University of Groningen
Identifying factors that
influence customer staying
time in Dutch shopping areas
MSc. Thesis - Marketing Management & Intelligence
Identifying factors that influence customer staying
time in Dutch shopping areas
Ron Laan
University of Groningen Faculty of Economics and Business
Department of Marketing
PO Box 800, 9700 AV Groningen, The Netherlands MSc. Thesis Marketing Management & Intelligence
January 2017
Telephone number: +316 30 77 63 64 E-mail address: laanron@hotmail.com
Student Number: S2886847
Supervisors University of Groningen
Management Summary
The shopping environment is changing rapidly. Many retail stores are filing for bankruptcy, consumers can shop on Sunday’s nowadays and are also increasingly shopping online. According to research, it has been shown that increasing the staying time of customers in the shopping street will result in higher sales. There is therefore a paramount need to understand the behavior of customers in the shopping environment. Thus this study examines the factors influencing the staying time of customers in the shopping street.
First, a literature review was done and many factors were being considered regarding the influence on staying time. Eventually a multiplicative model was chosen and 24 variables were being considered: 5 traffic variables, 10 weather condition variables, 6 days-of-the-week variables, and variables regarding the closure of the department store V&D, the customer confidence index, regarding a fun fair event in a city center. Data concerning these variables of four different cities in the Netherlands were obtained. After the description, assumption and pooling analyses, four bootstrapped regression analyses were done to examine the influences of these variables on the staying time of customers in these cities.
The results of this study showed that there are some variables significantly influencing the staying time in three or all cities, e.g. traffic, the closure of the V&D, and the barometric pressure. However, these results could not be generalized, because the influence was different for each city, i.e. it was either positive or negative. Only one variable was significant in all four cities (Wednesday effect), but this influence also had a different direction for each city. Some other interesting findings concerned the weak influence of the weather conditions on the staying time and the mixed influence of a fun fair in a city. In conclusion, not a single influence could easily be generalized.
Preface
Before you lies the thesis: ‘Identifying factors that influence customer staying time in Dutch shopping areas’, that was written to fulfill the graduation requirements of the Master
Marketing Management and Intelligence Program at the University of Groningen. I have been working on this thesis from September 2016 until January 2017.
I would like to take this opportunity to thank Jaap Wieringa for the support and
professional guidance of this research paper. He provided great and constructive feedback and interesting insights in how to perform the multiplicative model well. I also want to thank the employees of CityTraffic for helping me with the dataset concerning the staying time of customers and the traffic in the four cities used for this study. A final thanks goes out to my family, friends and my girlfriend. Even when things get tough, they helped me with staying motivated and enthusiastic.
I enjoyed working on this thesis and I really hope that you enjoy reading it.
Table of contents
1. Introduction 5
2. Literature Review 8
2.1 Definition 8
2.2 Antecedents of Staying Time 8
2.3 Consequences of Staying Time 11
2.4 Conceptual Development 12
2.4.1 Examination of Variables 13
3. Data and Variable Operationalization 20
3.1 Variable Selection 20 3.1.1 Conceptual Model 23 3.2 Model Selection 23 3.3 Model Assumptions 25 3.3.1 Assumption Definitions 25 4. Data Analyses 28 4.1 Descriptive Analysis 28 4.2 Pooling 30 4.3 Variable Transformation 31 4.4 Model Fit 32 4.4.1 Assumption testing 33 4.5 Results 37 5. Discussion 42
5.1 Discussion of the variables 42
6. Conclusions 46
6.1 Limitations 47
References 49
Appendix 53
5 1.Introduction
“Shopping on Sunday is getting more popular” (NOS, 2014)
“Plan for smaller shopping streets against lack of occupancy” (NOS, 2015) “The shopping street is becoming empty, the web shop benefits” (NOS, 2016)
As these headlines of the main Dutch public news website NOS suggest: offline shopping is changing in an increasingly faster pace. Not only are retailers allowed to open their store on Sunday in many different municipalities in the Netherlands, many organizations are also still struggling with the online shopping environment and competition, which has a huge impact on retail and consumer behavior in the streets. One of the main retail developments in the past year in the Netherlands is the bankruptcy of Dutch general department store Vroom & Dreesmann (V&D), which had over 62 stores in many cities in the Netherlands and Belgium. The departing of this iconic store resulted in an increase in retail vacancy of 350.000 m2 in the Netherlands. In cities like Utrecht and Amstelveen it even resulted in a percentage increase from 3 % to 16% and from 1% to 24% respectively (Colliers, 2016). Additionally, many other retail shops failed to adapt to the new retail environment in time and filed for bankruptcy between 2014 and 2016, including Schoenen Reus (206 stores), McGregor (150 own stores and 200 shop-in-shops), MS Mode (130 stores), and Free Record Shop (177 stores) (ANP, 2016). As a consequence, this may eventually have led to less customers staying in the shopping street.
Reasons why e.g. V&D went bankrupt was its indistinctiveness and not joining the online channels early enough like most other retailers did. The growth of online shopping implies that many customers are increasingly shopping more online and therefore do not shop the traditional way (Sorescu, et. al, 2011; Shankar, et. al, 2016). According to a research of the Dutch governmental institution Statistics Netherlands or Centraal Bureau voor de Statistiek (CBS), the amount of web shops in the Netherlands increased six fold in the period of 2007-2015 from more than five thousand to nearly 30 thousand. Consequently there has been a decrease of physical stores in the same period; the amount reduced with 8,2% from 80 thousand to approximately 74 thousand (CBS, 2016).
6 traditional brick-and-mortar stores. While many customers first searched for inspiration, products, and adventure in the shopping streets, they do now on their cell phone or computer (Pauwels et. al, 2011), resulting in less traffic and time spent in the street. Knowing precisely how customers behave while shopping is therefore important for municipalities and retailers to attract the customers to the shopping street and to survive in the highly dynamic and competitive markets.
One of the most important factors of customer shopping behavior is shopping or staying time. It has already been researched to have a great impact on customer purchases and retail sales (Feldman & Hornik, 1981; McDonald, 1994; Zhuang et. al, 2005). According to Ailawadi & Keller (2004) a pleasing atmosphere encourages customers to visit a store more often, stay longer, and as a consequence, buy more. The amount of customers in a certain place also has a significant influence on the time spend shopping (Eroglu & Harrell, 1986; Oakes & North, 2008; Byun & Mann, 2011), which consequently results in differences in sales. However, not only human factors are of influence, also daily factors may affect staying time, like certain weather conditions (Bahng & Kincade, 2011). When the weather is perceived as bad (e.g. high precipitation, low temperature) customers will not stay long to shop in a certain place, but search for a product and purchase it online. In conclusion, weather in this sense also has an impact on staying time and correspondingly on the total sales.
The aforementioned closure of many retail stores in the Netherlands and other factors are therefore a critical trend to examine as this may result in a lower staying time and lower operational results. Prolonging a customer’s time in a commercial shopping area is thus crucial for business and knowing the factors influencing the staying time is of great importance for both retailers and the municipalities.
For this thesis, the antecedents of staying time are therefore examined. Using external data and data regarding the behavior of customers in the shopping street factors are identified which have a positive or negative significant influence on a customer’s staying time. Therefore a model is made regarding multiple variables. The research question thus is:
7 The results of this study can be used to gain new insights in customer shopping behavior in the commercial shopping area under different conditions. This is important, as organizations and municipalities can then effectively allocate marketing actions accordingly. For example, when there is a significant effect on a certain day regarding the staying time of customers, retailers can use their marketing budget on those days where the highest staying time can be achieved, as this will eventually lead to more sales. In addition, many researches have already been conducted in terms of influences on the staying time of customers in a certain area (Feldman & Hornik, 1981; Eroglu & Harrell, 1986; McDonald, 1994; Zhuang et. al, 2005; Oakes & North, 2008; Bahng & Kincade, 2011). However, there hasn’t been a single paper regarding a complete set of factors like the ones being considered in this paper. Therefore, the new insights are an addition to existing scientific literature regarding how customers behave in the shopping streets according to the different factors.
Several interesting findings are for example that the closure of the V&D and the amount of traffic had mixed significant influences on the amount of seconds customers stayed in a shopping area. Additionally, maximum temperature did not have a huge significant influence on all the four cities, stating that e.g. the weather forecasts may be ignored while performing marketing campaigns.
8 2. Literature review
To understand why the retail environment is changing, an examination must be made regarding staying time. After that multiple factors are thoroughly analyzed in order to come up with possible antecedents. As stated before, this study is of explorative nature, meaning that many factors which are analyzed may not eventually be considered.
2.1 Definition
First, a definition of staying time is given. According to (Zhuang et. al, 2005) staying time concerns the number of minutes a shopper stayed in a shopping street. Breazeale & Lueg (2011) include the following definition in that instrument: “Shopping means time spent specifically
looking for and examining products or services. This does not include time spent just messing around on the Internet or hanging out at the mall”. Yiu & Ng (2008) on the other hand defined
it just as the time spent by the shopper in a shop. Staying time and shopping time are both interrelated and most often used synonymously. And while there are many different definitions regarding the construct, for this research however, the following is maintained:
- Staying time is the time in minutes within a day that a customer stays within a specific area. This could be e.g. a commercial street, a mall, a square or a city center -
In the following part, first the antecedents of staying time are examined. After that in 2.3, the consequences of an increasing or decreasing staying time of customers in a commercial shopping street were analyzed.
2.2 Antecedents of Staying Time
In history much research have been conducted regarding staying time, its antecedents and its influences. First, the antecedents are examined. Feldman & Hornik (1981) made a framework regarding the use of time for customers and came up with four components: time structure, resource availability, activity availability, and personal characteristics (see Figure 1).
9 The first component is ‘time structure’ and consists of four parts: work at job, necessities, homework, and leisure. Within a range of time, these factors should be considered. So when an individual uses 18 hours of a day for the first three factors, the remaining 6 hours can then be used for leisure. Reducing then for example the necessities component by one hour (e.g. less sleep) may result in an increase of the time for leisure of one hour. Shopping in this component may be different for each person. One may define shopping as leisure time, while another individual works as a mystery shopper and therefore uses its time shopping as ‘work at job’. The second component of the allocation of time is ‘resource availability’. There are three resource types. The first is the temporal aspect of activities, which is about the frequency and duration of them. The second is the economic dimension, i.e. the monetary costs. The third concerns the spatial dimension, which is about the location of the considered activities. In conclusion, every individual uses three types of resources: time, money, and space. When e.g. a customer has no money available, he/she may choose differently regarding shopping than when the customer has plenty of money. The same counts for space. When e.g. a H&M store is far away for a customer, he/she may eventually still travel to that particular store, even though there are many other clothing stores in the customer’s neighborhood. However, when time and money are scarce, the customer is likely to choose for stores nearby.
The third component of the framework is ‘activity availability’. Customer choices are different when certain activities cannot be performed. For example, when on a Sunday the stores in a customer’s home city are all closed, that customer cannot shop then for clothes in the shopping streets, and is likely to perform a different activity, like recreational activities or online shopping. Customers only consider those activities that will fulfill their needs; he/she will not waste time doing something unless utility from that specific activity is gained. Another example is when there is a fun fair in a city. When this is the case, the activity availability construct also plays a role, as only at this specific time a fun fair can be visited. Alternatively when there is none, a different activity is being performed. A last example is when there is a perception of bad weather, e.g. high precipitation and low temperature. In this way, the activity of going to the beach to swim and sunbathe is not ‘available’ and therefore, individuals will do other activities instead.
10 personal characteristic. Therefore, shopping motivations of customers is part of this component. A person who goes e.g. shopping with others for the enjoyment of being social and bonding may spend more time shopping than one who shops just because it is necessary. Two kinds of shopping motivations exist: utilitarian and recreational/hedonic. The former is about shopping being task-related and rational (Batra & Ahtola, 1991). Hedonic shopping is similar to the task orientation of utilitarian shopping, but the task is related to hedonic fulfilment, like amusement or adventure (Babin et. al, 1994).
So part of the fourth component are the shopping motivations. According to Arnold & Reynolds (2003) there are six types of hedonic shopping motivations. Adventure shopping concerns shopping for sheer excitement, adventure and the feeling of being in another world. Value shopping is about the enjoyment of hunting for bargains, due to the individual being a competitive achiever. Role shopping relates to the enjoyment that individuals derive from shopping for other people. Idea shopping concerns the need to keep up with new trends and to learn from this. Gratification shopping is about shopping for stress relief and to improve one’s mood. The last type of shopping motivation is social shopping, which is about the enjoyment people get from shopping with other people. Research has found that individuals having one of these six kinds of hedonic shopping motivations spend more time shopping and are also more impulsive in buying products than those individuals who have a utilitarian shopping motivation (Bellenger and Korgaonkar, 1980). The shopping motivation of a customer is therefore also influence the allocation of time.
11 Several other researchers have tried to find antecedents of staying time. For example, Ailawadi & Keller (2004) and Turley & Milliman (2000) state that a pleasing atmosphere encourages customers to visit a certain store more often, stay longer, and eventually buy more. The response of the atmosphere in this sense varies along three dimensions: arousal, pleasantness, and dominance (Mehrabian & Russell, 1974) and consequently the level of these dimensions says something about an individual’s purchase behavior. Additionally, Forsythe & Bailey (1996) and Swinyard (1993) found that shopping enjoyment was positively related to shopping intentions and time. The former study examined the influence of shopping experience of customers and (socio)demographic variables on the time spent shopping. A mediation effect was also predicted, with perceived time poverty being the mediator. This effect was not found to be significant. However, shopping enjoyment correlated significantly with the number of minutes spent shopping. The authors stated that this may be due to a positive relationship between enjoyment for shopping and information search. Customers who like shopping are eager to read and learn about certain products and therefore spent more time looking for the best product. Therefore, the assumption can be made that a positive mood of customers has a (in)direct influence on a customer’s staying time.
Another study regarding staying time of from Martin et. al (2012). They focused in their study on the effect of accidental interpersonal touch (AIT). Two groups were being considered: one where multiple customers were touched by male and female confederates and one where there was no touch at all by another person. The former group reported more negative product beliefs, brand evaluations, had a lower willingness to pay, and eventually spend less time shopping in the specified area than the latter group. This effect was stronger when the AIT was done by a male for both men and females. The authors claim that this effect may occur because customer want to distance themselves from the strangers, which results in avoidance behaviors.
In conclusion, according to these aforementioned frameworks and theories, the staying time of customers is then a function of different aspects concerning internal and external factors. In the next part, the consequences of staying time are being considered using scientific literature.
2.3 Consequences of Staying Time
12 customers is that it has a positive significant influence on sales of both food and non-food products, i.e. the longer a person shops, the more likely he/she is to buy more products (Mehrabian & Russell, 1974; Bagozzi, 1986; McDonald, 1994; Sherman, et. al, 1997, Zhuang et. al, 2005).
In a study about the influence of store environment on consumer purchase behavior, Sherman et. al (1997) found several interesting significant relations. In their research they tried to find what influences the levels of pleasure and arousal of customers. In a second part, the consequences of high levels of these constructs were examined. The results showed that social factors in a store had a positive influence on the level of pleasure of customers, but not on the level of arousal. Ambiance factors on the other hand showed positive significant values for arousal, but not for pleasure. Last, they found that design factors positively influenced pleasure, but contrary to their expectation had a negative significant influence on arousal. The second part of their study showed that the level of pleasure had no significant influence on the number of items purchased nor the time spent in store. Arousal on the other hand did, and this significance was positive. More importantly, the results showed that a higher value of both constructs lead to more money spent in shops, which is an interesting finding for retailers and municipalities. When customers are aroused, they are more likely to spend more time in the shopping area and consequently spend more money.
Zhuang et. al (2005) researched the same influence in their study. Respondents in a shopping mall in the US and China were asked questions regarding their shopping intention and shopping habits and processes, e.g. shopping time, frequency of mall visits and number of stores visited. In addition, data was obtained regarding the customer satisfaction, sociodemographic variables, and shopping area perceptions. The dependent variable in their study was a dummy variable concerning a buy/no buy. Results showed that there were many variables significantly influencing purchases. Some of these were staying time, frequency of visits, and total satisfactions. Additionally, staying time increased the probability of purchases of both food and non-food products, i.e. the more time the customer stayed in a shopping area, the more likely he/she is in buying a product.
2.4 Conceptual development
13 to have an influence on the dependent variable, in this case ‘staying time’. In the coming section multiple factors are elaborated using existing literature.
2.4.1 Examination of variables Traffic
The amount of people in a certain space has already been proved to have a significant effect on behavior (Eroglu & Harrell, 1986; Oakes & North, 2008; Byun & Mann, 2011). According to Meng & Kang (2015), a greater crowd density leads to a higher subjective loudness, which in turn decreases the acoustic comfort. Crowd density in this sense is expressed as the number of people in e.g. a shopping street divided by the area of the shopping street. The amount of traffic in a street is therefore correlated with the crowd density of that street. When traffic exceeds a certain limit, comfortability of the customers decreases, which may lead to a lower saying time (Byun & Mann, 2011; Forsythe & Bailey, 1996, Swinyard, 1993). Therefore, the prediction is that the total traffic of customers in a certain week decreases the average staying time significantly.
Additionally, the assumption can be made that when customers already spend quite some time shopping on t-1 or t-2, they will most likely not spend the same amount of time on day t due to limited amount of resources (e.g. time or money) (Feldman & Hornik, 1981). This could also be for the anticipation of high staying time in the near future. When customers believe other customers will shop a long time on t+1 or t+2, the assumption is then that the customer will spend much time shopping on time t too, because it will be crowded on the anticipated days. More crowd leads to lower mood and therefore the customer will shop more on time t in order to prevent himself from having a lower mood (Byun & Mann, 2011; Forsythe & Bailey, 1996, Swinyard, 1993). Lagged and lead variables are therefore also included in the model to explain influences from the past or from the future. When considering the aforementioned information, the assumption then is that when traffic is high on t-1 and t-2, staying time on time t will be lower. Consequently, when there is an anticipation for high staying time, i.e. higher at time t+1 and t+2, staying time on time t would be higher.
Weather conditions
14 weather may be perceived as very good, e.g. when the temperature is high and the sun is shining, customers may eventually not choose to shop, but to do recreational activities instead. Researchers have found a relationship between weather and mood; the type of weather has an influence on how individuals feel (Howarth & Hoffman, 1984; Keller et. al, 2005). Humidity, precipitation, temperature and hours of sunshine had the greatest effect on mood. For example, when the number of hours of sunshine increased, anxiety decreased. On the contrary, when it rained, anxiety and sleepiness increased. These are unfortunate conditions for retailers, as customers are not in the mood then come and stay long in a shopping street. When considering these conditions, the assumption can be made that weather conditions has an influence on a customer’s mood and therefore its staying time (Forsythe & Bailey, 1996, Swinyard, 1993). Barometric pressure already accounts for the amount of sunshine and rain throughout the day, i.e. low pressure is commonly associated with clouds and rain, whereas high pressure is associated with sunny and clear weather (Ahrens, 2000). Therefore, only barometric pressure and maximum temperature are examined.
Moreover, according to Bahng & Kincade (2011) there is strong evidence that a change in temperature has an impact on the sales of seasonal garments. In their research they focused on the increase or decrease of temperature in a certain season and the sales of apparel and found a negative correlation in the Fall and Winter seasons (i.e. a decrease in temperature led to a higher sale of clothes for these seasons). This states that also past data should be taken into consideration. Moreover when there will be good weather in the coming days this may result in a lower staying time in time t, as customers anticipate on the better weather conditions. The assumption can be made then that when the weather was really bad for t-1 or t-2, the staying time may be prolonged when the weather is perceived as good at t. In a model, lag and lead variables should also be added to account for this effect.
Shopping Street diversity and quality
15 than in a shopping street with a low diversity, as customers may perceive the more diverse and unique shopping street higher in quality than the one which is just the same as other shopping streets (Ailawadi & Keller, 2004; Turley & Milliman, 2000).. This may result in customers choosing Amsterdam then over other surrounding shopping areas, which will eventually result in more traffic and shopping time.
According to Borgers & Vosters (2011), exclusive and specialized stores are preferred over the well-known national chains. This means that a diverse and unique shopping supply consisting of multiple unique brands/stores is rated higher in quality than those which are ‘like the other streets’. Additionally, those shopping areas with a higher rated quality are consequently rated in a higher popularity and will therefore attract more traffic (González-Hernández & Orozco-Gómez, 2012). So when there is a higher diversity in a shopping street, this will eventually lead to a higher quality and popularity of the shopping area/and may thus lead to a higher staying time of customers.
Online sentiment
16
Individual data
As aforementioned, much research has already been conducted in history on the influences of personal characteristics and data on the staying time of that person (e.g. Saito et. al, 2010, see 2.2). Certain demographic variables were already being found significant regarding the staying time of customers in a shopping area. Therefore, it is necessary to also account for the individual effects in the proposed model in this study, e.g. age, sex and shopping purpose.
Store closures
As mentioned in the introduction, the closing of the iconic retail store V&D may have consequences for the staying time of customers in a shopping street. Instead of getting all items in one warehouse, the closure may leave customers visiting more stores to fulfil their shopping needs, meaning that they stay longer in the shopping street. However, the closure of a big store in the city center makes the visiting of it less attractive, and customers may for example search then online and shop offline (Pauwels et. al, 2011; (González-Hernández & Orozco-Gómez, 2012). The motivation for customers to stay longer in a shopping street is therefore lower, as part of the shopping process has already been done online, but also the quality of the shopping street decreases. In conclusion, the closure of a big store in the shopping street should thus be examined in order to find significant values. This may be relevant for municipalities and retailers to find out what the effects were of a big store closure on the shopping behavior of customers.
Days of the Week
Regarding the days of the week, there are several factors to take into consideration. In the first part of the coming section, the scientific literature regarding the effect of certain days of the week was given. After that, Sunday Shopping was being examined.
17 parts of the day. They found for example that the number of children had a significant and positive effect on the amount of shopping trips on Tuesday, Friday, and Saturday. Also, the higher the income, the greater the probability to make more shopping trips on Thursday. Additionally, individuals experience different emotions during certain days in a week. Ryan et. al (2011) focused on the effects of a weekend versus a weekday and work versus nonwork experiences on the emotional state of individuals. They wanted to research if the mood in the weekend and nonwork time was significantly better than when it was a weekday. Also, they tried to find if the level of mood for individuals who worked was higher than of those who did not work. Results showed that both hypotheses were true for both genders. As mood has a positive effect on staying time (Forsythe & Bailey, 1996, Swinyard, 1993), the assumption can then be made that customers shop longer in the weekends than during the week. The framework of Feldman & Hornik (1981) can complement this by stating that in the weekend, the time structure is different for people as they have less ‘time at work’ and therefore more ‘leisure time’, e.g. time for shopping. The differences in weekdays in the specific context of this study have not been researched extensively in history and for this study therefore the examination of this will be of explorative nature.
Another interesting factor is Sunday Shopping. The Dutch Winkeltijdenwet or Store Time Law is a law enacted in 1996 concerning the opening hours of stores. It states that stores can only be open between 6:00 in the morning and 22:00 in the evening, excluding Sundays and Christian holidays, e.g. Good Friday, Ascension Day or during Christmas. However, some stores are granted exemption and may still open during these days for the purpose of public health institutions, traffic and transport (e.g. gas stations, train stations, and airports), and the sales of newspapers and magazines. In addition, the opening of cultural institutions, sport complexes, and the sale of certain goods during Ramadan, All Saints’ Day and All Souls’ Day is also allowed. The law also states that municipalities can permit stores to open a maximum of twelve Sundays per year. These are the so-called ‘Koopzondagen’, or Shopping Sundays. If municipalities want to open more than the allowed amount of Sundays, then they can use the tourism provision, provided that there is substantial tourism in the community.
18 implemented the Shopping Sunday on a weekly basis, the distribution of the shopping time may differ throughout the week than when retailers are only open for the first six days of the week.
Consumer Confidence Index
Gordon et. al (2013) researched the impact of the macroeconomic environment on the price sensitivity of consumers. The result of the study was that it works countercyclical, i.e. price sensitivity increases when the macroeconomic environment weakens. It may seem that when consumers have less confidence in the current and future economy and in their own financial situation (i.e. weakening of the macroeconomic environment), they probably spend less time shopping, as a customer’s mood may be lower than when the economy is better (Forsythe & Bailey, 1996, Swinyard, 1993). The CCI in this sense may be considered to be part of the ‘resource availability’ component of the time allocation framework of Feldman & Hornik (1981), i.e. a weaker macroeconomic environment means less monetary resources. Therefore when using the customer confidence index of a country, one could look for its influence on the staying time of customers in the shopping areas. The assumption therefore in this study is that when the CCI increases, the staying time of customers also increase.
Fun Fair event
In most Dutch cities, many events take place. Rotterdam has for example its marathon, Nijmegen has its Four Day marches and Amsterdam has Sail. Respectively, they attract 925.000, 1,5 million and 2,3 million visitors to these cities annually, attracting also visitors to the city center (NOS, 2015; NN 2016; De Gelderlander, 2016; NOS, 2015). Commercial shopping streets are situated mostly in the heart of the city and thus will most likely be visited by these people. According to Turley & Milliman (2000) and Sherman et. al (1997) shopping time and spending increases when a customer’s arousal increases, meaning that during these ‘arousing’ events of entertainment and hurdles of people shopping time may increase evidently. Byun & Mann (2011) add that when there is high human crowding due to an exciting event, like a ‘Black Friday’, it will create positive emotions, meaning that customers want to stay longer in a certain area. Considering the aforementioned theory about the previous factor ‘traffic’, this means that there may be an interaction effect of the amount of traffic and such an event at time t on the staying time of customers.
19 to take into consideration, like the Summer Fun Fair or the Cheese Market. However, because these two events are completely different, a generalization cannot be made, i.e. one cannot say that events in general increases staying time then.
20 3. Data and Variable Operationalization
For this thesis, the various influences of factors are examined when focusing on the time customers spent in a shopping street. This is done by using different kinds of external data. In the previous section, multiple variables were examined according to scientific literature. In the coming section, the data acquisition, variable selection and model selection are explained. 3.1 Variable selection
Data regarding the staying time of customers (dependent variable) and the total amount of traffic are obtained from CityTraffic, a Dutch research company based in Amsterdam specialized in shopping areas. They have developed a unique method to count the traffic and its behavior in shopping streets within a city with the help of Bluetooth- and Wi-Fi signals. The data are accurately measured every half an hour, every day of the year. Not only does the system count the amount of people, it also counts the amount of unique visitors, how many times a visitor returned, how long they stayed in the shopping area and what their shopping path is. Due to the data received from the systems, the company can help cities or municipalities to gain insights in the development of the shopping traffic. Additionally, the data can be used to predict future behavior of shoppers or to measure the impact of e.g. city events, construction projects in the area, or marketing campaigns.
While privacy is a hot topic nowadays, CityTraffic assures that all data is anonymized, as the data is not and cannot be linked to specific (socio-)demographic information of the individual. Even though the system relies on Media Access Control addresses (MAC-address), which is a unique identification number, the data which is given to clients is completely anonymous. Due to this, individual customer data cannot be obtained and therefore used. In addition, the data in this analysis is aggregated on daily level, not on customer level, which makes it harder to combine, interpret and generalize the results. Therefore, only aggregated data regarding the staying time and traffic is used.
21 to this, the staying time of before and after the closure of V&D can be calculated. And third, the sociodemographic and geographical data regarding e.g. number of inhabitants, number of non-Dutch inhabitants, average number per household, income and location in the country were the most diverse in these four cities. For example, the number of inhabitants is as of the beginning of 2016 122.915, 107.822, 520.704 and 66.268 respectively (CBS, 2016). In this way, the results can be generalized, as it is about different sizes of municipalities.
Regarding the data of the weather conditions, the national data of the Koninklijke Nederlands Meteorlogisch Instituut (KNMI) is used. The KNMI is a scientific institute of the Dutch government and is specialized in climate science, seismology and meteorology. The institute receives daily data from 38 different weather stations throughout the country, e.g. the wind flow velocity, maximum temperature, amount of sunshine in minutes and precipitation (KNMI, 2016). As mentioned before, only the barometric pressure and maximum temperature is used (Ahrens, 2000). Therefore, this information is retrieved from the website and implemented in the data set. In addition, the lead and lagged effects are easily obtained by shifting the data points in the data set.
The Customer Confidence Index is obtained from the Dutch Central Bureau for Statistics. This governmental organization calculates the Consumer Confidence Index, the consumer’s perception regarding the economic climate, and the consumer’s willingness to buy per month by interviewing Dutch residents about the past and coming 12 months. The questions are about the financial situation of the individuals and the perception of the general economic development of the Netherlands. The results from this survey is also implemented in the data set together with the other variables. As these survey results are per three months, the data points are implemented and spread to make it fit into the daily data set.
22 Additionally, six dummy variables were made for the examination of the days-of-the-week effect. However, due to the fact that the data in this study does not start before the first of January 2015, examining the effect of the implementation of a Sunday Shopping is impossible. For example, two of the cities in this study, The Hague and Leiden, have assigned the touristic label to their city center, meaning that they can already be open 52 Sundays in a year since 2005. Therefore, this variable was not used in this study.
Additionally, as there is no data available regarding the perceived quality of a shopping street and the diversity of city centers, this variable was also not taken into consideration for this study. The same accounts for the data regarding online sentiment. This is shown by an example. Google Trends is one of the online analytical tools to find trends and the development of trends regarding certain search terms. The higher the number of searches of a specific term, the higher the index of the search term in that period. This is freely usable for everybody with access to the web. Because Google Trend Analytics uses indices over time, it is difficult to compare specific search terms, because you don’t obtain the real values of the number of searches. Therefore, you cannot compare e.g. the search terms ‘Amsterdam’, ‘Kalverstraat’, or ‘Amsterdam shopping’ with each other regarding the true values. In addition, one needs to consider what search terms to implement. As the shopping area is not solely on one street, it is difficult to check which search term to use. Due to this, it is difficult to retrieve and use data from Google Trends. Other data software regarding online sentiment can be obtained. However, this has many downsides. For example, emotional polarity computation as used in Li and Wu (2010) is quite time-consuming and very costly. This accounts for most of the data software regarding this issue and therefore, online sentiment cannot be used and examined in this study.
3.1.1. Conceptual Model
23
Figure 2: Conceptual Model
3.2 Model selection
A model must meet five criteria in order to have high validity: simple, evolutionary, complete, adaptive and robust (Leeflang et. al, 2015). A simple model starts quite simple and eventually more independent variables are added when they are improving the predictability of the model.
Complete means that the model is a good representation of reality and therefore accounting for
24 function, an increase of 1 will not only increase the sales by the coefficient of the advertisement variable, but also with the display variable, due to the multiplication of the function. This means that all variables in the model will interact with each other. For the model in this study, interaction effects are also likely to occur, between e.g. the weather conditions and the amount of traffic. Therefore, the multiplicative function is an effective function to use.
Additionally, the parameters of the multiplicative model are estimated by using weighted least squares and are interpreted as elasticities (Wittink et. al, 1988). This is convenient, as elasticities are often of primary interest.
The multiplicative model that is used for this study can be seen here below. Estimating this model will result in 24 parameters for four cities. In addition, it meets the five abovementioned criteria; new variables can be added, it is simple and evolutionary in nature, complete, and the results may be robust and meaningful.
STit= θi∏[ 4 i=1 𝑇𝑖𝑡𝛽𝑖1𝑇 𝑖𝑡−1 𝛽𝑖2𝑇 𝑖𝑡−2 𝛽𝑖3𝑇 𝑖𝑡+1 𝛽𝑖4𝑇 𝑖𝑡+2 𝛽𝑖5 𝑀𝑇𝑡𝛽𝑖6𝑀𝑇 𝑡−1 𝛽𝑖7𝑀𝑇 𝑡−2 𝛽𝑖8 𝑀𝑇𝑡+1𝛽𝑖9𝑀𝑇 𝑡+2 𝛽𝑖10 𝑃𝑡𝛽𝑖11𝑃 𝑡−1 𝛽𝑖12𝑃 𝑡−2 𝛽𝑖13𝑃 𝑡+1 𝛽𝑖14𝑃 𝑡+2 𝛽𝑖15 𝐶𝐶𝐼𝑡𝛽𝑖16] 𝛽𝑖17𝑆𝐶𝑖𝑡𝛽 𝑖18 𝐹𝐹𝑖𝑡 𝛽𝑖19𝛿1𝛽 𝑖20 𝛿2 𝛽𝑖21𝛿3𝛽 𝑖22 𝛿4 𝛽𝑖23𝛿5𝛽 𝑖24 𝛿6 𝜀𝑖𝑡
Where i = 1, 2, 3, 4 is the shopping area index, t = 1,… 670 is the number of observations. 𝑆𝑇𝑖𝑡 Staying time in minutes at shopping street i at time t
𝑇𝑖𝑡 Daily amount of people on average on time t in shopping street i
𝑇𝑖𝑡−1 Daily amount of people on average on time t-1 in shopping street i
𝑇𝑖𝑡−2 Daily amount of people on average on time t-2 in shopping street i
𝑇𝑖𝑡+1 Daily amount of people on average on time t+1 in shopping street i
𝑇𝑖𝑡+2 Daily amount of people on average on time t+2 in shopping street i
𝑀𝑇𝑖𝑡 Maximum temperature at time t in shopping street i
𝑀𝑇𝑖𝑡−1 Maximum temperature at time t-1 in shopping street i
𝑀𝑇𝑖𝑡−2 Maximum temperature at time t-2 in shopping street i
𝑀𝑇𝑖𝑡+1 Maximum temperature at time t+1 in shopping street i
𝑀𝑇𝑖𝑡+2 Maximum temperature at time t+2 in shopping street i
𝑃𝑖𝑡 Barometric pressure at time t in shopping street i
𝑃𝑖𝑡−1 Barometric pressure at time t-1 in shopping street i
𝑃𝑖𝑡−2 Barometric pressure at time t-2 in shopping street i
𝑃𝑖𝑡+1 Barometric pressure at time t+1 in shopping street i
𝑃𝑖𝑡+2 Barometric pressure at time t+2 in shopping street i
𝐶𝐶𝐼𝑡 Consumer Confidence Index at time t
𝑆𝐶𝑖𝑡 Store closure dummy variable at shopping street i
𝐹𝐹𝑖𝑡 Fun Fair event dummy variable at shopping street i
𝛿𝑑 Dummy variable for day 𝛿, where 1 = Monday, 2 = Tuesday, 3 = Wednesday, 4 = Thursday, 5 = Friday,
6 = Saturday and 0 = Sunday.
𝜀𝑖𝑡 Error term at period t at shopping street i
𝛽𝑖1, 𝛽𝑖2, 𝛽𝑖3, 𝛽𝑖4, 𝛽𝑖5, 𝛽𝑖6, 𝛽𝑖7, 𝛽𝑖8, 𝛽𝑖9, 𝛽𝑖10, 𝛽𝑖11, 𝛽𝑖12, 𝛽𝑖13, 𝛽𝑖14, 𝛽𝑖15, 𝛽𝑖16, 𝛽𝑖17, 𝛽𝑖18, 𝛽𝑖19 𝛽𝑖20 𝛽𝑖21 𝛽𝑖22 𝛽𝑖23 and 𝛽𝑖24 are
25 The results of the model show us which variables have a significant influence on the dependent variable staying time, in which direction this influence is (positive/negative), and also how strong this influence is. For example, 𝑃𝑖𝑡−2 being significant and positive means that the
barometric pressure t-2, i.e. high amount of sunshine and low amounts of clouds and precipitation of two days in the past, will result in a higher staying time of customers on time t. Marketers can then focus their marketing expenditures on those days that result in a higher customer staying time in order to obtain a higher sales. Another example is e.g. 𝑆𝐶𝑖𝑡. When this
variable is e.g. significant and negative, it means that the closure of V&D has a negative influence on the staying time of customers in a shopping area. Municipalities can then use this result by thinking for example about how to fill the gap in the shopping street.
For the functionality, the abovementioned equation will be transformed. The logarithms were used to make the multiplicative model in a linear model. It formalized the way variables in the specified model are related (Leeflang et. al, 2015). This is done in the chapter 4.3.
3.3 Model assumption
There are several assumptions that need to be checked in order to find the true quality of the model. A violation of these assumption will lead to wrong estimates of the variance of the parameters, i.e. wrong conclusions of the significance of the effects. These assumption are:
1. No correlation between predictor variables Test for multicollinearity
2. 𝑉𝑎𝑟(𝜀𝑡) = 𝜎2 Test for homoscedasticity
3. 𝐶𝑜𝑣(𝜀𝑡𝜀𝑡′) = 0 𝑓𝑜𝑟 𝑡 ≠ 𝑡′ Test for autocorrelation 4. 𝜀𝑡 is normally distributed Test for normality
In the coming section, definitions of the assumptions are given. After that, they are tested in Chapter 4 and if they are violated, transformation of the data set or variables are needed. When this is eventually done, the regression analyses of the four cities were done.
3.3.1 Assumption definition Multicollinearity
26 large variances in sampling distributions and thus unstable coefficients (Leeflang et. al, 2015; Goldberger, 1991). One can examine the multicollinearity of a model by checking the VIF scores as seen in the collinearity diagnosis. This can be found in SPSS while performing a linear regression model. When the values are higher than 5, multicollinearity exists in the model. To overcome this, a change in variables needs to be made by either deleting or recoding them to different variables.
Normality
This assumption concerns the normal distribution of the residuals (Leeflang et. al, 2015). Normality can be detected when looking at a histogram regarding the residuals and also the Shapiro-Wilk and Kolmogorov-Smirnov tests. These tests exist to identify possible disturbances of normality and should therefore be done on the residuals of all the four cities. In this case, one is not looking for significance of these tests, as the null-hypothesis states that the error term is normally distributed. When there is non-normality of the residuals, bootstrapping can be applied in order to overcome this problem.
Homoscedasticity
Instead of the variance of scalar observations, this assumption concerns the covariances of vector observations. To examine the assumption of heteroscedasticity, a hypothesis should be tested, which is about the assumption that the error term in the specified model is homoscedastic, i.e. exhibiting the same variance over time in all cases (Leeflang et. al, 2015). When there is a violation of the assumption, not all observations have thus equal variances of the disturbance term, i.e. 𝑉𝑎𝑟(𝜀𝑡) ≠ 𝑉𝑎𝑟(𝜀𝑡′). To test this hypothesis, the Goldfield-Quandt (1965) test can be performed. The null-hypothesis of homoscedasticity is 𝐻0: 𝛿1 = 𝛿2 = ⋯ =
27
Autocorrelation
This assumption focusses on the residuals, as these can have systematic patterns over time. The assumption is this that 𝐶𝑜𝑣(𝜀𝑡𝜀𝑡′) = 0 𝑓𝑜𝑟 𝑡 ≠ 𝑡′. Some of the covariances, or maybe all of them, will be non-zero for different points in the data range. When this happens it means that the residuals have a first-order autocorrelation. The problem therefore is that the model does not capture dependencies cross-sectionally over time. The result is then that there is a dependence of the disturbance term and that the efficiency of ordinary least squares (OLS) estimation is reduced (Leeflang et. al, 2015). The violation of this assumption can be found by plotting the residuals over time and by doing the Durbin-Watson test. The latter is based on the variance of the differences between two successive disturbances. According to Field (2009), values between 1,5 and 2,5 are relatively normal. Values which are outside of this range need to be examined and values below 1 and above 3 are a definite cause for concern. Values smaller than 1,5 indicate that there is a positive autocorrelation, while values larger than 2,5 indicate a negative autocorrelation.
To solve an issue of positive or negative autocorrelation, one should account for serial correlation of disturbances. This can be checked by recoding the values of all the continuous variables in the data set by the using the following formula:
28 4. Data analysis
In this chapter, the data is analyzed. First, the descriptive statistics, missing values and outliers are being examined. After that, a pooling analysis is done to determine whether the aggregation can be done using a unit-by-unit or a pooled model. Then the model fit and assumptions are examined. In the end, the significance and betas of the variables are analyzed.
4.1 Descriptive Analysis
In order to get a feeling of the dataset, a descriptive analysis is done. The staying time was transformed beforehand from a time variable to a number variable, i.e. the values of staying time were multiplied by 86400 to obtain the staying time in seconds. The total data set contained 2315 data points per variable for the four cities combined. As can be seen in table 1, the maximum values, minimum values, mean and standard deviation of staying time, traffic, V&D and Fun Fair of the four cities are given. There were some missing values due to the lead and lagged variables, but these were ignored as they would not affect the results significantly. No imputation therefore took place. In addition, there were several outliers. However, due to the nature of the data, they were not deleted. When deleting outliers, valuable information may be lost. Therefore, they were still taken into consideration for this study.
Date Range
Staying Time Traffic V&D
Value = 1
Fun Fair
Value = 1 Min. Max. μ σ Min. Max. μ σ
The Hague 1-1-2015 – 31-10-2016 1025 5251 3957 593,21 33461 104179 64816 12471,70 199 74 Leiden 1-1-2015 – 31-10-2016 3764 6830 5234 396,59 23773 226016 71157 15296,16 199 23 Alkmaar 1-1-2015 – 30-09-2016 2763 5868 4245 504,34 5937 89730 25522 8423,73 161 35 Bergen op Zoom 1-12-2015 – 31-10-2016 2436 5451 3599 434,43 2914 47949 17299 7239 191 15
Table 1: descriptive analysis of the four cities
29 When looking at traffic, Leiden received the most people in its city center per day on average, while Bergen op Zoom received the least, 71157 and 17299 respectively. However, the traffic in Leiden had the highest standard deviation (σ = 15296,16). When looking at the data set, this does not seem unnatural, as the city received some days only around 25000 people, whereas other days the traffic was more than 200.000, e.g. during the largescale yearly event Leidens Ontzet. The lowest traffic mean and standard deviation was in Bergen op Zoom (μ = 17299, σ = 7239). Also, this is not peculiar, as this is the smallest city of the four.
Figure 3: Histogram of spending time of the four cities
30 when the day is coming near. As many customers already bought a lot of presents when Sinterklaas is getting closer, the sudden drop may be explained by the fact that not much shopping has to be done after the festivity. Values returned to normal around Christmas time. Last, Bergen op Zoom had the same development as Alkmaar with Sinterklaas; high around the event and lower after. In addition, the values seem constant and relatively stable compared to other cities.
Considering the dummy variables, all four cities have a fairly amount of days where the V&D was closed. The Hague had most days regarding a fun fair (74) while Bergen op Zoom had only 15 days of fun fair. The latter is possibly due to the data set range, as Bergen op Zoom also had the least amount of data points.
In table 2 the descriptive values of the weather variables (barometric pressure and maximum temperature) and the CCI are shown. These are for all cities the same. The value of temperature in the data set is 0.1 degrees Celsius, so the range of the maximum temperature is -0,3 until 34,4 degrees Celsius. The mean over the data set is 15,4 degrees Celsius. As stated by Ahrens (2000), only the barometric pressure and maximum temperature can be used when looking at the influences of the weather variables. To check if these do not correlate, an analyses was done between the two. It was not significant (p = 0,106), stating that these variables can be used. The CCI ranges from a negative value -6 to a very positive 12. The overall mean over the period is 3,03 with a standard deviation if 4,47.
4.2 Pooling
There are multiple ways to treat data when doing regression analyses, but for this study two will be contemplated. The first way is the unit-by-unit method, which means that every city is being examined independently. The second method is called pooling. When talking about pooling, it refers to combining information or data. Pooled models allow the accounting of multiple entities (Wittink et. al, 1988). For this study, it means that the data is treated as one whole data set, i.e. doing regression analysis on the four cities together instead of doing it per
Min. Max. μ σ
Maximum Temperature -3 344 154,05 64,45
Barometric Pressure 9764 10379 10161,92 64,46
CCI -6 12 3,03 4,47
31 city. To check whether pooling is allowed or not, a hypothesis should be tested. This hypothesis is about testing if the parameters across the cities are homogeneous. If the null hypothesis is then rejected, it means that the estimation of the parameters should be done through a unit-by-unit model, which will result in city-specific parameters. If not, then pooling should be done. The test which was done is the Chow Test, and this test specifically can detect neglected parameter heterogeneity (Zietz, 2006).
𝐹 = (𝑆𝑆𝑅𝑝𝑜𝑜𝑙𝑒𝑑− 𝑆𝑆𝑅𝑢𝑛𝑝𝑜𝑜𝑙𝑒𝑑)/(𝑑𝑓𝑝𝑜𝑜𝑙𝑒𝑑− 𝑑𝑓𝑢𝑛𝑝𝑜𝑜𝑙𝑒𝑑) 𝑆𝑆𝑅𝑢𝑛𝑝𝑜𝑜𝑙𝑒𝑑/𝑑𝑓𝑢𝑛𝑝𝑜𝑜𝑙𝑒𝑑
To obtain the values of the Residual Sum of Squares (SSR) for both pooled and unpooled data, two regression analysis were done; one when everything is pooled, i.e. treat the data set as a whole, and one of all individual cities, i.e. treat the whole data set as four data sets. The SSR of the four cities individually will then be added together to obtain the unpooled SSR. The degrees of freedom (df) were obtained the same way. Here below are the values.
𝑺𝑺𝑹𝒑𝒐𝒐𝒍𝒆𝒅: 981763617,3 𝑺𝑺𝑹𝒖𝒏𝒑𝒐𝒐𝒍𝒆𝒅: 𝑆𝑆𝑅𝑇ℎ𝑒 𝐻𝑎𝑔𝑢𝑒+ 𝑆𝑆𝑅𝐿𝑒𝑖𝑑𝑒𝑛+ 𝑆𝑆𝑅𝐴𝑙𝑘𝑚𝑎𝑎𝑟+ 𝑆𝑆𝑅𝐵𝑒𝑟𝑔𝑒𝑛 𝑜𝑝 𝑍𝑜𝑜𝑚 81891712,06 + 70165376,89 + 112520479 + 38368269,94 = 302945837,89 𝒅𝒇𝒑𝒐𝒐𝒍𝒆𝒅: 2274 𝒅𝒇𝒖𝒏𝒑𝒐𝒐𝒍𝒆𝒅: 𝑑𝑓𝑇ℎ𝑒 𝐻𝑎𝑔𝑢𝑒 + 𝑑𝑓𝐿𝑒𝑖𝑑𝑒𝑛+ 𝑑𝑓𝐴𝑙𝑘𝑚𝑎𝑎𝑟+ 𝑑𝑓𝐵𝑒𝑟𝑔𝑒𝑛 𝑜𝑝 𝑍𝑜𝑜𝑚 641 + 641 + 610 + 307 = 2199
The Chow-test provides then the following F-value, which has to be used in order to check for significance:
𝐹 = (981763617,3 − 302945837,89)/(2274 − 2199)
302945837,89/2199 =
678817779,41/75
137765.27 = 65,70
For these values of the F-statistic and the degrees of freedom, the Chow Test is highly significant (p=0,000) and thus the null hypothesis is rejected. This means that pooling definitely cannot be used and therefore the unit-by-unit model was used for this study.
4.3 Variable transformation
32 values which are negative in the dataset need to be examined and transformed. There are two variables which contain negative values: maximum temperature and CCI. When looking at the maximum temperature, there is only one negative value, which is a -3 (-0,3) degrees Celsius on one day. Log-transforming this value will result in a missing value. The most logical transformation to do is using Kelvin instead of Celsius. The values of the temperature variable is in 0,1 degrees Celsius, which means that a value of 2730 was added to all values. Eventually this made all the values in the data set positive.
The other variable, CCI, has many negative values, which means that consumers had a very low confidence on a specific day regarding the future of the economy and their household spending. To resolve the negative values problem, a constant was added to these values, so that all the values were positive. As a consequence, the values when doing a log-transformation will also be positive then. The minimum value for CCI is -6, so a constant of 7 was added to all data points of this variable to make it positive. This will resolve the problem. However, when examining the results, extra attention must be paid in generalizing the results, as these may of course be inaccurate due to the addition of the constant.
The transformed model is shown here below. All four cities are examined and analyzed according to the unit-by-unit model.
ln(𝑆𝑇𝑖𝑡)= ln(𝜃)+ 𝛽𝑖1ln(𝑇𝑖𝑡)+ 𝛽𝑖2ln(𝑇𝑖𝑡−1)+ 𝛽𝑖3ln(𝑇𝑖𝑡−2)+ 𝛽𝑖4ln(𝑇𝑖𝑡+1)+ 𝛽𝑖5ln(𝑇𝑖𝑡+2) + 𝛽𝑖6ln(𝑀𝑇𝑡)+ 𝛽𝑖7ln(𝑀𝑇𝑡−1)+ 𝛽𝑖8ln(𝑀𝑇𝑡−2)+ 𝛽𝑖9ln(𝑀𝑇𝑡+1)+ 𝛽𝑖10ln(𝑀𝑇𝑡+2) + 𝛽𝑖11ln(𝑃𝑡)+ 𝛽𝑖12ln(𝑃𝑡−1)+ 𝛽𝑖13ln(𝑃𝑡−2)+ 𝛽𝑖14ln(𝑃𝑡+1)+ 𝛽𝑖15ln(𝑃𝑡+2) + 𝛽𝑖16ln(𝐶𝐶𝐼𝑡)+ 𝑆𝐶𝑖𝑡ln(𝛽𝑖17) + 𝐹𝐹𝑖𝑡ln(𝛽𝑖18)+ 𝛿1ln(𝛽𝑖19) + 𝛿2ln(𝛽𝑖20) + 𝛿3ln(𝛽𝑖21) + 𝛿4ln(𝛽𝑖22) + 𝛿5ln(𝛽𝑖23) 𝛿6ln(𝛽𝑖24) + +ln (𝜀𝑖𝑡) 4.4 Model fit
In SPSS all the variables, excluding the dummy variables, were being log-transformed by taking the natural logarithm of each value. In this way, a linear regression analysis can be done, while it is being considered as a multiplicative model. The results of the four cities can be seen in Appendix 1.
33 four cities can be seen in table 3. The four regression models as a whole were all significant (p=0,000). The Hague had the highest F-statistic and Bergen op Zoom the lowest.
The adjusted R-square is an adjustment or modification of the regular R-square. This modification is done according to the number of terms in the models. While the regular R-square increases when more predictors are being added to the model, the adjusted R-R-square only increases when the new predictor variable improves the model more than by chance would be expected. When looking at the adjusted R-square, Leiden, Alkmaar and Bergen op Zoom all have a similar value. In Leiden for example, approximately 30% of the variation in staying time is explained by the variables that are indicated in the model. The Hague has the highest value for the adjusted R-square; .535.
In order to check the real quality of the model, the assumptions regarding regression analyses should be examined: multicollinearity, heteroscedasticity, non-normality and autocorrelation. When these assumptions are violated, several transformations must be done, which influences the significance and coefficients of the variables. Therefore, first these assumptions were examined. After that, regression analyses were done and interpreted.
4.4.1 Assumption testing
The first assumption checked is homoscedasticity. As aforementioned, this can be detected by doing to Goldfield-Quandt test (1965). However, SPSS does not support this test, so a scatterplot was made of all the four cities regarding z*pred and z*presid (see Figure 4). In SPSS, a line is drawn to find if there is a linear relationship between the values of the standardized residuals and the predicted values of the standardized residuals. As can be seen, the lines are flat and show that there is no positive or negative linear relationship between the two. This means that there is no heteroscedasticity in the data.
F-Statistic p-value Adj. R-square Number of significant variables p < 0,05 p < 0,1 The Hague 32,843 ,000 .535 8 10 Leiden 13,006 ,000 .302 9 9 Alkmaar 10,505 ,000 .265 9 12 Bergen op Zoom 7,693 ,000 .327 8 11
34 Figure 4: Residual scatterplots to identify heteroscedasticity
35 in all cases are not normally distributed. A reason why this may be is that number of outliers in the data set. However, the outliers cannot be deleted or ignored, as then valuable information regarding these days will be lost. One way to overcome this problem is to use bootstrapping. For the final model, bootstrapping was therefore done to account for the non-normality in the model data.
Figure 5: Histograms of the normal distributions of the residuals for the four cities
36 residuals are all positive or negative. Bergen op Zoom for example has a lot of positive unstandardized residuals in the beginning of the total time period. The same accounts for Alkmaar in the middle of the data range; first there exists a large part with mostly positive unstandardized residuals and after that a large part with solely negative unstandardized residuals. Therefore, autocorrelation might exist, but this needs to be checked and calculated using the Durbin-Watson statistic (Durbin, 1970). This is based on the difference between two successive disturbances. As can be seen in Appendix 1, the Durbin Watson value for The Hague is 1,405, for Alkmaar it is 0,920, for Bergen op Zoom it is 1,000 and for Leiden it is 0,723. This means that for all models there is a positive autocorrelation and therefore this needs to be resolved.
Figure 6: Time plots of unstandardized residuals of the four cities
37 residuals and the lagged values of these were found. All the dependent and independent variables have been recoded then in the following way:
𝑙𝑛(𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟)𝑁𝑒𝑤 = 𝑙𝑛(𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟) − [𝑃𝐶𝑖 ∗ 𝑙𝑛(𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟)𝑡−1], where 𝑃𝐶𝑖 is the Pearson’s correlation between the unstandardized residuals and the lagged unstandardized residuals of the original model.
The PC for The Hague was 0,297, for Leiden it was 0,638, for Alkmaar 0,540 and for Bergen op Zoom 0,497 (see Appendix 3). All these four correlations were highly significant (p=0,000). Therefore, these values are used in the formula to create new variables for the new model. New regression analyses were done afterwards and the new Durbin Watson statistics were obtained. For the four cities they were now respectively 2,139 & 2,160, 2,172 and 2,224, which means that all the values are within the threshold range of 1,5 and 2,5 (see Appendix 4). Therefore, there is no violation of the assumption anymore and this means that the model does not exhibit serious serial correlation. However, the interpretation of the results should be done carefully due to the change of the values of the data.
The next and last assumption which was tested is multicollinearity. In appendix 4, the collinearity diagnosis and the VIF scores after correcting for autocorrelation and bootstrapping can be seen. There are no large VIF scores, so this means that there is no serious multicollinearity between variables. This assumption is therefore not violated.
4.5 Results
For the analysis and interpretation of the results, Appendix 4 is used, which contains the bootstrapped regression analyses of the new variables corrected for autocorrelation. In the coming part, each variable is analyzed individually. A summary of the significance of the model variables can be seen in table 4.
Traffic
In the multiplicative regression model, there were five variables regarding traffic; traffic on the current day, 1 and 2 days before, and 1 and 2 days after that specific day. These are interpreted for all cities individually.
38 10% level (p = 0,068). All had a positive influence on the staying time of customers. The influence of traffic on the current day was very large (β = 0,260), while the traffic of two days ago was smaller (β = 0,076). Traffic of t+1 had a smallest beta (β = 0,065). In Leiden three of the traffic variables were significant, which were traffic on the same day, t+1 and t-1. The influence for the first was positive (β = 0,112), while the latter two were negative (β = -0,035 and β = -0,070 respectively. In Alkmaar, only the traffic on the current day had a significant influence and this was positive (β = 0,093), stating that more traffic will result in a longer staying time. Lastly, in Bergen op Zoom, only the traffic of t+1, t-1 and t-2 were significant (β = 0,032, β = 0,029 and β = 0,030 respectively).
Weather conditions
There were two types of variables used in the multiplicative model regarding the weather conditions, namely maximum temperature and barometric pressure. Both variables consisted of the current day data and data regarding the lagged and lead effects of two days. For each city, the variables are examined and interpreted.
For The Hague, an increase of the maximum temperature had a strong negative effect on the staying time of customers (β = -1,051). Also, t+1 and t+2 of the maximum temperature had a strong significant influence. However, anticipation effect for one day was positive, while for two days this was negative (β = 1,358 and β = -1,224 respectively). Regarding the barometric pressure, only the ones on day t and day t-2 were significant and both had a negative effect on the amount of staying time in the shopping street (β = -2,588 & β = -1,840 respectively). When looking at the results for Leiden, no single variable regarding the maximum temperature was significant at the 5% level, only t-1, which was positive (β = 0,416). However, the barometric pressure for time t, t-1 and t+1 were significant. The former was strongly negative, while the latter two were strongly positive (β = 0,737 and β = 1,584 respectively).
Last, Alkmaar had four significant influences on the staying time of customers regarding the weather variables. These were anticipation of warm weather in two days (β = 1,513) and warm weather of two days ago (β = 0,74), and high barometric pressure in the last two days (β = -1,885 and β = 1,317).
CCI
