Does religious priming make people tip more generously?

(1)

Does religious priming make

people tip more generously?

Case study of New York City’s taxi industry

Master’s thesis

University of Groningen

Faculty of Economics and Business

Author: Leonie Stuve s3409112 Date: June 5th 2020 Supervisors: G.J. Romensen A.L. Schippers Course code: EBM877A20 Abstract

This paper estimates the effect of religious priming on tipping behaviour, where attending a religious service is used as priming instrument. The study is situated in Manhattan and uses a unique combination of TLC taxi data and a manually constructed dataset on places of worship. By estimating difference-in-differences analyses, the research finds no evidence of attend-ing a religious service on taxi riders’ tippattend-ing behaviour for Protestantism, Catholicism, and Judaism. The paper discusses potential causes of this result and suggests ideas for future research.

(2)

i

List of Figures

4.1 Places of worship included in the analysis . . . 12

4.2 Difference-in-Difference estimation. . . 13

5.1 Division of observations across religions . . . 16

5.2 Average tipping percentages by religion . . . 18

B.1 Boxplot tipping percentage of the control and treatment group . . . 39

B.2 Kernel density of trip time before the (counterfactual) service . . . 40

(5)

iv

List of Tables

4.1 Treatment and control days . . . 13

4.2 Duration services . . . 14

5.1 Descriptive statistics cleaned taxi dataset . . . 15

5.2 Descriptive statistics combined dataset . . . 17

6.1 Descriptive statistics of the treatment and control group prior to the (counterfactual) service . . . 20

6.2 Results difference-in-difference analysis . . . 21

6.3 Results difference-in-differences analysis for Judaism . . . 22

6.4 Results difference-in-differences analysis for Catholicism . . . 23

6.5 Results difference-in-differences analysis for Protestantism . . . 24

6.6 Results difference-in-differences analysis with a radius of 30 meters . . 25

6.7 Results difference-in-differences analysis for Buddhism with a time range of 10 minutes before and after the service . . . 27

A.1 Christianity: Excluded dates. . . 36

A.2 Judaism: Excluded dates. . . 37

A.3 Budddhism: Excluded dates. . . 37

A.4 TLC taxi data cleaning process . . . 38

C.1 Results difference-in-differences analysis for Catholicism with a time range of 15 minutes before and after the service . . . 41

C.2 Results difference-in-differences analysis for Catholicism with a time range of 10 minutes before and after the service . . . 42

C.3 Results difference-in-differences analysis for Protestantism with a time range of 15 minutes before and after the service . . . 43

C.4 Results difference-in-differences analysis for Protestantism with a time range of 10 minutes before and after the service . . . 44

C.5 Results difference-in-differences analysis for Judaism with service time of 2 hours on Friday evening . . . 45

C.6 Results difference-in-differences analysis for Judaism with a time range of 15 minutes before and after the service . . . 46

C.7 Results difference-in-differences analysis for Judaism with a time range of 10 minutes before and after the service . . . 47

(6)

1 Introduction

Religious values are related to topics such as morality, love, respect, and kindness. One may therefore be inclined to think that religion fosters morality, loyalty and generosity. Yet, existing research shows there is a discrepancy between self-reported prosociality and measured prosociality in lab experiments (Shariff, 2015). Studies on the connection between religion and economics extend the theory by adding "nonmarket", taking beliefs and group interactions into account (Iannaccone, 1998).

However, measuring religiosity and demonstrating a causal impact is inherently difficult due to the possibility of reverse causality. Religion is correlated with un-observed characteristics and people choose their religion (Bryan, Choi, & Karlan, 2018), which makes religion endogenous. Benjamin, Choi, and Fisher (2016) em-phasise that effects of economic behaviour on religion could be due to correlation rather than causation. Iannaccone (1992, p.1475) even states that "nothing short of a (probably unattainable) "genuine experiment" will suffice to demonstrate religion’s causal impact".

To address this concern, recent research focusses on religious priming as opposed to religion. In religious priming, subjects are exposed to religious affiliations in a controlled setting (Shariff, Willard, Andersen, & Norenzayan, 2016). This allows researchers to measure the effect of religious priming on economic outcomes. Ben-jamin et al. (2016) find differences between Protestants and Catholics. Protestants tend to give more to the public good while Catholics decrease their contribution after being exposed to religious affiliations. The authors find no effect on generosity for both Protestants and Catholics. While Benjamin et al. (2016) use a lab setting, Bryan et al. (2018) use an exogenous change in religion which generates religiosity. Subjects are randomly chosen and invited to attend weekly Christian values and theology training. The authors conclude that the programme increased income, while there was no effect found for life satisfaction, food security, consumption, and labour supply.

The current paper analyses the effect of religion in a natural field setting using a quasi-experimental design. The study attempts to measure the effect of attending a religious service on taxi riders’ tipping behaviour in New York City. Tipping is voluntary because it takes place after enjoying the taxi ride or after dining at a restaurant (Lynn & Grassman, 1990). Naturally, there can be strategic motives to tipping. However, tipping in the taxi industry is generally a non-repeated interaction because customers cannot choose their taxi drivers (Ge, 2018). Therefore, in the current setting, leaving a tip is viewed as a generous action from the individual.

(7)

times. The combination of these two datasets allow for answering unique questions on the relation between religion and economic outcomes, using a natural field setting rather than a lab setting (Benjamin et al., 2016) or imposed religiosity (Bryan et al., 2018).

Following Ge (2018), a difference-in-differences analysis is performed. Ge (2018) estimates the effect of emotional cues after a New York Knicks game on taxi riders’ tipping behaviour. By performing a difference-in-differences analysis, using attending a religious service as quasi-experiment, it is possible to estimate the effect of religious priming on tipping behaviour.

The results from the difference-in-differences analyses do not show an effect of religion on tipping behaviour for Protestants, Catholics, and Jews. This is line with the work of Benjamin et al. (2016), where the authors indicate insignificant effects on generosity for Catholics and Protestants. Other literature suggests that religion is associated to increased within-group prosocial behaviour (Norenzayan & Shariff, 2008; Orbell, Goldman, Mulford, & Dawes, 1992; Pichon, Boccato, & Saroglou, 2007), where Orbell et al. (1992) and Norenzayan and Shariff (2008) find that individuals act less prosocial towards people outside of their own group. The absence of a treatment effect can be consolidated with this hypothesis to the extent that attending a religious service does not cause individuals to behave more generous towards taxi drivers.

This research contributes to the literature as it connects religious priming to tip-ping behaviour, where a quasi-experimental design is used to measure the effect of religious priming. The study estimates the effect of attending a religious service on tipping behaviour for Protestantism, Catholicism, and Judaism. Various studies use a meta-analysis (Ahmed & Salas, 2011; Shariff & Norenzayan, 2007), while others primarily analyse Christianity and its denominations (Benjamin et al., 2016; Bryan et al., 2018; Orbell et al., 1992). Benjamin et al. (2016) also run an experiment for Ju-daism, but the priming instrument does not cause strong religious salience in Jewish individuals. The current paper adds to the literature on both tipping behaviour and religious priming as it studies whether beliefs affect prosocial behaviour. Although the study concerns small donations, measured as tipping behaviour, it is a step to-wards understanding prosocial behaviour which can be used by charity organisations to increase donations.

The remainder of the paper is structured as follows. The first section reviews related literature. The second section elaborates on the research objectives. The third section discusses the research design. The fourth section introduces the dataset. The fifth section presents the results and an exploratory analysis. The sixth section discusses the results in relation to the literature. The last section concludes.

2 Related literature

(8)

2.1 Religion

McKay and Whitehouse (2015) discuss the definition of prosociality, which will be used in this paper. They argue that prosociality are "voluntary behaviors which benefit others at a personal cost" (McKay & Whitehouse, 2015, p.450).1This section briefly discusses prosocial behaviour and altruism in relation to religion.

Using a theory of helping, Jackson and Esses (1997) conclude that religious indi-viduals want to help others who threaten their values (e.g. homosexuality). These "value-threatening" individuals do not always want their help. As a result, religion is associated with higher within-group prosocial behaviour but can be discriminating against individuals outside their group. This means that religion decreases prosocial behaviour. In contrast, Batson (1983) argues that humans have a natural tendency to show compassion for their blood relatives only. He proposes that religion connects people by using language that causes people to show compassion for individuals in the religious circle, which is a far larger group of people than their blood relatives. Therefore, religion increases, rather than decreases, prosocial behaviour.

Furthermore, Allport (1966) distinguishes two orientations for religion: extrinsic orientation and intrinsic orientation. Extrinsic orientation indicates that religion is not incorporated in the individuals’ way of life, but used to gain social standing and ap-proval for their behaviour. In contrast, intrinsic orientation is aimed at the "unification of being, takes seriously the commandment of brotherhood, and strives to transcend all self-centered needs" (Allport, 1966, p.455). There is a third religious orientation: religion as an open-ended quest (Batson et al., 1989). This type of orientation is more responsive to the desires of the person in need. After conducting experiments, Batson et al. (1989) conclude that both the extrinsic and intrinsic orientation have egoistic motives. Individuals with an extrinsic orientation act prosocial to avoid punishments, while individuals with an intrinsic orientation behave in this manner to obtain social or self-rewards for doing the right thing. The latter conclusion is in contrast with the definition of intrinsic motivation as proposed by Allport (1966). According to Batson et al. (1989), the quest orientation to religion is the only orientation without egoistic motives. The work of Batson et al. (1989) relies on correlations between prosocial behaviour and religious orientation. Andreoni (2006) adds to egoistic motives by discussing the warm glow theory. The author explains that humans are inherently moral, they gain positive utility from helping others and they enjoy doing what is right. By undertaking such activities, humans are relieved from guilt. Gaining posi-tive utility, or avoiding negaposi-tive utility, is called a "warm glow". This is an example of altruism with an egoistic motive, which is also called "impure altruism" (Andreoni, 2006).

The discussion above on the relation between religion, prosocial behaviour and altruism provides a starting point for the question whether religion affects generosity. This question is difficult to analyse because there are many unobserved factors affecting religion, which makes it difficult to infer causality (Benjamin et al., 2016; Bryan et al., 2018). Therefore, this study focusses on religious priming rather than religion. Recent literature researches the effect of religious priming as this is more quantifiable compared to estimating the effect of religion on economic behaviour.

1_{A caveat in this definition is that there exist behaviours which benefit a few people but can harm a}

(9)

2.1.1 Religious priming

Religious priming is defined as the presentation of religious stimuli temporarily affecting a person’s response. The individual can be aware of the stimulus, provided that he is "unaware of the effect of the stimulus on the measured response" (Shariff et al., 2016, p.28). Using religious priming as an indicator of religion provides researchers with the possibility to measure its impact on prosocial behaviour because it is measured in a controlled environment. Benjamin et al. (2016) perform laboratory experiments to assess whether religion affects economic behaviour. The authors measure the subjects’ choices when a religious identity is made salient compared to when it is not. They find that Protestants increase their contribution to public goods, while Catholics decrease their contribution. In the dictator game, they show that Christianity has no effect on generosity. This conclusion contradicts the findings of Shariff and Norenzayan (2007), who suggest that religious priming leads their participants to behave more generous and prosocial. The result for atheists remains inconclusive, possibly due to the fact that they do not believe in gods, and therefore, not in a higher power watching over them to enforce religious norms. It is interesting that the results of Benjamin et al. (2016) differ from results found by Shariff and Norenzayan (2007) because they use the same priming instrument. These studies use a sentence unscrambling task as priming instrument. Subjects receive five words, they have to drop one word and form a sentence out of the remaining four words. Benjamin et al. (2016) estimate the effect for different religions whereas Shariff and Norenzayan (2007) perform a meta-analysis. Benjamin et al. (2016) run a regression on the full sample, which includes Judaism and atheists, and find a very small, positive but insignificant effect. The positive effect is in line with Shariff and Norenzayan (2007), but the effect is insignificant.

Horton, Rand, and Zeckhauser (2011) conclude that religious priming positively influences cooperation in the prisoner’s dilemma among believers only, whereas Ahmed and Salas (2011) find a positive, significant effect on prosocial behaviour for atheists, as well as a positive effect for Christians.

Similarly, Pichon et al. (2007) detect stronger prosocial behaviour when, predom-inantly Catholic, participants were primed with positive religious affiliations. The authors note that one cannot conclude that all religious priming leads to increased prosocial behaviour because negative religious images are associated with negative behaviours (e.g. authoritarian education styles in the family).

Orbell et al. (1992) and Norenzayan and Shariff (2008) extend the discussion on prosocial behaviour by considering group dynamics. Both studies conclude that religious individuals behave more prosocial towards members of the same faith. Orbell et al. (1992) run an experiment in Utah and Oregon. They find a positive correlation between increased cooperative behaviour and church attendance among mormons in Utah. However, this is a very specific effect and cannot be confirmed for mormons in Oregon and neither for non-mormons in both locations. These results concern correlations so there is not necessarily a causal effect.

(10)

2.1.2 Religious values

Not all religions share the same beliefs, which can cause differences in how people interact with one another. This study compares tipping behaviour across different religions, where the main focus is on Judaism and Christianity. In Christianity, the denominations Protestantism and Catholicism are distinguished. As Judaism, Protestantism and Catholicism have many of the same sacred texts, these religions share many similar values (Benjamin et al., 2016). This section elaborates on aspects where these religions differ and how this can affect tipping behaviour.

Supernatural punishment

Johnson and Krüger (2004) study why individuals cooperate when they are not related by blood and when there are almost no reputation effects. They argue that people continue to cooperate due to the fear of supernatural punishment, in the present life or the afterlife. It is a powerful enforcement mechanism, irrespective of whether the threat is real. The mechanism is highly efficient because free-riders are always caught and punished (Johnson & Krüger, 2004). Adding to this argument, Protestants have a higher costs of defaulting on contracts than Catholics (Blum & Dudley, 2001). It is argued that Catholics can ask for forgiveness from the priest for their sins, whereas Protestants do not follow the same tradition. Protestants also have a stronger belief in life after death (Bekkers & Wiepking, 2011b) which strengthens the threat of supernatural punishment. Therefore, Protestants have a stronger incentive to act prosocial.

Charitable giving

Charitable giving relates to tipping behaviour as it is also a measure of generosity. Monsma (2007) argues that individuals, in whom religion is salient, contribute more to religious causes than less religious individuals. It is acknowledged that the re-lationship for secular causes is mixed and more research is needed in that regard. Further, Bekkers and Wiepking (2011b) find that Protestants are more likely to donate to charity and show more self-reported prosocial behaviour because they believe themselves to be more religious compared to Catholics. This is measured on beliefs such as life after death, God, and predestination. Moreover, Protestants believe less strongly in free will compared to Catholics. Overall, people can have different reasons to donate money such as upholding their values, altruistic motives or increasing their reputation (Bekkers & Wiepking, 2011a).

Emphasis on the collective

(11)

2.2 Tipping behaviour

Tipping behaviour is interpreted as a measure of generosity because an individual voluntarily leaves additional money after receiving the service.

Azar (2010) distinguishes two main reasons for tipping: strategic motivations and psychological motivations. First, tipping with a strategic motive is aimed at ensuring good service when the customer comes back. Second, tipping with a psychological motive involves considerations such as reducing negative feelings or increasing positive feelings. Negative feelings can arise from not following social norms or when not tipping is perceived as unfair, which results in feelings such as guilt and embarrassment. Positive feelings stem from impressing others and improving the agent’s self-image in terms of kindness and generosity. These feelings deliver positive utility to the agent.

In addition to strategic and psychological motivations, tipping can be influenced by the mood of individuals (Cunningham, 1979; Ge, 2018). As clear as these moti-vations seem, Nisbett and Wilson (1977) argue that people are not always able to identify why they make certain decisions. Consequently, it is interesting to evaluate literature on factors affecting tipping behaviour, which will be applied to the taxi industry in New York City.

The remainder of this section first elaborates strategic motives for tipping be-haviour. This is followed by a discussion on the equity theory. Third, the section examines the relevance of social norms in tipping behaviour. Fourth, the section reviews the importance of group size. Lastly, it concludes by discussing mood effects.

2.2.1 Strategic behaviour

Lynn and Grassman (1990) do not find evidence in favour of tipping with a strategic motive. Similarly, Azar (2010) concludes that individuals leave tips with psychologi-cal motives rather than strategic motives for future interactions. Both of these studies concern tipping behaviour in restaurants, where the authors note that individuals do not anticipate to attend the same restaurant again. In the context of the taxi market, repeated interactions are also unlikely (Ge, 2018). Therefore, it is implausible that people tip the taxi driver guided by a strategic motive.

2.2.2 Equity theory

Equity theory proposes that individuals feel distressed when participants in a rela-tionship do not provide the same inputs (Lynn & McCall, 2000). At first glance, it seems similar to tipping with a strategic motive. However, the equity theory is aimed at providing the same inputs to maintain equity in a transaction rather than tipping to increase future service.

(12)

by engaging in conversation with passengers or choosing not to interact with the passenger if he feels the passenger is not interested. Therefore, according to the equity theory, it is possible that customers tip more in order to fairly compensate the taxi driver for providing a service of high quality.

2.2.3 Social norms

The social norm in the United States indicates to leave a tip of 10-20% (Ge, 2018), which means that it is likely that most tips will be in this range. Haggag and Paci (2014) use TLC trip data for 2009 to study the effect of default tipping options on tipping behaviour of taxi riders. They find that setting higher default tipping options leads to higher tips. However, there is a limit to this effect. If the default options are too high, tips are reduced. They argue that there are three mechanisms that can explain this phenomenon. First, customers may be rationally inattentive. Second, the default tip could signal a social norm to unfamiliar customers. Third, deviating from the status-quo, signalled by the default options, provides the individual with negative utility. These results confirm that people tip to adhere to the status-quo. Deviating from this norm can lead to social pressure and feelings of guilt and embarrassment. Lynn and Grassman (1990) interview guests as they leave a restaurant in order to gather data on tipping behaviour. The results suggest that tipping is consistent with the social norm in the US. The study is conducted in one restaurant and the results indicate correlations rather than a causal effect.

Azar (2004) also finds evidence that people tip to adhere to social norms for both the taxi industry and tipping at restaurants. He points out that if people do not derive further benefits, in addition to avoiding social disapproval, the social norm would disappear over time. Clearly, the norm has sustained so there have to be other benefits to leaving a tip. Azar (2004) suggests that these additional benefits could be in the form of increased self-image, in terms of kindness and generosity. These results indicate that the social norms among groups of (religious) individuals will most likely influence their tipping behaviour. Additionally, the urge to help people may be stronger in religious individuals. It should be noted that 38% of Manhattan’s population is estimated not to be affiliated with a religion, about 40% of the population is Christian and the division between Protestantism and Catholicism is approximately equal (Jones, 2016). Therefore, the control group, as well as the treatment group, includes religious individuals.2Hence, social norms are unlikely to lead to differences in tipping behaviour between these two groups.

2.2.4 Group size

The literature is divided on whether group size has an effect on tipping behaviour. Lynn and Grassman (1990) and Cunningham (1979) find no significant effect of group size on the tip. Freeman, Walker, Borden, and Latane (1975) find a negative effect, whereas Conlin, Lynn, and O’Donoghue (2003) find a positive effect.

The negative effect of group size on tipping is explained by the theory of "diffusion of responsibility" (Freeman et al., 1975). Freeman et al. (1975) argue when a group of people goes out to eat, they each feel less responsible for the check as the responsibility is mentally divided among all individuals present. It is predicted that people dining alone leave a larger tip than people dining in groups. The results are consistent with this theory. However, the authors note that their results are correlations which means

(13)

that there are multiple effects which could be driving the result. For example, people in groups could be less generous or get worse service than people dining alone.

In contrast to Freeman et al. (1975), Conlin et al. (2003) find a positive effect of group size on tipping after controlling for quality of service and bill size. The result is inconsistent with the "diffusion of responsibility" theory as people will leave a larger tip in groups when they have received good service. This is in accordance with the alternative explanation by Freeman et al. (1975) that people dining in groups receive worse service.

Overall, there is no clear expectation whether group size influences tipping be-haviour. Most of these results were identified using correlations. Therefore, it is ambiguous whether group size has a causal effect on tipping behaviour. As previ-ously discussed, dining in a restaurant is perceived to be a much more heterogeneous service compared to taking a taxi ride. The "diffusion of responsibility", as proposed by Freeman et al. (1975), is more likely to hold than the positive effect found by Conlin et al. (2003).

2.2.5 Positive emotional triggers

In addition to strategic behaviour, altruistic behaviour, social norms and group size, tipping can be influenced by emotional triggers (Ge, 2018). Cunningham (1979) was the first to connect weather variables to social behaviour in people. He argues that sunshine has a positive effect on people’s mood. The findings suggest that outdoor sunshine level is positively related to tipping at the restaurant. The author acknowl-edges that the study is of correlational nature and there is uncertainty surrounding the effects caused by sunshine. This uncertainty is confirmed by Flynn and Greenberg (2012), who do not find a significant effect of weather variables on tipping behaviour. In addition to weather variables, there are other ways to induce a good mood. Strohmetz, Rind, Fisher, and Lynn (2002) perform an experiment where a customer receives the unexpected gift of candy along with the bill in a restaurant. They find a positive effect of receiving candy on the tip. The authors acknowledge that this is most likely due to the norm of reciprocity than a good mood effect. Therefore, the proposed explanation is more consistent with the equity theory than an effect due to positive emotional triggers.

(14)

3 Research objectives

The literature discussed shed light on prosocial behaviour, religious priming, and tipping behaviour. As this study concerns the relation between religion and tip-ping behaviour, the focus is on one construct of prosocial behaviour in particular: generosity. The current paper attempts to answer the following research question:

"What is the effect of religious priming on tipping behaviour?"

This question is investigated using a case study of the taxi industry in New York City, in which the effect of religious priming on taxi riders’ tipping behaviour is assessed. In the current context, attending a religious service is viewed as religious priming. Most studies (Benjamin et al., 2016; Norenzayan & Shariff, 2008; Pichon et al., 2007) use lab experiments. Shariff (2015) points out that such experiments may fail to create genuine religious situations. The current paper does not perform an economic experiment in a lab setting. Rather, it uses a natural field setting by focussing on religious priming in religious services and assesses the effect on tipping behaviour. Therefore, it is highly likely that individuals are exposed to a genuine religious environment. The approach is quasi-experimental as subjects are not randomly assigned to the treatment group. During the religious services, individuals are reminded of their religious values which can make their religious identity more salient to them. Individuals do not adjust their behaviour as a result of being monitored because the taxi data do not follow individuals over time. These characteristics validate the use of religious services as religious priming.

Furthermore, following Conlin et al. (2003) and Ge (2018), the outcome variable is measured as tipping percentage.1 Using tipping percentage is appropriate because most people determine their tip by thinking in terms percentages of the check (Conlin et al., 2003). It is a measure which shows the tip in relation to the total amount paid. This accounts for the fact that a higher check will have higher absolute tips.

Literature reviewed in section 2 indicates an ambiguous effect of religious prim-ing on tippprim-ing behaviour. On the one hand, religion is associated with an increase in within-group prosocial behaviour, while individuals act less prosocial towards individuals outside their religious group (Batson, 1983; Jackson & Esses, 1997; Noren-zayan & Shariff, 2008; Orbell et al., 1992). On the other hand, religious beliefs regarding supernatural punishment and the emphasis on the collective can increase prosocial behaviour towards individuals both within and outside of the religious group. To conclude, the net effect of religious priming on tipping behaviour remains ambiguous. Using a case study of the taxi industry provides a way to examine the overall effect and perform subgroup analyses for Protestantism, Catholicism and Judaism.

1_{Tipping percentage is defined as follows: (tipping amount / total amount)*100.}

(15)

4 Research Design

In order to study the research objectives defined in the previous section, the field setting and research design are elaborated upon in this section. The study focusses exclusively on the yellow taxis in Manhattan in 2013.

4.1 Field setting

The field setting is intended to provide context in which yellow taxis operate and how the dataset on places of worship has been collected and constructed.

4.1.1 Taxi market

The release of the dataset on taxi rides sparked an interest from academics in tip-ping behaviour. The dataset contains detailed information on individual taxi rides on variables such as GPS coordinates where customers have been picked up and dropped off, amount of the tip, trip time, and trip distance. Various studies have been published on topics such as default tipping behaviour (Haggag & Paci, 2014), tipping behaviour related to emotional cues (Ge, 2018), productivity improvements in the taxi industry (Haggag, McManus, & Paci, 2017), and potential information leakage of the federal reserve measured by taxi rides between the federal reserve and commercial banks (Finer, 2018).

The Taxi and Limousine Commissions (TLC) regulates the taxi industry and for-hire vehicle industry in New York City (Taxi & Limousine Commission, n.d.). According to its regulations, a car must be certified with a medallion to be allowed to carry passengers. There were 14,437 medallions in 2014 (Taxi and Limousine Commission, 2016). Furthermore, Manhattan was the most important borough for the yellow taxis in 2014 because 90.3% of the taxi pick-ups were there (Taxi and Limousine Commission, 2016). In 2014, most of the drivers were from Bangladesh (23.1%) and Pakistan (13.2%), whereas only 5.9% of the drivers were born in the United States (Taxi and Limousine Commission, 2016).

In New York City, it is mandatory for taxis to have a Passenger Information System (PIM) (Haggag & Paci, 2014). In 2013, there were two companies providing the systems (Taxi and Limousine Commission, 2012). These monitors collect information on variables such as pick-up and drop-off locations, the amount paid, and the number of passengers in the car. The monitors are subject to default tipping behaviour. When a customer pays for the fare, the monitor gives tipping suggestions which differ across companies providing the systems (Haggag & Paci, 2014). Default tip suggestions can lead to higher tips when the defaults are set accordingly, for example 15%, 20%, or 30% (Haggag & Paci, 2014). An analysis of default tipping behaviour is beyond the scope of the current study.

(16)

charged for rides ending in New York City, rides during peak hours and rides in the evening (Ge, 2018).1

4.1.2 Places of worship

A dataset on places of worship and service times is needed to study the effects of religious priming on tipping behaviour. As such a dataset did not exist yet, the dataset had to be constructed manually.

The foundation of the dataset is a publicly available list compiled by Homeland Infrastructure Foundation-Level Data (2018). The list includes places of worship in the United States based on data from 2007-2009. Given this time frame, the list is accurate under the assumption that there are no new places of worship opened between 2009 and 2013. The dataset contains, among other variables, coordinates of places of worship, addresses, religion and denomination. More specifically, the following religions are included: Buddhism, Christianity, Islam, Judaism, and Hinduism. The list contains 384 places of worship in Manhattan.2 The service times of religious places were not included. Therefore, the service times had to be manually retrieved from websites of the respective organisations.3

After collecting all service times, the Islam and Hinduism were removed from the study. Hinduism only had two temples in the dataset which is insufficient to draw conclusions. The service times for the Islam were very similar on two consecutive days, which implies that it is not possible to construct a control group. Consequently, mosques have been removed from this study. Some places of worship were either closed or service times were unavailable, the resulting dataset contains 187 places of worship. Figure 4.1 shows the places of worship considered in the analysis. It is clear from the figure that the analysis concerns Manhattan and that the places of worship cover the entire borough. The vast majority of the places of worship are Christian churches. There are 147 Christian places of worship, which can be divided in Catholic churches (78) and Protestant churches (69). Furthermore, there are 5 Buddhist temples and 35 synagogues in the dataset. Most places of worship have multiple services a day, which are all included in the analysis. As there are only a few Buddhist temples in the dataset, and there is little academic research on Buddhism available, these observations will be used for an exploratory analysis only. The main analysis therefore focusses on Judaism, Protestantism and Catholicism.

1_{A MTA tax of $0.50 is charged for all rides ending in New York City, a surcharge of $0.50 is charged}

when the ride takes place between 8 p.m and 6 a.m. When the ride is on a weekday from 4 p.m until 8 p.m an additional amount of $1.00 is charged (Ge, 2018). Lastly, there are surcharges, up to $3 for certain routes (Ge, 2018).

2_{The list does not include places of worship that are homes of religious leaders, religious schools or}

buildings that serve administrative purposes only (Homeland Infrastructure Foundation-Level Data, 2018).

(17)

FIGURE 4.1: Places of worship included in the analysis

Legend Protestant church Catholic church Synagogue Buddhist temple

4.2 Research setting

The research setting introduces the difference-in-differences method and the subgroup analyses. The section also discusses the design of the treatment and control groups.

4.2.1 Difference-in-differences methodology

Following Ge (2018), the paper uses a difference-in-differences method. The method is intended to measure the impact of an event on economic outcomes (Wooldridge, 2013, p.438). The present paper estimates the causal effect of attending a religious service on tipping behaviour. Hence, a difference-in-differences method is suitable in this case study. Ge (2018) compares tipping behaviour before and after a New York Knicks game using the previous day as control group. This paper uses the same method applied to religious services. A difference-in-differences analysis, as visualised in figure 4.2, compares trends in the treatment and control group. The treatment group attends the religious service and the control group does not. The difference between the expected outcome of the treatment group without intervention, which is based on the trend of the control group, and the actual outcome of the treatment group is the difference-in-differences estimator. It is visualised by the red line in figure 4.2. Therefore, important that the treatment and control group have similar trends (Angrist & Pischke, 2008, p.171).

(18)

Catholicism, and Judaism are carried out to investigate the treatment effect for each religion. The subgroup analysis for Buddhism is used for exploratory purposes only.

FIGURE4.2: Difference-in-Difference estimation. Figure adapted from Angrist and Pischke (2008).

4.2.2 Definition treatment and control group

The treatment and control days are defined in this section. The treatment day is the day of the service, whereas the control day is the day either before or after the service. Hence, this indicates a counterfactual service. Table 4.1 shows the treatment and control days.

For each religion, the most important day of the week is chosen as treatment day. In Christianity, Sunday is the most important day of worship. Saturday is chosen as control day because, compared to Monday, it is more similar in terms of traffic flows to the treatment day. For Judaism, Friday has been chosen as treatment day because it is, in addition to Saturday, the most important day of the week. As service schedules are more consistent over time on Fridays, this day has been chosen as treatment day. The control group is constructed based on taxi rides on Thursday because, compared to Saturday, this day is more similar to Friday.

TABLE4.1: Treatment and control days

Religion Treatment day Control day

Christianity Sunday Saturday

Judaism Friday Thursday

Buddhism4 _Thursday _Friday

When creating the dataset, it has been ensured that there are no services starting within half an hour on the control day compared to the treatment day. This is done to

4_{Buddhism is only used for exploratory analyses. Thursday has been chosen as treatment day as}

(19)

isolate the effect of attending a religious service on tipping behaviour. Furthermore, federal and religious holidays are excluded from the sample because service times vary year to year for these holidays so those data points are unreliable. Additionally, most of the services on religious holidays are not available yet as they are announced leading up to events such as Easter, Rosh Hashanah and Christmas. If the holiday was either on the treatment day or control day, both days are excluded. Tables A.1 -A.3 indicate which dates are excluded for each religion.

With respect to the service times, a service duration has to be assumed. These assumptions are shown in table 4.2, which are based on several forums online. The duration of a service can vary across religious institutions. For each service, the difference is taken between tipping percentage in taxi rides originating 0-30 minutes prior to the service and taxi rides originating 0-30 minutes after the service. This difference on the treatment day constitutes the first set of differences, while this difference on the control day represents the second set of differences.

TABLE4.2: Duration services

Religion5 Duration service

Christianity 60 minutes

Judaism 30 minutes

5 Data

The current paper utilises a combination of data sources: TLC taxi data and a manually constructed dataset on places of worship. This allows for the analysis of tipping behaviour in a unique setting. The previous section indicated the field setting and method, this section continues the discussion by elaborating on the process of cleaning the TLC taxi dataset and creating the places of worship dataset. Furthermore, it provides descriptive statistics of the TLC taxi dataset and the dataset on religion combined. The section concludes by discussing a preliminary analysis.

5.1 Cleaning TLC taxi dataset

The TLC taxi data is made publicly available through a Freedom of Informational Law (FOIL) request1_{. The data contains information on trip duration, number of}

passengers in the car, coordinates where passengers have been picked up and dropped off, payment method, fare amount, tolls, MTA taxation, surcharge, tip amount and total amount. The original dataset contains 173,179,759 observations for the year 2013. The current section elaborates on the cleaning process of the taxi data, which is summarised in table A.4.

The data is cleaned according to the approach of Ge (2018) and Haggag and Paci (2014). The following steps were taken; (1) all duplicate observations are dropped; (2) the variables pickup_datetime and dropoff_datetime are split in date and time

5_{The service duration for Buddhism is known, so there is no assumption necessary. The exploratory}

(20)

separately and then deleted in the interest of keeping the dataset as small as possible; (3) Trip distances of 0 or larger than 100 miles are dropped; (4) Trip durations of 0 or longer than 10800 seconds (3 hours) are dropped; (5) Observations where the tip was at least five times as large as the fare amount are dropped; (6) Payment types other than card are dropped; (7) MTA tax values of anything other than $0.50 are deleted; (8) Surcharges of larger than $3 are dropped; (9) Fare amounts that do not coincide with the base amount of $2.50 and increments of $0.50 are deleted as well.

In step 6, the sample is restricted to payments by card because tips are recorded for such transactions only. All transaction recorded as "cash", "no charge", "un-known", "dispute" or "voided trip" are removed from the dataset. This action re-moved 78,445,015 observations. Furthermore, steps 7 through 9 are taken because surcharges are not allowed to be larger than $3, MTA tax has to be $0.50 according to TLC regulations (Ge, 2018). The fare amount starts at $2.50 and increases by $0.50 for every one-fifth mile covered above 12mph or every 60 seconds in slow traffic (Taxi and Limousine Commission, n.d.). Data points that fail to meet these criteria are noise. After performing these operations, there are 92,512,258 observations left in the dataset. Table 5.1 shows the summary statistics for the cleaned dataset. The mean of tipping ratio (15.01%) is in line with the social norm to leave a tip of 10-20% (Ge, 2018).

TABLE5.1: Descriptive statistics cleaned taxi dataset

Variable N Mean Max Min S.D.

Tipping percentage2 92,512,258 15.0096 99.3846 0 5.0709 Number of passengers 92,512,258 1.6867 255 0 1.3753 Trip time in seconds 92,512,258 797.3093 10,787 1 558.3412 Trip distance in miles 92,512,258 3.1199 99.2000 0.01 3.4214 Tip amount ($) 92,512,258 2.4793 888.1900 0 2.2522

5.2 Combining taxi data with data on religion

The cleaned taxi dataset is combined with data on places of worship. Using ArcGIS, taxi rides originating within a 50 meter radius of places of worship are selected. The 50 meter radius is chosen to reduce the likelihood that there is noise in the data, i.e. taxis that do not carry passengers leaving a religious service. Additionally, some GPS coordinates are not accurately measured so there might be some noise in the range of 10-30 meters. Such deviations can occur when the entrance, or exit, is on the other side of the building than indicated by the GPS coordinates. When choosing a range of 50 metres, this variation is not a concern. After selecting the taxi rides based on location, stata was used to select the relevant taxi rides based on date and time of pick-up. These are taxi rides originating within 30 minutes of the start of a (counterfactual) service and taxi rides that leave from the place of worship within 30 minutes after the service ends. The result is a pooled cross section dataset, consisting

2_{Following Ge (2018), tipping percentage = (tip / total amount)*100.}

(21)

of 40,380 observations, which allows for the difference-in-differences analysis. The remainder of the section provides descriptive statistics and a preliminary analysis.

5.3 Descriptive statistics

The dataset consists of 40,380 observations, where Buddhism, Judaism, Protestantism and Catholicism are distinguished as religious. As previously noted, Buddhism will only be used for exploratory purposes, whereas the remaining religious are used for difference-in-differences analyses. This section provides additional insight into the dataset by presenting its descriptive statistics.

First, observations are not evenly distributed across religions. As shown in figure 5.1, Christianity accounts for 80% of the observations, of which about a third were assigned to a Protestant church and two-thirds to a Catholic church. About 15% of the taxi rides were near a synagogue and only 5% of the observations were assigned to a Buddhist temple.

FIGURE5.1: Division of observations across religions

(22)

TABLE5.2: Descriptive statistics combined dataset

Religion Variable N Mean Max Min S.D.

Total Tipping percentage 40,380 15.0884 85.1064 0 4.9114 Number of passengers 40,380 1.7409 6 1 1.4249 Trip time in seconds 40,380 696.2034 5,888 3 446.4214 Trip distance in miles 40,380 2.5598 48.4 0.02 2.7087 Tip amount ($) 40,380 2.1311 70 0 1.7765 Judaism Tipping percentage 5,906 14.9639 60.6061 0 4.4607 Number of passengers 5,906 1.1733 6 1 1.4292 Trip time in seconds 5,906 763.8744 4,800 45 527.7535 Trip distance in miles 5,906 2.2915 48.4 0.03 2.5082 Tip amount ($) 5,906 2.2177 30 0 1.6692 Protestantism Tipping percentage 10,506 15.2420 85.1064 0 4.9949 Number of passengers 10,506 1.7781 6 1 1.4565 Trip time in seconds 10,506 673.6551 5,168 3 449.5235 Trip distance in miles 10,506 2.6755 25.9 0.04 2.8091 Tip amount ($) 10,506 2.1469 37.95 0 1.8330 Catholicism Tipping percentage 21,757 15.0881 71.0660 0 4.8830 Number of passengers 21,757 1.7291 6 1 1.4082 Trip time in seconds 21,757 681.2465 5,888 40 451.8953 Trip distance in miles 21,757 2.5763 32.74 0.02 2.7122 Tip amount ($) 21,757 2.0882 70 0 1.7664

5.4 Preliminary analysis

The preliminary analysis is intended to give an indication of results based on the dataset. The difference-in-differences analyses compare the treatment and control groups both before and after the service. Figure 5.2 indicates average tipping percent-ages of the treatment and control group both before and after the service. The figure shows the tipping percentage for Catholicism, Protestantism and Judaism.3

First, figure 5.2 shows that individuals leaving close to a Catholic church tip less after a religious service while the control group tips more. Although this pattern can indicate a treatment effect, the effect is likely to be insignificant due to its small size.

3_{Tipping percentages are based on taxi rides leaving within a 50 meter radius of a place of worship.}

(23)

The decrease in tipping behaviour for the treatment group is only 0.06 percentage points, which is the same size as the increase in tipping behaviour for the control group. Second, tipping behaviour after a service decreases for Protestantism. This decrease is slightly larger for the treatment group compared to the control group. The decrease is 0.05 percentage points for the treatment group, whereas the decrease for the control group is only 0.01 percentage points. Compared to Catholicism, the effect for Protestantism is even more likely to be insignificant as both the treatment and control group exhibit decreased average tipping percentages after the service. Third, for Judaism, tipping percentages increased after a religious service for both the treatment and control group. Similar to Catholicism and Protestantism, the effect is small. As this observation holds for both the treatment and control group, this effect is presumably not due to attending a religious service. The econometric specification and estimated treatment effects are presented in the next section.

FIGURE5.2: Average tipping percentages by religion

Legend Treatment group before service Treatment group after service Control group before service Control group after service

6 Method and results

This section presents the econometric specification of the difference-in-differences analysis. It also shows the results of the full sample as well as the results of the subgroup analyses. It then discusses the robustness tests and the exploratory analysis of Buddhism. The section focusses on interpreting the results, but the interpretation of the findings in relation to the literature is postponed to the discussion.

6.1 Econometric specification

(24)

where Yijtrepresents the passenger’s tipping ratio for ride i, place of worship j and

time t; Tj is a dummy variable equal to 1 when the ride took place after the service

and 0 if the ride took place before the service; Grouptis a dummy variable equal to 1

when the taxi ride took place on the day of service and 0 if it was on the control day;

Xijtrepresents the following ride-level characteristics: ride distance, ride duration,

and the number of passengers. It also includes dummy variables for morning and afternoon, and dummy variables for each month; eijtis the error term.

The control variables ride distance, ride duration, and number of passengers are chosen because these factors can affect tipping behaviour. For example, if a ride takes longer, while keeping ride distance constant, this can be due to a traffic jam. The perceived quality of service can be lower, which decreases the tip. In order to account for such circumstances, ride-level characteristics are added as control variables. Similar to Ge (2018), dummy variables for time of day (morning, afternoon, evening) and dummies for months are added to control for possible time effects, for example time specific shocks if people tip more in a certain month.

In the difference-in-differences analysis, ˆβ3is of the coefficient of main interest

because it estimates the effect of attending a religious service on tipping behaviour. The estimated effect is the average treatment effect (Wooldridge, 2013, p.441) of attending a religious service on tipping percentage. As discussed in the research objectives, the net effect of religious priming is ambiguous prior to estimation. If ˆβ3

is positive and significant, attending a religious service increases tipping percentage, whereas a negative estimate of β3indicates that religious priming leads to a decrease

in tipping percentage.

(25)

TABLE6.1: Descriptive statistics of the treatment and control group prior to the (counterfactual) service

Group Variable N Mean Max Min S.D.

Treatment group Tipping percentage 8,557 15.0819 71.066 0 4.9527 Number of passengers 8,557 1.7530 6 1 1.4310 Trip time in seconds 8,557 674.3648 5,168 48 457.0583 Trip distance in miles 8,557 2.6262 34.61 0.04 2.7945 Control group Tipping percentage 9,634 15.0858 83.5410 0 4.9707 Number of passengers 9,634 1.7126 6 1 1.4049 Trip time in seconds 9,634 699.492 4,800 40 470.6202 Trip distance in miles 9,634 2.5343 32.74 0.03 2.6713

6.2 Results full sample

(26)

TABLE6.2: Results difference-in-difference analysis

Dependent variable Full sample

Tipping percentage 1 2 3 4 Dummy service 0.0485 0.0704 0.0719 0.0720 (0.0680) (0.0676) (0.0676) (0.0676) Group -0.0039 -0.0232 -0.0234 -0.0243 (0.0736) (0.0731) (0.0732) (0.0732) Treatment* -0.0849 -0.0976 -0.0969 -0.0974 after service (0.0985) (0.0976) (0.0976) (0.0976) Number of -0.0470*** -0.0468*** -0.0468*** passengers (0.0172) (0.0172) (0.0172) ln(triptime) -0.7629*** -0.7460*** -0.7543*** (0.0670) (0.0695) (0.0699) ln(distance) -0.2797*** -0.2937*** -0.2856*** (0.0581) (0.0600) (0.0603) Constant 15.0858*** 20.1800*** 20.0244*** 20.2971*** (0.0505) (0.4020) (0.4245) (0.4372) Observations 40,380 40,380 40,380 40,380 R-squared 0.0000 0.0196 0.0197 0.0204 Adjusted R-squared -0.0000 0.0195 0.0195 0.0199 Overall F-statistic 0.62 102.43 77.53 34.55

Dummies months No No No Yes

Dummies time of day No No Yes Yes

The dependent variable is tipping percentage. Group is defined as 1 for the treatment group and 0 for the control group. Dummy service is defined as 1 for time ranges after the service and 0 for time ranges before the service. Dummies for time of day are dummy variables indicating morning, afternoon and evening taxi rides. Robust standard errors are used. *, **, and *** indicate the significance levels 10%, 5% and 1%, respectively.

6.3 Results by religion

Benjamin et al. (2016) study the effects of religious priming on economic behaviour for atheists, Judaism, Protestantism, and Catholicism. They find no effect of religious priming on generosity for any of the religions. The current paper also distinguishes Judaism, Protestantism, and Catholicism. This section focusses on the results from the difference-in-differences analysis. Further connections to the literature are reviewed in the discussion.

6.3.1 Judaism

(27)

after the service, for both the treatment and control group. Moreover, this effect is positive in both the treatment and control group. Overall, it is likely that there is no effect of attending a Jewish service on taxi riders’ tipping behaviour.

TABLE6.3: Results difference-in-differences analysis for Judaism

Dependent variable Subsample: Judaism

Tipping percentage 1 2 3 4 Dummy service 0.0681 0.0527 0.0488 0.0467 (0.1739) (0.1732) (0.1728) (0.1728) Group -0.2707 -0.2864 -0.2884 -0.2822 (0.1877) (0.1865) (0.1864) (0.1865) Treatment* 0.0000 0.0287 0.0336 0.0438 after service (0.2496) (0.2483) (0.2481) (0.2484) Number of -0.1070** -0.1077** -0.1047** passengers (0.0441) (0.0441) (0.0439) ln(triptime) -0.6045*** -0.6076*** -0.6079*** (0.1620) (0.1618) (0.1625) ln(distance) -0.1059 -0.1048 -0.1123 (0.1469) (0.1468) (0.1471) Constant 15.0574*** 19.1947*** 19.2195*** 19.7245*** (0.1303) (0.9990) (0.9968) (1.0328) Observations 5,906 5,906 5,906 5,906 R-squared 0.0009 0.0115 0.0117 0.0139 Adjusted R-squared 0.0003 0.0105 0.0104 0.0109 Overall F-statistic 1.69 9.71 7.38 4.10

The dependent variable is tipping percentage. Group is defined as 1 for the treatment group and 0 for the control group. Dummy service is defined as 1 for time ranges after the service and 0 for time ranges before the service. Robust standard errors are used. *, **, and *** indicate the significance levels 10%, 5% and 1%, respectively.

6.3.2 Catholicism

(28)

TABLE6.4: Results difference-in-differences analysis for Catholicism Dependent variable Subsample: Catholicism

Tipping percentage 1 2 3 4 Dummy service 0.0505 0.0788 0.0796 0.0786 (0.0906) (0.0898) (0.0900) (0.0900) Group 0.0191 -0.0241 -0.0240 -0.0287 (0.0966) (0.0989) (0.0990) (0.0991) Treatment* -0.1114 -0.1105 -0.1118 -0.1110 after service (0.1136) (0.1322) (0.1322) (0.1322) Number of -0.0310 -0.0.0305 -0.0319 passengers (0.0233) (0.0233) (0.0233) ln(triptime) -0.8642*** -0.8652*** -0.8689*** (0.0931) (0.0954) (0.0960) ln(distance) -0.2602*** -0.2672*** -0.2529*** (0.0820) (0.0840) (0.0846) Constant 15.0748*** 20.7609*** 20.6633*** 20.9049*** (0.0671) (0.5560) (0.5750) (0.5907) Observations 21,5756 21,5756 21,5756 21,5756 R-squared 0.0000 0.0226 0.0226 0.0236 Adjusted R-squared -0.0001 0.0223 0.0222 0.0228 Overall F-statistic 0.37 61.91 46.67 21.34

6.3.3 Protestantism

(29)

TABLE6.5: Results difference-in-differences analysis for Protestantism

Dependent variable Subsample: Protestantism

Tipping percentage 1 2 3 4 Dummy service -0.0101 0.0266 0.0318 0.0279 (0.1389) (0.1388) (0.1389) (0.1387) Group 0.0372 0.0655 0.0600 -0.0666 (0.1461) (0.1455) (0.1457) (0.1456) Treatment* -0.0390 -0.1178 -0.1135 -0.1104 after service (0.1960) (0.1946) (0.1946) (0.1946) Number of -0.0479 -0.0489 -0.0487 passengers (0.0344) (0.0343) (0.0344) ln(triptime) -0.6593*** -0.6868*** -0.6828*** (0.1435) (0.1483) (0.1490) ln(distance) -0.4185*** -0.4005*** -0.4051*** (0.1183) (0.1196) (0.1207) Constant 15.2402*** 19.7439*** 19.9684*** 20.2697*** (0.1028) (0.8495) (0.9406) (0.9618) Observations 10,506 10,506 10,506 10,506 R-squared 0.0009 0.0216 0.0217 0.0228 Adjusted R-squared -0.0003 0.0210 0.0209 0.0210 Overall F-statistic 0.05 29.08 22.20 9.94

6.4 Robustness tests

The robustness tests are intended to adapt some of the assumptions underlying the analysis. The assumptions concern the radius and time frame, before and after a service, that is used to select taxi rides into the dataset. This section challenges these assumptions.

6.4.1 Radius

(30)

TABLE6.6: Results difference-in-differences analysis with a radius of 30 meters

Dependent variable Subsample: Full sample Tipping percentage 1 2 3 Dummy service 0.2347 0.3086 0.3353 (0.5521) (0.5467) (0.5459) Group 0.6094 0.5591 0.5768 (0.5274) (0.5274) (0.5255) Treatment* 0.1133 0.0541 0.0156 after service (0.7173) (0.7136) (0.7065) Number of -0.0704 -0.0712 passengers (0.1221) (0.1221) ln(triptime) -0.5241 -0.6272 (0.5529) (0.5733) ln(distance) -0.1924 -0.0980 (0.4150) (0.4258) Constant 14.5124*** 18.0526*** 18.7336*** (0.4086) (3.3190) (3.5154) Observations 642 642 642 R-squared 0.0067 0.0162 0.0172 Adjusted R-squared 0.0020 0.0069 0.0048 Overall F-statistic 1.41 1.34 1.05 Dummies time of day No No Yes

The dependent variable is tipping percentage. Group is defined as 1 for the treatment group and 0 for the control group. Dummy service is defined as 1 for time ranges after the service and 0 for time ranges before the service. There were insufficient observations to estimate month dummies. Robust standard errors are used. *, **, and *** indicate the significance levels 10%, 5% and 1%, respectively.

6.4.2 Time frame

Time frames are decreased to 10 and 15 minutes, whereas the main analysis uses a 30 minute time frame before and after the service. The section discusses the results for Christianity and Judaism separately.

Christianity

Tables C.1 and C.2 show the results when reducing the time frames to 15 and 10 minutes for Catholicism. The treatment effect is insignificant in both specifications and the control variable for distance of the ride turns insignificant. Tables C.3 and C.4 show the results for Protestantism. These results show that the control variable of number of passengers is significant in the specification using a 15 minute time frame only. The control variables for trip time and trip distance are of similar magnitude and significant in all specifications. Finally, similar to the main analysis, the treatment effect is insignificant at any conventional significance level.

Judaism

(31)

reducing the time frames to 10 and 15 minutes. These adjustments do not affect the results of the analysis to a large extent. The treatment effect is insignificant and the control variables are of similar magnitude compared to the main analysis.

To conclude, the results confirm the preliminary analysis that there is no causal effect of attending a religious service on taxi riders’ tipping behaviour in Judaism, Protestantism and Catholicism. The possible causes of these findings are discussed in more detail and related to the literature in the discussion section.

6.5 Exploratory analysis

The exploratory analysis is intended to analyse the dataset in more detail. The section first discusses results for Buddhism, which are an interesting addition to the research but not the main focus due to the lack of academic literature on Buddhism. Second, the section presents ideas for further research on this topic.

6.5.1 Buddhism

The dataset includes taxi rides near Buddhist temples. Buddhist services have an advantage over the services of Judaism and Christianity because the times at which the services end are known. This can greatly increase the accuracy of identification of religious individuals because there is no assumption on the duration of the service necessary. Although there are only five Buddhist temples in the dataset, there are 2,211 observations available. This is sufficient to perform the difference-in-differences analysis.

Interestingly, as shown in table 6.7, there is a significantly negative treatment effect when the time frames before and after the service are reduced to 10 minutes. Attending a Buddhist service leads to a decrease in tipping percentage of 1.28 per-centage points. The overall F-statistic is insignificant for the first two columns and significant at the 5% level for the third and fourth column. As shown in table C.8, a similar estimate of -1.08 is found when the time range is set to 15 minutes. The overall F-statistic is insignificant for the first two columns, significant at the 5% level in the third column and significant at the 10% level in the fourth column. The negative treatment effect is consistent with the theory that religious priming only increases within-group prosocial behaviour and that individuals tend to act more hostile to-wards people outside their group (Jackson & Esses, 1997; Norenzayan & Shariff, 2008; Orbell et al., 1992). However, the literature is mostly written on Protestantism, Catholicism, and Judaism. There is not much literature available on Buddhism, which makes it difficult to explain the negative treatment effect. These results warrant more extensive research in the future.

6.5.2 Areas for further research

There are more analyses possible in the context of religious priming and economic behaviour. Although these are interesting areas of research, these could not be performed given the current dataset. This section introduces two extensions that can be performed in the future.

(32)

TABLE 6.7: Results difference-in-differences analysis for Buddhism with a time range of 10 minutes before and after the service

Dependent variable Subsample: Buddhism

Tipping percentage 1 2 3 4 Dummy service 0.8061 0.8246 1.0039* 1.0110* (0.5126) (0.5195) (0.5228) (0.5216) Group 0.8827 0.9141 0.8975 0.9117 (0.5793) (0.5774) (0.5680) (0.5700) Treatment* -1.1971 -1.2423* -1.2532* -1.2752* after service (0.7562) (0.7508) (0.7412) (0.7547) Number of -0.1530 -0.1755 -0.1881 passengers (0.1292) (0.1298) (0.1293) ln(triptime) -0.1680 -1.1111* -1.1490* (0.5406) (0.6732) (0.6727) ln(distance) -0.0745 0.7026 0.7338 (0.4665) (0.5929) (0.5825) Constant 14.0403*** 14.7106*** 16.6046*** 16.8423*** (0.4127) (1.1946) (1.4638) (1.5446) Observations 809 809 8090 809 R-squared 0.0035 0.0056 0.0189 0.0333 Adjusted R-squared -0.0002 -0.0018 0.0091 0.0100 Overall F-statistic 1.04 0.93 2.02 1.58

As most organisations indicate service times of these events close to the holiday, data on service times could not be collected.

Second, there are two churches on Roosevelt island which are more remote com-pared to other parts of Manhattan. This could decrease the noise in the data, which improves identification of religious individuals. There are insufficient observations available to accurately estimate the treatment effect.1 The absence of observations can indicate that individuals might not take a taxi to leave the service or that services in these churches were not crowded. The analysis for these churches could be interesting on religious holidays, assuming churches will be crowded. Note that Roosevelt Island is more remote than the remainder of Manhattan. Therefore, individuals living on the island may choose not to take a taxi to church. Overall, the above mentioned sug-gestions can extend the current study but an even more elaborate dataset is needed to perform these analyses.

(33)

7 Discussion

The previous section discussed whether religious priming affects tipping behaviour. As explained in section 3, prior to running the analyses, the net effect is ambiguous. The current study focusses on tipping behaviour as a measure of prosociality. There-fore, the main focus is on generosity rather than on other aspects of prosociality such as helpfulness. This section evaluates results in context of the literature and considers limitations to the research.

Results indicate there is no significant treatment effect of attending a religious service on taxi riders’ tipping behaviour for Protestantism, Catholicsm, and Judaism. These results are in line with Benjamin et al. (2016) as the authors find an insignificant effect of religious priming on generosity, for both Protestantism and Catholicism. It is possible that this is the true effect of religious priming on tipping behaviour. However, the results differ from studies suggesting that religious priming increases prosocial behaviour (Ahmed & Salas, 2011; Pichon et al., 2007; Shariff & Norenzayan, 2007). Benjamin et al. (2016) also run a regression on the full sample which, in addition to Christianity, includes Judaism and atheists. The authors find a small, positive but insignificant effect. This is not in line with the current study as results for the full sample, presented in table 6.2, show a small, negative and insignificant effect.1 Results can deviate from the literature because the priming instrument is different. According to Shariff (2015), lab experiments may not create genuine religious situations. This is no concern in the current study because it uses religious services as religious priming. These are genuine religious situations rather than lab experiments. Therefore, this is a possible explanation of why results differ from studies that find a positive effect of religious priming on prosocial behaviour. The current study focusses on generosity, which is only one aspect of prosocial behaviour. This means that extending the definition, by including aspects such as helpfulness, could alter the conclusions with respect to prosociality.

Another potential explanation for the differences in results is that the hypothesis of increased within-group prosocial behaviour holds. This hypothesis is suggested by Norenzayan and Shariff (2008), Orbell et al. (1992), and Jackson and Esses (1997).2 If the hypothesis holds, religious individuals are more prosocial towards members of their own religion. It is reasonable to assume that taxi drivers are not viewed as part of their community as they did not attend the same service. Moreover, the majority of taxi drivers are from Bangladesh and Pakistan (Taxi and Limousine Commission, 2016), which are countries where over 90% of the population is Muslim (Pew Research Center, 2015). The absence of a treatment effect could indicate that religious individuals are not more or less generous towards taxi drivers compared to the control group, who did not attend a religious service.

Furthermore, the literature overview in section 2 reviews possible differences between religious values. It discusses that Protestants are more likely to have a stronger incentive to behave more prosocial than Catholics due to higher costs of

1_{The full sample includes Judaism, Buddhism, Protestantism, and Catholicism. It does not include}

Does religious priming make people tip more generously?