000 7.444 · 10−1 9.444 · 10−1 Marketing TV 1.359 · 10−7 2.664 · 10−8 0.000 8.347 · 10−8 1.884 · 10−7 Channel Direct −1.980 1.012 · 10−1 0.000 −2.179 −1.780 Organic 9.287 · 10−1 1.001 · 10−1 0.000 7.315 · 10−1 1.125 Referral 3.055 1.025 0.003 1.037 5.073 Social −6.448 · 10−1 2.639 · 10−1 0.015 −1.164 −1.125 · 10−1 Season Holiday 2.680 · 10−2 3.709 · 10−3 0.000 2.130 · 10−2 3.590 · 10−2 Winter 1.564 · 10−2 4.715 · 10−3 0.001 6.357 · 10−3 2.492 · 10−2 Spring −2.649 · 10−2 5.509 · 10−3 0.000 −3.734 · 10−2 −1.564 · 10−2 Other Data_day 1.229 · 10−3 4.139 · 10−4 0.000 4.140 · 10−4 2.044 · 10−3 Total_visit 2.090 · 10−6 1.332 · 10−7 0.000 2.352 · 10−6 1.827 · 10−3
Table 11: Table of content, Final Model: Sessions
To be able to compare the magnitude of the effect that each explanatory vari-able has on the response varivari-able the regression coefficients are standardized. These coefficients represent the mean change in the response variables standard deviation per standard deviation change in the explanatory variables.
TV Direct Organic Referral Social 0.162 -0.984 0.521 0.085 -0.079 Holiday Winter Spring Data Total_visits 0.244 0.135 -0.185 0.196 0.152 Table 12: Standardized β -coefficients, Final Model: Sessions
The graphs for the Sessions Final Model indicates that none of the fundamental assumption has been violated.
Figure 10: Normal QQ-plot, Final Model: Sessions
Figure 11: Histogram of Resid-uals, Final Model: Sessions
Figure 12: Scale-Location plot, Final Model: Sessions
4.2
Final Model: Transactions
Transactions = β0+ β1(T V ) + β2(Direct) + β3(Organic) + β4(Email) +
...β5(Holiday) + β6(W inter) + β7(Autumn) + β8(Data_day)
...+ β9(T otal_visit)
As for the Sessions model, the goodness of fit and the adjusted R2have decreased marginally. Anyhow, now all the explanatory variables contribute significantly to the model. The degrees of freedom has increased and the standard error is the same.
R2 Adjusted R2 Degrees of freedom Residual standard error 0.797 0.791 265 0.019
Table 13: Goodness of fit, Final Model: Transactions
Estimate Std.Error p-value lower upper (Intercept) 4.238 · 10−1 4.127 · 10−2 0.000 3.426 · 10−1 5.051 · 10−1 Marketing TV 1.337 · 10−7 2.630 · 10−8 0.000 8.196 · 10−8 1.855 · 10−7 Channel Direct −1.304 1.017 · 10−1 0.000 −1.504 −1.104 Organic 8.413 · 10−1 1.014 · 10−1 0.000 6.416 · 10−1 1.041 Email 6.139 · 10−1 1.701 · 10−1 0.000 2.791 · 10−1 9.488 · 10−1 Season Holiday 2.461 · 10−2 3.471 · 10−3 0.000 1.778 · 10−2 3.144 · 10−2 Winter 2.618 · 10−2 5.114 · 10−3 0.000 1.611 · 10−2 3.624 · 10−2 Autumn 2.477 · 10−2 4.664 · 10−3 0.000 1.558 · 10−2 3.395 · 10−2 Other Data_day 2.688 · 10−3 3.032 · 10−4 0.000 2.091 · 10−3 3.285 · 10−3 Total_visit −1.812 · 10−6 1.348 · 10−7 0.000 −2.078 · 10−6 −1.547 · 10−6
Table 14: Table of content, Final Model: Transactions
TV Direct Organic Email Holiday 0.175 -0.712 0.517 0.103 0.230 Winter Autumn Data Total_visits
0.249 0.190 0.471 -0.526
The graphs for the Transaction Final Model indicates that the fundamental assumption for Ordinary Least Squares are met.
Figure 13: Residuals vs. Fitted values, Final Model: Transactions
Figure 14: Normal QQ-plot, Final Model: Transactions
Figure 15: Histogram of Resid-uals, Final Model: Transactions
5
Discussion
5.1
Model Accuracy
The final models have both relatively high coefficients of determination, 0,839 and 0,797 for Sessions and Transactions respectively. The value is considered quite high, however there exist other factors, that have been excluded due to data unavailability or due to unquantifiability, that most likely have high impact on the proportion of smartphone users. For instance, the models do not take the websites user experience and user interface into account. However, the user interface would only be considerable if many different websites would be compared or if the comparisons were made on several versions on the same website, which was not the case in this thesis. Quantifying these factors in a suitable way and including them in the models would most likely generate even higher coefficients of determination, especially for the Transactions model where these factors is assumed to have high explanatory power. These factors will be further discussed in section 5.3.
The data for the models was only collected for the year of 2016. The reason to this choice of short period was to minimize the impact of the increasing trend of smartphone usage. Considering observations ranging over several years, the trend would most likely be the only variable with substantial impact. After taking changes on the mobile website into consideration the data set was re-duced to only include data points from the middle of March 2016 to the end of December the same year. It was done to avoid interfering effects from the old mobile website which could lead to inaccurate and obsolete results. However, the choice of period and reduction of data points also resulted in that the final models contained relatively few observations, i.e. 264 observations each. The few observations and the fact that the models only took one specific year into consideration could have a negative impact on the models accuracy.
The increasing trend of smartphone usage is essential for the proportion of smartphone users at websites, this factor is however hard to quantify. In an attempt to take the trend into consideration, the average amount of data used per smartphone subscription per day was included in the model to represent the trend. However, this data might not display the trend totally accurate. Firstly, the data only take the amount of data used per smartphone subscription in Telia’s network in consideration, not giving a totally accurate picture of the data usage per smartphone in Sweden overall. Secondly, the data also includes subscriptions for mobile broadband, which not necessarily have to be connected to a smartphone, instead these subscriptions are often used in tablets
and desktops while traveling. These subscriptions only made up for a tiny amount of the subscriptions (less than 5 %) so they were considered to be negligible, however, this is signalling that this data may not be perfectly suitable to represent the increasing trend of smartphone usage. Lastly the amount of data used per smartphone subscription do not only have to do with how much time the subscription user spends online. It also has to do with what the user do while online. If the user for instance downloads a lot of heavy programs, streaming movies or listen to a lot of music, that would also result in a higher amount of data usage.
5.2
Explanatory Variables
In this part of the discussion all the statistically significant explanatory variables were analyzed through interpretation of the results of the final model for both Sessions and Transactions.
Marketing
TV- commercial is the only marketing variable that, according to the regres-sion models, have a statistically significant impact on both the proportion of smartphone sessions and the proportion of smartphone transactions. The stan-dardized β -coefficient is taking a positive value, 0.162 and 0.175 for sessions and transactions respectively.Thus the hypothesis in section 3.2.2 is confirmed. Increasing the amount of money spent on TV-commercial would therefore posi-tively affect the proportion of smartphone users at the site. It is reasonable since it can be assumed that a smartphone is more easily accessible than a desktop while watching TV.
Channels
Among the channel variables that were included in the final models, Direct and Organic had most influential effect on the proportion of the smartphone users for sessions as well as for transactions.
The β -coefficients for Direct are negative for Sessions and Transactions. A negative β -coefficients could at first be confusing, but this result indicates that there are more visitors coming from other devices than smartphones from this channel. Thus, it will affect the proportion of smartphone users in a negative manner.
The β -coefficients for Organic are 0,521 and 0,517 for session respectively trans-action. The interpretation could be that it is more common to search for a branded key word than to type the URL direct in the browser on the
smart-phone.
Furthermore, Referral and Social have a small impact on Sessions. While Re-ferral is conducing to a higher proportion of smartphone users, Social is giving a smaller proportion. For Transactions the variable Email had a small posi-tive impact, indicating that promotion emails could increase the proportion of smartphone transactions.
Season
Seasonal changes do in fact influence the proportion of smartphone users. The hypothesis, presented in Section 3.2, that Summer should have the highest pos-itive impact on the proportion of smartphone users is however not supported by the models. Instead both models show that Winter is the time of the year that the proportion of smartphone users is the highest, both for sessions and transactions. Moreover, is Spring statistically significant for sessions, indicat-ing that durindicat-ing this season the proportion of smartphone users decline. For transactions, Autumn is statistically significant indicating that the number of smartphone users raise during this season, however, not as much as during win-ter.
Since this thesis only look at observations from 2016, from the middle of march to the end of December, the first part of the year, also defined as winter is not taken into consideration: This leads to that the seasons come in the order spring, summer, autumn and winter. The variable considering the data usage per smartphone takes the arising trend of smartphone usage in account. How-ever, by looking at the standardized coefficients a pattern can be seen, that is that the number of smartphone users increase with the seasons over the year and hence, even more rapidly than what the variable Data_day portrays. Since there is an arising trend of smartphone usage, this could in fact be the expla-nation to why winter is considered to be the season that have the most positive effect on the proportion of smartphone users.
Other variables
The β-estimate for the variable Data_day, which represents the rising trend of smartphone usage, is positive for both models indicating that a larger average amount of mobile data results in a larger proportion of smartphones users. It is reasonable and verifies that the amount of mobile data indeed has a lot of influence on the proportion of smartphone users. However, it is important to bear in mind that the data used in these regression models only takes the average amount of data used per smartphone subscription in Telia’s network in consideration, possibly not giving a totally accurate picture.
The number of total visitors is the only variable where the two regression models standardized coefficients differ in sign. In the model for sessions the coefficient is slightly positive, indicating that more visitors lead to a higher proportion of smartphone users. For Transactions on the other hand, the coefficient is negative, indicating that the proportion of smartphone users going through to the final booking site decline with the number of total visitors. The result highlights the challenge of conversion explained in section 1.2. More and more customers are using their smartphone going online shopping, but businesses struggle to keep up the conversion rates through this device. And even though customers might return to the website to finalize their purchase later on, while at their desk, there is always a chance that the customer forgets about the purchase or finds another offer somewhere else. Hence, the companies who can sell the customers what they want straight away will close sales, while the ones who makes the customers wait until they get back to their desks risk to lose out.
5.3
Other Variables Affecting Smartphone Usage
When e-commerce was released a new level of convenience was brought to cus-tomers. However, some businesses had a hard time adopting the substitution of human interaction and chose not to adapt to the new paradigm shift putting their business at risk. Consequently, new opportunities occurred, but also nu-merous obstacles to overcome. In a similar way m-commerce is upcoming with new opportunities and obstacles. In the literature "Securing Transactions and Payment Systems for M-commerce" Madan argues that there are two factors that have high impact on m-commerce transactions (2016). As m-commerce emerges these factors need to be properly analyzed by the companies in order for them to stay competitive. These factors are the mobile website’s user expe-rience and the level of certainty that customers feel while shopping through the smartphone.
User interface/ User experience
One reason leading to that customers not finalizing their purchases could be that the company’s mobile website is not enough well-adapted for smartphone users. For instance search functions and click-through processes might not be prop-erly designed for a small touch-screen. It could be reason enough for potential customers to instead turn to a competitor with a mobile-first approach where the customers can go through with their booking or purchase straight away; not having to wait until they get back to their desktop to finalize their booking.
Applications, commonly referred to as apps enables better user experience and is another way to access company’s value offerings through a smartphone. Apps enables better user experience than mobile websites in terms of layout and de-sign since apps don’t need browser features as address bars and refresh buttons and could be an alternative or complement to the company’s mobile website. Apps also offers functionalities not available through mobile websites. A com-pany could for instance reach customers in real time and send push notifications to attract attention and encourage them to use the app. It could be a decisive tool to keep up the retention rates. Moreover, Apps promotes loyalty among customers since the customer has the company’s icon permanently exposed on their smartphone menu, constantly reminding them of the company. It will increase the possibility that customers first turn to the company and its app whilst looking for a product in its range.
Uncertainty
According to Madan do some customers feel unsecure when shopping through the smartphone. It consorts with the findings in DIBS annual report regarding e-commerce, saying that people feel that it is more uncertain to do purchases through a smartphone than through a desktop. (DIBS, 2016). There is only space for a limited amount of information on a smartphone, resulting in less information on mobile websites. Consequently, companies need to think about what information they decide to display on their mobile websites. To make the customer feel as secure as possible using the mobile website the company need to analyse what key features and key information that is of most importance for the customers, displaying that information.
Another part of the online shopping process that tend to make customers feel unsecure is the payment processes. According to Fiserv research 81% of online shoppers would choose security over convenience (2017). Some customers do not like to give out their card details. As mentioned before in section 1.1, new and simpler ways of mobile payment methods are giving an increasing reliance on smartphone payments, not forcing customers to give out their card details. For instance, swish has now more than 5 million users in Sweden (Swish, 2016) and is now avaliable on multiple web shops online (Swish, 2017). By offering these type of mobile payments, online businesses can reach out to a larger group of customers that already have built up their trust to these type of smartphone payments. It simplifies the check-out process through smartphones, and helps to overcome the problematic with customers tendencies to not go through with smartphone transactions.
6
Conclusion
The main aim of this study was to determine what factors that affect the pro-portion of smartphone sessions as well see the propro-portion of smartphone trans-actions on Flygresor.se’s website. The answers to the targeted questions are presented below.
• What factors affect the proportion of smartphone users at the booking website Flygresor.se?
Sessions Model
The amount of TV-commercial is the only factor from the category Marketing that have significant impact on the proportion of smartphones sessions. It has a positive impact on the proportion indicating that investments in TV-commercial will lead to more smartphone users.
The factors from the category Channel that have a significant positive impact on the Sessions model are Organic and Referral, while the channels Direct and Social have a negative impact on the proportion smartphones sessions.
The seasonal effects that could be spotted on the sessions model were that the proportion smartphone user increases during Winter, and decreases during spring. Swedish Holidays affect the proportion positive.
The amount of megabytes and the total number of visitors had both positive impact on the proportion of smartphone Sessions.
Transactions Model
Similar to the sessions model, TV-commercial was the only marketing variable with significant influence on the proportion smartphones transactions. The β-value indicates that TV-commercial have positive impact on the propor-tion.
The channel factors that were of relevance for the transactions model were Organic, Email and Direct, where Organic and Email had positive impact and Direct had negative impact.
Also the transactionst model showed seasonal changes, where both Autumn and Winter had significant positive impact on the proportion. However, the effect of Winter was greater than Autumn.
The amount of data used per smartphone subscription had a positive impact on the proportion smartphone transaction. The total number of visitors on the
website had a negative impact, meaning that desktop users in larger extent tend to click-through to the final booking sites.
• Do the factors affect the proportion of smartphone transactions or just the proportion of smartphone sessions?
There is only one significant factor of the ones analyzed that affect the pro-portions differently, namely The number of total visitors, where the Sessions Model got a positive value and the Transactions Model got a negative value. The result confirms the problem stated, that visitors coming from a smartphone device tend to not go through with their purchases.
The analysis shows however not any differences in the remaining variables, hence any variables that only had a substantial impact on transactions were not iden-tified. Thus, the variables chosen may not be the most vital in the customers decision of clicking through to the final booking website. It is therefor assumed that there exist other factors which have a greater impact on smartphone users tendency to finalize a booking.
References
Apple (2007). Apple Reinvents the Phone with iPhone. url: https : / / www . apple.com/pr/library/2007/01/09Apple- Reinvents-the- Phone- with-iPhone.html.
Chaouchi, Hakima and Maryline Laurent-Maknavicius (2010). Wireless and Mo-bile Network Security: Security Basics, Security in On-the-shelf and Emerging Technologies. London, United Kingdom: ISTE Ltd.
Cho, Yoon C. (2008). “Assessing User Attitudes Toward Mobile Commerce In The U.S. Vs. Korea: Implications For M-Commerce CRM”. In: Journal of Business Economics Research 6.2, pp. 91–101. doi: http://dx.doi.org/ 10.19030/jber.v6i2.2394.
Chong, Alain Yee-Loong (2013). “Predicting m-commerce adoption determi-nants: A neural network approach”. In: Expert Systems with Applications 40.2, pp. 523–530. doi: https://doi.org/10.1016/j.eswa.2012.07.068. ComScore (2017). U.S. Cross-Platform Future in Focus. url: http : / / www .
comscore.com/Insights/Presentations- and- Whitepapers/2017/2017-US-Cross-Platform-Future-in-Focus.
DIBS (2016). Svensk E-handel, DIBS årliga rapport om e-handel. url: http: / / www . dibspayment . com / sites / corp / files / files / SE / NEH _ SE _ 2016 _ WEB.pdf?_ga=1.4078647.687843159.1480431860.
Ericsson (2017). Innovations with impact. url: https://www.ericsson.com/ en/networked-society/innovation/innovations-with-impact.
Fiserv (2017). Expectations and Experiences | Household Finances. url: https: //www.fiserv.com/resources/expectations- experiences- household-finances-1702.pdf.
Flygresor.se (2017). Flygresor.se. url: https://www.flygresor.se/information. Google (2000). Google Launches Self-Service Advertising Program. url: http:
//googlepress.blogspot.se/2000/10/google-launches-self-service. html.
