The relationship between generalized trust and an individual’s
self-rated current health state
- A comparison of survey and experimental trust measures –
Manon Wehlmann
University of Amsterdam, Faculty of Economics and Business
A Master’s Thesis Submitted in Partial Fulfilment of the Requirements
for the Degree of:
Master of Science in Economics
University of Amsterdam
Under the Supervision of Dr. Gönül Dogan
October 2013
ABSTRACT
Generalized trust is a concept that received a lot of attention within the last two decades. Its investigations reach from simple experimental designs on an individual basis to sophisticated designs on a nation-wide level. In 1998 Barefoot et al. were the first to link high levels of trust to functional health and longevity. With the help of an experiment they measured individuals’ trust levels. The relationship between generalized trust and health was then revived when the term social capital was established. In social capital studies it was common practice to use survey measures of trust to address the relationship of generalized trust and health outcomes in society or the individual level. Studies, both experimental or survey investigations, found a positive relationship between generalized trust and an individual’s health state. However, none of these studies ever investigated the correlation of survey and experimental trust measurements with a third variable in the frame of one study. This research aims to close this gap in literature. To do this, this research uses a data set from the German socio-economic panel that conducted a trust experiment with a sub-sample of survey respondents. Investigating whether both measurement types reveal the same results, these measures are further used to give a statement on the relationship of generalized trust and an individual’s self-rated current health state.
1 INTRODUCTION
In a large body of studies manifold positive effects of generalized trust have been addressed. On the macro level, it has been shown that generalized trust leads to economic growth (Knack and Keefer, 1997) and lowers corruption (Zak and Knack, 2001; Guiso, Sapienza and Zingales, 2004). On the micro level, it has been shown that generalized trust leads to better interpersonal relationships in romantic relationships and friendships (Rusbult, Kumashiro, Coolsen and Kirchner, 2004). The term “generalized trust” describes the form of trust an individual places into unknown others (“strangers”) and has to be distinguished from other forms of trust, such as trust individuals put into family members or into organizations (Roth, 2009). In general generalized trust is regarded as a construct with positive effects on everyday aspects of life, which led to a fast growing literature on the topic in the last two decades. As generalized trust is a latent variable, there is no consensus on how to measure it correctly. However, two forms of measurement established themselves, which will be fundamental to the present research: (1) Experimental trust measures and (2) survey trust measures.
A well-known experimental trust measure for instance is a value gained by a game called the Trust Game (Berg, Dickhaut and McCabe, 1995) or the Gift Exchange Game (Fehr, Kirchsteiger and Riedl, 1993). The higher the trust value, the more trust the participant is said to have in other people (“strangers”). The most common survey question addressing trust is often taken from large surveys such as the American General Socical Survey (GSS) and is therefore called GSS-trust in the present research. The question reads as follows: “Generally speaking, would you say that most people can be trusted or that you can’t be too careful in dealing with people?” The participants in these surveys can answer this question in a binary way by either choosing “Most people can be trusted” or “Can’t be too careful”. But later on this way of trust measure has been critically called into question (Gabriel et al., 2002; Miller and Mitamura, 2003), which led to a development of an extended pool of trust questions in surveys consisting of more than one single question asking about a person’s attitude towards trust. In general both measurement types exist next to each other. Although it is generally believed that experimentally determined trust behaviour towards strangers provides a more convincing trust measure as this way of measurement captures beliefs and preferences of the respondent instead of being just “attitudinal”, none of the measurement types established itself to be the better method to capture actual trust behaviour.
The first attempt to validate the GSS-trust measurement with experimentally gained trust measures was conducted by Glaeser et al. (2000) who made use of two experiments and the GSS-trust question for their study. They found no correlation between the GSS-trust question and actual trust behaviour. Following two years later, Fehr et al. (2002) did a similar investigation, but instead of using just one trust question, they examined the correlation between different trust questions in a survey and an accompanying trust experiment. In contrast to Glaeser et al. (2000) they found a significant correlation of their survey and experimental trust measures. As the present research uses the same data source as Fehr et al.’s study (SOEP data set), their study serves as the basis for the present research. The SOEP survey involves randomly selected participants from households in Germany, thus the representativeness of the sample is ensured. Further the data set includes socio-economic and demographic background information of the participants. Fehr et al.’s (2002) results revealed that survey and experimental trust measures correlate well and that questions addressing people’s past trusting behaviour and direct questions about people’s trust are the best predictors of trusting behaviour. I expect to see a similar picture in the present data set (from 2003), but as Fehr et al. (2002) used data of a pretested data set, the number of observations differs quite a lot and some statistics have to be tested again. Also, their investigation is limited in so far as they compressed several questions into single factors, which will be elaborated in the present research.
Studies that addressed the correlation of generalized trust and a third variable of interest often made use of the GSS-trust questions as it is an easy accessible variable within a large data set. Next to earlier attempts by Hibbard (1985) or Grace and Schill (1986) to relate trust and health, Barefoot et al. (1998) were the first to link high levels of generalized trust to functional health, psychological well being and longevity. It was the time when the term “social capital” was established that brought back the connection of generalized trust and health. Social capital is known to be a multi-dimensional concept covering various indicators such as adherence to norms or participation in social networks that are said to “improve the efficiency of society by facilitating coordinated actions” (Putnam, 1993, p. 167). Among the indicators, trust is considered to be the most important (Coleman, 1990; Zak and Knack, 2001). In these studies, generalized trust was always surveyed with the GSS-trust question.
While evidence of many studies persuade that generalized trust is positively related to better health outcomes (e.g. Wang et al., 2009; Barefoot et al., 1998; Lindström, 2004), none of these studies ever tried to cross-validate experimental and survey measures of generalized trust within one study and to compare the effects of the different measurement types. All of
these studies either chose experimental or survey measures of trust. The present research therefore aims to investigate the following question: Which trust measure correlates better
with an individual’s current health state – a survey or an experimental trust measure?
In order to find an answer to this question, this research follows the model by Fehr et al. (2002) and first tests which survey questions correlate best with actual trust behaviour. In a second step the survey and experimental measures of trust are then examined regarding their correlation with an individual’s self-rated health status. The goal of the present research is to assess and compare the correlation of survey and experimental trust measures with an individual’s self-rated health state. Thus the present research extends the approach initiated by Glaeser et al. (2000) and Fehr et al. (2002) by not only validating survey trust measures with experimental trust measures but by comparing the correlations of survey and experimental trust measures with a third variable, the participant’s self-rated current health state.
2 THEORETICAL FOUNDATION
2.1 What is trust?
What is trust? One thing that can be said for sure is that due to its manifold positive correlations it certainly plays an important role in human relationships and in economy. As early as 60 years ago social scientists and psychologists discovered its importance in society (e.g. Barber, 1983; Mayer et al., 1995; Simmel, 1950). As Rotter (1967) described it, trust is “one of the most salient factors in the effectiveness of our present complex social organization” because the “survival of any social group depends upon the presence or absence of such trust” (p. 651). Simmel characterized trust with these profound words: Trust is “one of the most important synthetic forces within society” (1950, p. 318) and “without the general trust that people have in each other, society itself would disintegrate” (1990, p. 178). His work shaped substantial works on trust by various authors who took trust as a basis for different phenomena. For instance Luhman (1979) considered trust to be the basis for a reduced social complexity whereas Gambetta (1988) considered trust to be important for cooperation. Coleman (1990) on the other hand took it as a basis for individual risk-taking behaviour and Putnam (1995) interpreted trust to be the most important element of social capital.
During the last two decades there has been a lot of empirical research on trust which helped developing experimental tools to measure trust (Fehr et al., 1993; Kiyonari and
Yamagishi, 1996) as well as identifying its determinants (Yamagishi et al. 1999; Chaudhuri and Gangadharan, 2007; Bohnet et al., 2008). In addition survey measures of trust in panel data sets and its usefulness has been explored (e.g. Glaeser et al., 2000; Fehr et al., 2002; Ermisch et al., 2007). This broad research led to various intriguing results on the role of trust, which justifies the extensive research in this field. Trust is considered to not only have a dominant part in on-going interpersonal relationships, but also to be decisive for economic growth (LaPorta et al., 1997; Knack and Keefer, 1997; Reis et al., 2000). This makes trust indispensable in so far that in its absence implicit contracts no longer can exist which can further lead to an end of relationships or even worse, a breakdown of the whole economic system (Arrow, 1972; Zak and Knack, 2001).
Many attempts to capture the fundamental essence of trust have been performed but yet a concise and universally accepted definition of trust could not be agreed on. Here, Coleman’s (1990) definition of trust is followed, which can be described as a behaviour-based definition of trust. Two traits are characteristic for his definition of trust: Firstly the trustor voluntarily places resources in the free disposal of another person (trustee) without having control over the latter’s actions. Secondly the trustor sees a potential gain in trusting the other person, which builds the incentive to trust. The incentive is such that if the trustee is trustworthy, the trustor is better off than if trust were not placed, and worse off if the other person is not trustworthy and trust were placed (Fehr, 2002).
Literature mainly distinguishes three forms of trust: (1) generalized trust, which is also referred to as “trust in strangers” and is thus trust put into unknown others, (2) “thick trust”, which is generated by family networks and (3) “institutional trust”, which refers to trust or confidence that people put in institutions or organizations (Roth, 2009). Most of the results on the role of trust that have been found are based on either experimental measurements of trust or well known survey questions addressing “trust in others”. Both measurement types are commonly referred to as measuring “trust in strangers”. As the present research aims to give an opinion on the role that “generalized trust” plays in predicting an individual’s self-rated health, the other two forms of trust will be neglected. To avoid any confusions, when it is referred to trust in this research, “generalized trust” is meant and none of the other forms of trust.
2.2 How to measure generalized trust? Evidence from surveys and laboratory experiments
One might claim the best way to capture actual trust behaviour is running a behavioural experiment such as the gift exchange game developed by Fehr, Kirchsteiger and Riedl (1993) or the trust game (Berg, Dickhaut and McCabe, 1995). Although it took experimental economics several decades before it got accepted as a tool to enhance the understanding of human behaviour, it certainly provides some advantages over theoretical reasoning of empirical facts. One advantage for example is the high control level of the environment in which participants make their decisions. This makes data replication easy. Furthermore the control of exogenous variation of variables makes causal interpretations clear and reliable. On the other hand laboratory experiments are often criticized for the self-selection problem that could likely arouse due to the composition of the subject pool that most commonly consists of students. Another point worth mentioning is that experiments are often claimed to lack external validity because they are too artificial whereas the real world is more complex (Schram, 2005). But often these experimental investigations are expensive and not available which is why the use of survey measures of trust became popular. In surveys the subject recruitment usually follows a randomized manner, which is why the representativeness of the sample can be ensured. A main disadvantage of survey data on the other hand is that answers to sensible questions about people’s values may be biased due to social desirability. Moreover different respondents can interpret words in different ways, which complicates analyses.
The first attempt to get an accurate estimate of an individual’s trust level was performed by Rotter in 1967. He developed a questionnaire using a Likert format that resulted in the Interpersonal Trust Scale. It consisted of different questions concerning the survey respondent’s trust towards family, friends, colleagues, teachers, etc. and thereby addressed various social objects. A high score indicated trust for a great variety of these social objects. Additionally demographic data on participants were collected that showed that scale scores in the questionnaire were significantly related to positions in religion, family and socioeconomic level.
Nowadays the most common measurements of trust are taken from major databases such as the American General Social Survey (GSS) or the World Values survey (WVS). In these surveys trust is measured with the statement “Generally speaking, would you say that most people can be trusted or that you can’t be too careful in dealing with people?” The participants in these surveys can answer this question in a binary way by either choosing
“Most people can be trusted” or “Can’t be too careful”. The behavioural relevance of GSS-trust has been called into question. Firstly it has been criticized that respondents have to choose between trust and caution instead of choosing between trust and distrust or between cautious and incautious behaviour (Yamagishi et al., 1999). It has been shown that a questionnaire measuring trust and caution separately leads to different results than solely asking the GSS-trust question that does not disentangle trust and caution. Secondly, different societies tend to interpret the GSS-trust question differently (Gabriel et al., 2002), which results in problems when comparisons between countries are drawn. This problem arouses because people have different interpretations for the term “most people” (Reeskens and Hooghe, 2008). While some societies primarily understand “most people” as acquaintances others connect this expression to strangers. Thirdly Miller and Mitamura (2003) showed that survey respondents answering the GSS-trust question are very likely to agree with both statements. The authors reason that when answering this question, survey respondents either use their own experience and behaviour or they imagine how they would behave in situations that involve social risk. Thus it is not only the belief about others’ trustworthiness that is consulted when answering this question, but also the respondent’s preferences towards social risk. As a solution Miller and Mitamura suggested one-dimensional questions that address trust and distrust separately with answer categories on a 7-point Likert scale ranging from “not at all” to “complete trust”. Nevertheless the GSS-trust question has been used to draw important conclusions. Putnam (1993) for example rank-ordered countries by their trust levels on the basis of GSS-trust or Robinson and Jackson (2001) concluded that trust among Americans declines. But doubts about interpretation, validity and reliability of GSS-trust query these conclusions.
2.3 Health measurement
Conducting thorough and expensive medical examinations can be used to assess an individual’s health state. But these procedures are long and expensive. Easier ways that are equally effective were established that facilitate assessing an individual’s health state. Among these the most prominent method is to employ questions about health in a survey. Herewith questions about present diseases can be addressed, but what is quite more important, that one simple question can capture an individual’s health state quite precisely. This single question simply asks the respondents to answer the following: “In general, how would you describe your current health?” Thereby the respondents can choose between five answer categories
ranging from “very good” to “bad”. This method is called the concept of self-rated health (SRH) as the individual indicates his health state based on his own judgement. Although SRH is a subjective measure of health, it has been widely used in literature and studies have validated it as a substitute for more objective health measures like for example blood pressure or present diseases (e.g. Gold, Franks and Erickson, 1996; Poortinga, 2006; Carlson, 2004; Jylhä et al. 2006). The reliability and validity of SRH have been well-established (Idler and Benyamini, 1997).
2.4 The relationship of trust and health
Related to other evidence that positive attitudes and interpersonal orientations such as religiousness (Oxman et al., 1995) or optimism (Scheier et al., 1989) can have beneficial effects, Barefoot et al. (1998) used Rotter’s Interpersonal Trust Scale to address the potential beneficial effects of trust. Next to earlier attempts by Hibbard (1985) or Grace and Schill (1986) to relate trust and health, Barefoot et al. (1998) were the first to link high levels of trust to functional health, psychological well being and longevity in a sample of 100 men and women aged between 55 and 80 years in a longitudinal study. In order to find these results, they combined scores from the Interpersonal Trust Scale with self-rated health and psychological well being from follow-up questionnaires. Their study strongly supports the claim that trust promotes self-rated physical health. However, their study consists of only 100 participants aged over 55, which first of all does not allow giving any results for the younger generation and secondly is a rather small sample size.
Recently Schneider et al. (2011) deepened the examination through which mechanisms trust effects health. They demonstrated that in a relationship “strong trust inhibits anxiety and depression, which in turn promotes physical health” and that “conversely, […] weak trust promotes anxiety and depression, which in turn is harmful to physical health” (p. 669). As in Barefoot et al. (1998), their trust measurement, defined as trust in a romantic relationship, was measured with the Interpersonal Trust Scale (Rempel et al., 1985). Health measurements on the other hand were split into physical and mental health, the first measured with the Cohen and Hoberman Physical Health Scale (Cohen and Hoberman, 1983) and the latter measured with the Derogatis Psychological Adjustment Scale (Derogatis, 1994). This gave of course a more detailed measurement of health compared to the method used by Barefoot et al. (1998) who relied on self-rated health. However, this study only considered trust between two individuals that are in a relationship and hence investigated the trust form “thick trust”
mentioned above. Since the trust level between two individuals that know each other for a longer time is certainly different from a trust level between two strangers, it remains unclear whether similar effects could be replicated in a study that considers singles and couples. But the manifold positive correlations of trust mentioned earlier are mostly results from studies that investigated generalized trust (trust put into a “stranger”) and not the level of trust an individual puts into acquaintances or partners.
Evidence that an individual’s health state is strongly influenced by social and personal resources, determinants of health are of considerable interest in health research (Ahnquist et al., 2012). It has been shown that medical care alone is not the only intervention that can be used to improve a population’s health state but that social disparities play a crucial role for an individual’s health state (Bertelsmann Stiftung, 2012; Danis and Pesce, 2012). Factors like education, social background, gender or chances on the labour market are such sources of disparities. The concept of social capital has been widely used to address these and other important health determinants. Social capital is known to be a multi-dimensional concept covering various indicators such as adherence to norms or participation in social networks that are said to “improve the efficiency of society by facilitating coordinated actions” (Putnam, 1993, p. 167). Among the indicators trust is considered to be the most important (Coleman, 1990; Zak and Knack, 2001). In the last two decades it has been shown in a large body of studies that social capital is positively related to a person’s health status (e.g. Lindström, 2004; Phongsavan et al., 2006). At first it has been shown on an aggregate level that average levels of generalized trust are related to various health measures in a state (Kawachi et al., 1999). Later this investigation has been extended on an individual level. High levels of social capital have been shown to positively correlate with lower mortality (Wilkinson, 1996; Kawachi et al., 1997) and better general health (Lindström, 2004; Veenstra, 2005). However, trust measurements in these studies are mostly kept simple because they rely on the attitudinal GSS-trust question and are not cross-validated with experimental trust measurements. Moreover, depending on the conceptualization of social capital and the demographic characteristics of the subjects or nations investigated, the strength of the relationship between social capital and individual self-rated health varies (Kennelly, O’shea and Garvey, 2003; Poortinga, 2006).
Based on the results of the studies listed above there exists strong evidence that there is a positive relationship between trust and an individual’s health state. However, as became clear this relationship so far has only been investigated in limited environments. The present
research therefore aims to extend the existing studies on the relationship between generalized trust and individual health by considering the shortcomings of former studies.
2.5 Combining survey and behavioural trust measures
As it became clear in the abstracts above, neither laboratory experiments nor survey measures of trust by themselves seem sufficient to give a fully reliable trust measure. Surveys usually lack the validity of trust measurements, whereas experiments lack external validity and representativeness. In the following a brief overview of studies combining survey and experimental trust measures is presented and the usefulness of this approach is elaborated.
Glaeser et al. (2000) were the first to combine experimental and survey trust measures. In their study they aimed to validate answers of a survey trust measure with experimentally gained trust measures. In addition to filling out a questionnaire containing demographic, socio-economic and trust data, the participants took part in two behavioural trust experiments following three to four weeks later. For their sample consisting of 189 Harvard students of an introductory economics course they found that GSS-trust (survey trust measure) is not correlated with actual trusting behaviour (experimental trust measure) observed in the experiments, but that it predicts people’s trustworthiness. Similar results have been replicated by several other studies (e.g. Ashraf et al., 2003; Ermisch et al., 2007; Haile et al., 2008).
Fehr et al. (2002) on the other hand found opposing results. As a basis they used the SOEP data set from 2002 containing - next to other socio-demographic questions - six different and relevant questions addressing trust and fairness (Appendix 1). Via factor analysis they built five factors out of these various questions: (1) Belief that people are fair, (2) belief in trustworthiness of others, (3) trust in others and institutions, (4) frequency of past trustful behaviour and (5) whether a person has benefited from the generosity of a stranger in the past. Additionally the survey was accompanied by a modified trust experiment that was conducted with parts of the survey respondents that served as an experimental measure for trust and trustworthiness. In total 429 participants agreed to take part in the experiment. The authors found their survey measurement of belief in trustworthiness of others (=trust in strangers) and past trusting behaviour to significantly correlate with the experimental trust measure, but no survey measure to be a good predictor for trustworthiness. In order to find these results they conducted ordered probit regressions with the factors and the experimental trust or trustworthiness measures. Also, they checked the raw correlations between the answers and behaviour via Spearman Rank Correlations, which supported their results from
the ordered probit regressions. Further their study is very informative concerning the construction of future survey research, as their results made clear that questions do no necessarily need to directly ask about a person’s belief in the trustworthiness of others to get a good indicator for trust in strangers but that the frequency of past trustful behaviour correlates even better with experimental trusting behaviour. Similar results have been replicated by for instance Vyrastekova and Garipikati (2005) or Bellemare and Kroeger (2007).
Next to different survey measures of trust used in the studies of Glaser et al. (2002) and Fehr et al. (2002), Sapienza et al. (2007) identified two differences between the studies to cause conflicting results: The degree of homogeneity and the degree of mutual knowledge among the subjects in the sample. While Glaser et al. (2000) had a subject pool with undergraduate students from Harvard, Fehr et al. (2002) had a subject pool composed of randomly selected representative households in Germany. Compared to anonymous and independent households in a survey, students know each other better and are relatively similar as they share the same stage of life. Hence the degree of homogeneity and mutual knowledge in the student sample is much more pronounced.
Although the design of the studies investigated by the above authors differs, the main message communicated is the same: Combining survey data with behavioural experiments first of all can be used to cross-validate behavioural and survey data and secondly allows to investigate socio-economic and demographic determinants of experimental behaviour. Furthermore it is an excellent tool to overcome weaknesses of both methods.
The results of the studies also made clear that there is no consensus about whether standard survey questions about trust, such as the GSS-trust question, or improved questions addressing trust measure general trust or account for trustworthiness. Hence it is important to identify those questions from the present survey that are best to predict trusting behaviour and to evaluate them in correct dimensions as the subject pool can differ in important determinants of trust. As the present research uses the same data source as Fehr et al. (2002), but a different year with a different pool of trust and fairness questions, the analytical approach resembles the one in their study.
Although it is generally believed that experimentally determined trust behaviour towards strangers provides a more convincing trust measure, Naef and Schupp (2009) and Fehr (2009) have shown that an extended scale of survey trust measures, in contrast to the GSS-trust question, is significantly affected by risk and social preferences, just like experimental trust measures. Thus, contrary to the general opinion that survey trust measures only represent the belief about the trustworthiness of others, they actually do capture beliefs and preferences.
Hence, the survey trust measures in the present research are consistent with the idea of Miller and Mitamura (2003) that “people derive their answers to trust questions from introspecting on their own likely behaviours in situations requiring trust” (Fehr, 2009; p. 250). The extended pool of trust questions, as will be used in the present research, is thus more likely than the GSS-trust question to capture beliefs and preferences of the respondents and thus to correlate more with experimental trust measures than GSS-trust.
The advantages of combining survey and experimental trust measures will be used to address the relationship between an individual’s generalized trust level and his self-rated current health state. The hypothesis to be investigated for the present research reads as follows: Which trust measure correlates better with an individual’s current health state – a
survey or an experimental trust measure?
This section pointed out why trust is important and how it is related to an individual’s self-rated health status. Various problems arousing from different measurement techniques were mentioned, which will be elaborated in the following section.
3 METHOD
3.1 Trust and trustworthiness: Experimental and survey measures
This paper is based on a data set from the German Socio-Economic Panel, a representative longitudinal survey of private households in Germany that comprises about 12,061 households and 22,611 individuals. The data set includes socio-demographic information about the participants from a personal questionnaire and additional information extracted from an experiment that was executed within the household survey. Since 1984 the research institute ‘TNS Infratest Sozialforschung’ conducts the survey on an annual basis. The research institute is well known in Germany and enjoys the reputation of being reliable and trustworthy. It sends interviewers to households in order to fill in the questionnaire together with the survey respondent. The personal questionnaire is made for individuals above the age of 16 and gathers information about e.g. income, employment, education or health. Pretested in 2002, a pilot study explored whether it is able to integrate an incentivized economic experiment in the SOEP survey. Developed by Fehr, Fischbacher, Schupp, von Rosenbladt and Wagner (2002), the experiment has first been executed in the survey year 2003 under the name “give and take”. For financial reasons the experiment was carried out
with a randomly selected representative sample that summed up to a total of 1,432 participants. Half of them were randomly assigned to be in the role of “player 1”, the other half in the role of “player 2”. After all questions have been answered the interviewer explained the experiment with the help of written instructions to the survey respondent (Appendix 2). Participation in the experiment was voluntarily but 95% agreed to it. Information about the average time necessary to carry out the experiment is not given.
The advantage of using survey respondents in this case is that they usually do not interact with each other, which facilitates an anonymous and independent allocation of two respondents. However, to ensure that there was no interaction at all, only one person per household was allowed to participate in the experiment. The experiment itself represents a sequentially played one-shot Prisoner’s Dilemma game where one player 1 and one player 2 are matched. Both participants are initially endowed with ten Euros. The first-mover, player 1, indicates on his decision sheet (Appendix 3) how much to transfer to player 2. This can be any integer amount between one and ten Euros. The experimenter doubles the transfer, that is, if player 1 sends x Euros, player 2 receives 2x Euros. After player 1 made his decision, he also indicated his expectation about the second-mover’s transfer on his decision sheet. Player 2 gets informed about the first-mover’s transfer with a hand written number on his decision sheet (Appendix 4). He then indicates his transfer to player 1. As with the first-mover, this could be any integer amount between zero and ten. Before it reaches player 1, the experimenter doubles the amount sent by player 2. This design results in the following payoffs: 10-x+2y gives the total payoff for player 1 and 10+2x-y the total payoff for player 2. Hence the payoff could vary from 0 to 30 Euros. After a player made a decision, he folded his decision sheet, put it into an envelope and closed it. The players made their decisions anonymously and privately, the interviewer would not know about the participant’s decision. When all surveys and interviews had been executed the first- and second-mover were matched, individual earnings were computed and checks were sent by mail to the participants. The pairwise matching process followed ex-post, after both players have made a decision. Based on previous knowledge about the distribution of first-mover actions from the pilot experiment an ex-ante distribution of first-movers could be determined. Each player 2 got confronted with one randomly chosen first-mover action. Consequently this design allowed the interviewers to conduct the experiments simultaneously with both player types and to match them ex-post without distorting the usual survey process. Furthermore, this type of implementation did not require any additional administrative or logistical organization. A
detailed analysis of how to implement a sequential pairwise experiment into a survey can be found in Fehr et al. (2002).
In the present research the observed experimental behaviour of the participants will be used as experimental measurements of trust and trustworthiness. The amount transferred by the first-mover thereby serves as a measurement of trust placed by the first-mover into the second-mover. The amount sent by the second-mover will serve as a measurement of trustworthiness.
Next to behavioural data about trust and trustworthiness gathered from the experiment, the personal questionnaire collected additional information about trust and fairness attitudes of the participants. In total the personal questionnaire comprised 134 questions (Appendix 5) among which the trust and fairness questions were answered at the beginning (questions 3 – 8). In the following the relevant questions from the personal questionnaire are described in detail.
Question 3: What is your opinion on the following three statements? a) On the whole one can trust people
b) Nowadays one can’t rely on anyone
c) If one is dealing with strangers, it is better to be careful before one can trust them
Question 3 had four answer categories: “Totally agree” (coded as 1), “agree slightly” (coded as 2), “disagree slightly” (coded as 3), and “totally disagree” (coded as 4).
Question 4: Do you believe that most people…
− Would exploit you if they had the opportunity − or would attempt to be fair towards you?
Question 5: Would you say that for most of the time, people…
− Attempt to be helpful?
− or only act in their own interest?
Question 7: How often does it occur that, …
a) You lend your friends your personal belongings (i.e. CD’s, books, car, bicycle)?
b) You lend your friends money?
c) You leave the door to your apartment unlocked?
Question 7 consisted of five answer categories: “Very often” (coded as 1), “often” (coded as 2), “sometimes” (coded as 3), “seldom” (coded as 4), and “never” (coded as 5).
Question 8: Have you ever profited from the generosity of a person, who you had not previously met?
The first goal of the present research is to identify those questions from the trust and fairness questions that are best to predict actual trusting and trustworthy behaviour as exhibited in the experiment. But as has been shown by several studies (e.g. Bellemare and Kroeger, 2007 or Fehr et al., 2002), other questions than those directly addressing trust in others are better predictors for actual trusting behaviour. Which and if any questions will serve as survey measures of trust and trustworthiness will be seen in further analysis. Some of the weaknesses in the construction of survey measures of generalized trust that have been addressed earlier in part 2 and are overcome here shall be looked at again.
Trust, as well as trustworthiness, are both latent concepts that are very hard to measure since they are not easily observed. Based on evidence that the construct of the GSS-trust question measures both, trust and caution, the SOEP decided to split this question up into two parts in order to disentangle theses constructs. Question 3a) asks about the general trust in people and question 3b) refers to the reliability of others. Addressing trust and trustworthiness separately avoids the disentanglement of both constructs afterwards and ensures that the interpretation of each question is clear. Next to this it has been criticized that the term “most people”, as has been used in the GSS-trust question, can be interpreted differently and thus lead to different results. For this reason the SOEP added another question (question 3c) to the block that intends to measure generalized trust, which results in an extended version of the trust scale. This question directly addresses “trust in strangers”, which should avoid misunderstandings and different interpretations. It will be interesting to see whether this question serves as a good predictor for actual trust-in-strangers-behaviour as exhibited in the
experiment since it directly addresses “trust in strangers”. Moreover the SOEP asked the survey respondents to indicate their agreement (or disagreement) on a four-point Likert scale ranging from “agree fully” to “disagree fully” for every single question addressing generalized trust (see question 3). This way the survey respondent did not have to choose between trust and caution but could express his agreement (or disagreement) with each statement separately. This extended trust scale design meets the requirements that have been criticized in the design of the GSS-trust question.
3.2 Health measure
As for trust and trustworthiness, health is a variable that is hard to measure by observation. In the present research information about the individual health state of a respondent are solely taken from the survey. The study employs a self-rated general health measure, which can be found in the personal questionnaire (question 98 in Appendix 5). SRH was assessed on a five-point Likert scale on which respondents rated their current health as “very good” (coded as 5), “good” (coded as 4), “satisfactory” (coded as 3), “poor” (coded as 2), or “bad” (coded as 1). Although SRH is a subjective measure of health, it has been widely used in literature and studies have validated it as a substitute for more objective health measures like for example blood pressure or present diseases (e.g. Gold, Franks and Erickson, 1996; Idler and Benyamini, 1997; Poortinga, 2006; Carlson, 2004; Jylhä et al. 2006). Thus SRH serves as a measurement for the individual current health state and represents the dependent variable in the present research.
3.3 Stability of trust and health
The present data set comprises only one-time information about a respondent’s trusting behaviour, which is observed through the respondent’s behaviour in the experiment and his answers on the trust questions. As this research is not equipped with follow-up surveys in a timely manner, it is impossible to test the stability of trusting behaviour over time. However, Naef and Schupp (2009) compared respondents’ answers of 139 participants on trust questions in a distance of six weeks and showed that survey trust, similar to the present trust measurement in the SOEP survey, is only moderately stable over time. Thus the present research only works with temporary measurements of trust. Equivalent to the trust measurement, the study only employs one-time information about a respondent’s current health state. SRH also has been shown to be only moderately stable over time regardless of
any perturbations in health status (e.g. Perruccio et al., 2010). Hence, both measurements encompass the same time horizon regarding stability.
3.4 The relevance of different questions as a predictor for trusting and trustworthy behaviour
In examining the relevance of the different questions from the personal questionnaire for trusting and trustworthy behaviour this research follows the example of Fehr et al. (2002). This means that the questions are partly summarized to factors, for instance questions 3a), 3b) and 3c) representing one factor called “belief in trustworthiness of others”. Using the trust and fairness questions presented above, this approach results in five factors. The composition of the factors is presented in Figure 1. For each factor a Likert scale is constructed that indicates how intense an individual agrees (or disagrees) with the factor. The higher a factor the more an individual believes in other people’s fairness, helpfulness and altruism (questions 3, 4 and 5), the more close friends he indicates to have (question 6), the more he has shown trusting behaviour in the past (question 7), and the more he has spontaneously benefited from other people’s generosity in the past (question 8). An overview over all descriptive characteristics of each item and the factors is given in Appendix 6. Necessary modifications for an appropriate coding of the items and a factor analysis that confirms the one-dimensionality of the items that are summarized to one factor can be found in Appendix 7 and Appendix 8.
As the present research is dealing with ordinal dependent variables, the following analytical approach is used: The factors are first investigated regarding their raw correlation with the dependent variable. For this purpose the Spearman Rank correlations are used as this research is dealing with ordinal variables that have a monotonic relationship. Then an ordered probit regression will show which factors are the best predictors for actual trusting and trustworthy behaviour as could be observed in the experiment. The ordered probit regression analysis is used because it is suitable for modelling with a categorical dependent variable where the differences between the ordinal categories are discerned (McKelvey and Zavoina, 1975). For example, the model does not assume that the difference between a trust value of 5 and 6 in the experiment is the same as the difference between a trust value of 9 and 10, given a unit change in the explanatory variable. With this model qualitative differences between different trust values are captured. Also, unlike other models this multivariate approach is especially useful as it accounts for interdependencies among explanatory variables. The ordered probit model allows estimating the statistical significance and direction of the
relationship each explanatory variable has to each trust level as well as the marginal effects of each relationship (to be found in later analysis).
On the one hand working with factors simplifies the analysis because there are fewer items to interpret. But on the other hand it poses risks if something is overseen, as factors consisting of more items might lead to false conclusions if items influence the dependent variable in opposite directions. Therefore a detailed analysis of single items follows the investigation of the factors.
Figure 1
Composition of factors
Spearman Rank correlations of factors and first- & second-movers’ transfers
The relevance of different questions as a predictor of trusting behaviour can in a first step be assessed examining the raw correlations between the respondents’ answers and the respondents’ behaviour in the experiment. Therefore the Spearman Rank correlations between answers to the questions and actual behaviour in the experiment are presented in Table 1. Regression (1) shows the relationship between the first-movers’ transfers and the five factors, regression (2) displays the relationship between the second-movers’ transfers and the five factors. What can be seen is that questions about the belief in trustworthiness of others (question 3, ρ = 0.127) and questions about past trustful behaviour (question 7, ρ = 0.178) are by far the best predictors for the first-movers’ transfers and thus for trusting behaviour in the experiment. Although both factors show a rather low correlation with the experimental measurement of trust both are highly significant at the 1% level. Also the question about having benefited from a stranger in the past (question 8) co-varies significantly with trusting behaviour, but only on the 5% level and with an even lower correlation coefficient (ρ =
Factor 1: Belief in trustworthiness of others
Questions 3a), 3b) and 3c)
Factor 2: Belief about people's fairness and helpfulness
Questions 4 and 5
Factor 3: Number of close friends
Question 6
Factor 4: Frequency of past trustful behaviour
Questions 7a), 7b) and 7c)
Factor 5: Benefited from generosity of stranger in the past
0.087). In general one can observe that all factors co-vary positively with the first-movers’ transfers.
Table 1
Spearman Rank correlations of factors and first-mover and second-mover transfers
Spearman’s rho
(1) (2)
Factor for question 3:
belief in trustworthiness of others 0.127** 0.011
Factor for question 4 & 5:
belief about people’s fairness and helpfulness 0.001 -0.005
Factor for question 6:
number of close friends 0.012 0.013
Factor for question 7:
frequency of past trustful behaviour (e.g. lending money or car to friends) 0.178** 0.028
Factor for question 8:
question about whether person benefited from generosity of stranger in the past
0.087* 0.087*
Regression (1): Spearman Rank Correlations between first mover transfers and trust questions, N = 660 Regression (2): Spearman Rank Correlations between second mover transfers and trust questions, N = 668
* correlation coefficients that are significant at a 1% level or lower ** correlation coefficients that are significant at a 5% level or lower
Regarding regression (2) the results are not as striking as in regression (1). The only factor that co-varies significantly with the second-movers’ transfers and thus with trustworthiness is the question about whether a person has benefited from the generosity of a stranger in the past (question 8, ρ = 0.087). As in regression (1), the significance level for this question is at the 5% level. Except for the question about the belief about people’s fairness and helpfulness (questions 4 and 5), the remaining factors co-vary positively with the second-movers’ transfers albeit not at a significant level.
Ordered probit regressions of factors and first- & second-movers’ transfers
In a second step an ordered probit regression of first-mover and second-mover behaviour on the factors is presented in Table 2. Herewith the behavioural relevance of the questions about trust and fairness can be assessed. The table presents four different regressions among which the first two concentrate on trusting behaviour and the latter two on trustworthy behaviour.
Regression (1) shows that, equally to the results of the Spearman Rank correlations, questions about the belief in trustworthiness of others (question 3) and questions about past trustful behaviour (question 7) are the questions that correlate significantly with trusting behaviour in the experiment. Both are again significant at the 1% level. The other factors demonstrate no such correlation. This indicates that questions directly addressing trustworthiness in others as well as questions about past trustful behaviour do predict trusting
behaviour in the experiment. The results further suggest that the remaining factors are not relevant predictors for real trusting behaviour.
Table 2
The behavioural relevance of questions about trust and fairness with factors
1st movers transfer 1st movers transfer 2nd movers transfer 2nd movers transfer
(1) (2) (3) (4)
Factor for question 3:
belief in trustworthiness of others
0.30*** (0.045) 0.27*** (0.089) 0.047 (0.097) 0.063 (0.098) Factor for question 4 & 5:
belief about people’s fairness
-0.17 (0.11) -0.19* (0.10) -0.024 (0.11) -0.12 (0.11) Factor for question 6:
number of close friends
-0.008 (0.008) -0.003 (0.008) 0.004 (0.007) 0.007 (0.009) Factor for question 7:
frequency of past trustful behaviour (e.g. lending to friends)
0.24*** (0.055) 0.12** (0.058) -0.010 (0.063) 0.003 (0.06) Factor for question 8:
Question about whether person benefited from generosity of stranger in the past
0.097 (0.11) 0.12 (0.11) 0.24** (0.12) 0.26** (0.11) Belief about the transfer of the 2nd
mover
0.30*** (0.033)
Received from 1st mover 0.17***
(0.019)
Observations 660 668 660 668
Note: Ordered probit regression. Robust standard errors in parantheses *significant at 10%, ** significant at 5%, *** significant at 1%
Recalling to mind that the first-movers were asked to indicate on their decision sheet what transfer to expect from the second-mover, this measure can be used as a control variable for expectations. Controlling for the expected transfer, regression (2) presents correlations between the factors and trusting behaviour of the first-movers. As anticipated the expected transfer variable is highly significant (1% significance level). Furthermore the questions about past trustful behaviour (question 7) become less significant (10% level) while those about the belief in the trustworthiness of others (question 3) stay significant at the 1% level. The other factors still remain insignificant. What this regression suggests is that for given expectations about second-movers’ transfers those respondents who reported higher trust in strangers (question 3) are more likely to show more trusting behaviour in the experiment and hence transferred a higher amount of their initial endowment. The same applies for questions about past trustful behaviour (question 7). For given expectations about second-movers’ transfers those respondents who reported more trustful behaviour in the past are more likely to show more trusting behaviour in the experiment. Thus both questions 3 and 7 are good predictors for actual trusting behaviour.
Looking at regression (3) only one question can be considered to have a significant impact on the second-movers’ transfers, namely the question about whether a person has
benefited from the generosity of a stranger in the past (question 8). Even after controlling for the received amount from the first-mover (regression 4), this question stays significant at the 5% level. Thus for given amounts received from the first-movers, those respondents that indicated that they have benefited from the generosity of a stranger in the past are more likely to show more trustworthy behaviour in the experiment. This suggests that individuals that had positive experience with strangers in the past try to reward trusting behaviour.
Summing up these regression results identified two survey measures (questions 3 and 7) from the personal questionnaire that correlate significantly with trusting behaviour in the experiment and one survey measure (question 8) that correlates significantly with trustworthy behaviour. Except for the fact that this data set reveals a significant correlation between question 8 and trustworthy behaviour, the results resemble those of Fehr et al.’s study (2002). These preliminary results suggest that these survey measures can be taken as good predictors for trust and trustworthiness, respectively.
Ordered probit regressions of items and first- & second-movers’ transfers
In the beginning of this section it was mentioned that compressing single items to factors might lead to false interpretations because items could influence the dependent variable in different directions. Whether this is the case in the present research will be shown by an ordered probit regression that considers each item separately. Although the factor analysis in Appendix 8 confirmed the one-dimensionality of the items that are compressed to one factor, a detailed analysis of all items will give further information. The results of the regressions are presented in Table 3.
What can be seen from regression (1) is that indeed, questions 3 and 7 still play an important role as predictor for trusting behaviour but not necessarily every single item of the questions. Questions 3a) “on the whole one can trust people” and 7b) “lend money to friends” for example become insignificant. Furthermore 7b) now shows a negative correlation to the first-movers’ transfers. Thus questions about lending money to friends have to be treated differently than lending personal belongings like books or cars to friends (question 7a)). Also, questions 3b) and 3c) become less significant (5% vs. 1%). It can be seen that directly addressing trust in other people, as exhibited in question 3a), does not predict real trusting behaviour. Rather, it appears that other questions prove to be helpful in predicting actual trust behaviour.
Table 3
The behavioural relevance of questions about trust and fairness with single items
1st movers transfer 1st movers transfer 2nd movers transfer 2nd movers transfer
(1) (2) (3) (4)
Question 3a)
On the whole one can trust people
0.02 (0.07) -0.00 (0.07) -0.02 (0.08) -0.03 (0.08) Question 3b)
Nowadays one can’t rely on anyone
0.13** (0.06) 0.15** (0.06) -0.02 (0.06) 0.02 (0.06) Question 3c)
If one is dealing with strangers, it is better to be careful before one can trust them
0.13** (0.06) 0.11 (0.07) 0.08 (0.06) 0.06 (0.06) Question 4)
Belief about other people’s fairness
0.09 (0.09) 0.02 (0.09) 0.06 (0.09) 0.01 (0.09) Question 5)
Belief about other people’s helpfulness
-0.26*** (0.09) -0.21** (0.09) -0.06 (0.09) -0.12 (0.04) Question 6)
Number of close friends
-0.01 (0.009) -0.00 (0.01) 0.01 (0.01) 0.01 (0.01) Question 7a)
Frequency of past lending behaviour regarding personal belongings 0.15*** (0.04) 0.09 (0.05) -0.00 (0.04) -0.02 (0.04) Question 7b)
Frequency of past lending behaviour regarding money -0.08 (0.06) -0.09 (0.06) -0.01 (0.06) 0.00 (0.06) Question 7c)
Frequency of leaving door to apartment unlocked
0.12*** (0.04) 0.07* (0.03) -0.00 (0.04) 0.02 (0.04) Question 8)
Whether respondent has benefited from the generosity of a stranger in the past
0.11 (0.11) 0.12 (0.12) 0.25** (0.12) 0.27** (0.11)
Belief about the transfer of the 2nd mover 0.29***
(0.03)
Received from 1st mover 0.17***
(0.02)
Observations 660 660 668 668
Note: Ordered probit regression. Robust standard errors in parantheses *significant at 10%, ** significant at 5%, *** significant at 1%
What is striking now is that question 5) “people attempt to be helpful or only act in own interest” suddenly becomes highly significant (1% level) while question 4) “people try to exploit you or attempt to be fair” stays insignificant and changes the sign from negative (see Table 2) to positive (see Table 3). Thus compressing questions 4 and 5 into one single factor “belief about people’s fairness and helpfulness” does not seem appropriate. First of all the correlations signs lead in different directions and secondly, taken separately, the significance level changes, at least for one of the items. This suggests that fairness and helpfulness, as exhibited in questions 4 and 5, should be regarded separately when trying to assess the relevance for the first-movers’ transfers.
Controlling for the expected transfer shows an even clearer picture (see regression 2 Table 3). As expected, the variable for the expected transfer stays highly significant. Question 3b) shows no change in significance. For given expectations about second-movers’ transfers those respondents that indicated to rely on other people are more likely to trust more in the experiment. For questions 7a) and 7c) one can observe a decrease in significance, from the
1% to the 10% level. This suggests that for given expectations about second-movers’ transfers those respondents that reported a higher frequency of past material lending behaviour of personal belongings (7a) or that indicated to leave the door to their apartment more frequently unlocked (7c) are more likely to send a higher amount to the second-mover. Surprisingly question 3c) is now insignificant, which indicates that a question directly addressing the cautiousness towards strangers is not a good indicator to predict someone’s actual trust behaviour towards strangers. Question 5) is still significant at 5% and thus relevant to predict someone’s trusting behaviour. This indicates that for given expectations about the second-movers’ transfers those respondents that reported to think of others to be helpful are more likely to transfer less on average to the second-mover. This conclusion might sound a little strange at first, but obviously helpfulness does not align with trust. Probably social desirability causes people to indicate that they think of others to be helpful. However, when it comes to a real trust situation, helpfulness is not decisive to place trust into a stranger.
Now looking at regression (3) again only question 8 shows a significant effect (at the 5% level) on the second-movers’ transfers. Controlling for the amount received from the first-mover does not change this relationship (regression 4). This implies that the detailed ordered probit regression does not reveal any news concerning the predictors for the second-movers’ transfers.
Summing up the these results the detailed ordered probit regression led to four single survey measures that are relevant predictors for trusting behaviour. These are the questions addressing the reliability of others (3b), the helpfulness of others (5), the frequency of past lending behaviour of material belongings (7a) and the frequency of leaving the apartment door unlocked (7c). The best survey measure to best predict trustworthiness is still only question 8, asking whether a person has benefited from the generosity of a stranger in the past. These results are further supported by the Spearman Rank correlations, which can be found in Appendix 9. Taken together the results from both the ordered probit regression (Table 3) and the Spearman Rank correlations (Appendix 9), two questions can be identified to be the best predictor for trust and trustworthiness, respectively. Instead of asking directly about the trust someone places into strangers as is done in question 3c), a relevant question to predict someone’s actual trusting behaviour is to ask about the reliability of others (question 3b). Likewise a good predictor for trustworthy behaviour is to ask about whether the person has benefitted from the generosity of a stranger in the past but not asking about his own trustworthy behaviour in the past. This analysis has shown that it was worth examining a detailed regression of the single items before getting to superficial conclusions.
In the following the answers to questions 3b) and 8) are used as survey measures to assess the correlation of trust and trustworthiness with an individual’s self-rated health state. Whether these survey measurements will reveal the same results as the experimental measures of trust and trustworthiness will be seen in further analyses.
4 DESCRIPTIVE STATISTICS AND EXPERIMENTAL RESULTS
4.1 Comparison of sample and overall SOEP wave
A comparison between the sub-sample taking part in the trust experiment and the whole SOEP wave from the year 2003 shows that the sub-sample is representative as for German citizens. The distribution of important demographic characteristics is displayed in Table 4. The small differences between the two groups are insignificant. The sample was balanced in gender, 657 men and 671 women participated. The age ranged from 18 to 92. About 20% of the participants exceeded the retirement age (=65 in 2003). 62% of the sample were married. Only 7.3% were unemployed and only 15% were living alone. About 15% did not finish high school. Likewise Figure 2 shows that the differences in self-rated health states between the sample and the overall SOEP wave are insignificant. The good health state was rated most frequent in both the sample and the overall SOEP respondents with over 40%.
Table 4 Representativeness of sample Experiment sample 2003 N = 1,328 Experiment sample distribution 2003 (in %)
Distribution whole wave SOEP 2003 (in %) N = 22,611 Gender Women 671 0.5053 0.5188 Men 657 0.4947 0.4812 Age 18-19 47 0.0354 0.0462 20-29 157 0.1182 0.1358 30-39 207 0.1559 0.1880 40-49 300 0.2259 0.2051 50-59 203 0.1529 0.1571 60-69 229 0.1724 0.1571 70-79 148 0.1114 0.0809 80-89 36 0.0276 0.0264 90-100 1 0.0008 0.0032 Origin East 313 0.2364 0.2620 West 950 0.7175 0.6959 Abroad 61 0.0461 0.0422 Marital Status
Married (Separated) 26 0.0196 0.0176 Single 302 0.2274 0.2422 Divorced 78 0.0587 0.0667 Widowed 95 0.0715 0.0630 Household size 1 person 208 0.1566 0.1341 2 persons 501 0.3773 0.3631 3 persons 244 0.1837 0.2051 4 persons 266 0.2003 0.2023
5 and more persons 109 0.0821 0.0953
Employment status
Employee 1157 0.8712 0.8708
Self-employed 74 0.0557 0.0597
Unemployed 97 0.0730 0.0694
Education
Less than high school 202 0.1521 0.1689
High school 871 0.6559 0.5776
More than high school 210 0.1581 0.2031
Figure 2
Self-rated health state in sample and all SOEP respondents 2003
4.2 Behaviour in the experiment
The design of the experiment within the SOEP survey resembles the sequential structure of the gift exchange game (Fehr, Kirchsteiger and Riedl, 1993) or the trust game (Berg, Dickhaut and McCabe, 1995). A standard version of the trust game basically captures the enforcement problems such as the effort enforcement problem in the labour market due to incomplete information (Johnson and Mislin, 2011). Likewise the experiment within the household survey represents a sequential trade under incomplete contract enforcement. Although the game differs slightly from the original game of Berg, Dickhaut and McCabe (1995), theoretical assumptions for these experiments are the same. Assuming both players to
0% 10% 20% 30% 40% 50%
[5] Very Good [4] Good [3] Satisfactory [2] Poor [1] Bad
Rel at ive fr equ en cy
Self-rated health state
Experiment Sample, N = 1,328 SOEP 2003, N = 22,611
be rational and selfish and that this is common knowledge, the Nash equilibrium would equal a situation in which neither of the players transfers any positive amount to the other player. But contrary to standard theoretical predictions subjects in these games show considerable willingness to cooperate, although there is no direct benefit from making monetary sacrifices in a one-shot game (Johnson and Mislin, 2011). In the survey experiment, for each player the Euros owned by the other player are worth twice as much for himself, thus both participants are better off by sending the entire endowment. But the first-mover has to trust player 2 to be willing to send the entire endowment and player 2 has to be trustworthy. Economic theory postulates that an individual’s decision to trust is influenced by risk preferences, social preferences and expectations, which leads to different results (Johansson-Stenman et al., 2005). The behaviour observed in the experiment from the present research also appears to be different from theoretical predictions. The graphs in Figures 3, 4 and 5 illustrate these observations.
In Figure 3 the distribution of the first- and second-movers’ transfers is displayed. It can be seen that transfers of five Euros were the ones that occurred most frequent in each role of player types (first-movers: 43%, second-movers: 40%). This is not very surprising as in many replications of the trust game it is about 50% of the initial endowment that is sent most frequent (Johnson and Mislin, 2011). For the first-movers, four and ten Euros were the second most frequent transfers with both 12% whereas for the second-movers zero and ten Euros occurred second most frequent with 13% and 10%, respectively. A transfer of zero in the position of the first-mover occurred in only 7% of the cases. Comparing the frequencies of zero transfers between and second-movers, the data indicate a higher trust level for first-movers. Also, the transfers of zero to four Euros occurred less frequent in the position of players 1 (about 30%) compared to the second-movers (about 35%). This suggests overall higher trust levels of players 1.
The average amount sent by the first-mover was 5.10 Euros, which is slightly above the half of the senders’ initial endowment. The average amount sent by the second-mover was 4.75 Euros and thus 3.5% lower than the amount transferred by the first-mover. The expected amount returned equals on average 4.02 Euros. Thus the first-movers’ expectations have been exceeded; on average they received 7% Euros more than expected. In Figure 4 the average transfer of the first-movers is presented as a function of their beliefs about transfers of the second-mover. The graph indicates a strong relationship between the first-movers’ transfers
and their beliefs about the second-movers’ transfers. The higher the first-movers’ transfer, the higher the believed transfer of the second-mover.
Figure 3
Distribution of first-movers’ and second-movers’ transfers
Figure 4
Average transfer of first-movers’ as a function of first-mover’s belief about the transfer of the second-movers 0% 10% 20% 30% 40% 50% 0 2 4 6 8 10 Rel at ive fr equ en cy Transfers First-mover transfer Second-mover-transfer 0 1 2 3 4 5 6 7 8 9 10 0 2 4 6 8 10 Fir st -m over s' tr an sfer s
First-movers' belief about the second-movers' transfer
Figure 5
Average transfer of second-movers’ as a function of the first-mover’s transfer
Concerning the behaviour of the second-mover, Figure 5 displays the average transfer of the second-movers depending on what they received from the first-mover. The graph suggests that there is a positive relationship. The higher the amount received from the first-mover, the higher the amount transferred back to the first-mover. Interestingly the average transfer of the second-mover differs depending on whether the amount sent by the first-mover falls below or above five Euros. In the range from zero to four Euros, the average of the second-movers’ transfers are always above the initial receipt of the first-movers and above five Euros the average back transfers are always below the initial receipt. Hence, with an increasing amount received from the first-movers the average transfer back increases, but with a decreasing rate.
The results in Figures 3, 4 and 5 suggest that trust and trustworthiness are positively related to each other and that expectations influence the decision to trust. As with the survey measurements of trust and trustworthiness (see Tables 2 and 3), the experimental measurements of trust and trustworthiness are equally influenced by expectations and the amount received by the first-mover.
0 1 2 3 4 5 6 7 8 9 10 0 2 4 6 8 10 Sec on d-m over s' tr an sfer s First-movers' transfers
5 EMPIRICAL RESULTS
5.1 (Social) Determinants of health
Previous studies on health have shown that many socio-economic characteristics are important determinants for an individual’s health state. It has been proven that not only healthy behaviour such as physical activity or being a non-smoker lead to a better health state, but that these factors are further influenced by social indicators such as education, income or age, which lead to differences in health states (Bertelsmann Stiftung, 2012; Ahnquist et al., 2012). In the present study the most prominent socio-economic characteristics are used as control variables. Among these are age, gender, marital status, nativity, employment status, income and level of education (e.g. Prus, 2011; Giatti et al., 2010). Age is divided into three categories: Aged 18-44, 45-64 and 65+. Married, single and separated/divorced/widowed are the categories to cover the marital status. Nativity is categorized into West Germany, East Germany and people from abroad. The respondents were asked to indicate their highest school level, which was categorized into three levels: less than high school, high school and more than high school (college or university). Employment was simply recorded by asking whether the respondents are officially registered as unemployed at the Employment Office and the income was corrected for the members living in the household (adjusted household income).
Furthermore other questions from the survey that are thought to be important determinants for health are taken as additional control variables. These are for example physical activity, social intercourse with friends or relatives, worries about crime, the economic situation in general or worries about the own financial situation. Both physical activity and social intercourse with friends or relatives are measured with a five-point Likert scale, indicating that the respondents “often”, “at least once a week”, “at least once a month”, “seldom” or “never” take part in active sports or meet with friends, relatives or neighbours. Regarding the health care system in Germany, one faces a mandatory system. It is divided into a statutory health insurance (standard) and a more expensive private health insurance. As the health care system in Germany is compulsory there are only very few that are not enrolled and those are mostly the very rich people (German Federal Statistical Office, 2004). In the present survey only five respondents indicated to have no health insurance, which accounts for less than 1% of the total sample.
