• No results found

The impact of the Covid-19 pandemic on the sales and returns of an online retailer

N/A
N/A
Protected

Academic year: 2021

Share "The impact of the Covid-19 pandemic on the sales and returns of an online retailer"

Copied!
61
0
0

Bezig met laden.... (Bekijk nu de volledige tekst)

Hele tekst

(1)

The impact of the Covid-19 pandemic on the sales and returns of an online retailer

by

Rick Vondeling

A thesis submitted for the degree of MSc in Marketing Intelligence University of Groningen

Faculty of Economics & Business Department of Marketing

(2)

Abstract

(3)

1. Introduction

Covid-19 has the world in its grip. In most parts of Europe, the pandemic broke out in full force in March 2020. At that time the number of new cases became greater than those in China and the World Health Organization (WHO) began to consider Europe the active centre of the Covid-19 pandemic (Fredericks, 2020). Cases by country across Europe had doubled over periods of typically 3 to 4 days, with some countries showing doubling every 2 days (Roser et al, 2020). While infections soared, hospitals were flooded with Covid-19 patients. This led to some disastrous situations where some patients could not be treated, due to limited capacity. The Covid-19 virus is mainly spread through respiratory droplets when someone is in close contact with someone who has Covid-19 (CDC, 2020). To relieve the healthcare sector and counter the spread and fatalities of the virus, authorities imposed restrictions on their populations to decrease close contact between people. Depending on the pace of the spread of the virus, countries

imposed different restrictions. For example, in Italy where infections and fatalities soared a complete lockdown was employed. This lockdown restricted the movement of the population except for necessity, work and health circumstances (Lowen, 2020).In Sweden on the other hand, people were less restricted in their activities because the virus was relatively under control there (Claeson et al., 2020). In the period of the pandemic, retail sales shifted from traditional brick and mortar to online channels. In April 2020, total retail sales in the EU diminished by 17.9%, while sales through online channels grew 30% compared to April 2019 (Eurostat, 2020). It is not clear however, which factors caused this shift to online shopping.

1.1. Sales

During the pandemic, stringent measures may encourage or necessitate consumers to stay at home. These measures possibly limit most options for consumers of shopping in a physical environment. During such times, online retailers provide an alternate channel for customers to still shop products, when these cannot be bought in physical stores. Online channels are in a position to capitalize on store closures, travel bans, and stay-at-home orders that catapult. These online retailers stand to gain a greater share of consumer wallets, at the expense of department stores and specialty retailers that rely heavily on in-store traffic. Consumers may also choose to voluntarily stay at home, because of fear of getting infected by the virus. As explained before, the virus is mainly spread when people are in close with each other. Due to the absence of an effective vaccine, people are at risk if they expose themselves to the virus. Consumers may therefore choose to do their shopping more from home than in physical stores. Fear of the virus may also invoke impulsive shopping behaviours that stimulate customers to do more purchases in offline as well as online channels. The first objective of this research is to examine the sales of an online retailer. We will analyze if a disruption took place and if that disruption can be

attributed to restrictive governmental measures, fear of the virus or some other factor.

Research question 1: How has the Covid-19 pandemic impacted the sales of the online retailer? 1.2. Product returns

(4)

generally have less than 10% of their products returned (Rudolph, 2016). These returned products constitute a major cost factor for online retailers in the form of lost sales and reverse logistics. A study by Apriss Retail (2020) showed that consumers in the US returned

approximately 41 billion worth of products that they ordered online in 2019. The share and amount of returned products may have become even more problematic during the pandemic. An increase in sales is likely to also result in an increase of returned products. If the amount of returned products grows at the same rate as sales grow, or even faster, companies may have to handle double or triple the amount of usual returned products. If retailers are not prepared for this, such situations can certainly cause problems. Furthermore, both stringent measures and fear of the virus may have shifted a lot of consumers from offline to online channels. These

customers may be inexperienced at evaluating products that are presented online. This can lead to incorrect expectations of products. If the ordered product fails to meet the expectation of the customer, it is more likely to be returned. Besides incorrect expectations, products that were bought due to impulsiveness during the pandemic may also invoke negative emotions at

customers that result in more product returns.The second objective of this research aims to test if the stringency of measures and fear of the virus have impacted the returns of the online retailer. Research question 2: How has the Covid-19 pandemic impacted the product returns of the online retailer?

1.3. Country and product category

Of course the situation is not the same for all retailers. According to a research by Deloitte (2020) pure online retailers experienced 45% in growth during the pandemic, whereas

multichannel companies with physical footprint are challenged to protect their revenues. These retailers may have seen an uplift in their online sales but not nearly enough to compensate for disappearing store traffic. Not all retailers can sell their products easily online or transfer their business to online channels. Furthermore, certain types of products may become more in demand than others. Fear of the virus may cause hoarding behavior in food or other essential products. Social distancing and lockdown measures resulted in the closing of offices, gyms and social activities. When spending most of their time at home during lockdown, consumers may develop different needs for products. Hygienic products that prevent infection of the virus and products that will help them function and stay entertained at home (e.g. toys, fitness products) are likely to more in demand during lockdown. Cultural differences between countries may also alter the response of consumers to the Covid-19 pandemic. Most aspects of consumer behavior are culture-bound (De Rooij, 2020). The retailer in this research operates in 14 countries around Europe and it is likely that consumers of different countries responded different to the pandemic. A research by McKinsey (2020) showed significant differences between consumers purchase intent and orientation across several countries in Europe. For the reasons stated above it is important to take product category as well as country into account. These variables are included in the research, to test if the impact of the Covid-19 differs across categories and countries. 1.4. Practical Relevance

(5)

Covid-19 on sales and returns can be key for business survival. Quantifying the extent to which sales and returns may be impacted should help businesses to be more prepared and ultimately be more profitable during a health crises like the Covid-19 pandemic. Although these pandemics are a relatively rare occurrence, they can prove to be vital for business growth and survival. For some businesses the pandemic might push them to bankruptcy. Others are forced to transform their business model to ensure survival. Companies that handle these crises successfully, may have started the path to long term success. For example, the SARS epidemic in Asia is known for functioning as a catalyst for the online retail boom in China (Pronk, 2020). Several articles have focused on how the SARS pandemic spurred growth opportunities Alibaba and JD.com (e.g. Zheng, 2020). These current online retail giants, who are the second and third largest online retailers globally after Amazon (Levy, 2019), used the closing of China during the SARS outbreak to turn their business models fully to the then primitive setup of online retail.

Understanding changing customer needs and behaviors that develop during a pandemic is vital for business profitability.

1.5. Scientific Relevance

This paper builds on and contributes to the growing amount of literature measuring the financial impacts of Covid-19. Recent papers concerning the financial impact of Covid-19 use high-frequency transaction data, such as credit and debit card data, to analyze aggregate

consumer spending during the pandemic (see e.g. Golec et al., 2020; Chetty et al., 2020). These studies show how consumers respond to the pandemic, for instance by adjusting their

expenditures for certain industries and sectors. However, they do not reveal how consumers’ orientation may change across different product categories due to Covid-19 related factors. This study hopes to provide new insight into consuming behavior, by analyzing what product

categories benefited most and least from the pandemic. Furthermore, prior studies limit their focus on the impact of the Covid-19 pandemic on just the purchases of consumers. They do not include other consumer behaviors in the customer purchase process, such as the return of bought products. Like explained before, there good reasons why consumers may return more products during the pandemic and it is important to research this subject. This research aims to test if this is indeed the case. By doing so,

1.6. Approach

To address the research questions at hand, the following steps will be taken. In the next section a brief overview of existing literature and empirical evidence regarding the topics at hand will be presented. Relevant theories will be discussed in order to explain the reasoning behind the hypotheses that will be tested in this research. In the third section we will discuss the study design and provide some insights in the data using descriptive analysis. In the fourth section, the results from an empirical study conducted in cooperation with the European online retailer will be presented. Transaction data ranging van January 1st 2019 to September 30th 2020 will be analyzed by means statistical modelling. We will use data from the University of Oxford and the European Centre for Disease Prevention and Control (ECDC) to quantify the stringency of measures and fear of the virus. In the fifth section of this paper the results of these analyses will be discussed. Finally, the last section of this paper discusses the results and deals with

(6)

2. Conceptual Framework 2.1. Literature review 2.1.1 Covid-19

This section will summarize important findings of other studies on the impact of the Covid-19 and SARS epidemcs on retailing. The research of Golec et al., (2020) measures the impact the Covid-19 pandemic on aggregate spending behavior. This study distinguishes the economic effects of voluntary responses to Covid-19 from those attributable to government lockdown measures. They found that the amount of new hospitalized Covid-19 cases has a strong and statistically significant negative effect on the change in physical transactions by consumers. This finding holds even for sectors that were subject to restrictions due to the lockdown measures. The drop in physical transactions could be partly offset by an increase in online transactions. Their analysis shows that the incidence of Covid-19 has a large positive correlation with online grocery spending. This study pointed out a further avenue of research on the subject of which product categories are most impacted by fear. This research will try to provide more insight in this subject. Chetty et al. (2020) has examined the effect of executive lockdown orders on changing consumer spending in the US. They found that spending fell sharply in most states before formal state closures. Moreover, states’ reopenings had little immediate impact on economic activity. They conclude that health concerns are the core driver of reductions in spending, rather than government-imposed restrictions. Goolsbee et al. (2020) examined to what extent the reduction in economic activity was due to government restrictions or to people’s voluntary decision to stay at home. They perform an analysis of customer store visits on a county level. By comparing consumer behavior within the same commuting zones but across counties with different government restrictions, they find that lockdown orders account for only a modest share of the decline in economic activity. They state that although overall consumer traffic fell by 60%, legal restrictions explain a decline of only 7%.

Chen et al. (2020) find a similar result to that of Goolsbee et al. (2020) and for China. They study the drop in card and QR scanner transactions through UnionPay. They also find that the effect on consumption is stronger in cities that have had more Covid-19 cases. More

specifically, they argue that in the 20 cities that received the highest inflow of Wuhan residents (the epicentre of the Covid-19 outbreak), consumption decreased by 12% more than in other cities in their sample. For cities reporting zero cases (as of late March), the decrease in offline consumption was 13% less than for cities with positive Covid-19 cases in the same time period. They find that a doubling of the number of infected cases at the city-level was associated with a 2.8% greater decrease in offline consumption.

A research by Ahmed et al. (2020) concluded that multiple factors, such as fear of complete lockdown, the scarcity of goods and panic buying have had a compelling and affirmative influence on the sharp swings of impulse buying patterns during the pandemic. Among employed survey participants, spending more time at home seems to lead to impulsive buying behavior. Aftonbladet (2020) found that during the pandemic consumption of cheaper “unnecessary” products has increased, despite the overall reduced consumption. This was

(7)

even more notable in countries where lockdowns were undertaken.Due to technological advances and massive e-commerce growth, online impulsive purchasing has become a widely spread phenomenon. According to the estimations, online impulsive buying accounts for 40% of all online consumer expenditure (Liu et al., 2013).

2.1.2. SARS

Research has found similar developments in consumer behaviour during the SARS (SARS-CoV-2) epidemic in Asia in 2002-2003. SARS is, like COVID-19, from the family of coronaviruses that are associated with the common cold. Jung et al., (2016) a found that the outbreak of the SARS epidemic caused a substantial disruption in consumer expenditures.

Consumers spent less in traditional shopping channels, like brick and mortar stores, but increased their expenditures in online channels. Forster et al. (2005) found similar consumer behavior during the SARS epidemic. Using transaction data from credit card expenditures, Forster et al. (2005) found that there was a clear rise in online shopping in response to the spread of the SARS virus. The demand for online shopping appeared was closely related to the spread of the

infection, rising most rapidly when infections were growing quickly and slowing again when infections tapered off.

Very few articles have covered the subject of Covid-19 and product returns. The research done in these studies was mainly from a supply chain perspective, for example the strain on reverse logistics (De Angelis, 2020).To summarize, research so far has been able to conclude that consumer spending behavior changes substantially during a pandemic. Some industries and channels benefit from the pandemic whereas others are negatively impacted. In nearly all cases consumers dramatically cut down expenditures at traditional shopping channels, like brick and mortar stores, but spend more in online channels. Furthermore, research found that this change in consumption behavior is not fully attributable to restrictive measures. Consumers are also

voluntarily choosing to stay at home in self isolation. Combined with literature on the drivers of sales and product returns, this research adds to existing literature concerning the financial impact of Covid-19 in the following ways:

i) Providing insight into which product categories consumers shifted their expenditures to during the Covid-19 pandemic

ii) If this increase in sales is attributable to restrictive measures, fear of the virus or another factor

iii) How consumers product returning behavior is affected by the Covid-19 pandemic

2.2. Theoretical Framework

(8)

2.2.1. Consumer channel selection

Before making purchases, customers gather information of the potential purchase. Customers weigh up benefits and costs for alternatives for different channels (Keeney, 1999; Shih, 2004). Motivational factors account for these perceived costs and risks. These motivational factors are associated with convenience, risk and service preferences (Chiang et al., 2006;

Montoya-Weiss et al., 2003). Convenience orientation characterizes customers, who regard shopping as a rational problem-solving process. It is important to these customers to acquire the sought-after product with a minimum investment of time, physical effort, and mental effort (Schröder et al., 2008). The risk aversion motive refers to perceived risk, such as customer uncertainty as to the negative consequences of a purchase and the significance of these consequences. Concerning service, customers that are moved my this motive may value the option of expert advice about offered products. Service oriented customers may base their decision on customer the quality of service offered in each channel. Channel conflicts in

multichannel systems can occur in case alternative means of reaching customers compete with or bypass existing channels (Balasubramanian, 1998; Steinfield, 2004), which potentially leads to the cannibalization of sales from one channel by another. In other words, if the benefits and costs of one channel outweigh those of another channel, channel conflicts may occur. This potentially leads to the cannibalization of sales from one channel by another.

2.2.2. Product returns

Customers’ decisions to purchase and return a product are based on their level of

expectations about the product’s performance and the uncertainty surrounding these expectations (Minnema et al., 2016). The Confirmation/Disconformation (C/D) framework proposed by Oliver (1980) can be used to explain the return behaviours caused by unsatisfactory purchases. According to this framework customers’ satisfaction is impacted by the expectations and perceived performance of products. Expectations refer to the attributes or characteristics that a person anticipates or predicts will be associated with an entity such as a product or service service. Perceived performance refers to a person’s perceptions of the actual performance of the product or service. Following the (C/D) framework, individuals experience negative

disconfirmation when a product performs below the expectation, and consequently

(9)

2.2.2. Impulsive buying as a coping mechanism

Impulse buying behavior can be induced by external stimuli. External events, which have the potential to threaten safety, heighten the compensatory processes and behaviors to alleviate stresses. To relieve the pains and stress of the external event, consumers turn to shopping as a coping mechanism, as observed by many researchers (Arndt et al. 2004; Maheswaran & Agrawal 2004). In the face of disasters, consumers tend to shop more than the usual, exercising excessive panic buying. This was observed during the September 11th attacks on the World Trade Centre, as the US President strongly urged Americans to do so (Arndt et al. 2004). When the hurricane Katrina hit the American coast in 2005, feelings of loss of control linked to the event, resulted in increased levels of stress (Kennett-Hensel et al. 2012). This phenomenon led to undesirable buying behaviors, such as impulsive and compulsive buying (Sneath, Lacey & Kennett-Hensel 2008). As a consequence of the disastrous storm, consumers participated in the purchasing of utilitarian goods, such as tools and material for repairs, personal care items and clothing. Post event, many consumers participated in impulsive hedonic purchasing, such as designer clothes and accessories, in order to reduce stress levels and escape the negative effects of the disaster (Kennet-Hensel et al. 2012). The choice on the type of products purchased, utilitarian or hedonic, is based on the attached shopping experience, as it is known to support people regulate distress and restore a feeling of control over the situation (Rick et al., 2014). According to Addo et al. (2020), not only external stimuli are involved, but also internal stimuli related to personal

emotions. Impulse buying could help people deal with low confidence, negative emotional states and negative mental thinking.

2.3. Conceptual model and hypotheses 2.3.1. Sales model

Stringency of measures

Stringent measures can limit the offline channel options for consumers. During the pandemic, retailers that excelled in their ability to sell through their online channels were in a position to capitalize on store closures, travel bans, and stay-at-home orders. These companies stand to gain a greater share of consumer wallets, at the expense of department stores and specialty retailers that rely heavily on in-store traffic. For these reasons we hypothesize that the stringency of measures has a positive relationship with daily sales.

H1a. The stringency score of measures has a positive relationship with the daily sales of the online retailer.

ICU patients

(10)

because consumers may indulge in impulsive purchases to cope with the virus. Accordingly we postulate the following hypothesis:

H2a. Fear of the virus has a positive relationship with the daily sales of the online retailer

Moderating effects of product category and country

The positive effects of both Covid-19 related factors in sales will not likely be consistent across product categories. Social distancing and lockdown measures resulted in the closing of offices, gyms and other social activities. When spending most of their time at home during lockdown, consumers may develop different needs for products. Hygienic products that prevent infection of the virus and products that will help them stay entertained at home (e.g. toys, fitness products) are likely to more in demand during lockdown. Other categories, such as travel related products and formal wear may experience declines in sales. We hypothesize that the strength and direction of the effects is moderated by product category.

H1b. The positive effect of the stringency of measures on sales is moderated by ones product category

H2b. The positive effect of fear of the virus on sales is moderated by ones product category Consumers across countries may cope differently with the pandemic. This may be represented in their purchasing behavior. This research will not go into depth about the cultural differences between countries and the impact of those differences on consumer behavior. What this research will do, is test if the effect of the Covid-19 variables was consistent across

countries. To do this, the following hypotheses are tested.

H1c. The positive relationship between the stringency of measures and daily sales is moderated by country

H2c. The positive relationship between fear of the virus and daily sales is moderated by country

Control variables

We included day of week, month and year as control variables in our model to account for time trend and variation. We also included a Black Friday in the control variables, to control for outliers in sales for this day. Figure 1 shows a graphical representation of the

(11)

Figure 1. Conceptual sales model

2.3.2. Returns model

Stringency of measures

Like discussed earlier, stringent measures and Covid-19 outbreaks may have shifted consumers from traditional to online channels. Consumers that would usually shop in offline channels may be inexperienced in shopping online. This inexperience can lead to forming incorrect expectations of products that are ordered from online channels. Following the C/D framework explained in section 2.2.2. products that do not live up to the expectations, can leave customers unsatisfied. This is likely to result in a higher probability of product return. For this reason it is expected that the stringency of measures has a positive impact on the probability of return.

H3a. The stringency of measures has a positive impact on the probability of product return.

Fear of the virus

Fear of the virus may have moved consumers from physical to online channels. Like explained above, these new customers may be inexperienced at evaluating products. This could ultimately lead to dissatisfaction and product returns. Furthermore, fear of the virus may have also invoked impulsive shopping behavior during the pandemic. Customers may buy products to distress and restore a feeling of control over the situation. As discussed in the theoretical

framework, impulsive purchases are more likely to be returned. For this reason we postulate the following hypothesis:

(12)

Moderating effects of product category and country of customer

Previous research has stated the importance of product category in product returns. A major difference between categories is the degree to which it is difficult for customers to assess the fit between the product and their own preferences (fit uncertainty; Hong and Pavlou 2014). Thus, it is critical to first understand how the category of the product might affect the product return rate. Consumers during the pandemic may have indulged in impulsive buying more in some categories than others. We will test if there is a moderating effect of product category on the relationship between the Covid-19 variables and probability of return.

H3b. The positive effect of stringency on the probability of return is moderated by product category

H4b. The positive effect of fear of the virus on the probability of return is moderated by product category

We also want to test if the effects of stringency and ICU patients on the probability of return is moderated by country. Again, this research will not go into detail into the factors explaining different behavior between countries. We will test if the country of customer moderates the relationship between the Covid-19 variables and the probability of return.

H3c. The positive effect of stringency on the probability of return is moderated by the country of the customer

H4c. The positive effect of fear of the virus on the probability of return is moderated by the country of the customer

The online retailer in this research not only sells its products through its own channels, but also through the channels of other online retailers. This is comparable to a retailer selling its products through its own website but also through Amazon for instance. By doing this the retailer can benefit not only from its own customer base, but also the visitors of other online retailers. Products that were sold through these channels are labeled as marketplace products. Products may be displayed or described in a different way than on the own website of the online retailer. As a consequence people can form different expectations of the product, ultimately leading to different return rates. This research also wants to test if the marketplace attribute impacts a products probability of return.

H5a. Marketplace products have a higher probability of return than regular products

We also want to test if the effect of the marketplace attribute on probability of return is consistent across categories. Like explained before, major differences exist between product categories and the degree to which it is difficult for customers to assess the fit between the product and their own preferences.

(13)

Control variables

Payment method

Previous research has stated the importance of consumers’ payment methods when it comes to returning. Nowadays there less risks for consumers, because their purchases are protected if they pay with credits card. Other payment options allow consumers to pay after delivery. Research has found that customers who pay after delivery are about twice as likely to return an item they purchased than those who pay in advance (Asdecker, 2015). The reason for this phenomenon is that when returning products, customers who pay after delivery do not have to make sure that their money is refunded. Instead, they simply never transfer the money in the first place. The variable payment method is also included in our model to control for payment method returning variation.

Price

When it comes to price, customers are more critical in evaluating products for expensive products. Higher priced products are therefore are more likely to get returned when a product lacks fit (Hess and Mayhew, 1997). This is confirmed in a research by Anderson et al. (2009). They found that lower priced products had lower return rates compared to higher-priced

products, all else being equal. For these reasons it is important to take product price into account and product price is included as a control variable in our conceptual model.

Time

We again included day of week, Black Friday, month and year as control variables in our model to control for day, month and year variation. Figure 2 shows a graphical representation of the interrelationships between variables.

(14)

3. Study design

3.1. Quantitative study

To answer the research questions at hand we will make use different sources of data. The goal of this research is to describe and predict information for the sales and products probability of return. We will describe how variables are conceptualized and how data is collected for these variables in order to quantify them. Later in this section, we will describe the data. After that, we will use statistical models to test for possible relationships between variables, which allows us to draw conclusions from the data. These models are simplified, mathematically-formalized way to approximate reality and to make predictions from this approximation. We want to find out what explains and predicts the dependent variable, which in our case are:

1. Sales

2. The probability of product return

In statistical modeling, regression analysis is a set of statistical processes for estimating the relationships between a dependent variable and one or more independent variables. In order to explore the relationships between our independent and dependent variables, we will use multiple regression methods. Different types of data require different types of techniques for analyzing the data. Further in this section we describe the conceptualizations of important variables and explain what data we used to conduct this study. After that we describe the methodology for modelling the different variables.

3.2. Data

3.2.1. Retailer data

The data for this study came from an online retailer that sells exclusively through its online channels. The retailer sells products across data 13 categories in 14 countries across Europe. In section 3.4 the different categories and countries are listed and their descriptive statistics discussed. The retailer offers around 300.000 different products across approximately 250 web shops. These products are focused around four main markets. These are; Home & Interior, Sports & Outdoor, Dining and Toys & Hobby. The company also sells their products through channels of other online retailers. In the dataset these are called marketplace products. We use a dummy variable to make a distinction between regular and marketplace products. The retailer originated in 2004 in the Netherlands. After a few successful years the company

(15)

ordered, the payment method used, if it was a regular or marketplace product and if the product was returned. More in depth descriptive analysis on the data will be discussed in section 3.4.

3.2.2. Stringency data

In order to conceptualize and quantify the stringency of measures, we use the stringency index of the University of Oxford. The University of Oxford collects data on the stringency of measures per day per country. Data is collected from public sources by a team of over one hundred Oxford University students and staff from every part of the world (University of Oxford, 2020). This stringency index is a composite measure of nine indicators measured on an ordinal scale, rescaled to vary from 0 to 100. The data from the indicators is aggregated into an index, reporting a number between 1 and 100 to reflect the level of government action on the topics in question. These nine indicators are:

1. School closing 2. Workplace closing 3. Cancel public events 4. Restrictions on gatherings 5. Close public transport 6. Stay at home requirements

7. Restrictions on internal movement 8. International travel controls 9. Public information campaigns

This index is simple average of the individual component indicators. This is described in the equation below where k is the number of component indicators in an index and Ij is the sub-index

score for an individual indicator:

Stringency of measures is thus defined in this research as the stringency index score per country per day.

3.2.3. Hospital data

(16)

deadly, which is the case for Covid-19. The amount of fatalities would also be a possibility to measure fear, but this number likely underestimates the amount of people that that experience serious illness through the virus. Data from the European Centre for Disease Prevention and Control (ECDC) will be used as the source of data for ICU patients across countries. The ECDC is an agency of the European Union aimed at strengthening Europe's defences against infectious diseases. The core functions cover a wide spectrum of activities, including epidemic intelligence. The ECDC collects information about Intensive Care Unit (ICU) admission rates and current occupancy for Covid-19 by day and country. Fear of the virus is in this research thus measured as the amount of Covid-19 infected ICU patients per day, per country.

3.3. Methodology 3.3.1. Sales

The first objective of this research is to analyze how the daily sales are impacted. The orders in our data were grouped by country, category and day to create a new variable: daily sales. If no sales were observed for a certain combination of day, category and country then a 0 was noted. Because we have 13 different categories and 14 different countries in our dataset, we have a total of 182 observations per day. The data ranges from January 1st 2019 to September 30th 2020. This is a total of 639 days. This leaves us with 639 * 182 = 116298 observations. Our

dependent variable daily sales is a continuous variable. This means that it can take on an unlimited number of values between the lowest and highest points of measurement. The most common method to predict a continuous variable is linear regression, in which we estimate the line that most closely fits the data according to a specific mathematical criterion. For example, the method of ordinary least squares (OLS) computes the unique line (or hyperplane) that minimizes the sum of squared differences between the true data and that line (or hyperplane).

Our data has a corner solution, this means that an observed value cannot be higher and/or lower than a certain threshold. In our case the data has a corner solution at 0, because we an observation cannot have negative sales. An observation can take any value between 0 and infinity. In our dataset a total of 63091 observations take the value of the corner solution (0) and 53207 observations are above this value. A distribution of these observations is visualized in Appendix B. The correct values for these observations are observed, but values above 0 are observed in a selected sample that is not representative of the population. For example, for our retailer around 40% of all revenue comes from the Netherlands, whereas less than 1% of revenue is generated from Portugal. A lot more 0’s are observed for Portuguese observations than for observations from the Netherlands. Because we are dealing with data that has a corner solution, we cannot estimate a model using OLS. OLS regression leads to inconsistent parameter estimates because, in linear regression y = xβ + ε there is no way to guarantee that all predicted values are positive.

The statistical model that is appropriate to use for a context with a corner solution is the so called standard censored Tobit model or type I Tobit model. Below, the mathematical

(17)

For a corner response variable we are interested in E(y| x) or E(y| x,y > 0) as the latent variable usually has no sensible interpretation in this context, this will be explained later. The expected value of y is derived in the following way:

Where λ(c) = 𝜙(c)

Φ(c) is called inverse Mills ratio.This is the probability of being uncensored (observation being > 0) multiplied by the expected value of y given y is uncensored. This model can be estimated with Maximum Likelihood: for what β-values is it most likely the model matches the data. The Tobit coefficient ("beta") estimates the linear increase of the latent variable for each unit increase of your predictor. The latent variable is identical to the observed variable for all observations that are above the threshold. It also measures the linear increase of your predictor on your response for all observations above that threshold, just like an OLS coefficient. For this reason, we cannot interpret these Tobit coefficient estimates of a regressor as giving the marginal impact of that regressor on the mean value of the observed regressand. The marginal effects at the average observation are instead calculated by multiplying these beta coefficients by the probability of being uncensored:

(18)

the models will be compared using several model evaluation metrics, such as Akaike’s

Information Criterion (AIC) and Bayesian Information Criterion (BIC). Lastly the Tobit models will be compared to a regular OLS models, to test if the Tobit models fit the data better. A definition of all variables that were included in these models is presented in appendix A. 3.3.2. Probability of product return

The second objective in this study is to analyze what variables impact the probability of return, and how they impact the probability of return. In the case of product return, the dependent variable has a binary outcome. Observations can be classified in two groups (0: no return, 1: returned products). A linear model will not suffice, because dependent variable Yi is binomial

and can only take two values. A linear function would not fit the data well because predictions of

Yi could go below 0 and above 1. These values of 0 and 1 are arbitrary. The important part is not

to predict the numerical value of Yi, but the probability that a product is returned or not, and the

extent to which that probability depends on the predictor variables. The binomial logistic regression model is intended for binary classification problems. The logistic regression predicts the probability of an instance belonging to class 0 or 1. It is a linear algorithm and assumes a linear relationship between the input variables and the output variables. Instead of fitting a straight line or hyperplane, the logistic regression model uses the logistic function to map the output of a linear equation between 0 and 1:

We call the term in the log() function "odds". This is the probability of event divided by probability of no event and wrapped in the logarithm it is called log odds. By applying the exp() function to both sides of the equation, we can figure out how the prediction changes when one of the features of xj is changed by 1 unit:

For a binary feature, one of the two values is the reference category. Changing the feature xjfrom the reference category to the other category changes the estimated odds by a

(19)

3.4. Descriptive Analysis

In this section we will perform descriptive analysis on various variables included in models. Descriptive analysis is an important step for conducting statistical analyses. It gives us an idea about the distribution of our data and enables us to identify possible associations among variables. This makes us more ready to conduct further statistical analysis.

3.4.1. Covid-19 variables

In the figures below, the stringency index scores over time per country are visualized (Figure 3.), as well as the amount of ICU patients over time per country (Figure 4.) In both these figures, the year 2019 is excluded from the X-axis, because for both variables all countries observed only 0’s in 2019.

Table 1. Descriptive statistics for Stringency index score and ICU patients

Variable Min. 1st Quantile Median Mean 3rd Quantile Max

Stringency Index Score 0.00 0.00 0.00 18.50 44.44 93.52

ICU patients 0 0 0 122.30 11 7019

Figure 3. Amount of ICU patients over time by country

(20)

Italy (red) was the first country that exceeded the amount of a thousand Covid-19 ICU patients. Following and eventually exceeding Italy in the amount of Covid-19 ICU patients was France (yellow). On April the 8th in France, the maximum amount of Covid-19 ICU patients in our dataset was reached, which was 7018. We observe that quickly after reaching their maximum amount, the amount of Covid-19 ICU patients in countries like Italy and France decreased at a very fast rate. For countries such as the Netherlands (pink) and Belgium (light blue), who never reached a maximum amount of Covid-19 ICU patients more than 1500 patients, the decline in patients was much more gradual. From this graph we can also observe that around July 2020, pretty much all countries had regained control over the amount of Covid-19 ICU patients. No country had more than 500 ICU patients at this time. However, around September 2020 the amount of ICU patients started to increase again.

Figure 4. Stringency index scores over time by country

In Figure 4 we observe that during the time that the amount of Covid-19 ICU patients soared in most countries, countries started applying stringent measures. This is only logical as authorities tried to counter the spread of Covid-19 with this measures, to relieve pressure of the ICU units. Italy (red) was the first country in the dataset that exceeded the threshold of a thousand patients and Italy was also the first to apply stringent measures, followed by France (yellow). The maximum observed stringency score of measures was also observed in Italy, which was 93.52. In this graph we can see that before April, every country had applied measures. Ireland was the last country to impose measures, on March 26th. The stringency scores of most countries decreased gradually after peaking in April. This is in line with the developments in the amount of ICU patients, which also started to decrease around this time. After applying

(21)

3.4.2. Daily Sales

Table 2. Descriptive statistics for daily sales

Variable Min. 1st Quantile Median Mean 3rd Quantile Max

Daily Sales 0.00 0.00 0.00 316.20 182.90 37540.50

Table 2 shows the descriptive statistics of the daily sales variable. As stated in section 3.3.1. a lot of 0’s are observed, that is why the first three values in the table are 0. A distribution of this variable is visualized in Appendix B. The maximum amount of sales that was observed was 37,540.50 for the category Toys & Hobby in the Netherlands on June 24, 2019. In Table 3 the sales per country and category are presented. Most of sales come from the Netherlands (44.03%), Belgium (20.86%) and Germany (15.16%). The sales from these three countries account for 80.05% of all sales. As for product categories, Garden (18,02%), Home & Interior (17,31%), Sports Cycling (15,86%) and Toys & Hobby (14,38%) constitute the biggest portion of sales. When visualizing sales, we chose to plot weekly sales instead of daily sales. The weekly sales lines are less influenced by outliers and provide clearer overall picture than daily sales, because a high number of lines are plotted.

Figure 5. Weekly sales over time by product category

Table 3. Sales per country and category

Category Absolute amount of total sales % of total sales Country Absolute total amount of sales Relative amount of total sales Custom 391,548 1,06% Austria 681,406 1,85%

Tools & DIY 468,043 1,27% Belgium 7,762,377 20,86%

Electronics 281,949 0,77% Czech Republic 66,425 0,18%

Photography & Optics 1,824,192 4,96% Denmark 1,265,763 3,44%

Household 2,803,709 7,62% Estonia 176,296 0,48%

Toys & Hobby 5,288,956 14,38% Germany 5,575,761 15,16%

Office supplies 34,312 0,09% Finland 72,054 0,20%

Beauty & Lifestyle 664,938 1,81% France 1,266,316 3,44%

Outdoor 3,884,510 10,56% United Kingdom 1,602,048 4,36%

Garden 6,626,965 18,02% Ireland 755,190 2,05%

Sports 2,304,905 6,27% Italy 164,586 0,45%

Sports Cycling 5,832,821 15,86% The Netherlands 16,189,780 44,03%

Home & Interior 6,366,588 17,31% Portugal 111,664 0,30%

(22)

In Figure 5 the weekly sales per category are plotted over time. From this figure we can clearly observe that categories like Garden (orange), Sports cycling (red) and Sports (light blue) experienced heightened demand starting in April 2020. From Figure 3 and 4 we know that this was around the same time that countries applied stringent measures and ICU patients increased. These categories experienced a big increase in sales also in the summer of 2019, around July. The pandemic may have enlarged this increase and shifted it a couple of months forward in 2020. The category that experienced the highest peaks in the summer of 2019 was Toys & Hobby (grey). This category also saw an increase in sales in April, 2020. Home & Interior was the most popular product category in the winter (dark green line). This category did not

experience an increase like Garden or Sports cycling during the pandemic, but levels of sales remained pretty stable. After the period of very high sales in March 2020 – June 2020, we can also see that for a lot of categories sales dipped to their lowest point at the end of the figure, this was in September 2020. At this time the amount of ICU cases had already dipped down and were starting to gradually increase again. Measures were less stringent than in the beginning of the pandemic, but were still active.

(23)

In Figure 6 the weekly sales over time by country are visualized. From this figure it becomes clear that most sales are coming from the Netherlands (pink line) followed by Belgium (light blue) and Germany (orange). When looking at 2019, we see that sales from the

Netherlands, Belgium and Germany peaked during the summer in June and July and gain near the end of year, which is probably due to holiday season. We also see that sales in pretty a lot of countries started trending upwards in March 2020 and peaking in April or May. For the

Netherlands and Germany the amount of weekly sales did not, or just exceed the levels of July 2019. For countries such as Belgium and France (yellow) sales clearly did exceed levels that were achieved in 2019. From these graphs it becomes clear that in the same that countries imposed measures and Covid-19 patients soared, sales of the online retailer trended upwards. In the next chapter we will test if there is a relationship between the Covid-19 variables and sales. 3.4.2. Returns data

In this section the data on products returns will be described and visualized. In the dataset a total of 772.035 products were ordered, of which 53.327 products were returned, leaving an average return of 6,91%. This is a relatively low percentage compared to what earlier research has found for other online retailers. Table 4 show the absolute amount of returns and the return percentage per country and category.

(24)

From the statistics in Table 4 we see that Outdoor products have highest rate of return (11,73%). This percentage is substantially higher than every other category, the second biggest rate of return is 8,71% for Sports cycling products. Toys & Hobby and Office supplies have the lowest rates of return, 3,77% and 3,57%. Between countries, products ordered from Germany have the highest return rate (10,23%). The highest amount is again substantially higher than other groups. The second largest return percentage is 7,61% for products ordered by Austrian consumers. Ireland has the lowest rate of return with 2,98%. The most absolute amount of returns came from the countries which also generated the most sales, which is logical.

In Figure 7 the absolute amount of returns and return percentage over time are visualized. From this graph we can see that the absolute amount of returns started increasing in March 2020, but not exceeding amounts that were reached in late 2019. The amount of ordered products grew faster than the amount of returned products, because the relative amount of returned products returned to its lowest point in this timeframe. So, when stringent measures were imposed and ICU patients increased, the relative share of returned products decreased. This interesting, because we hypothesized earlier that the Covid-19 variables would have a positive impact on the probability of return. In the next chapter we will test if there is a relationship between these variables and if there is, what kind of relationship.

Category Absolute amount of product returns % Return percentage Country Absolute amount of product returns % Return percentage Custom 800 5,64% Austria 781 7,61%

Tools & DIY 752 7,10% Belgium 7,563 4,72%

Electronics 411 6,94% Czech Republic 47 3,66%

Photography & Optics 1,409 5,90% Denmark 1,292 5,13%

Household 4,890 4,77% Estonia 126 4,96%

Toys & Hobby 3,316 3,77% Germany 11,608 10,23%

Office supplies 41 3,57% Finland 44 5,41%

Beauty & Lifestyle 998 5,97% France 971 5,50%

Outdoor 8,310 11,73% United Kingdom 1,473 5,56%

Garden 4,295 4,55% Ireland 418 2,98%

Sports 4,275 8,15% Italy 117 4,91%

Sports Cycling 13,361 8,71% The Netherlands 27,636 7,38%

Home & Interior 10,469 7,59% Portugal 75 3,45%

(25)

Figure 7. Weekly absolute amount of returns and return percentage over time

(26)

Figure 8. Comparison between return percentages for marketplace and non-marketplace products per category

Chapter 4. Results 4.1. Daily sales models

The results of our models are presented below. A table including standard deviations of variables is included in appendix C. The predictor variables in these models included many categorical variables, such as category and country. This means that we had to use dummy variables to estimate our model using a reference level. The reference level that was used in the sales models is presented in table x. The constant in the models are the prediction of the

reference level. The intercept of the reference level is €2,519.44. The effects of coefficients represent a change in sales relative to the reference level.

Table 5. Reference level for sales models

Category Country Day Month Year Black Friday

(27)

Table 6. Sales model output

Dependent variable: Daily Sales in €

Type of model

Tobit type 1 OLS Tobit type 1 Marginal

regressions regression effects

Null model With Control Variables

With Direct Effects

With interaction effects (Full model)

(28)
(29)

4.1.1. Comparison between models

The Tobit model that includes all effects will be used for interpretation of the effects of the independent variables on our dependent variable. This model scored the best on our model evaluation metrics. We can see that the censored regression with interaction effects has the best AIC and BIC scores. These are an unbiased estimate of the model prediction error MSE (Mean Squared Error). The lower these metrics, the better the model. For the Tobit model there is no conventional measure of R2. This is because the standard linear regression model estimates parameters by minimizing the residual sum of squares (RSS), whereas the Tobit model maximizes the likelihood function. The likelihood ratio test (LRT) is a statistical test of the goodness-of-fit between two models. A relatively more complex model is compared to a simpler model to see if it fits a particular dataset significantly better. Adding additional parameters will always result in a higher likelihood score. However, there comes a point when adding additional parameters is no longer justified in terms of significant improvement in fit of a model to a particular dataset. After conducting likelihood ratio tests, the full Tobit model proved to be significantly different from than its nested equivalents. This means that the worth of additional parameters, in this interaction effects, was justified.

PT*ICU_patients -0.290 0.596 0.149 SE*ICU_patients -0.564*** -0.514*** -0.129 AT*Stringency -3.670*** -6.791*** -1.695 BE*Stringency 0.633 1.022 0.255 CZ*Stringency -3.214*** -8.736*** -2.181 DE*Stringency 0.865 0.733 0.183 DK*Stringency -3.359*** -5.623*** -1.404 ES*Stringency -2.874*** 2.898* 0.723 FI*Stringency -3.342*** -30.816*** -7.692 FR*Stringency -2.398*** -3.839*** -0.958 UK*Stringency -1.117** -1.302* -0.324 IE*Stringency -1.970*** -1.169 -0.292 IT*Stringency -2.536*** 0.675 0.169 PT*Stringency -2.888*** -1.410 -0.352 SE*Stringency -2.645*** -4.699*** -1.173 logSigma 7.385*** 6.952*** 6.947*** 6.914*** Observations 116,298 116,298 116,298 116,298 116,298 R2 0.414 Adjusted R2 0.413 Log Likelihood -504,581.200 -460,423.900 -460,043.900 -933,320.300 -457,774.200 Akaike Inf. Crit. 1,009,166.000 920,939.900 920,183.900 1,866,837.000 915,744.400 Bayesian Inf. Crit. 1,009,186.000 921,384.400 920,647.800 1,867,784.000 916,691.500

Residual Std. Error 740.056 (df = 116201)

F Statistic 854.807

*** (df = 96;

116201)

(30)

4.1.2. Interpretation of effects

For interpreting effects of the Tobit model, we need to look at the marginal effects at the average observation, because this is representative of our population. The coefficients of the Tobit model are based on observations > 0. These coefficients are multiplied by the Inverse Mills ratio (probability of being > 0) to get to the marginal effects. The marginal effects are displayed in the last column of Table 6. We see that both stringency and ICU patients have a significant positive impact on daily sales. This confirms hypothesis 1a and 2a. The marginal effect for Stringency is greater (1.217) than that of ICU patients (0.252), but we have to keep in mind that the maximum value of stringency is 100, whereas the amount of ICU patients could mount up to hundreds or thousands.

Looking at the effects of the control variables, we see that from the days of the week variables Monday has the most positive effect on daily sales (64.844). Saturday has the most negative impact (-195.097). We also observe that Black Friday has a substantial positive impact on sales of 339.234. For the month variables we see that June has the greatest positive impact on daily sales (243.782). As for the effects product categories, we see that the effect is most positive for Home & Interior (88.135) and the most negative for Office supplies (-654.562). This seems logical as most sales were generated by the Home & Interior category and the least Office supplies, as is stated in section 3.4.2. For country effects we see that all are negative, this is logical as the reference level is the Netherlands and most sales are generated from this country. The effect of Finland was the most negative (-897.699).

In the table below, the marginal effect for stringency on sales is calculated by category. This is done by adding up the marginal effect of stringency and the marginal effect of the interaction term for stringency and category x. For example, the marginal effect of stringency is 1.217. The marginal effect for stringency*Tools & DIY is -1.790. Adding these two up leads to a value of -0.573, which would be the marginal decrease in sales for category Tools & DIY for every increase in a value of stringency. This same procedure is applied to stringency and country in Table 8. As well as ICU patients and category, and finally to ICU patients and country in Table 9. And Table 10. If an interaction term was not significant, the overall effect will be the same as the direct effect.

Table 7. Marginal effects for stringency and interactions between stringency and product categories

Category Direct effect of

stringency

Interaction effect between Stringency and Category i

Total effect (sum of direct effect and interaction effect)

Direction and magnitude of total effect on sales

Toys & Hobby 1.217 No interaction

(reference level)

1.217 Positive

Custom 1.217 Not significant 1.217 Positive

Tools & DIY 1.217 -1.790 -0.573 Negative

Electronics 1.217 -1.361 -0.144 Slightly negative

Photography & Optics 1.217 -1.494 -0.277 Negative

(31)

Office supplies 1.217 -3.702 -2.485 Very negative

Beauty & Lifestyle 1.217 -1.543 -0.326 Negative

Outdoor 1.217 -1.119 -0.098 Slightly negative

Sports 1.217 -1.094 0.123 Slightly positive

Garden 1.217 3.005 4.222 Very positive

Sports cycling 1.217 1.753 2.970 Very positive

Home & Interior 1.217 -1.468 -0.251 Negative

From the results we can confirm hypothesis 1b, the effect of stringency of measures on sales is moderated by product category. The effect is very positive for the product category Garden and Sports cycling, while the effect is negative for categories such as Office supplies and Tools & DIY. This is visualized in Figure 9. A stringency score of 70 would lead to an increase of 295.54 in daily sales for product Category garden, holding all other variables constant. For office supplies on the other hand, this will lead to a decrease of 173.95, holding all other

variables constant. The intercept of the categories in the figures below is calculated by adding up the intercept and the categorical effect. For example the intercept of Garden at a stringency score of 0 would be adding up the constant of the model and the categorical effect of garden. This is 2,519.44 - 13.107 = 2506,34.

Figure 9. The moderating effect of category on stringency

Table 8. Marginal effects for stringency and interactions between stringency and countries

Country Direct effect of

stringency

Interaction effect between Stringency and Country i

Total effect (sum of direct effect and interaction effect) Direction and magnitude of total effect on sales Netherlands 1.217 No interaction (reference level) 1.217 Positive Austria 1.217 -1.695 -0.478 Negative

(32)

Czech Republic 1.217 -2.181 -0.964 Negative

Germany 1.217 Not significant 1.217 Positive

Denmark 1.217 -1.404 -0.187 Negative

Estonia 1.217 0.723 1.940 Very positive

Finland 1.217 -7.692 -6.475 Very negative

France 1.217 -0.958 0.259 Positive

United Kingdom 1.217 -0.324 0.893 Positive

Ireland 1.217 Not significant 1.217 Positive

Italy 1.217 Not significant 1.217 Positive

Portugal 1.217 Not significant 1.217 Positive

Sweden 1.217 -1.173 0.044 Slightly positive

Based on the results, we can confirm hypothesis 2b, the effect of stringency of measures on sales is moderated by country. We observe that the effect is negative for countries such as Austria, Czech Republic, Denmark and very negative for Finland. Other countries show positive effects of stringency on sales, with Estonia having the strongest effect. The moderated effect of country is visualized in Figure 10.

Figure 10. The moderating effect of country on stringency

Table 9. Marginal effects for ICU patients and interactions between ICU patients and product categories

Category Direct effect of

stringency

Interaction effect between Stringency and Category i

Total effect (sum of direct effect and interaction effect)

Direction and magnitude of total effect on sales

Toys & Hobby 0.252 No interaction

(reference level)

0.252 Very positive

Custom 0.252 -0.0848 0.1672 Positive

(33)

Electronics 0.252 -0.151 0.101 Positive

Photography & Optics 0.252 -0.095 0.157 Positive

Household 0.252 -0.094 0.158 Positive

Office supplies 0.252 -0.076 0.176 Positive

Beauty & Lifestyle 0.252 -0.112 0.140 Positive

Outdoor 0.252 -0.131 0.121 Positive

Sports 0.252 -0.069 0.183 Positive

Garden 0.252 -0.072 0.180 Positive

Sports cycling 0.252 -0.079 0.173 Positive

Home & Interior 0.252 -0.090 0.162 Positive

From these results we can confirm hypothesis 2b, the effect of stringency of measures on sales is moderated by country. This is visualized in Figure 9. Although the direction of the effect of ICU patients is consistent across countries (positive effect in all countries) the magnitude of the effect differs per country. The results indicate that Toys & Hobby are impacted the strongest by the amount of ICU patients.

Figure 11. The moderating effect of product category on ICU patients

Table 10. Marginal effects for ICU patients and interactions between ICU patients and countries

Country Direct effect of ICU

patients

Interaction effect between ICU patients and Country i

Total effect (sum of direct effect and interaction effect) Direction and magnitude of total effect on sales Netherlands 0.252 No interaction (reference level) 0.252 Positive

Austria 0.252 0.780 1.032 Very positive

(34)

Czech Republic 0.252 Not significant 0.252 Positive

Germany 0.252 -0.147 0.105 Slightly positive

Denmark 0.252 0.337 0.589 Positive

Estonia 0.252 Not significant 0.252 Positive

Finland 0.252 Not significant 0.252 Positive

France 0.252 -0.126 0.126 Slightly positive

United Kingdom 0.252 -0.125 0.127 Slightly positive

Ireland 0.252 1.247 1.499 Very positive

Italy 0.252 -0.155 0.097 Slightly positive

Portugal 0.252 Not significant 0.252 Positive

Sweden 0.252 -0.129 0.123 Slightly positive

We observe similar results for country as a moderating variable as we did with product category. The effect remains consistent across countries, but the effect is much stronger in some countries than others. The positive effect of ICU patients on sales is very strong for Ireland and Austria, with a marginal increase of more than 1 in sales for every ICU patient. In other countries the effect is sometimes 10 to 15 times smaller. This confirms hypothesis 2c, the effect of ICU patients on sales is moderated by country. This is visualized in Figure 12.

Figure 12. The moderating effect of country on ICU patients

4.2. Probability of product return models

(35)

the full model not provide substantially more performance for the model. These terms are penalized heavier in the BIC evaluation metric. These models were, like the sales models, estimated with a reference level. The reference level that was used in the models is presented in Table 10. When we say full model, we refer to the second model that does include interaction effects. The first model includes just the direct effects, we refer to this as the nested model.

Table 10. Reference level for sales models

Category Country Day Month Year Black

Friday

Marketplace Payment Method

Toys & Hobby NL Wednesday April 2020 No No iDEAL

The results of these models is presented in Table 11. There are multiple ways to assess the impact of the independent variables on the dependent variable. A positive and significant β estimate for the independent variable leads to an increase in the probability of observing Y = 1. In our case Y = 1 stands for a product being returned. A negative and significant β estimate leads to a decrease of the probability of observing Y = 1, which means a decrease in the probability of a product being returned. The estimated intercept of the reference level is similar for both models. By taking the exponent of estimated β’s we can map the estimated (change in)

probability. This shows that the probability of return for our reference level is 2.59% in the first model and 2.58% in the second model. Overall, we see very similar findings across both models.

Table 11. Logistic regression results for modelling probability of return

Dependent variable: Probability of return Logistic regression with control variables and

direct effects

Logistic regression with all variables

Variable β coefficient Exponent of β coefficient β coefficient Exponent of β

coefficient

Constant -3.655*** 0.0259 -3.657*** 0.0258

Stringency -0.002*** 0.998 -0.001 0.999

ICU patients -0.00003** 0.999 -0.0001* 0.999

Product price 0.007*** 1.007 0.007*** 1.007

Product price hat -0.00001*** 0.999 -0.00001*** 0.999

Day of week Friday 0.037** 1.038 0.036** 1.037

Day of week Monday -0.003 0.997 -0.003 0.998

Day of week Saturday 0.0001 1.000 -0.0001 0.999

Day of week Sunday -0.001 0.999 -0.001 0.999

Day of week Thursday -0.006 0.994 -0.006 0.994

Day of week Tuesday 0.026 1.027 0.027 1.027

Black Friday -0.192** 0.825 -0.196*** 0.822

Month September 0.135*** 1.144 0.083*** 1.087

Month November -0.049* 0.951 -0.096*** 0.908

Month December -0.120*** 0.887 -0.166*** 0.847

(36)

Month May 0.036 1.037 -0.017 0.983 Month March -0.141*** 0.869 -0.161*** 0.851 Month June 0.020 1.021 -0.036 0.964 Month July 0.021 1.022 -0.029 0.971 Month January -0.099*** 0.906 -0.133*** 0.876 Month February -0.068** 0.935 -0.097*** 0.908 Month August 0.064*** 1.066 0.011 1.011 Year 2019 -0.096*** 0.908 -0.069*** 0.933 Category Custom 0.479*** 1.615 0.209** 1.233

Category Tools & DIY 0.717*** 2.048 0.694*** 2.002

Category Electronics 0.656*** 1.927 0.921*** 2.511

Category Photography & Optics 0.490*** 1.632 0.475*** 1.608

Category Household 0.321*** 1.379 0.345*** 1.412

Category Office supplies 0.059 1.061 -0.503* 0.605

Category Beauty & Lifestyle 0.508*** 1.662 0.742*** 2.099

Category Outdoor 1.130*** 3.095 1.158*** 3.184

Category Garden 0.199*** 1.220 0.139*** 1.149

Category Sports 0.882*** 2.417 0.993*** 2.699

Category Sports cycling 0.923*** 2.517 0.922*** 2.513

Category Home & Interior 0.743*** 2.103 0.782*** 2.186

Country AT -0.203*** 0.817 -0.156*** 0.856 Country BE -0.564*** 0.569 -0.504*** 0.604 Country CZ -0.751*** 0.472 -0.669*** 0.512 Country DE 0.087*** 1.091 0.059*** 1.061 Country DK -0.366*** 0.694 -0.410*** 0.663 Country ES -0.344*** 0.709 -0.393*** 0.675 Country FI -0.779*** 0.459 -0.825*** 0.438 Country FR -0.260*** 0.771 -0.270*** 0.763 Country UK -0.302*** 0.740 -0.162*** 0.850 Country IE -0.836*** 0.433 -0.846*** 0.429 Country IT -0.413*** 0.662 -0.698*** 0.498 Country PT -0.664*** 0.515 -1.166*** 0.312 Country SE -0.828*** 0.437 -0.877*** 0.417

Payment Method Afterpay 0.990*** 2.693 0.975*** 2.651

Payment Method American Express 0.121 1.129 0.123 1.131

Payment Method Bank Transfer 0.033 1.034 0.026 1.026

Payment Method Diners -1.223 0.294 -1.253 0.286

Payment Method EPS -1.796* 0.166 -1.804* 0.165

Payment Method Giropay -0.287 0.751 -0.280 0.759

Payment Method Klarna 0.958*** 2.605 0.956*** 2.602

Payment Method Maestro 0.470** 1.601 -0.451** 1.571

Payment Method Mastercard 0.262*** 1.299 0.253*** 1.287

Payment Method Mr. Cash 0.163*** 1.177 0.150*** 1.161

Payment Method Invoice -0.771*** 0.463 -0.778*** 0.459

Payment Method PayPal 0.256*** 1.292 0.255*** 1.291

Payment Method Sofort 0.121** 1.128 0.118** 1.126

Payment Method Trustly 0.396 1.486 0.247 1.226

Payment Method Visa 0.186*** 1.204 0.177*** 1.194

Payment Method Carte Bleue -0.223 0.800 -0.205 0.814

Payment Method Dexia 0.252 1.287 0.172 1.187

Payment Method KBC -0.514*** 0.598 -0.565*** 0.569

Is Marketplace 0.558*** 1.748 0.675*** 1.963

Marketplace*Custom 0.163 1.176

(37)

Marketplace*Electronics -0.726*** 0.484

Marketplace*Photography & Optics -0.030 0.971

Marketplace*Household -0.188*** 0.829

Marketplace*Office supplies 1.054*** 2.868

Marketplace*Beauty & Lifestyle -0.655*** 0.519

Marketplace*Outdoor -0.067 0.935

Marketplace*Sports -0.084 0.919

Marketplace*Garden -0.008 0.992

Marketplace*Sports cycling -0.143*** 0.867

Marketplace*Home & Interior -0.105** 0.901

Custom*ICU patients -0.0001 0.999

Tools & DIY*ICU patients 0.00004 1.000

Electronics*ICU patients -0.0004 0.999

Photography & Optics*ICU patients 0.0001* 1.000

Household*ICU patients -0.0002*** 0.999

Office supplies*ICU patients -0.007 0.993

Beauty & Lifestyle*ICU patients -0.0003* 0.999

Outdoor*ICU patients -0.0001** 0.999

Sports*ICU patients -0.0001** 0.999

Garden*ICU patients -0.00004 0.999

Sports cycling*ICU patients -0.00004 0.999

Home & Interior*ICU patients -0.00002 0.999

Custom*Stringency 0.003** 1.003

Tools & DIY*Stringency -0.003* 0.996

Electronics*Stringency 0.005** 1.005

Photography & Optics*Stringency 0.0005 1.000

Household*Stringency 0.002** 1.002

Office supplies*Stringency 0.014 1.014

Beauty & Lifestyle*Stringency -0.0001 0.999

Outdoor*Stringency -0.001 0.999

Sports*Stringency -0.004*** 0.996

Garden*Stringency 0.002** 1.002

Sports cycling*Stringency 0.001 1.001

Home & Interior*Stringency -0.001 0.998

Referenties

GERELATEERDE DOCUMENTEN

Therefore probably the Covid-19 pandemic represents one of the most significant environmental changes in the modern marketing history, which could potentially have a profound impact

Afgelopen zijn er wel wat meer toeristen geweest dus we willen die eigenlijk vasthouden zorgen dat die volgend jaar weer komen dus we gaan kijken of we de toeristen die dit jaar

Economy private labels have been advanced as a potentially effective defensive tool for traditional grocery retailers against the rise of hard discounters like ALDI

Customers purchasing with competitors as an effect of media contact with Company X do not increase secondary demand and thus the competitor demand increases as a result of a decrease

Since virtue characteristics (i.e. healthy and wholesome) are ascribed to sustainability labels, they can reduce guilt by making the negative consequences of vice products

H1a: The exposure to offline (i.e. print, radio, television and folder) - and online advertisement (i.e. banner advertisement) has a positive effect on sales in general... H1b:

[r]

Multiple studies show that older adults engage in various self-regulation strategies aimed at continuously maintaining or restoring person- environment fit (e.g., Kooij et al., 2020