The Social Norms of Corruption in the Field: This is NOT Africa
By: Marleen Troost (11377399)
Supervisor: Nils Köbis
University of Amsterdam, 15 August 2018
Abstract
Corruption marks a major challenge around the world. Although many corrupt practices such as bribery are outlawed in all national codes of law (Mungiu-Pippidi, 2006) and international conventions (Olken & Pande, 2012), stark differences in its (perceived) prevalence exist (Fis-man & Golden, 2017). To explain this gap between legal norms and actual behavior, the current corruption literature emphasizes the importance of social norms, the unwritten rules that guide behavior (Bicchieri, 2016). Although lab research suggests that short prompts that target social norms can reduce corrupt behavior in a bribery game (Köbis et al., 2015), no field evidence exists for the link between information about the behavior of others influencing a) people’s own beliefs and b) corrupt behavior. Conducting a lab in the field study in South Africa we find that a descriptive social norms prompt targeted at the perceived frequency of bribe trans-actions in the region successfully lowered participants’ perceived frequency levels of this type of transaction as well as the levels of actual bribes offered in a simple bribery game. We were however unable to establish a mediating effect of the perceived bribe frequency on the ultimate decision whether to offer a bribe or not.
Introduction
Corruption is a persistent and widespread problem that marks a major challenge around the world. According to standard survey data most countries report that corruption remains a rela-tively common practice. This means that the majority of people on this earth must regularly face the fundamental dilemma that corruption represents namely; going along with it and reap-ing the sought after and often much needed benefit or do the right threap-ing and suffer the conse-quences.
Simply because of the fact that individuals stand to gain from corrupt behavior, once it has taken hold, it proves to be very difficult to root out. Turning more specifically towards bribery, bribes are considered ‘a species of reciprocity’ (Noonan, 1988). Reciprocating behav-ior is known to be deeply rooted in the evolutionary make-up of human beings as it proves essential for establishing long-term cooperation, which can set successful individuals or groups apart from lesser ones (Axelrod, 1981). Therefor depending on the culture and context it might even be socially disapproved of not to reciprocate a bribe. Moreover, a bribe transaction is not directly harmful and the consequences often not immediately apparent, therefore, people en-gaging in it wrongfully believe to be part of a win-win situation. In reality, a bribe is never a win-win situation, there is always a third party, not directly involved in the transaction, that loses or loses out; be it the person that cannot pay the bribe and is thus forced to forego the service, or an entire nation because systemic corruption has impeded the functioning of both their government and economy.
The costs of corruption are therefore not limited to the sum of bribes paid or funds embezzled but rather magnified by the negative effect it has on growth (Fisman and Svensson, 2007), social cohesion and welfare (Chan et al., 2006), and widening inequalities within a country (Gyimah-Brempong, 2002) to name just a few. When corruption becomes endemic the very governmental bodies that have been put in place to monitor corruption become corrupted themselves (Rothstein and Uslaner, 2005). At this point government jobs start attracting dis-honest people, perpetuating the problem (Rema and Wang, 2017). When this occurs, countries are said to have fallen into a corruption trap, at this point only a ‘Big push’ can lift countries out of this trap (Collier, 2000). Proposed interventions include an injection of funds, top-down sanctions and enforcement as well as transparency and information. Usually large costs are associated with such remedies.
Whereas all countries are uniform in outlawing corrupt practices including bribery (Mungiu-Pippidi, 2006), across countries stark differences in the levels of corruption prevail (Fisman and Golden, 2017). These differences between regions can be understood and ex-plained by various theoretical models that are based on the concept of frequency dependent equilibria (Andvig and Moene, 1990). As the term suggests the dominant strategy, bringing about the highest level of utility, is dictated by the expected frequency of each option being chosen by others. Naturally, it follows that depending on varying levels of expectations multi-ple equilibria can come about. To illustrate how a frequency dependent equilibrium might
ap-skipping a red light or speeding. In most countries such a traffic offence results in a hefty fine in the unlucky event that you are caught. In this example you are unlucky indeed and a police officer asks you to pull over. While you wait for the police officer to approach you consider your options; one is to simply accept the fine the officer is about to give you another option could be to slip the officer a banknote instead, in an attempt to avoid the fine. When considering this option what would be your main concern? Most likely you will try to estimate the chances of the police officer accepting your offer and in doing so, your best measure would be the local frequency of such a transaction. If this scenario played out in a low corruption country like The Netherlands you would know that bribery transactions are rare if not non-existent, offering a bribe would most likely only get you in more trouble. However, should you find yourself in this same situation in a high corruption setting like South Africa, knowing bribery is much more common practice, offering a bribe might just make you better off.
The described behavior, namely where we make a decision based on first considering what we expect others to do, can be classified as contingent behavior (Schelling, 1978) and forms the basis for the formation of a descriptive norm. Literature in sociology and philosophy has long since made the distinction between two types of norms namely ‘descriptive norms’ being; what we expect others to do in a given situation and ‘injunctive norms’ being; what we believe others think we ought to do. In the field of Economics these same concepts exist and are referred to as empirical and normative expectations (Paprzycka, 1999; Goffman, 1963). We will use the terms descriptive and injunctive norms to refer to these expectations throughout this thesis. Previous research has pointed out the important role that social norms play in indi-vidual decision-making (Fehr and Fischbacher, 2004; Cialdini, 1991; Goldstein et al., 2008) decisions regarding corruption are no exception (Bicchieri, 2006; Köbis, 2015).
Building on this recent work we conduct a lab in the field study investigating the impact of a descriptive norm prompt at three levels of the individual decision-making process. Firstly, we look at the effect of a norm prompt on the participants’ beliefs about the frequency of cor-rupt behavior of others. We further observe the effect on actual behavior by looking at bribery levels in a corruption game as well as the truthful reporting of a die roll. Lastly, we test for effects on corrupt behavior outside of the lab by tracking stock discrepancies. We find evidence that a widely distributed message targeting the perceived frequency of a bribery transaction has the desired effect in lowering both perceived frequency levels as well as bribery rates. We find no such effect for behaviors that are not directly prompted by the message.
The first section of this thesis reviews the background of our research, in this section we position our experiment within the current anti-corruption literature, explain the use of a descriptive norm prompt and look at corruption in the context of South Africa and its public service. We continue by outlining the methodology and experimental setup in the second sec-tion. The third part describes our data analysis followed by the presentation of our results ac-companied by a discussion of the measured results in section four. The last section concludes and presents policy implications and suggestions for possible future studies.
Part I: Background
LiteratureDefinition of corruption and bribery
Generally speaking, corruption can be defined as ‘the abuse of entrusted powers for private gain’1. From the various classifications of corruption, ordinary citizens tend to suffer the most
from the abuse of entrusted power by low- level public servants; often bribes are required to gain access to basic goods or services that should be gratuitously provided such as healthcare, education and protection. This occurs especially in places where income is low, and the state is fragile. According to Noonan (1984) a given society ‘has at least four definitions of a bribe – that of the more advanced moralists; that of the law as written; that of the law as in any degree enforced; that of common practice.’
Descriptive and injunctive norms
By acknowledging these different definitions within a society Noonan highlights early on the importance of people’s ideas of right and wrong (injunctive norms) and common practice (de-scriptive norms) on the (moral) assessment of bribing and accepts the fact that these norms might be separate from the norm dictated or enforced by law.
Whereas injunctive norms tend to fluctuate less and often correspond closely to the norm pre-scribed by the law, descriptive norms are known to show greater variation between societies and tend to be more malleable (Barr and Serra, 2010). Because of these qualities descriptive norms have been identified as ‘one of the most promising ways to fight corruption’ (Rothstein, 2000).
Moving to more recent work Bicchieri and Xiao (2009) further demonstrate, using various Dictator game experiments that when descriptive and injunctive norms are in conflict, which is generally the case with corruption, it is the descriptive norm subjects have of the choices made by others’ facing the same situation, that best predict the decisionmaker’s eventual choice. This shows that people are guided by what is considered the ‘normal’ thing to do, es-pecially in situations that people are themselves unfamiliar with. In other words when faced with imperfect information other peoples’ choices might be informative and for this reason guide behavior (Banerjee, 1992)
The causal effect of descriptive norms
Building on these findings Köbis et al. (2015) conduct three further studies in a laboratory setting. Their most important results show that there is an association between the perceived descriptive norm and subsequent corrupt behavior and that prompting participants with short statements containing information about applicable descriptive norms successfully influences the perceived frequency of the corrupt act and with that also the decision to engage in corrupt behavior. With this result they establish a causal link that seems to be especially strong when low levels of corruption are prompted which highlights the potential to use such prompts to ‘nudge1’ people towards less corrupt behavior. The authors do note that such messages are
‘especially effective in contexts where people do not have first-hand experience or where it is falsely believed that corrupt decisions are frequent’. Whereas the false belief with regards to the frequency of corrupt transactions is likely to be applicable to countries where corruption exist, not having first-hand experience in corrupt transactions seems to raise some question as to the external validity of results found in the laboratory in the context of corruption.
External validity of laboratory studies
Armantier and Boly (2013) explore this issue of external validity by conducting the same ex-periment in three different settings, the first being a laboratory exex-periment in a developed coun-try (Canada) followed by a lab experiment in a developing councoun-try (Burkina Faso) as well as a field experiment in that same country. The results from the laboratory and the field study in Burkina Faso showed remarkable similarities, especially the probability that a participant chooses to ‘accept a bribe’ in any given treatment is almost identical between both environ-ments. In the laboratory experiment in Canada the direction and magnitude of several treatment effects are the same as in that same setting in Burkina Faso, after controlling for observable differences between subject groups. Both results provide some evidence albeit not definitive
Accepting the external validity of laboratory experiments on corruption allows for a unique position to study corrupt behavior at a microlevel which is otherwise hard to achieve consider-ing its secretive and illegal nature.
A lab in the field study on corruption
For this exact reason field studies on the topic of corruption have been scarce. An exception is a recent experiment conducted by Benoît-Falisse and Leszczynska (2015) amongst public serv-ants in Burundi using a lab in the field set-up similar to the current design described in the next section. Their focus was however on the effect of a professional identity message on bribe taking and subsequent reciprocating behavior. Their results indicate that messages reminding participants of good governance as well as their professional identity have no effect on bribe acceptancy rates, they do however significantly impact behavior that results from the bribe as fewer public servants divide resources unequally, in other words these messages decrease the reciprocation rate of bribes paid. While focusing on the bribe acceptancy and service delivery of public officials the study leaves a need for a focus on the bribe offering side of the transaction from the citizens perspective.
Field studies on the impact social norms on other undesired behaviors
Further field studies have been conducted on the impact of descriptive norms on other unde-sirable behaviors besides corruption. In a natural field experiment Hallsworth et al. (2014) es-tablish that social norm and public goods messages first of all increase the moral cost involved in the decision of paying one’s taxes and demonstrate further that descriptive norms have a significantly larger effect on increasing tax payment rates than injunctive norms. This result is in line with Bicchieri’s earlier findings. In other field studies social norms have been shown to nudge consumer behavior in the right direction for example when it comes to energy conser-vation (Allcott, 2011) littering (Cialdini et al., 1990) and towel re usage as part of an environ-mental conservation program (Goldstein et al., 2008).
The importance of framing
In general framing effects have been known to play a big role in decision making. Research led by Kahneman and Tversky (1989) shows that preferences can be context dependent in such cases and even a small change in context can alter or even reverse a person’s preferences. Therefore, to successfully manipulate a social setting careful attention needs to be paid towards the frame of a social norms message. Predicting how a given context will be interpreted, which cues might be viewed as salient and how these cues relate to specific norms is essential
(Bic-Contribution of current study
Awareness raising campaigns are amongst the most cost-effective strategies to discourage cor-rupt behavior of citizens and public officials alike. However, little is known or demonstrated with regards to their effectiveness outside of the lab. To fill that gap and provide first insight into the effectiveness on social norms messages, we present a lab in the field experiment that will be conducted in South Africa. In it, we test whether a social norms message presented on a poster in close proximity to a mobile lab a) effectively influences participants’ beliefs about the corrupt behavior of others and consequently b) reduces their willingness to engage in brib-ery in a corruption game and c) break an ethical norm outside the lab. We frame this message as a decreasing trend in bribe behavior of citizens.
Public service in South Africa The legacy of apartheid2
Understanding the dynamics of corruption in South Africa requires an introduction to the coun-try’s recent socio-history. South Africa is a nation recovering from a history that created severe inequality amongst its inhabitants. During the era of ‘apartheid’ financial accountability and transparency were not a priority for the ruling minority. Having shifted towards a democracy in 1994 the new regime is having a hard time implementing a set of rules and structures that do promote these qualities but in addition to that also struggles with the concepts of justice and redistribution (Koelble, 2018). The African National Congress (ANC) has implemented black economic empowerment (BEE) strategies in an effort to achieve the Constitutional right of equality by increasing the broad-based participation of blacks in the economy. When it comes to anti-corruption policies however South Africa is commitment rich but appears to be imple-mentation poor as the government has failed to take a unified stance against corruption to date. Many previously disadvantaged people feel that now is ‘their time to feed’ according to an overused African saying and ‘get their share’. Rightfully so recent years has seen the rise of a black middle class which is a step towards decreasing the inequality levels between black and
2 Apartheid was a system of institutionalised racial segregation that existed in South Africa from 1948 until the
white however the gap between rich and poor South Africans keeps growing as demonstrated by the consistently high Gini coefficient3 that reached 0.65 in 2014 (World Bank, 20174). Data on corruption in South Africa
The latest Africa Survey or Afrobarometer from Transparency International (2015) highlighted South Africa as the worst performer with 86% of participants indicating that corruption had recently increased. In 2013 83% of respondents indicated they felt the police force was corrupt to extremely corrupt, followed by political parties at 77% and in third place public officials and civic servants with 74%. The above-mentioned numbers include the broad spectrum of corrupt acts of which bribery forms a part. Recent years have seen an improvement in reported per-centage of bribes paid. Whereas in 2013 this was still at 48% the most recent survey indicates a decrease towards 7% recent bribes paid. There thus seems to be a decreasing trend in bribe levels countrywide. Reported bribe levels between provinces however still shows a great vari-ance, KwaZulu-Natal where we will conduct our study ranks third highest within the country with 30% of respondents reporting bribes.
This is Africa
A common concept in current South African popular culture is ‘This is Africa’ (TIA), this statement is often referred to when things are not the way they are supposed to be or do not go according to plan. In this way potholes, power outages, tardiness, pace of business and even questionable business ethics are excused and accepted as TIA. Interestingly this simple concept helps people cope with the inefficiencies that haunt the country but at the same time formulates a norm of tolerance towards these factors, corruption included.
Part II. Methodology Setting
The experiment took place between July 16,2018 and the 4th of August of the same year. We
set up a mobile lab in the most frequented shopping center of Manguzi. Manguzi is a small rural town near the border with Mozambique, situated in the province of Kwazulu-Natal. Par-ticipants to the study were visitors of the two main shops in the center being the Cambridge Food Store and ThembiNkosi Pharmacy, making up a varied sample of the local community.
3 The Gini coefficient measures inequality. The coefficient of 0.64 means that poorest 20% of the South African
Community members mostly live in informal settlements constructed on a piece of land ap-pointed to them by local chiefs and have to catch a ride on the back of a ‘bakkie5’ into town.
Due to the time and effort associated with a visit to town, shopping patterns at this shopping center are therefore estimated to be monthly visits. Most people in this region refer to them-selves as belonging to the Tsonga ethnic group. Whereas Xitsonga is the language generally spoken within this group in and around Manguzi isiZulu is the more commonly spoken lan-guage. Most people understand and speak some English, however only the more educated will have the English proficiency to conduct this survey in. To accommodate the biggest possible group the participants in the survey are thus offered the choice to either take the survey in English or isiZulu. This decreases any possible selection bias, increases the likeliness of all instructions being properly understood and it also decreases the risk of a foreign language mit-igating the emotional response of the participants (Costa et al, 2014). In order to ensure cohe-sion between both vercohe-sions of the survey the English text was translated into isiZulu by an experienced translator, after this someone with a very basic understanding of isiZulu (like most of our anticipated participants) translated the survey back to English. This confirmed that both versions of the survey had an identical interpretation (refer to Appendix B).
Payment
Participants received the payment of their earnings from the experiment within 48 hours after completing the survey via an Instant Money Transaction as clearly stated in the survey instruc-tions. An Instant Money Transaction sends a voucher for a certain value to a cell phone number via SMS. PIN numbers to redeem the voucher were sent manually after that. Participants could redeem this voucher without any identification by entering their received voucher number and PIN at any ATM or Money Market counter. Instructions were formulated carefully to stretch the fact that a participant’s identity would not be known at any point in time and cell phone numbers being the only would be deleted upon successful payment. Additionally, participants were provided with an e-mail address printed on a slip of paper, in order to have a point of contact should something go wrong during any part of the payment process in order to reduce doubts or uncertainty about being paid out.
In addition to a participation fee of R25, - (then €1,50), participants received between R25, - to R45, - (then ± €1,50 to €2,73). In May 2018 a new minimum wage level of R20, -
was introduced, for Domestic workers this level is at R15, - and for Extended Public Works Programme workers the minimum remuneration amounts to R11, - per hour (Govt SA6).
There-for the minimum amount that could be earned equals payment There-for roughly 2-5 hours of work based on current minimum wage levels while bearing in mind that many participants might not be employed.
Experimental structure
The main source of data in the present study is a survey, upon entering the mobile lab this survey will be conducted on a tablet or mobile device and consists of three sections. In part one, we elicit social norms (both descriptive and injunctive) using the incentivized norms as-sessment method put forth by Krupka and Weber (2009).
In part two, we use a simplified version of a bribery game proposed by Köbis et al. (2018) in their recent study ‘A Market for Honesty’. In brief, the game models a bribery transaction between a public official and a private citizen as a social dilemma.
In part three, participants determine their final pay by a die-rolling task (Shalvi, Dana, Hand-graaf, & De Dreu, 2011) that allows for misreporting.
Outside of the lab, we will also collect field data to track embezzled stock in the local phar-macy. Below each part is specified in detail.
Structure for incentivized norms elicitation
This study follows the incentivized norms elicitation method suggested by Krupka and Weber (2009). In this method a scenario is described to participants in which ‘Person A’ has various choices. Participants then judge each presented choice on a 4-point Likert scale as being “very socially inappropriate” (1), “somewhat socially inappropriate” (2), “somewhat socially appro-priate” (3), or “very socially approappro-priate” (4). To counter the fact that the term ‘socially (in)ap-propriate’ might not often be used in the community of interest, a rephrasing of is added to each description using the term ‘acceptability’ to avoid unnecessary effort from the participant to understand the available options and leave no room for confusion.
6
Participants are rewarded when their evaluation matches that of another randomly selected subject. By rewarding participants in this way, they are incentivized to reveal their belief about the collectively shared idea of appropriateness with regard to each choice as opposed to their own personal judgement. Hence, it uses a coordination game to assess social norms.
In the current study, three scenarios are described of which one is a bribery transaction between a citizen and a police officer. In this scenario the citizen is caught speeding and faces the di-lemma to accept the deserved fine or offer a bribe instead. For the sake of simplicity, the amount of choices deviates from the original format used by Krupka and Weber and has been limited to two of which only one will be judged (see the instructions text for more details). For each scenario, the participants’ beliefs about the frequency of the given course of action is assessed first. Second, they are asked to indicate their beliefs about the social appropriateness of this behavior. Finally, participants provide their personal appropriateness rating of the given behavior. For the first two questions each of the participants’ answers is matched with that of a randomly selected other participant. When both evaluations match each participant will earn an additional R5, - for each question. Question 3 is not incentivized, participants are simply asked to give their honest opinion on their personal rating of appropriateness.
Structure of the bribery game
We consider a game with 10 players; each player will make a decision first as a private citizen and then as a public official. Participants are randomly assigned to a group of 10 players after they have made their choices in the survey. From those 10 players 5 will furthermore be ran-domly assigned the role of citizen, the other 5 the role of public official. Since decisions are paired N=5 for the payoff function.
The citizen needs to obtain a certificate (V) that has a certain value to him/her. The citizen has two choices s/he can either apply for the certificate and pay an application fee (f) or pay a bribe (b) to avoid the application fee and still obtain the certificate but retain a higher net value. On the other hand, the public official also has two choices, he can process the application and earn his wage (W) or he can earn additional money by accepting a bribe (b).
However, for each successful bribery transaction a small loss ‘a’ will be incurred by all 10 players establishing a social dilemma within the game. The payoff structure resembles a pris-oner’s dilemma where all players are better off not bribing but each have the dominant strategy being ‘to bribe’. For the public official the decision to ‘accept the bribe’ is dominant regardless
associated with this choice, apart from the small loss incurred by all 10 players should the decision result in successful bribery transaction. The citizen however suffers a cost if the bribe is not accepted by the official and therefor the dominant strategy is determined by the expected behavior of other participants in the role of public official, simulating a frequency dependent equilibrium as in a real-life bribery decision. At the same time this cost increases the im-portance of anticipating the expected behavior of other participants. In terms of framing of the game citizens and public officials are labelled as such, however the bribe is framed as a ‘side payment’ to reduce social desirability concerns.
Our main interest lies within the citizens choice to offer a bribe or not, as this is the behavior that is targeted with a descriptive norm prompt.
Payoff for the private citizen
The earnings of the private citizen depend on three factors. Namely, their own decision to either pay the ‘application fee’ or to offer a ‘bribe’.
When choosing to pay the bribe, the decision of the public official (B) they are paired up with, will furthermore determine if the bribe transaction is successful. Should the bribe transaction fail due to the decision of the official ‘not to accept the bribe’ the citizen will be forced to pay the application fee and incur an additional cost (c). Finally, for each successful bribery trans-action all 10 players will suffer a loss. The sum of these losses will be deducted from the value determined in the first two steps. To conclude, the payoff for the private citizen P(A, B,Nb) depends on the citizen’s bribe choice A, the decision of the public official B to bribe or not and the number of other successful bribe pairs Nb.
Therefore, if the citizen chooses not to bribe the public official (A=0), the citizen’s payoff is:
Π(0, 𝐵, 𝑁𝑏) = 𝑉 − 𝑓 − (𝑎 ∗ 𝑁𝑏)
If instead the citizen chooses to bribe the public official (A=1), the citizen’s payoff is:
Where B=0 when the public official rejects the bribe and B=1 when the public official accepts the bribe.
Payoff for the public official
In turn the earnings of the public official also depend on three factors. First of all, their own decision to ‘accept a bribe’ or ‘not to accept a bribe’. When opting to accept the bribe, the decision of the private citizen they are randomly paired up with, will again determine if the bribe transaction is successful. Should the bribe transaction fail the public official will simply earn his wage without receiving an additional bribe payment. Lastly, for each successful brib-ery transaction all 10 players will suffer a loss. The sum of these losses will be deducted from the value determined in the first two steps. In summation, the payoff for the public official P(B,A,Nb) depends on the public officials bribe choice B, the decision of the private citizen A to bribe or not and the number of other successful bribe transactions Nb.
So, if the public official is not offered a bribe or if he/she chooses not to accept a bribe (B=0), the official’s payoff is:
Π(0, 𝐴, 𝑁𝑏) = 𝑊 − (𝑎 ∗ 𝑁𝑏)
If instead he/she is offered a bribe and accepts the bribe (B=1), the official’s payoff is:
Π(1, 𝐴, 𝑁𝑏) = 𝑊 + 𝑏 − 𝑎 ∗ (𝑁𝑏 + 1)
Structure of data collection on stock loss
Interviews with local stakeholders revealed that the local pharmacy has been dealing with an imbalance between sold stock captured in the till system by employees and actual stock sold. It has been identified that whencustomers purchase certain items they do not require a receipt. As a consequence, employees frequently do not capture the sale of the product in the system. The customers do pay for these items and the value is then removed from the till at a later stage. Measures have been taken to curb this practice which has seen a reduction in stock discrepan-cies, yet some imbalance remains. This stock loss can stem from embezzlement (i.e. employees pocketing the goods) or bribery (i.e. employees selling the goods underpriced to customers).
During the course of the experiment the pharmacy manager (blind to the purpose of the study) records the daily sales and actual stock levels of 6 products that have been classified as high potential items namely; the morning after pill, gout mix, pregnancy test, Ektoban, Flumix and Rash cream. By focusing on only a few items that are known to ‘disappear’ this data can be recorded without any other pharmacy staff knowing about it which could otherwise be cause for altered behavior.
Structure of the die rolling task
The selected payment method only allows for payments in multiples of R10, - which is not compatible with our current earnings structure. To overcome this challenge participants deter-mined if their earnings would be rounded upwards or down by a self-reported digital die roll. The roll of the die was randomized and shown in a pre-recorded video, after which the partic-ipant was asked to tick one of two boxes. By ticking the first box particpartic-ipants indicated that they rolled a 1, 2 or a 3 and their earnings would be rounded off downwards to the closest multiple of ten. Alternatively, by ticking box number 2 they could indicate that they had rolled a 4, a 5 or a 6 which would result in their pay off being rounded upwards instead. Only the indication of the die roll determined the participants final earnings.
Treatments
The experiment consists of two treatments namely a baseline treatment and a poster treatment. In the baseline treatment no posters were put up and the lab-in-the-field experiment was con-ducted as outlined above. During the poster treatment the perceived descriptive social norms around bribery transactions in the area was targeted. Throughout the town of Manguzi posters were put up reading the message “Less and less people from KwaZulu-Natal pay bribes” (isi-Zulu: ‘Siyehla isibalo sabantu abafumbathisayo KwaZulu-Natal’) a trend based on surveys conducted by Transparency International (Global Corruption Barometer 2013/2017). The de-cision to formulate the descriptive norm message as a trend is based on the notion that speci-fying an exact percentage can successfully adjust the perception of people with a high initial level of perceived bribe frequency downwards, however in contrast it can have the adverse effect on the people whose initial perception of bribe rate levels is low, leading them to adjust their frequency perception upwards (Schultz et al. 2007) also referred to as the ‘boomerang effect’ (Clee and Wicklund 1980).
Posters in both languages were placed on signposts next to the road as well as in other places where often pamphlets are displayed (refer to Appendix C). The poster was also dis-played in near proximity to our field lab at a place where participants waited their turn. With this approach we hoped to ensure that most participants during our treatment period would read the message on the poster. By distributing the poster widely, we aimed to diffuse the link be-tween the poster and our experiment as such a link might compel participants to please the experimenter by giving the answer they think a researcher is looking for. In relation to this concern it is important to mention that it is not unusual for anti-corruption messages to be circulated via billboards, pamphlets or other forums in South Africa.
Hypotheses
Provided that participants have been exposed to the descriptive social norm advocated in our treatment condition, our hypotheses for each section are as follows:
With regards to the norms assessment we expect the poster to have a negative impact on the participants’ beliefs about the frequency of bribery. That is, we anticipate a decrease in the estimate participants give of the amount of people that would choose to ‘offer a bribe to avoid a fine’ during the treatment period as opposed to the control condition. In terms of the assess-ment of participants’ injunctive norms we expect the judgeassess-ment of the level of appropriateness of others and of self to remain unaffected between treatment conditions. Moving on to expected behavior in the bribery game, our hypotheses states that in the role of citizen participants will offer less bribes when exposed to a descriptive norms prompt. To a lesser extent we also expect choices made in the role of public official to be affected and that we will also see a decrease in ‘bribes accepted’ even though this behavior is not directly targeted by the poster. Finally, in our stock analysis we expect to see a drop in missing stock during the poster treatment.
Control questions
• Time of payment: Being aware that salaries are paid out at different times during the month for various occupations in South Africa, we will assess on what day of the month people receive their income. With this information we can control for time of the month effects. The provided payment date gives us furthermore an indication as to what sector the participant is employed in or if they instead depend on social grants.
• Visibility of the poster: To ensure that the participant has been exposed to the proposed treatment two further control questions are included. Namely if the participant has
noticed a poster within the last week that reminds him or her of a topic in this survey and if s/he can recall any of the words mentioned on said poster.
Exclusions 1. Understanding
In order to ensure that participants understand the tasks test questions were included in the study. For the incentivized norms assessment there were two very basic test questions and for the bribery game three. Upon a wrong answer in the first round, participants received a detailed explanation of the correct answer and then got a chance to answer the question again. Partici-pants that failed to answer the question correctly upon second trial were excluded from the part of the experiment the test question was applicable to.
2. Manipulation check
Based on the assessment of whether participants saw the poster and were able to reproduce some of its content, or not. Data will be analyzed in two ways. First, the analysis will be re-stricted to participants who indicate that they recognized the poster and recalled at least some part of its message. Second, we will analyze all data and not apply this restriction. The reason for this choice is that participants might not consciously recall having seen the poster but none-theless be influenced by its content.
For each part of our data analysis exclusions will be mentioned.
Recruitment
An independent research assistant (RA1) was hired to recruit participants and oversee the gen-eral logistics of the experiment. He received R10, - for each recruited individual and to promote quality over quantity received a bonus of R500 at the end of the study based on the quality7 of
all entries. Recruitment was only conditional on the participant being 18 years and older as well as being able to read and write either isiZulu or English. Throughout the entirety of the study the participants anonymity was ensured and at no point where they asked for their names. Participants filled out the survey in individual cubicles6, before they were handed the tablet or
mobile device the research assistant entered a password to unlock each survey that was only
7 By quality we mean individually given answers, without repeat participation, where people take at least 15
known to him, to avoid multiple entries from the same participant8. Payment via an Instant
Money Transaction also guaranteed total anonymity during this final step of the study. Careful attention to anonymity was important given the sensitive nature of the topic that is central in the experiment to encourage participants to make spontaneous decisions without fearing any consequences or moral judgement.
Timeline of treatment
Baseline treatment - In the first two weeks of the experiment the research assistant conducted
the study and recruited participants as outlined above.
Poster treatment - At the start of the third week another research assistant (RA2) placed posters
on strategic locations throughout town9. During this period RA1 continues to conduct the
lab-in-the-field experiment the same as before. Both research assistants were blind to the hypoth-esis and purpose of this study.
Ethics statement
The Ethics Committee Economics and Business at the University of Amsterdam approved the study design in June 2018.
Part 3: Empirical Analysis Descriptive statistics
The collected data sample from the experiment consists of 261 participants of which 57,85% was female, the average age within the sample amounted to 30,16 years with a standard devi-ation of 8,37. From our 261 observdevi-ations 137 formed part of the baseline treatment and the remaining 123 participants made up the observations in the poster treatment. In analyzing the descriptive statistics no participants were excluded. In terms of education most people in our
8This was a concern as participants were likely to have multiple SIM-cards to overcome connectivity issues in
the region.
9The locations of the posters are marked on a map of the town attached in the Appendix C, pictures with examples
sample namely 68,3% stated their highest qualification as ‘Matric’. Matriculation exemption is equivalent to a Western high-school degree and constitutes 14 years of education. From out sample 13,4% proceeded to receive tertiary education at a College or University, the remaining 18,3% never finished high school and as a result do not hold any qualification or degree. From the solicited dates upon which monthly income is received we can conclude that the level of unemployment in our sample is high, at 53,6% more than half the participants do not have a job, a small fraction does however receive income from social grants. This number is roughly in line with the non -metropolitan unemployment rate in KwaZulu-Natal of 47,95% in the first half of 2017 (StatsSA, 201810). In South Africa people working in the private sector are paid
on the 25th of each month whereas public workers receive their income on the 15th. The basic socio-demographic data from this experiment therefore indicates that from the employed part of our sample 83,5% work in the private sector, the remaining 16,5% are employed by the government. Within the sample 120 people chose isiZulu as their preferred language, the re-maining 141 participants filled in the survey in English. Table 1 shows the descriptive statistics reported separately for the control and treatment group that will be used as control variables. The important result to note in this table is that there are no significant differences between the means of both groups for most of the variables, with the exception of the dummy variable ‘none’. This largely supports our claim that selection during both treatments was random. A possible explanation for the larger part of the treatment group indicating that they do not re-ceive an income on any date of the month could be related to the timing of our treatment period. Treatment took place in week three. Whereas during the first two weeks our research assistant had to actively convince and recruit participants to take part in our survey, as the study pro-gressed, and more and more people received payment from earlier sessions, participants came to our mobile lab out on their own accord. The people most likely to do so are the ones that have no work responsibilities and a greater need for income. This type of selection bias is hard to avoid when participants are required to sign themselves up to take part in the study.
Table 1 – Basic socio-demographic indicators, by treatment (1) baseline mean s.d. (2) treatment mean s.d. difference P-value Gender (female)d 0.59 0.49 0.56 0.5 0.03 0.664
Age (in years) 31.12 8.37 29.1 8.28 2.02 0.051
Education grade 8d 0.04 0.02 0.04 0.02 0.003 0.890 grade 10d 0.12 0.32 0.17 0.38 -0.053 0.226 matricd 0.72 0.45 0.64 0.48 0.078 0.178 colleged 0.09 0.28 0.08 0.27 0.007 0.841 universityd 0.04 0.19 0.07 0.26 -0.036 0.198 Income date 1std 0.12 0.33 0.1 0.3 0.027 0.485 15thd 0.09 0.29 0.06 0.23 0.038 0.425 25thd 0.44 0.5 0.35 0.48 0.091 0.133 Noned 0.34 0.48 0.5 0.5 -0.157 0.01 Survey Language isiZulud 0.45 0.5 0.47 0.5 -0.015 0.807 Englishd 0.55 0.5 0.53 0.5 0.015 0.807 Observations 137 124
P-values are based on t-tests. d indicates the use of a dummy variable
Part 4. Results and Discussion
This part presents the empirical results of our experiment. It starts with reporting the impact of the descriptive norm prompt on five outcome variables being the perceived frequency of bribe transactions, social and personal perceptions of the appropriateness of such a transaction, miss-ing stock levels and lastly the truthful reportmiss-ing of a die roll. It then presents the main results from the private citizens binary decision to offer a bribe or not. A binary logistic regression tests the effect of the poster treatment on this decision in two forms, first based on the actual implementation of the treatment and second on the basis of self-reported exposure to the poster as outlined earlier. Finally, a Bootstrapped mediation analysis explores whether perceived de-scriptive norms mediate the link between treatment and the bribery choice in the corruption game as an outcome variable.
Taking part in an experiment was a completely new experience for most if not all of the par-ticipants. Much consideration was given to keeping the experiment and its instructions as sim-ple as possible however still a large number of participants did not manage to answer all test
questions correctly even upon the second trail. Based on the earlier stated exclusion criteria 42 participants were excluded from part 1 of the experiment (norms assessment) and 32 from part 2 (bribery game), as the incorrect answers given were an indication that they had not properly understood the principles applicable to that particular section of the experiment.
Another phenomenon occurred on the 30th of July, which was intended to be the first day of
our treatment period. On this day we observed a coordination of answers across all parts of our survey. Considering that these answers did not reflect the participants true beliefs and actions a further 23 participants were excluded from each part11.
The impact of descriptive norm prompt on: 1. Perceived frequency of a bribe transaction
The distribution of the perceived frequency of ‘a bribe being offered to avoid a fine’ is repre-sented in figure 1. A great density is observed in the baseline treatment around the option in-dicating that 7-8 out of 10 people are believed to ‘offer a bribe to a police man to avoid a fine’ when given the opportunity, this density seems to be redistributed more towards to lower op-tions of 0-2 and 5-6 out of 10 people in that same situation in the poster treatment.
A mean comparison of the perceived bribe frequency between treatments is tabulated in table 2, in it we see a decrease of 9.17% in the poster treatment as opposed to baseline levels. Based on a t-test we reject the null hypotheses that perceived bribe frequency levels are equal between treatments in favor of the one-sided alternative that a descriptive norm prompt signif-icantly lowers the perceived frequency levels at a 5% level significance level (P-value= 0.049).
11Upon further investigation employees from a nearby shop had compared the earnings some had received in the
first two weeks of our experiment (during the baseline treatment). The remaining staff had agreed upon the answers that seemed to pay-off the most, resulting in coordinated entries. Participants were paid out for these entries but only after conclusion of the experiment, in an attempt to discourage this behavior. No further ‘collusion’ was observed.
Figure 1 – Histogram of density of the perceived frequency of ‘bribes being offered to avoid a fine’ between treatments on a 5-point Likert scale indicating low – high frequencies12.
2. Social and personal perceptions of the appropriateness of a bribe transaction
As our hypotheses with regards to the impact of a descriptive norm prompt on perceived in-junctive norms predicts we observe no such change in both the social and personal acceptability ratings of the behavior ‘to offer a bribe’ (refer to table 3 and 4 in Appendix A).
Table 2 – Mean comparison between treatments of the perceived frequency of a person ‘offer-ing a bribe to avoid a fine’.
Observations (N) Mean Standard error s.d. Perceived frequency
baseline 108 3.49 0.12 1.27
treatment 88 3.17 0.15 1.41
t-test (1-sided: >0) t= 1.6618 P-value= 0.049 df=194
3. Missing stock levels
Comparing missing stock levels of six representative products that often disappear between treatments gives us a positive change meaning that more products were embezzled during the poster treatment which is in direct opposition to our hypothesis. However, this effect does not prove to be significant (refer to table 7 in Appendix A). Considering that we specifically tar-geted bribe behavior in our descriptive norms message and not corrupt behavior in general, it could be that the message was simply not salient enough in terms of association with the dis-played behavior. Assuming the missing stock is on account of embezzlement by the employee without the customer being aware, potential benefits of this specific act of corruption depend less on the perceived frequency of the behavior by others and more on the risk of being caught. What we could interpret from the stock level data was that whenever stock was replenished, reported sales of the product in question dipped below the daily average. The actual number of added stock was often impossible for the manager to determine, creating a window of oppor-tunity for fraudulent behavior. Missing stock levels also tended to spike over weekends when there was less supervision and thus a smaller risk of being caught. These assumptions were made by simply eyeballing the data. All in all, missing stock levels were relatively low over the observed period of four weeks, making it hard to conduct a statistical analysis of signifi-cance.
4. Truthful reporting of a die roll
The final mean comparison of the effect of the descriptive norms prompt on the honest report-ing of a die roll did also not lead to any significant differences between treatments (refer to
observations only 16 people misreported their die roll, what’s more 6 out of those 16 people misreported their die roll to their own disadvantage resulting in most cases to a reduction of their final earnings. This insinuates that some form of misunderstanding was present as to the consequences of their decision. Considering only the part of the sample that had a reason to be dishonest13, only 8.6% misreported their die roll14, the total percentage of people misreporting
a die roll and choosing to pay a bribe was even smaller at 1.7% (10 and 2 out of 117 respec-tively). These low numbers are remarkable to say the least especially when we consider that there were no potential costs or externalities associated with the misreporting of the die, par-ticipants could only gain. What we conclude from this is that people’s attitudes and subsequent behavior towards dishonesty and corruption through bribery are very different. Some partici-pants stated that they ‘needed money’ as their reason for choosing to offer a bribe but would continue to truthfully report a die roll of 3 or lower. In fact, reporting a die roll of 4 and above proved in most cases to be more profitable than cooperating in a bribe transaction. A possible explanation for this observed behavior is that people felt that the die roll was not completely anonymous and that it could be determined by the experimenters if they had lied or not however a similar concern should have been present during the decision to bribe as well, which high-lights different attitudes. People who did not choose to offer a bribe were four times more likely to misreport a die roll. This supports decisions observed by Shalvi et al. (2015) where people attempt to strategically maintain a positive self-image by balancing ‘good’ and ‘bad’ choices. Judging from the comments given afterwards participants choosing to offer a bribe felt a greater need to explain their decision than did non-bribers and mostly referred to private circumstances and the need for extra income (refer to figure 3).
Bribe decision
The next section of our results constitutes the main focus of this data analysis. To repeat our hypotheses, we expect the fraction of citizens ‘offering a bribe’ to be lower in the poster treat-ment where participants were exposed to a descriptive norms message compared to baseline levels. In order to test this hypothesis a binary logistic regression was conducted where the options ‘bribe’ and ‘no bribe’ comprised the dummy coded dependent variables and the cate-gorical treatment variable the predictor. Table 3 shows the results. The negative value of the
13The distribution of participants choosing to offer a bribe of having no reason to misreport their die roll because
coefficient as well as an odds ratio of 0.55 (<1) indicates a decrease in the log likelihood that a citizen will choose to bribe when exposed to a descriptive norms prompt. The effect becomes larger when socio-demographic control variables are introduced and larger still when we only consider the effect of the norms prompt on people that can recall part of the message. Effects across the board are significant at a 5%
Table 3 – Binary logistic regressions of a private citizens decision to bribe using coefficients.
Logit
(1) Logit (2) Logit (3) Logit (4) Poster treatmentc1 -0.599** (0.334) -1.148** (0.391) Poster seenc2 -.698** (0.392) -1.092*** (0.421) Constant -0.860*** (0.208) -1.493 (1.305) -0.504 (0.380) -1.751 (1.282)
Controls no yes3 no yes3
Observations (N) 206 206 206 206
c1- categorical variable for groups poster treatment and baseline c2- categorial variable for groups poster seen and poster not seen
3- we refer to table 5 in appendix A for a more details on the regressions involving controls. Significance levels of 10/5/1% are indicated by */**/***
The described analysis was repeated using a categorial variable for the participants indication of having seen the poster on the condition that they could recall some of its message, or not. In both categories where participants were either subconsciously or consciously exposed to the message stating that ‘Less and less people in KwaZulu-Natal pay bribes’ we observe a signif-icant decrease of the log likelihood that the person in the role of citizen will choose to pay a bribe and reduces it to almost half the initial value. Figure 3 shows that the initial bribe levels in the baseline treatment were remarkably low compared to levels recorded in various other studies on the topic of corruption especially for citizens. Only 27.7% of citizens chose to offer a bribe and 40.1% of officials chose to accept a bribe in this condition. The fraction of partici-pants in their decision as a citizen to ‘offer a bribe’ and as a public official to ‘accept a bribe’ decreased to 20.8% and 29.7% respectively during the poster treatment. Notably, these bribe
rates led to only 12 successful bribe transactions15 during the course of the experiment. Where
the maximum loss of a given group of 10 players was R4, -.
Figure 2 – Bar chart showing the percentage of participants choosing to offer/accept a bribe in each role between treatments
The large difference between the proportion in observed bribe levels between roles is not sur-prising. Per our bribery game design, a public official faced no risk or additional costs should no bribe offer be extended by the randomly matched citizen. For citizens the risk of a cost was applicable, there for the net gain of a bribe being accepted was R3, - versus the potential loss of R5, - in the event that the bribe was rejected. This difference in decision parameters is likely central in explaining the observed gap as losses are known to loom larger than gains (Kahne-man and Tversky, 1979). To a lesser extent participants in the role of public official, seeing that the role was framed as such, can have displayed behavior they perceived as fitting for a that function.
Briefly returning to the initial low levels of bribing behavior, in relation to the added element of a social dilemma, many comments solicited after the bribe decision mentioned the ‘social costs’ associated with corruption as being the reason for them not to bribe. In some cases, participants formulated this social cost as part of the bribery game structure in others they 0,0% 5,0% 10,0% 15,0% 20,0% 25,0% 30,0% 35,0% 40,0% 45,0% citizen official
Percentage of participants choosing to bribe
baseline treatment
p<0.05
referred to real life experiences, such as the occurrence of crime. We will review the comments in more detail below.
Figure 3 depicts the main motivational reasons that participants gave for making their decision. These reasons could be roughly divided into 6 main categories for each of the two available choice options.
Figure 3 – Pie chart of reasons given by participants for making their decision as a citizen, divided into categories and shown separately for bribers and non-bribers
Roughly a quarter of people choosing not to offer a bribe, did not feel the need to explain their decision, this number was much lower for people who did choose to offer a bribe. Another quarter of non-bribers was deterred by law and the (il)legality of such a transaction both in terms of a fear for the consequences as simply wanting to follow rules and regulations. The rest stated they wanted to do the right thing (injunctive norm) where 11% specifically hoped this decision would have a positive effect on the community. Bribers seemed to motivate their decision by describing their private circumstances in the hopes of convincing themselves and others their choice was justified because they needed the additional money more. Another group stretched the efficiency that offering a bribe would create. The 9% that referred to a descriptive norm wrongfully believed that everyone else would do the same. Whereas we
Reasons for bribing
private circumstances (25%) pay off (24%) efficiency (20%) no explanation (18%) descriptive norm (9%) reciprocity (4%)
Reasons for not bribing
no explaination (26%) law (25%)
result (no need) (16%) injunctive norm (15%) social cost (11%) other (9%)
c'
a
b
cannot draw any concrete conclusions from this data, it does give us some additional insight into the motivation behind the decision.
Exploratory Mediation Analysis
In an effort to study the mechanisms through which the effect on bribe levels occurs, we present the results of an exploratory mediation analysis testing whether perceived frequency of corrupt behavior mediates the link between treatment and bribery choice in the surveyed corruption game. For this part of the analysis all people with either an incorrectly answered test question in the norms solicitation section or in the bribery game were excluded. The sample was there-fore reduced to 179 observations of which 94 were in the baseline treatment. Results indicated that after these additional exclusions the independent treatment variable was no longer a sig-nificant predictor of perceived bribe frequency, treatment was however still a sigsig-nificant pre-dictor of the citizens decision to bribe based on a one-tailed t-test, b = -.112, SE = .065 p < .05. These results in themselves do not support the mediational hypothesis. Approximately a mere 3% of the variance in bribe decision was accounted for by the predictors (R2 = .027). The indirect effect was tested using a bootstrap estimation approach with 1000 samples and the coefficient was not significant (Shrout & Bolger, 2002).
Figure 4 – Illustration of mediation pathways
The need to exclude nearly 32% of our total sample from this section of the analysis contributed to a decrease of statistical power resulting in non-significant a and b pathways.
Mediation Variable (perceived bribe
fre-quency) Independent Variable (Treatment category) Dependent Variable (Bribe decision) Residual 1 Residual 2
Additionally, the link between the ultimate decision to bribe and the perceived frequency of a bribe transaction is weakened by the fact that the context in each scenario was slightly different. Although both of them were similar in nature each involving a corrupt transaction with theft, meaning that national income was forfeited in exchange for the bribe (Schleifer and Vishny, 1993). The perceived benefit in each scenario differed. Fines tend to be large sums of money that most Manguzians simply cannot afford, even if they would be willing to pay the penalty. Paying a bribe might therefor be viewed as a necessity rather than a choice. In the bribery game however, receiving the certificate is a certain for the citizen, a bribe is thus less of a necessity. 16% of participants that decided not to pay a bribe to acquire their certificate in the bribery game actually indicated this was the main reason for their choice. If they had to forego the certificate by refusing to pay a bribe they might have chosen differently.
In our experimental lab in the field set up we cannot disconnect the experimental context from participants real life situations which allows for an omitted variable bias between scenarios. Whereas our lab in the field study improves the external validity of participants decisions we lose some elements of control.
Conclusion
Our lab in the field experiment directly observed the effect of a widely distributed descriptive norms prompt targeting bribe offers on the perceived frequency of the specified transaction and subsequently a citizen’s decision to engage in this behavior. Results show that the exposure both conscious- and subconsciously to this descriptive statement, significantly lowers the per-ceived frequency levels as well as the likelihood of a bribe being offered.
These results support earlier claims that sanitization and anti-corruption campaigns targeting descriptive norms can be a powerful tool to ‘nudge’ behavior in the desired direction. Results further demonstrate the importance of the message being salient towards behavior, as other behaviors however closely associated with bribery were not impacted.
Being unable to establish a mediating link of the perceived frequency levels on subsequent behavior more research is necessary to explore this possible link, while controlling for the per-ceived necessity of the bribe transaction.
The relatively low levels of bribes offered by participants as a citizen, supported by the moti-vation for this decision, shows that the people from Manguzi and surroundings are no longer willing to accept the status quo of inefficiencies and corrupt service delivery. Their willingness
concept stating that “This is NOT Africa” (TINA). Future research, where the stakes of a bribe decision are higher will have to point out how strong the support of this concept is.
Appendix A
Table 3 – Mean comparison between treatments of the perceived social acceptability of a person ‘offering a bribe to avoid a fine’
Variable Observations (N) Mean Standard
er-ror Standard de-viation Perceived frequency baseline 108 1.981 0.094 0.98 treatment 88 2 0.122 1.14
T-test (1-sided: >0) t= -0.122 P-value= 0.549 df=194 Welch’s T-test t= -0.120 P-value= 0.548 df=173.43
Table 4 - Mean comparison between treatments of the perceived personal acceptability of a person ‘offering a bribe to avoid a fine’
Variable Observations
(N)
Mean Standard er-ror Standard de-viation Perceived frequency baseline 108 2.074 0.109 1.133 treatment 88 1.909 0.122 1.141
T-test (1-sided: >0) t= -1.010 P-value= 0.157 df=194 Welch’s T-test t= -1.010 P-value= 0.157 df=187.585
Table 5 – Effect of the descriptive norm prompt on bribery levels
Logit (1) Logit (2) Logit (3) Logit (4) Treatment (poster) -0.599** (0.334) -1.148** (0.391) Poster seen -.698** (0.392) -1.092*** (0.421) Language (isiZulu)d 0.393 (0.390) 0.320 (0.390) Age 0.014 (0.024) 0.021 (0.024) Gender (male)d 0.190 (0.370) 0.098 (0.037) Highest Education Grade 10d -0.303 (0.864) -0.369 (0.861) Matricd 0.082 (0.780) 0.015 (0.775)
(0.935) (0.928) Universityd -1.463 (1.355) -1.463 (1.342) Income1 2 weeks after pdd 0.888 (0.642) 0.635 (0.620) 2 weeks before pd -0.241 (0.793) -0.160 (0.786) 1 week before pd2 -1.978*** (0.841) -1.853** (0.834) No pay dayd 0.548 (0.515) 0.539 (0.512) Constant -0.860*** (0.208) -1.493 (1.305) -0.504 (0.380) -1.751 (1.282)
Controls no yes no yes
N 206 206 206 206
d Refers to use of a dummy variable. 1.Categorial dummy variables reflecting the relative time between a
per-son’s monthly pay day and the date the survey was taken in weeks and an additional category for no income. 2. This category holds a significant negative relationship towards the bribe decision, which could be explained by the fact that participants did not expect to receive their earnings before pay day, decreasing their need to bribe.
Table 6 - Mean comparison between treatments the truthful reporting of a die roll
Variable Observations
(N) Mean Standard error deviation Standard Misreport die rolld
baseline 132 0.091 0.025 0.289
treatment 101 0.040 0.020 0.196
T-test (1-sided: >0) t= 1.536 P-value= 0.0630 df=231
d Dummy variable 0 if reported truthfully, 1 if not.
Table 7 – Mean comparison of missing stock levels (in items) between treatments
Variable Observations (N) Mean Standard
error
Standard deviation Missing stock
baseline 138 0.254 0.183 2.122
treatment 361 0.639 0.372 2.232
T-test (2-sided) t= -0.948 P-value= 0.344 df=172
1Note the low number of observations during our treatment, the treatment condition was only run for a week
(due to reaching max number of participants), after which the posters were removed and we switched back to baseline.
Appendix B
English instructions to survey (isiZulu instructions are available upon request).
Welcome and thank you for participating. This survey will take 15 – 20 minutes to complete. Please read the instructions carefully.
You receive a participation fee of R25, -. This participation fee is yours to keep inde-pendent on how you decide in the upcoming tasks.
In addition to this participation fee you can earn extra money. The amount you earn extra de-pends on your decisions in several tasks and the decisions of others in these tasks. The maxi-mum amount you can earn extra is R60, -
You will receive the total amount of money you earned via a Standard Bank Instant Money transaction to your cell phone number within 48 hours. You can collect your earnings at Spar by presenting the voucher number in your SMS and entering the release PIN commu-nicated in a separate SMS. We, therefore, will ask you for your cell phone number at the be-ginning of the study. Your mobile number will be deleted from our records after successful payment.
Important: We will not ask you for any other identifiable information. That means we will not ask you for your name etc.
The payment will be done by an external researcher from the University of Amsterdam. No-body in Manguzi will, therefore, be informed about your earnings and your decisions in the study. That means that all your decisions are fully anonymous.
The study has 2 parts:
- In Part 1 you have to answer 9 questions.
- In Part 2 you have to make 2 decisions in a task.
Both parts will be explained in more detail below. If you have any questions at any point of the study, do not hesitate to ask the research assistant.
Please fill out the following information:
• Cell phone number eg (072137****)
• Gender: Male/Female/Other prefer not to say
• Highest Education: Grade 8/Grade 10/Matric/College/University • Age
• On what day of the month do you receive your monthly income. Please select the closest option: None/1st/15th/25th
In this part, you will read 3 scenarios. Each scenario presents a situation where a person makes a decision between two options. You will be asked three questions regarding these options.
In Question 1 and Question 2, you have to guess the answer of another participant to that same questions. Question 3 asks you for your opinion.
This is an example, you do not have to answer any questions in this part.
‘Person A’ wins R1000 in the lottery. When the win is paid out at the kiosk, another person ‘Person B’ is also present. That person B did not win in the lottery. 'Person A’ has two options:
1. Person A can keep the win entirely OR
2. Person A can share half of the win with person B
Subquestion 1:
This question is about the frequency of people in and near Manguzi choosing option 1. There are five possible answers (see below).
a) 0 - 2 out of 10 people b) 3 - 4 out of 10 people c) 5 - 6 out of 10 people d) 7 - 8 out of 10 people e) 9 - 10 out of 10 people
We do not ask you to indicate how many out of ten persons choose '1: Keep the win entirely'. Instead, you have to guess what is the most frequent answer given by the other participants in the study.
Let's consider this example. You think that most other participants in this study will indicate that 9 or 10 Manguzians choose '1: Keep the win entirely'. In order for you to increase your chances of winning an extra R5 you have to the answer (9-10 people).
To calculate your earnings for this question, the computer will randomly select another participant from this survey. Your answer and the answer of this other participant will be matched.
If you and this other participant provided the same answer, you will each receive R5 extra. if you and this other participant provide different answers, you both receive R0 extra.
Subquestion 2:
This question is about how appropriate you think most other people would rate a choice. There are 4 possible answers:
