• No results found

The determinants of trust in digital messages

N/A
N/A
Protected

Academic year: 2021

Share "The determinants of trust in digital messages"

Copied!
42
0
0

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

Hele tekst

(1)

The determinants of trust in digital messages

Name: Nandi A. Oud

Student number: 10002425

Programme: Master Organization Economics

Track: Business Economics

Number of ECTS: 15

Name supervisor: Jeroen van de Ven

Date: 26-06-2015

Abstract

In organizations people have to deal with a lot of situations that involve trusting other people. The decision to trust or not to trust often depends on communication via digital messages. This research has investigated by making use of a survey if people are good at judging whether someone can be trusted via a digital message and how they judge this. If we look at the behaviour of people in the trust game, one of the findings is that people are sometimes good at predicting whether someone can be trusted. Another finding is that a digital message is trusted more often if it contains humour. People who can be trusted write a longer message and more often make a promise than people who can’t be trusted. Furthermore females trust more often than males and the decision to trust depends on their past experiences with trust.

(2)

2

1. Introduction

In organizations people have to deal with a lot of situations that involve trusting other people. For example because contracts are incomplete or there is no contract at all. The decision to trust or not to trust often depends on communication between the trustor and trustee. In a world of globalization and digitalization this communication often proceeds via digital messages. The aim of this research is to identify some determinants of trust in digital messages. Therefore this research investigates whether people are good at judging whether someone can be trusted via a digital message and how they judge this. This is interesting to investigate because if you know the determinants of trust, you can act more trustworthy.

The research question is studied by doing a survey. In this survey the messages that are send by people during the trust game from Charness and Dufwenberg (2006) are used. This research has found that if we look at the behaviour of people in the trust game, that people are only good at predicting whether someone can be trusted in the $7 condition. This research has also found that a digital message is trusted more often if it contains humour. People who can be trusted more often write a lengthier message and make a promise than people who can’t be trusted. Furthermore this research has found that females trust more often and that the decision to trust depends on their past experiences with trust. In the $5 condition the decision to trust negatively depends on the statement ‘whenever I trusted someone in the past, that trust was betrayed’ and it positively depends on the statement ‘I can be trusted’. In the $7 condition, the decision to trust negatively depends on the statements ‘whenever you trusted someone in the past, that trust was betrayed’ and ‘when dealing with strangers it is better to first take care before you start to trust them’.

However if we look at whether people can estimate if someone can be trusted, the results are less striking. There are only some significant effects found in the $7 condition. One of the findings is that people can predict whether someone can be trusted. Another finding is that people trust more often if a promise is made.

Clearly, there is a difference between how people estimate if someone can be trusted and what their behaviour in the trust game is. When people estimate whether someone can be trusted, they base their choice on less characteristics of themselves and the messages, than when we look at the behaviour of them in the trust game. One explanation is that they could simply have forgotten the characteristics of the messages, and through this, their estimate was less accurate than their behaviour.

(3)

3

In section two, the related literature is discussed. This section will explain what previous research has found so far concerning this topic. In section three, the methodology of this research will be explained. Furthermore, the hypotheses are presented. In section four, the results are presented and some limitations are discussed. In section five, the results are

discussed and a conclusion is reached. In the remainder the reference list and appendix can be found. “In this paper use is made of data of the LISS (Longitudinal Internet Studies for the Social sciences) panel administered by CentERdata (Tilburg University, The Netherlands).” (lissdata.nl).

2. Related literature

In this section the most important related literature will be described. It will explain what other research has found so far and how this research improves upon that. In neoclassical economic theory, it is often assumed that humans are homo economicus. The concept homo economicus holds that people maximize utility by making rational choices that are only in their self-best interest. However many research has pointed out already that this is not the case. In strategic games it is often found that humans behave as homo reciprocans (Gintis, 2000). A homo reciprocans cooperates if the other player cooperates and defects or even punishes if the other player does not cooperate. The decision to cooperate depends on the level of trust (Chaudhuri et al., 2002). But how do people know whether someone can be trusted?

There has been done a lot of research about trust in organizations in the field of

economics as well in the field of psychology. For example Mishra and Morrissey (1990) did a survey among managers about how they could build trust. The most important factor

according to managers is open communication. Loomis (1959) investigated the choices and perceived trust of participants in a trust game. He found that participants that could not communicate during this game were less likely to perceive trust than participants that could communicate.

In social dilemma’s, researchers often allow communication via (digital) messages. However the results about the influence of communication on cooperation, and thus trust, are mixed. Bracht and Feltovich (2009) found for example that sending a message has little effect on cooperation. However Bochet et al. (2006) found another result. They found that

(4)

face-to-4

face interaction. Maybe these differences in findings arise because different games are used in the two research papers.

This research is based on the research of Charness and Dufwenberg (2006). Charness and Dufwenberg (2006) have investigated the influence of communication on whether people trust and cooperate in a trust game. Charness and Dufwenberg (2006) found that if people make a promise they act more trustworthy. So communication can influence behaviour in this trust game. Ismayilov and Potters (2012) used the same game in their research but in their design, only 50% of the messages are delivered. They found that even if the message is not delivered, that people that make a promise act more trustworthy. People also act more

trustworthy if their message is delivered irrespective of whether they have made a promise or not.

This paper improves upon the research of Charness and Dufwenberg (2006) and Ismayilov and Potters (2012) because it also examines the determinants of trust in the messages. So this research takes a different point of view. It examines not if someone acts trustworthy but it examines if the opponent also sees whether that person can be trusted. In the research of Charness and Dufwenberg (2006), the beliefs of the players were also measured (they asked A players to guess the proportion of B players who chose roll). This is also done in the survey of this research. It is not enough to use the data from the Charness and

Dufwenberg (2006) paper to answer the questions for several reasons. The first is that in this research the same message is judged ten times by different participants. This will make the average decision to trust or distrust more reliable. Furthermore this research also examines the influence of gender and the past experiences with trust of the participants on the decision to trust.

Schniter and Sheremeta (2013) found that 16,6% of the people that send a message are distrusted and that 18,8% of the trusted people that make a promise, break their promise. However they use a different game from the game that is used in this paper. They use a

repeated game setting. However it is more interesting to see if people can predict in a one shot game whether someone can be trusted, because then people cannot build a trustworthy

reputation. This research will investigate if people can predict whether someone can be trusted via a digital message, and how they judge this.

Some people send a humorous message and others send a non-humorous message. Sternthal and Craig (1973) have investigated some literature about humour used in

advertising. Advertising is used to convince and persuade people. People try to do the same with their messages in the trust game of Charness and Dufwenberg (2006). Sternthal and

(5)

5

Craig (1973, p. 17) say that “A humorous context may increase liking for the source and create a positive mood. This may increase the persuasive effect of the message.” and “To the extent that a humorous context functions as a positive reinforce, a persuasive communication placed in such a context may be more effective.”. So it is also interesting to investigate whether people that a send digital message that contains humour are trusted more often.

Croson and Buchan (1999) have found a gender effect in their trust game. They found that female second-movers in a trust-game reciprocate more. Buchan et al. (2008) have also investigated how gender influences trust. They found that males trusted more often than females. However this probably stems from the fact that males have a more strategic viewpoint in this game. They also found that females are more trustworthy than males. However Croson and Buchan (1999) and Buchan et al. (2008) have used a different game than is used in this research. This raises the question whether there also is a gender-effect with respect to trust in this game.

Poppo et al. (2008) investigated the influence of the past and the future on trust building. They found that the past indirectly affects building trust. The expectations in the future are more critical to trust building. In this research, use is made of a one-shot game. So only the past can influence the decision of the participants. Therefore it is interesting to investigate if past experiences with trust influence the decision to trust in this game.

If you allow communication via digital messages, people can make a promise. The previous discussed research of Ismayilov and Potters (2012) found that people who make a promise act more trustworthy. This raises the question if the opponent also knows whether people that make a promise act more trustworthy. Therefore this paper also investigates whether people trust more if the opponent makes a promise than when the opponent does not make a promise.

According to Rockmann and Northcraft (2008) communication that involves

technology leads to less trust than face to face interactions, but trust varies with the richness of the communication channel. They investigated face-to-face interactions, interactions via video and interactions via a computer. Face-to-face interactions lead to more trust than interactions via video and interactions via video lead to more trust than interactions via a computer. Schniter and Sheremeta (2013) allowed people in their research to build new trust and to rebuild damaged trust because people played the game twice. Schniter and Sheremeta (2013) found that trustees who were distrusted in the first game used long messages. In this research we can see long messages as emails and short messages as WhatsApp- or

(6)

SMS-6

messages. The question is than whether emails are richer than WhatsApp- or SMS-messages and thus whether trust varies with the length of the message.

3. Methodology

In this section, the methodology of this research is described and the hypotheses are stated. In this research the messages from the Charness and Dufwenberg (2006) supplement are used in a survey. This paper contains 91 messages of which 81 are used in the surveys. Only 81 messages are used because ten people decided to send an empty message. Of these 81 messages, 38 are send by people in the $5 condition and 43 are send by people in the $7 condition. The messages from the $5 condition can be found in appendix 7.1 and the

messages in the $7 condition can be found in appendix 7.2. There are eight different surveys, four for both conditions.

Charness and Dufwenberg (2006) have investigated the influence of communication on whether people trust and cooperate in a trust game. There are two players who play this game, A and B. First A makes the choice to play in or out. Then player B makes the choice to roll or don’t roll (a die). Before the game starts, B can send a message to A. These messages are used in this research. If player A plays out, both players get $5 or $7, so there are two conditions. If player A plays in, the payoff depends on player B and/or the roll of a die. If player B plays don’t roll, player A gets $0 and player B gets $14. If player B plays roll, the payoffs are determined by the roll of a die. If the die comes up 1, player A gets $0 and player B gets $10. If the die comes up 2-6 player A gets $12 and player B gets $10. Note that player B does not know whether A has played in or out (Charness & Dufwenberg, 2006). The payoffs are presented in table 1:

TABLE 1: Payoffs in trust game of Charness and Dufwenberg (2006)

A gets B gets

A out $5/$7 depending on condition $5/$7 depending on condition

A in, B don’t roll $0 $14

A in, B roll, die = 1 $0 $10

(7)

7

We can think of player A and B as a principal and an agent. The principal offers the agent an incomplete contract (in) or offers no contract (out). Then the agent chooses to exert effort (roll) or exert no effort (don’t roll). However if the agent exerts effort, whether the project will succeed will depend on luck (roll of a die). Before the principal offers the contract, some communication will take place between the principal and the agent via a digital message.

First a pilot was done to see whether the participants would understand the questions in the questionnaire. The pilot was a success and nothing was changed to the questionnaires. Thereafter each survey was filled in by ten different participants. The participants were friends, family, colleagues and UVA students. The friends, family and colleagues were also students. The survey was distributed on paper and per email. The surveys that were

distributed on paper were distributed in the canteen of the UVA and in the computer space of the UVA. One of these surveys can be found in appendix 7.3. In every survey, the participants judged nine, ten or eleven messages.

During this survey the participants first read how the game works. Then they indicate for nine, ten or eleven messages whether they would choose in or out, what they think the chance is that person B will play roll and whether they think the message is funny or not. The choice to play in or out and the chance that person B will play roll are used as the measures of trust. Thereafter they fill in whether they are male or female. At last they fill in five questions about their past experience with trust. They can indicate this on a scale from one to seven. These last five questions are from the LISS Panel.

The Charness and Dufwenberg (2006) supplement also contains for every message if people made a promise, whether A played in and whether B played roll. This will also be used in the analysis. The last variable is the length of the message. This is both a binary variable and the length of the message in words. For the binary variable, the message is short if it contains 21 words or less and is long if it contains 22 or more words. This cut-off value is chosen because than, halve of the messages is long and the other halve is short.

Based on my research question, I will test some hypotheses. The research question is: Are people good at judging whether someone can be trusted via a digital message, and how do they judge this? The first part of the research question ‘Are people good at judging whether someone can be trusted via a digital message’, is tested in the first hypothesis. The second part of the research question, how people judge whether someone can be trusted via a digital message is tested in three other hypothesis. These hypotheses can be split up. On the one hand it is tested whether trust depends on a characteristic of the messages themselves, namely humour. This is covered by hypothesis two. On the other hand, it is tested whether

(8)

8

trust depends on the characteristics of the receiver of the message; this is tested by hypothesis three and four.

The first hypothesis is: People are good at predicting whether someone can be trusted via a digital message. It is expected that people are good at predicting whether someone can be trusted because people are often in one-shot situations in which they have to choose whether to trust or distrust someone. This hypothesis will be tested by doing a t-test. The null hypothesis is that the proportion of people that plays in is the same as the proportion of people that plays in for the messages where the sender plays roll. The alternative hypothesis is that the proportion of people that plays in is smaller as the proportion of people that plays in for the messages where the sender plays roll. This is tested for both the $5 condition and the $7 condition.

The second hypothesis is: If the message contains humour, this message is trusted more often than when it doesn’t contain humour. This is the humour hypothesis. This

hypothesis is tested by doing a t-test. The null hypothesis is that the proportion of people that plays in is the same as the proportion of people that plays in if they receive a humorous message. The alternative hypothesis is that the proportion of people that plays in is smaller than the proportion of people that plays in if they receive a humorous message. This is tested for both the $5 and $7 condition. Furthermore it is very likely that message-senders, who can be trusted, will more often make a promise and will write a lengthier message. This will also be tested.

The third hypothesis is: There will be a gender effect with respect to trust, females trust more often than males. This hypothesis is tested by doing a t-test. The null hypothesis is that the proportion of males that plays in is the same as the proportion of females that plays in. The alternative hypothesis is that the proportion of males that plays in is smaller than the proportion of females that plays in.

The last hypothesis is: If people have had negative (positive) experiences with trusting in the past, they will trust less (more). This hypothesis is also tested by doing a t-test. The null hypothesis is that the average rating of the five questions separately is the same for the people that play in and for the people that play out. The alternative hypothesis is that the average rating of the five questions separately is lower (higher) for the people that play out than for the people that play in.

Before testing the hypotheses, we have a look at the summary statistics to see whether there are differences between the data from this research and the data of the Charness and Dufwenberg (2006) paper. Thereafter some t-tests are done to compare the data from the two

(9)

9

different conditions. After the t-tests some regressions are carried out with in and b_guess as dependent variables. Logistic regressions are used when in is the dependent variable because this variable is binary and use is also made of continuous variables in these regressions. OLS-regressions are used when b_guess is the dependent variable, because this variable is

continuous. It is important to run some regressions to see how big the influence of the variables is on the choice to play in and the estimated chance that player B will play roll.

4. Results

In this section the results will be presented and some limitations are discussed. In table 2 the most important summary statistics can be found. Table 2 also explains the variables. The variables from the Charness and Dufwenberg (2006) paper are in capital letters, the variables from this research are in small letters.

(10)

10

TABLE 2: Variables list, summary statistics

Mean Std.

Dev. Min

Ma

x Explanation

A_IN 0,642 0,480 0 1 1 If A plays in

B_ROLL 0,593 0,492 0 1 1 if B plays roll

PROMISE 0,593 0,492 0 1 1 if B makes a promise in message

LENGTH 27,148 23,194 1 95 Length of message in words

LENGTHBIN 0,494 0,500 0 1 1 if amount of words > 22

in 0,584 0,493 0 1 1 if participant would play in

b_guess 50,608 29,151 0 100 Chance that B plays roll according to participant

humour 0,249 0,433 0 1 1 if participant thinks the message is funny

male 0,475 0,500 0 1 1 if participant is male

pastexp1* 3,137 1,405 1 7 Whenever the participant trusted someone in the

past, that trust was betrayed

pastexp2* 3,356 1,483 1 7 The participant thinks you can no longer trust

strangers

pastexp3* 6,067 0,749 3 7 People can trust the participant

pastexp4* 5,240 1,302 2 7

When dealing with strangers, the participant thinks it is better to first take care, before you start to trust them

pastexp5* 4,443 1,808 1 7 Participant regularly lends money to friends

*1= not agree, 7= agree

When looking at the summary statistics in table 2 we see that the mean proportion of people that plays in (A_IN) in the Charness and Dufwenberg (2006) paper is somewhat higher than the mean proportion of participants that plays in (in). We also see that people predict that the mean chance that someone will roll (b_guess) is 50,6% while the proportion of people that actually plays roll (B_ROLL) is 0,593, which is higher. Furthermore we see that the 80 participants did not use the lower bound of pastexp3 and pastexp4, so there is less variability in these measures. The summary statistics per condition are presented in table 3. It is

important to look at the summary statistics per condition to see whether there are striking differences between these two conditions. Some t-tests are done to see if the means from the variables for the two conditions differ. It is important to do these t-tests to see whether we can

(11)

11

analyse the two conditions at the same time or, if there are differences, if we have to analyse the two conditions separately.

TABLE 3: Summary statistics per condition

Mean $5 condition Std. Dev. $5 condition Mean $7 condition Std. Dev. $7 condition Amount of observati ons A_IN 0,790 0,408 0,512 0,500 81 B_ROLL 0,658 0,475 0,499 0,499 81 PROMISE 0,632 0,483 0,558 0,497 81 LENGTH 32,921 27,447 22,047 17,128 81 LENGTH BIN 0,579 0,494 0,419 0,494 81 in 0,563 0,497 0,602 0,490 810 b_guess 49,315 28,442 51,751 29,751 810 humour 0,266 0,442 0,235 0,424 810 male 0,524 0,500 0,433 0,496 80 pastexp1* 3,324 1,448 2,972 1,346 80 pastexp2* 3,221 1,389 3,474 1,553 80 pastexp3* 5,979 0,901 6,144 0,573 80 pastexp4* 5,263 1,285 5,219 1,318 80 pastexp5* 4,321 1,825 4,551 1,789 80

*1= not agree, 7= agree

If we look at table 3 we see that the proportion of people that have chosen in (A_IN) in the Charness and Dufwenberg (2006) paper is bigger in the $5 condition than in the $7 condition (2 sided p-value = 0,009). So from now on we analyse the two conditions separately because people react differently in the two conditions. This is quite logical because the people in the $5 condition have less money to loose. The proportion of people that have chosen roll (B_ROLL) in the Charness and Dufwenberg (2006) paper is not statistically significantly different between the two conditions. The proportion of players B that made a promise (PROMISE) in the Charness and Dufwenberg (2006) paper is also not statistically

(12)

12

significantly different between the two conditions. However the messages from the players B are lengthier (LENGTH) in the $5 condition than in the $7 condition (2 sided p-value = 0,004). Maybe this can be explained by the fact that people in the $5 condition have more to gain and try harder to convince the opponent by writing a longer message. The proportion of respondents that plays in (in) is not statistically different between the two conditions.

It is remarkable that there is a difference in the data from the Charness and

Dufwenberg (2006) paper but not in the data from the respondents of the survey. This may stem from the fact that the respondents of the survey had to judge nine, ten or eleven messages while the participants in the Charness and Dufwenberg (2006) paper only had to judge one message. If you judge one message, you have to choose in or out. However if you judge more messages, it is likely that you try both options, and maybe randomize between the two options.

The mean guess of the respondents concerning roll or don’t roll (b_guess) is also not statistically significantly different between the conditions. The proportion of respondents that is male (male) does not differ between the two conditions even as the amount of messages that are rated as humorous (humour). All past experience measures (pastexp) are also not statistically significantly different between the two conditions. From this we can conclude that the subject sample is the same between the two conditions.

In table 4 the p-values of several t-tests are presented. It is important to do some t-tests to investigate if the mean proportion of people that plays in differs if another variable varies. In this way we can see whether this variable has an influence on the choice to play in. In row one the two-sided p-value is used, because it is not clear what the direction of the effect will be. For all the other tests, the one-sided p-value is used.

(13)

13

TABLE 4: Two-sample t-test on the means using groups

Total $5condition $7condition

1. Mean ‘A_IN’ if ‘in’ 0 = mean ‘A_IN’ if ‘in’ 1

0,023 0,519 0,015 0,811 0,070 0,141 2. Mean ‘in’ if ‘B_ROLL’ 0 = mean ‘in’ if

‘B_ROLL’ 1 0,065 0,033** 0,044 0,795 0,163 0,000*** 3. Mean ‘b_guess’ if ‘B_ROLL’ 0 = mean ‘b_guess’

if ‘B_ROLL’ 1 5,547 0,004** 3,415 0,866 13,435 0,000*** 4. Mean ‘in’ if ‘PROMISE’ 0 = mean ‘in’ if

‘PROMISE’ 1 0,177 0,000*** 0,111 0,018** 0,240 0,000*** 5. Mean ‘in’ if ‘LENGHTBIN’ 0 = mean ‘in’ if

‘LENGTHBIN’ 1 0,032 0,180 0,053 0,847 0,120 0,006**

6. Mean ‘in’ if ‘humour’ 0 = mean ‘in’ if ‘humour’ 1

0,192 0,000*** 0,217 0,000*** 0,170 0,001*** 7. Mean ‘in’ if ‘male’ 0 = mean ‘in’ if ‘male’ 1 0,138

0,000***

0,159 0,001***

0,114 0,008***

The first number is the difference between the two means; the second number is the P-value. *Significant at the 10% level, **Significant at the 5% level, ***Significant at the 1% level

In the first row of table 4 we see that the mean proportion of people that has played in (A_IN) in the Charness and Dufwenberg (2006) paper is the same whether the participant plays in or out (in). So it looks like people are not good at predicting whether someone would play in or out. The correlation between ‘A_IN’ and ‘in’ is 0,023. This means that people from this sample do not choose in more often if people from the original sample chose in.

However in the second row of table 4 we see that the mean proportion of participants that plays in (in) does differ between the people that have played roll or don’t roll (B_ROLL) in the Charness and Dufwenberg (2006) paper. To be more precise, the proportion of people that plays in is bigger if people have played roll in the Charness and Dufwenberg (2006) paper, especially in the $7 condition. So it looks like people can predict whether the message-sender can be trusted in the $7 condition, but not in the $5 condition.

In the third row of table 4 we see that the mean prediction of the percentage of players that plays roll (b_guess) is different if the message sender played roll or don’t roll (B_ROLL). The average prediction is higher if the message sender has played roll than if the message sender has played don’t roll in the $7 condition. So in the $7 condition, people are again good

(14)

14

at predicting whether the message-sender can be trusted or not. People are good at predicting whether the message-sender can be trusted or not if they predict better than would be expected if we look at chances. So the first hypothesis ‘People are good at predicting whether someone can be trusted via a digital message’ is only confirmed for the $7 condition.

In the fourth row of table 4 we see that the mean proportion of participants that plays in (in) differs between the messages wherein a promise is made and wherein no promise is made (PROMISE). The proportion of participants that plays in is bigger if a promise is made than when no promise is made. In the fifth row of table 4 we see the mean proportion of participants that plays in (in) after reading a long or a short message (LENGTHBIN) only differs in the $7 condition. In the $7 condition the proportion of participants that plays in is bigger if the participants have read a long message. However the measure ‘LENGTH’ is more reliable because this uses the length in words so this will be discussed later.

In the sixth row of table 4 we see the mean proportion of participants that plays in (in) differs between reading a humorous or not-humorous message (humour). The proportion of participants that plays in is bigger after reading a humorous message in both conditions. This confirms the second hypothesis ‘If the message contains humour, this message is trusted more often than when it doesn’t contain humour’. However we should be cautious with interpreting this result because this could be an experimenter demand effect because in the questionnaire it was asked whether the participants found the message funny or not.

In the seventh row of table 4 we see that the mean proportion of participants that plays in (in) depends on whether the participants are male or female (male). Males play in less often than females in both conditions. So third hypothesis ‘There will be a gender effect with

respect to trust, females trust more often than males’ is confirmed. The most important foregoing results are summarized in figure 1.

(15)

15

FIGURE 1: The proportion of participants playing in for the variables B_ROLL, PROMISE, LENGTHBIN, humour and male for the $5 condition, $7 condition and in total

Notes. B_ROLL 1means that player B has played roll, B_ROLL 0 means that player B has played don’t roll. PROMISE 1

means that there is a promise in the message; PROMISE 0 means that there is no promise in the message. LENGTHBIN 1 means that the message is longer than 21 words; LENGTHBIN 0 means that the message is shorter than 22 words. Humour 1 means that the message is rated as funny, Humour 0 means that the message is rated as not funny. Male 1 means that the participant is male, Male 0 means that the participant is female

In table 5 the results regarding the participants’ past experiences with trust and whether they have played in or out are presented. Some t-tests are done to see if the mean of the past experience measures is different for the people that played in than for the people that have played out. 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0 1 0 1 0 1 0 1 0 1

B_Roll Promise Lengthbin Humour Male

$5 condition proportion of participants playing IN $7 condition proportion of participants playing IN Total proportion of participants playing IN

(16)

16

TABLE 5: Two-sample t-test on the means using groups

Total $5 condition $7 condition

Mean ‘pastexp1’ if ‘in’ 0 = mean ‘pastexp1’ if ‘in’ 1

0,340 0,000*** 0,238 0,056* 0,406 0,001***

Mean ‘pastexp2’ if ‘in’ 0 = mean ‘pastexp2’ if ‘in’ 1

0,067 0,263 0,275 0,972 0,397 0,005***

Mean ‘pastexp3’ if ‘in’ 0 = mean ‘pastexp2’ if ‘in’ 1

0,236 0,000*** 0,348 0,000*** 0,123 0,015**

Mean ‘pastexp4’ if ‘in’ 0 = mean ‘pastexp2’ if ‘in’ 1

0,184 0,024** 0,093 0,757 0,433 0,000***

Mean ‘pastexp5’ if ‘in’ 0 = mean ‘pastexp2’ if ‘in’ 1

0,242 0,030* 0,425 0,012** 0,095 0,300

The first number is the difference between the two means; the second number is the P-value. *Significant at the 10% level, **Significant at the 5% level, ***Significant at the 1% level

The mean on the statement ‘whenever you trusted someone in the past, that trust was

betrayed’ (pastexp1) is higher for the participants that played out than for the participants that played in in both conditions. The mean on the statement ‘you can no longer trust strangers’ (pastexp2) is higher for the participants that have played in rather than out in the $5 condition and lower for the participants that have played in rather than out in the $7 condition. This is a very remarkable result. You would expect that the mean of pastexp2 would be lower for the participants that have played in rather than out in both conditions. Probably both significant results are type I mistakes. The mean on the statement ‘people can trust me’ (pastexp3) is higher if the participant played in rather than out in both conditions. This is an interesting result because people who think they can be trusted themselves trust others more often. The mean on the statement ‘when dealing with strangers, it is better to first take care, before you start to trust them’ (pastexp4) is higher for the participants that play out than for the

participants that play in, but only in the $7 condition. The mean on the statement ‘I regularly lend money to friends’ (pastexp5) is higher for the participants that played out than for the participants that played in in the $5 condition. This is again a remarkable result and is again blamed on a type I mistake. This statement is probably the worst predictor for trust because nowadays many people use bankcards.

Before running regressions it is important to take a look at the data. In the $5 condition, the variables PROMISE, humour, male, pastexp1 and pastexp3 have been

(17)

17

identified by the t-test to possibly predict playing in or out. In the $7 condition, the variables PROMISE, humour, male, B_ROLL, LENGTHBIN, pastexp1, pastexp3 and pastexp4 have been identified by the t-test to possibly predict playing in or out. In table 6 the correlations between the independent variables in the $5 condition can be found, and in table 7 the correlations between the independent variables in the $7 condition can be found.

TABLE 6: Correlation Matrix independent variables $5 condition

PROMISE humour male pastexp1 pastexp3

PROMISE 1,000

humour -0,084 1,000

male 0,003 -0,166 1,000

pastexp1 -0,033 0,121 -0,311 1,000

pastexp3 -0,030 0,074 -0,128 -0,041 1,000

TABLE 7: Correlation matrix independent variables $7 condition

PROMISE humour male B_ROLL LENGTH pastexp1 pastexp3 pastexp4

PROMISE 1,000 humour -0,015 1,000 male -0,064 -0,052 1,000 B_ROLL 0,673 0,044 -0,033 1,000 LENGTH 0,443 0,148 -0,060 0,548 1,000 pastexp1 0,027 -0,001 -0,146 0,015 -0,031 1,000 pastexp3 0,077 -0,025 -0,130 0,023 0,009 -0,279 1,000 pastexp4 -0,012 0,033 0,162 0,017 0,032 -0,131 -0,116 1,000

If we look at table 6, we see one correlation that is bold. This means that there is a high correlation between the variables pastexp1 and male. If we look at table 7 we see four correlations that are bold. So there is a high correlation between PROMISE, LENGTH and B_ROLL. There is also a high correlation between pastexp1 and pastexp3. We have to keep that in mind when running the regressions because high correlations between independent variables cause imperfect multicollinearity. Imperfect multicollinearity will make the model less reliable. It will produce larger standard errors and thus type II mistakes. Furthermore, the coefficients are not well estimated.

(18)

18

Now it is important to run some regressions to see whether the variables that have been identified by the t-test as predictors for playing in are also significant altogether. In this way we can build a model to identify the determinants of trust in this game. During this regressions use is made of heteroskedasticity- and autocorrelation-consistent standard errors because the standard errors probably are correlated within participants. Furthermore with the use of regressions we can see how big the influence of the variables is on the choice to play in and how likely the participants think it is that player B will play roll. At last, it is important to run some regressions because then we can see if the variable ‘LENGTH’ also has an effect instead of ‘LENGTHBIN’.

We first have a look at the $5 condition, and thereafter we have a look at the $7 condition. In both conditions we first do (a) regression(s) with ‘in’ as the dependent variable. In this way, we can see whether the independent variables have an influence on the behaviour of people in the trust game. Thereafter, in both conditions, we do a regression with b_guess as the dependent variable. In this way, we can see how people estimate whether someone can be trusted. Trust is always a continuous variable, which translates into binary behaviour

(trust/don’t trust) in this game. It is important to see if there is a difference between the influence of the independent variables on these two dependent variables (in and b_guess). This is important because it would be very interesting if there is a difference between people estimating whether someone can be trusted, and the choice to trust or not to trust. The correlation between b_guess and in is 0.5. The results of the regressions for the $5 condition can be found in table 8.

(19)

19

TABLE 8: Determinants of ‘in’ and ‘b_guess’ in the $5 condition

(1) (2) PROMISE 0,587** (0,249) 0,144 4,711 (3,260) humour 0,973*** (0,291) 0,224 2,905 (4,827) male -0,681* (0,349) -0,165 -9,300 (6,190) pastexp1 -0,218** (0,105) -0,053 -0,935 (2,212) pastexp3 0,390* (0,210) 0,096 -0,147 (3,048) Constant -1,591 (1,402) 54,423*** (19,712) N 380 380 (Pseudo) R-squared 0,089 0,034

Notes. Dependent variable regression 1: in (1- in and 0-out). Dependent variable regression 2: b_guess. The table represents

coefficients from a logistic regression for specification 1. The table represents coefficients from a linear regression for specification 2. Heteroskedasticity- and autocorrelation-consistent standard errors are in parenthesis. The third number in regression 1 is the marginal effect. *Significant at the 10% level, **Significant at the 5% level, ***Significant at the 1% level.

We first have a look at the first logistic regression in table 8. Hence that we keep all other variables constant if we discuss the effects further on. For an effect to be economically significant, the cut-off value is 0,1. This value is chosen because if the marginal effect of a variable on ‘in’ is bigger than 0,1 or smaller than -0,1, this effect is quite important and big in economic terms.

If we look at the first regression we see that the predicted probability of playing in is 0,144 greater for the individual that reads a message wherein a promise is made than for an

(20)

20

individual that reads a message wherein no promise is made. This result is significant at the 5% level and is also economically significant. We also see that the predicted probability of playing in is 0,224 greater for the individual that rated the message as funny as for an

individual that rated the message as not funny. This result is significant at the 1% level and is also economically significant. The predicted probability of playing in is 0,165 lower if a male participant reads the message than when a female participant reads the message. This result is statistically significant at the 10% level, and is also economically significant. The predicted probability of playing in is 0,053 lower if people score one point higher on the statement ‘whenever you trusted someone in the past, that trust was betrayed’. This result is statistically significant at the 5% level but is not economically significant. The predicted probability of playing in is 0,096 higher if people score one point higher on the statement ‘people can trust me’. This result is statistically significant at the 10% level but is not economically significant.

However the Pseudo R-squared of this model is 0,089, which means that only 8,9% of the variance is explained by this model. This is a very small number. The R-squared of the second regression is even smaller, 0,034. If we look at that regression we see that none of the variables is statistically significant. So if the participants have to estimate how likely it is that player B has played roll, it is not clear how they judge this.

Clearly, there is a difference between how people estimate if someone can be trusted and what their behaviour in the trust game is. When people estimate whether someone can be trusted, they base their choice on less characteristics of themselves and the messages, than when we look at the behaviour of them in the trust game. One explanation for this difference to arise is that people first had to choose whether they would play in our out, and then had to estimate how likely it was that the message-sender could be trusted. They could simply have forgotten the characteristics of the messages, and through this, their estimate was less accurate than their behaviour. However this has to be investigated. The results of the regressions of the $7 condition can be found in table 9.

(21)

21

TABLE 9: Determinants of ‘in’ and ‘b_guess’ in the $7 condition

(1) (2) (3) (4) (5) PROMISE 1,090*** (0,244) 0,255 1,129*** (0,231) 0,264 8,245** (4,046) humour 0,877*** (0,326) 0,191 0,905*** (0,324) 0,196 0,729** (0,326) 0,162 0,805** (0,315) 0,177 4,122 (4,715) male -0,423 (0,261) -0,100 -0,430 (0,262) -0,102 -0,431* (0,247) -0,102 -0,455* (0,254) -0,108 -2,256 (5,689) B_ROLL -0,081 (0,300) -0,019 0,744*** (0,221) 0,175 8,341** (3,971) LENGTH 0,006 (0,010) 0,001 0,019** (0,008) 0,005 -0,030 (0,101) pastexp1 -0,337*** (0,100) -0,079 -0,340*** (0,099) -0,080 -0,297*** (0,099) -0,070 -0,315*** (0,102) -0,074 -3,631 (2,492) pastexp3 0,030 (0,182) 0,007 0,023 (0,185) 0,006 0,101 (0,180) 0,024 0,078 (0,179) 0,018 0,504 (3,604) pastexp4 -0,321*** (0,100) -0,075 -0,320*** (0,099) -0,075 -0,307*** (0,102) -0,073 -0,312*** (0,098) -0,074 1,246 (2,348) Constant 2,264* (1,309) 2,373* (1,307) 1,930 (1,331) 2,172* (1,292) 44,559 (32,995) N 430 430 430 430 430 (Pseudo) R-squared 0,120 0,119 0,086 0,092 0,010

Notes. Dependent variable regression 1, 2, 3 and 4: in (1- IN and 0-OUT), regression 5: b_guess. The table represents

coefficients from logistic regressions for regression 1, 2, 3 and 4 and from a linear regression for regression 5.

Heteroskedasticity- and autocorrelation-consistent standard errors are in parenthesis. The third number in regression 1, 2, 3 and 4 is the marginal effect. *Significant at the 10% level, **Significant at the 5% level, ***Significant at the 1% level

(22)

22

We now have a look at table 9. Hence that we keep all other variables constant if we discuss the effects further on. For an effect to be economically significant, the cut-off value is 0,1. This value is chosen because if the marginal effect of a variable on ‘in’ is bigger than 0,1 or smaller than -0,1, this effect is quite important and big in economic terms. The variables PROMISE and/or B_ROLL and/or LENGTH are left out of the second, third and fourth regression because these variables had no significant influence on playing in because of imperfect multicollinearity. When we look at the first regression of table 9 we see that the pseudo R-squared is 0,120. This means that 12% of the variance is explained by this model. The marginal effects will be discussed now; however we have to keep in mind that these can be biased because of imperfect multicollinearity.

If we look at the first logistic regression of table 9 we see that the predicted probability of playing in is 0,255 greater for the individual that reads a message wherein a promise is made than for an individual that reads a message wherein no promise is made. This result is significant at the 1% level and is also economically significant. The coefficient in the second regression does not change a lot by omitting B_ROLL and LENGTH.

The predicted probability of playing in is 0,191 greater for the individual that rates the message as funny as for the individual that rates the message as not funny. This result is significant at the 1% level and is also economically significant. The predicted probability of playing in is 0,100 lower for a male individual that reads a message than for a female individual. This result is not statistically significant. However if we look at the third and fourth regression, this variable is statistically significant at the 10% level. The non-significance probably stems from imperfect multicollinearity.

The predicted probability of playing in is 0,019 lower for an individual that reads a message from someone who played roll than for an individual that reads a message from someone who played don’t roll. This is a very remarkable result; we would expect that this coefficient would be positive. However if we look at regression four, this coefficient is positive. This remarkable result stems from imperfect multicollinearity. According to regression four the predicted probability of playing in is 0,176 higher for an individual that reads a message from someone who played roll than for an individual that reads a message from someone who played don’t roll. This result in significant at the 1% level and is also economically significant.

The predicted probability of playing in is 0,001 higher for every word that is added to the digital message. This result is not statistically significant. However there was a correlation between PROMISE, B_ROLL and LENGTH. If we look at regression three, PROMISE and

(23)

23

B_ROLL are left out of the regression. The predicted probability of playing in is then 0,005 higher for every word that is added to the digital message. This is significant at the 5% level but is not economically significant.

The predicted probability of playing in is 0,079 lower for people who score one point higher on the statement ‘whenever you trusted someone in the past, that trust was betrayed’. This result is statistically significant at the 1% level, but not economically significant. The variable pastexp3 is not significant in any of the specifications. This probably arises because people only chose 5, 6 and 7, so there is not enough variability in this measure. The predicted probability of playing in is 0,075 lower for people who score one point higher on the

statement ‘when dealing with strangers, it is better to first take care, before you start to trust them’. This result is also statistically significant at the 1% level, but not economically significant.

If we look at the fifth linear regression in table 9, we see that only the variables PROMISE and B_ROLL are statistically significant. If participants read a message wherein a promise is made, they judge it 8,2% more likely that the message-sender will play roll. This is statistically significant at the 5% level, and economically significant. If the message-sender has played roll, the participants judge it 8,3% more likely that the message-sender has played roll. This is significant at the 5% level and economically significant. So again, in the $7 condition, people can predict whether the message-sender can be trusted. The other variables are not statistically significant.

So again, there is a difference between how people estimate if someone can be trusted and what their behaviour in the trust game is. When people estimate whether someone can be trusted, they base their choice on less characteristics of themselves and the messages, than when we look at the behaviour of them in the trust game. The same explanation as discussed earlier applies here.

Now it is important to run a regression to see how the imperfect multicollinearity can be explained. If you think rational, it would be logical that people who write a message first think whether they will play roll or don’t roll, and then write a message. It is then most likely that people who are going to play roll will more often make a promise and write a lengthier message. The results of this regression can be found in table 10. Both conditions are tested at the same time.

(24)

24

TABLE 10: Determinants of playing roll

(1) (2) LENGTH 0,031*** (0,005) 0,007 0,031*** (0,005) 0,007 PROMISE 2,127*** (0,130) 0,478 2,146*** (0,130) 0,481 humour 0,242 (0,183) 0,056 Constant -1,598*** (0,191) -1,652*** (0,201) N 810 810 Pseudo R-squared 0,240 0,242

Notes. Dependent variable: B_ROLL (1- roll and 0-don’t roll). The table represents coefficients from logistic regressions.

Heteroskedasticity- and autocorrelation-consistent standard errors are in parenthesis. The third number is the marginal effect. *Significant at the 10% level, **Significant at the 5% level, ***Significant at the 1% level

If we look at the first logistic regression in table 10, we see that the predicted probability of playing roll is 0,007 times bigger for an individual that has wrote one word more. This is statistically significant at the 1% level but not economically significant. The predicted probability of playing roll is 0,478 bigger for the individual that made a promise. This is statistically significant at the 1% level and economically significant. So people can predict (in the $7 condition) whether someone can be trusted, and they do this by looking at if someone made a promise and how long the message is. If we look at the second regression, we see that the variable humour is not statistically significant. So people who write a humorous message are not more likely to play roll. However people do trust a humorous message more often.

According to Stock and Watson “the requirements for internal validity are that the (OLS) estimator is unbiased and consistent, and that standard errors are computed in a way that makes confidence intervals have the desired confidence level.” (2012, p. 355-356). The estimators of the causal effects in all previous regressions are most likely consistent. This means that the sample average will converge in probability to the population mean (Stock and Watson, 2012). The estimators of the causal effects can in all regressions be biased because of

(25)

25

omitted variable bias. Furthermore, the regressions in table 8 and 9 suffer from imperfect multicollinearity, which can also bias the estimators. However if we look at the variance inflation factor, the imperfect multicollinearity is not that severe. The variance inflation factor never exceeds 10. The standard errors are computed as heteroskedasticity- and

autocorrelation-consistent standard errors, so the confidence intervals have the desired

confidence level. So the internal validity is difficult to assess, but everything has been done to prevent the internal validity to be bad.

There is no need to worry about reverse causality because all characteristics of the messages, as well as the participants, were there before they made the choice to play in or out and before they indicated what the chance was that player B would play roll. However this research may suffer from omitted variable bias. This can bias the coefficients. There are a lot of variables that are not investigated, for instance spelling mistakes. It can be that the amount of spelling mistakes correlates with the amount of humour used. Spelling mistakes can lower the propensity to play in, while people play in more often if humour is used. The same can happen for the length of the message. If the message is lengthier, more spelling mistakes can be made. There is however not enough variability in spelling mistakes to investigate this. There can also be other variables that cannot be analysed. So omitted variable bias is always a problem. If we look at the example of spelling mistakes and humour, the humour measure is expected to be biased downwards. If we would add spelling mistakes as a variable, we expect the humour coefficient to go up.

Measurement error does not seem to cause any problems. The measurement error will most likely consist of random error; this does not bias the results, unlike systematic error. It could be for instance that some participants did not understand the game. However these participants will have randomized their answers in this case, which results in random error.

In this research it is taken into account that there can be unobserved heterogeneity in the data. Participants judged nine, ten or eleven messages. Therefore there can be

autocorrelation and heterogeneity in the decision to play in or out, the chance that B will roll and whether they think the message is funny or not. Therefore heteroskedasticity- and autocorrelation-consistent standard errors are used in the regressions. These

heteroskedasticity- and autocorrelation-consistent standard errors are clustered per participant. Misspecification of the estimated model can always be a problem. The coefficients are then biased and inconsistent. In this research logistic regressions are used because the

outcome variable is binary. This research has also done some linear regressions. There is no need to assume that the models are not correctly specified.

(26)

26

The external validity is also difficult to assess. In this research a trust game is used. We don’t know if the results can be generalized to for instance WhatsApp and email messages between employers and employees. The population studied and the population of interest differ from each other. Students often have different characteristics than people who are working. Students are younger for example. However we don’t know whether students react differently in these situations than people that are working. Both the institutional environment and the physical environment in which the survey was filled in differ between the population of interest and the population studied. From this we conclude that the external validity is low.

In conclusion, some patterns that are found in the t-test are more difficult to see in the regressions because of imperfect multicollinearity. If we look at the behaviour of people in the trust game, some hypotheses are confirmed. The first hypothesis ‘People are good at

predicting whether someone can be trusted’ is only confirmed for the $7 condition. Clearly, people play the game more random in the $5 condition, probably because they have less to loose, and more to gain. It looks like people base their decision in the $7 condition on the characteristics of the messages, for instance the length and whether a promise is made or not. People who can be trusted write a longer message and more often make a promise. The second hypothesis ‘If the message contains humour, this message is trusted more often than when it doesn’t contain humour’ is confirmed for both conditions. This is a remarkable result because people, who use humour, can’t be trusted more often. The third hypothesis ‘There will be a gender effect with respect to trust, females trust more often than males’ is also confirmed for both conditions. The last hypothesis ‘If people have had negative (positive) experiences with trusting in the past, they will trust less (more)’ is only weakly confirmed. In the $5 condition this hypothesis is confirmed for two out of five statements. Namely for the statements ‘whenever you trusted someone in the past, that trust was betrayed’ and ‘people can trust me’. In the $7 condition this hypothesis is also confirmed for two of the five statements. Namely for the statements ‘whenever you trusted someone in the past, that trust was betrayed’ and ‘when dealing with strangers, it is better to first take care, before you start to trust them’.

If we look at whether people can estimate if someone can be trusted, the results are less striking. People trust more if a promise is made in the $7 condition. The hypothesis ‘People are good at predicting whether someone can be trusted’ is only confirmed in the $7 condition. The other hypotheses are not confirmed if we look at whether people can estimate if someone can be trusted.

(27)

27

5. Discussion and conclusion

In this section, the results are discussed and a conclusion is reached. The data used in this thesis is obtained by doing a survey and analysing the messages in the Charness and Dufwenberg (2006) paper. The summary statistics can be found in table 2. There are no striking numbers except for the ones discussed in this section. We see that the mean proportion of people that plays in (A_IN) in the Charness and Dufwenberg (2006) paper is somewhat higher than the mean proportion of participants that plays in. We also see that people predict that the mean chance that someone will roll is 50,6% while the proportion of people that actually plays roll is higher (0,593).

The summary statistics per condition can be found in table 3. We see that the

proportion of people that choose in in the Charness and Dufwenberg (2006) paper is bigger in the $5 condition than in the $7 condition. This is quite logical because the people in the $5 condition have less money to loose. The messages from the players B are lengthier in the $5 condition than in the $7 condition. Maybe this can be explained by the fact that people in the $5 condition have more to gain and try harder to convince the opponent by writing a longer message.

In conclusion, some interesting results have been found. This research has investigated whether people are good at judging if someone can be trusted via a digital message and how they judge this. The first result is that this depends both on the payoffs and on which measure of trust we use. If people can gain more and lose less, they make the decision to trust more random. They can’t predict whether someone can be trusted. They base the decision to trust on if a promise is made in the message and whether humour is used. Females trust more often than males if people have less to lose and more to gain. Furthermore they base the decision to trust on their past experiences with trust. However foregoing results only apply if we look at the behaviour of people in the trust game. There are no results if we look at whether people can estimate if someone can be trusted.

If people have more to lose and less to gain, they are actually good at predicting whether someone can be trusted, both if we look at the behaviour of people in the trust game and if we look at whether people can estimate if someone can be trusted. If look at the behaviour of people in the trust game, people look at the length of the message, whether a promise is made, and if humour is used. However only lengthier messages and messages that contain a promise can be trusted more often. Females trust more often than males, and the decision to trust is based on their past experiences with trust. If we look at whether people can

(28)

28

estimate if someone can be trusted, people can also predict whether someone can be trusted. They base the decision to trust on whether a promise is made.

So employers can, in some situations, predict whether someone can be trusted. If you communicate via digital messages as an employee, it is recommendable to make promises, use humour and write a lengthy message if you want to be trusted. Female employers will trust you more often than male employers, but this also depends on their past experience with trust.

However there are some limitations to this research. The first limitation is that the participants make hypothetical choices and that they got no incentives. It could be that the participants would have filled in the questionnaire in a different way if the choices they made were coupled to incentives. Moreover, it could be that they would fill in the questionnaire more seriously if they would have gotten incentives.

The second limitation is that the questionnaire was very long. The participants were busy for 15 minutes on average. This can also be coupled to the previous limitation. If the participants would have gotten an incentive for filling in the questionnaire, maybe they would fill it in more seriously. The participants had to read very good to understand the game well and it looked like some participants didn’t read good enough to understand it. Moreover it could be that the questionnaire was too long for their attention to hold.

The third limitation is that in the questionnaire it was asked whether the message was funny or not. This could have led to experimenter demand effects. However this could go both ways; people could choose in ór out more often if they thought the message was funny. They did not know what the experimenter wanted to find.

The fourth limitation is that it is difficult to generalize these findings to the real world because students have filled in the questionnaire. It can be that students react differently than employers in these situations. It is also difficult to generalize these findings to the real world because in organizations you don’t play this game. In organizations people send each other digital messages wherein they could make a promise that would be too costly if they didn’t hold on to that promise. Moreover in organizations people can have contact via messages repeatedly. If something is not clear, they can ask for it.

The fifth limitation is that the characteristics of the messages and the participants are not exhaustively investigated. It could be that there are a lot more determinants of trust. For example it could be that spelling mistakes in messages lower trust.

The sixth limitation is that there can be a sequence effect in the answers from the respondents in the questionnaire. There were 8 different questionnaires, but all the

(29)

29

questionnaires had the same sequence. However the messages differed. It can be that the participants rated the first message as funny and the second one not because the second one was less funny. In further research the messages must be randomly assigned to each

questionnaire.

Another limitation is that it could be that there is sample selection bias. Only people who wanted to fill the questionnaire in filled it in. It could be that people who have filled in the questionnaire are nicer, and nicer people trust more.

The last limitation is that in the questionnaire, it was first asked whether the

participant wanted to play in or out, and thereafter it was asked what they thought the chance was that person B would play roll. We have found interesting differences between the two trust measures. But we cannot say if these difference stems from a sequence effect or if there is another explanation.

If this research would be carried out all over again it is important to change the order of both the messages and the question to play in or out and the chance that person B will play roll. It is also recommendable to give the participants an incentive for filling in the

questionnaire. Furthermore it is interesting to study more determinants of trust, for example the influence of spelling mistakes on trust. At last it is recommendable to let a separate group of participants rate the funniness of the messages. In this way there are no experimenter demand effects.

However it is even more interesting to investigate emails between employers and employees for example. In this way the sample is more representative. You could analyse the application emails that employees send to their future employers. If the employee is hired, you can analyse whether they exert low, medium or high effort after their probation. Furthermore the generalizability of the results is easier if you analyse emails between employers and employees.

(30)

30

6.

References

Bochet, O., Page, T., & Putterman, L. (2006). Communication and punishment in voluntary contribution experiments. Journal of Economic Behavior & Organization, 60(1), 11- 26.

Bracht, J., & Feltovich, N. (2009). Whatever you say, your reputation precedes you: Observation and cheap talk in the trust game. Journal of Public Economics, 93(9), 1036-1044.

Buchan, N. R., Croson, R. T., & Solnick, S. (2008). Trust and gender: An examination of behavior and beliefs in the Investment Game. Journal of Economic Behavior & Organization, 68(3), 466-476.

Chaudhuri, A., Sopher, B., & Strand, P. (2002). Cooperation in social dilemmas, trust and reciprocity. Journal of Economic Psychology, 23(2), 231-249.

Charness, G., & Dufwenberg, M. (2006). Promises and partnership. Econometrica, 74(6), 1579-1601.

Charness, G., & Dufwenberg, M. (2006). Supplement to ‘Promises and Partnerships’, Econometrica Supplementary Material, 74, 1-14.

https://www.econometricsociety.org/sites/default/files/ECTA5008SUPP_0.pdf Croson, R., & Buchan, N. (1999). Gender and culture: International experimental evidence

from trust games. American Economic Review, 386-391.

Gintis, H. (2000). Beyond Homo economicus: evidence from experimental economics. Ecological economics, 35(3), 311-322.

Ismayilov, H., & Potters, J. J. J. (2012). Promises as commitments.

Loomis, J. L. (1959). Communication, the development of trust, and cooperative behavior. Human Relations.

Mishra, J., & Morrissey, M. A. (1990). Trust in employee/employer relationships: A survey of West Michigan managers. Public Personnel Management.

Poppo, L., Zhou, K. Z., & Ryu, S. (2008). Alternative origins to interorganizational trust: An interdependence perspective on the shadow of the past and the shadow of the future. Organization Science, 19(1), 39-55.

Rockmann, K. W., & Northcraft, G. B. (2008). To be or not to be trusted: The influence of media richness on defection and deception. Organizational Behavior and Human Decision Processes, 107(2), 106-122.

(31)

31

Schniter, E., Sheremeta, R. M., & Sznycer, D. (2013). Building and rebuilding trust with promises and apologies. Journal of Economic Behavior & Organization, 94, 242-256. Sternthal, B., & Craig, C. S. (1973). Humor in advertising. The Journal of Marketing, 12-18. Stock, J. H., & Watson, M. W. (2012). Introduction to Econometrics: Global Edition. Pearson

Education.

Trautmann, S.T., Kuilen, Van De, G., & Zeckhauser, R.J. (2012) Betrayal, Beliefs, and the Demographics of Social Capital. LISS Panel, CentERdata/MESS project, number 77.

(32)

32

7.

Appendix

1. Messages from player B, $5 condition Charness & Dufwenberg (2006)

The first number is the session number, the second number is the message number. 1 1 Please choose In so we can get paid more.

1 2 Choose in, I will roll dice, you are 5/6 likely to get 2,3,4,5, or 6  $12. This way both of us will win something.

1 3 If you stay in, the chances of the die coming up other than 1 are 5 in 6 – pretty good. Otherwise, we’d both be stuck at $5. (If you opt out)

1 4 I have to do laundry tonight and I really don’t want to do it! But I don’t have any clean underwear left and I don’t want to go commando tommorrow. We’ll see what I decide tonight. This man acts funny doesn’t he? But he seems cool, he’s quite a character. All this mystery is kinda cool.

1 5 If you will choose “In”, I will choose to roll. This way, we both have an opportunity to make more than $5! 

1 7 If I roll a 2–6 (you’ll know when you receive the $, you will give $5.00 to a stranger.

_________________________________________ Sign here if you are so kind

Thanks.

You’ll still be gaining more than if I had chosen Don’t roll.

1 8 The fairest thing to do is if you opt “IN”. Then I will proceed to choose “roll.” That way you and I have 5/6 chances to make money for the both of us. That’s much better than just making $5 each. Increases both our chances. Thanks.

1 9 Choose In and I will Roll You have my word 1 10 Good luck

I do not know what I’m going to do, so I have no hints on how to advise you on choosing “in” or “out.” Though it would be beneficial for me to pick don’t roll and hope you pick “in”, I also like to give you a chance to gain some cash. Who knows? 1 11 What’s up? Good luck on your decision. Choose whatever. If you choose “out,” you

get only $10 total. If you choose “In,” you can get $17 total instead of only $10. 7 bucks is a lot of money!

(33)

33

1 12 Hey. OK I think that the best way for both of us to make a profit is for you to choose IN and for me to roll. That way we both make some money. There’s no point in me not rolling because that would give you and me less profit. So I’m a roller if you’re in .

1 13 take a risk

1 14 If you choose IN the first round and then I will choose Don’t Roll at first. I will get $14 but then after that I will choose roll each time after the 1st role. Chances are most likely you will get $12 and I will get only $10. I will the only take 7 rolls for you to get even with me. That way we both leave with a good amount of money. Hope you have a great evening and that this works out for both of us. 

1 15 If you choose in I’ll roll.

Why? If you choose out, we walk out with $10 each. If you choose IN & I choose IN then both of us coin. So it’s a compromise. By agreeing to this I guarantee myself more $ than risking you choose out. So if you choose out I get $10 ($5 diff.) if you choose in I get $15 vs. $19 ($4 diff.). that’s why

1 18 Choose “In” so we can both make some $$ What are the chances me rolling a 1? I’ll try my best.

2 1 I’m going to roll. 2 2 I’ll choose roll. 2 3 I will choose roll. 2 4 I’m going to choose roll 2 5 choose in, & I’ll roll.

2 6 You can have the 2 extra dollars. I’ll be nice and choose to roll. 

2 8 Hey, choose in and I will roll. You have to like your odds that I will roll a 2,3,4,5, or 6. 5/6 odds ain’t bad.

2 9 If you choose “In”, I’ll choose Roll and you’ve got a 5/6 chance of getting $12. 2 10 Stay IN, I really need the money.

2 11 If you choose IN, and I roll, the chances of our getting the most $ are very high. The likelyhood of my rolling a 1 is small compared to the chances of rolling a 2–6. So we both get cash.

2 12 Hi, well I’m going to Roll so you have at least a shot for more money. I hope it works out.

Referenties

GERELATEERDE DOCUMENTEN

Lasse Lindekilde, Stefan Malthaner, and Francis O’Connor, “Embedded and Peripheral: Rela- tional Patterns of Lone Actor Radicalization” (Forthcoming); Stefan Malthaner et al.,

A legal-theory paradigm for scientifically approaching any legal issue is understood to be a shared, coherent collection of scientific theories that serves comprehension of the law

Muslims are less frequent users of contraception and the report reiterates what researchers and activists have known for a long time: there exists a longstanding suspicion of

As I held her in my arms that last night, I tried to imagine what life would be like without her, for at last there had come to me the realization that I loved her-- loved my

Everyone in Charleston was so welcoming and the International Office was so helpful and organized events where we all as internationals got to meet each other and were matched

[r]

You go to one of the world’s great art exhibits looking forward to seeing human creation at its most beautiful and instead you experience human nature at its ugliest.. 2 I am

Because they failed in their responsibilities, they would not be allowed to rule any more (cf.. Verses 5 and 6 allegorically picture how the terrible situation